xref: /netbsd-src/sys/netinet/tcp_output.c (revision f876c1012ec8280866d3287842a50de849d25967)
1 /*	$NetBSD: tcp_output.c,v 1.129 2005/03/29 20:09:24 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 1.129 2005/03/29 20:09:24 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 
200 #ifdef IPSEC
201 #include <netkey/key.h>
202 #endif
203 
204 #ifdef notyet
205 extern struct mbuf *m_copypack();
206 #endif
207 
208 /*
209  * Knob to enable Congestion Window Monitoring, and control the
210  * the burst size it allows.  Default burst is 4 packets, per
211  * the Internet draft.
212  */
213 int	tcp_cwm = 0;
214 int	tcp_cwm_burstsize = 4;
215 
216 #ifdef TCP_OUTPUT_COUNTERS
217 #include <sys/device.h>
218 
219 extern struct evcnt tcp_output_bigheader;
220 extern struct evcnt tcp_output_predict_hit;
221 extern struct evcnt tcp_output_predict_miss;
222 extern struct evcnt tcp_output_copysmall;
223 extern struct evcnt tcp_output_copybig;
224 extern struct evcnt tcp_output_refbig;
225 
226 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
227 #else
228 
229 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
230 
231 #endif /* TCP_OUTPUT_COUNTERS */
232 
233 static
234 #ifndef GPROF
235 __inline
236 #endif
237 int
238 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
239 {
240 #ifdef INET
241 	struct inpcb *inp = tp->t_inpcb;
242 #endif
243 #ifdef INET6
244 	struct in6pcb *in6p = tp->t_in6pcb;
245 #endif
246 	struct socket *so = NULL;
247 	struct rtentry *rt;
248 	struct ifnet *ifp;
249 	int size;
250 	int iphlen;
251 	int optlen;
252 
253 #ifdef DIAGNOSTIC
254 	if (tp->t_inpcb && tp->t_in6pcb)
255 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
256 #endif
257 	switch (tp->t_family) {
258 #ifdef INET
259 	case AF_INET:
260 		iphlen = sizeof(struct ip);
261 		break;
262 #endif
263 #ifdef INET6
264 	case AF_INET6:
265 		iphlen = sizeof(struct ip6_hdr);
266 		break;
267 #endif
268 	default:
269 		size = tcp_mssdflt;
270 		goto out;
271 	}
272 
273 	rt = NULL;
274 #ifdef INET
275 	if (inp) {
276 		rt = in_pcbrtentry(inp);
277 		so = inp->inp_socket;
278 	}
279 #endif
280 #ifdef INET6
281 	if (in6p) {
282 		rt = in6_pcbrtentry(in6p);
283 		so = in6p->in6p_socket;
284 	}
285 #endif
286 	if (rt == NULL) {
287 		size = tcp_mssdflt;
288 		goto out;
289 	}
290 
291 	ifp = rt->rt_ifp;
292 
293 	size = tcp_mssdflt;
294 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
295 #ifdef INET6
296 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
297 			/*
298 			 * RFC2460 section 5, last paragraph: if path MTU is
299 			 * smaller than 1280, use 1280 as packet size and
300 			 * attach fragment header.
301 			 */
302 			size = IPV6_MMTU - iphlen - sizeof(struct ip6_frag) -
303 			    sizeof(struct tcphdr);
304 		} else
305 			size = rt->rt_rmx.rmx_mtu - iphlen -
306 			    sizeof(struct tcphdr);
307 #else
308 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
309 #endif
310 	} else if (ifp->if_flags & IFF_LOOPBACK)
311 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
312 #ifdef INET
313 	else if (inp && tp->t_mtudisc)
314 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
315 	else if (inp && in_localaddr(inp->inp_faddr))
316 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
317 #endif
318 #ifdef INET6
319 	else if (in6p) {
320 #ifdef INET
321 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
322 			/* mapped addr case */
323 			struct in_addr d;
324 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
325 			if (tp->t_mtudisc || in_localaddr(d))
326 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
327 		} else
328 #endif
329 		{
330 			/*
331 			 * for IPv6, path MTU discovery is always turned on,
332 			 * or the node must use packet size <= 1280.
333 			 */
334 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
335 			size -= (iphlen + sizeof(struct tcphdr));
336 		}
337 	}
338 #endif
339  out:
340 	/*
341 	 * Now we must make room for whatever extra TCP/IP options are in
342 	 * the packet.
343 	 */
344 	optlen = tcp_optlen(tp);
345 
346 	/*
347 	 * XXX tp->t_ourmss should have the right size, but without this code
348 	 * fragmentation will occur... need more investigation
349 	 */
350 #ifdef INET
351 	if (inp) {
352 #if defined(IPSEC) || defined(FAST_IPSEC)
353 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
354 			optlen += ipsec4_hdrsiz_tcp(tp);
355 #endif
356 		optlen += ip_optlen(inp);
357 	}
358 #endif
359 #ifdef INET6
360 #ifdef INET
361 	if (in6p && tp->t_family == AF_INET) {
362 #if defined(IPSEC) || defined(FAST_IPSEC)
363 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
364 			optlen += ipsec4_hdrsiz_tcp(tp);
365 #endif
366 		/* XXX size -= ip_optlen(in6p); */
367 	} else
368 #endif
369 	if (in6p && tp->t_family == AF_INET6) {
370 #ifdef IPSEC
371 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
372 			optlen += ipsec6_hdrsiz_tcp(tp);
373 #endif
374 		optlen += ip6_optlen(in6p);
375 	}
376 #endif
377 	size -= optlen;
378 
379 	/* there may not be any room for data if mtu is too small */
380 	if (size < 0)
381 		return (EMSGSIZE);
382 
383 	/*
384 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
385 	 * assumes that path MTU is the same for both ways).  this is only
386 	 * for silly window avoidance, do not use the value for other purposes.
387 	 *
388 	 * ipseclen is subtracted from both sides, this may not be right.
389 	 * I'm not quite sure about this (could someone comment).
390 	 */
391 	*txsegsizep = min(tp->t_peermss - optlen, size);
392 	/*
393 	 * Never send more than half a buffer full.  This insures that we can
394 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
395 	 * therefore acks will never be delayed unless we run out of data to
396 	 * transmit.
397 	 */
398 	if (so)
399 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
400 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
401 
402 	if (*txsegsizep != tp->t_segsz) {
403 		/*
404 		 * If the new segment size is larger, we don't want to
405 		 * mess up the congestion window, but if it is smaller
406 		 * we'll have to reduce the congestion window to ensure
407 		 * that we don't get into trouble with initial windows
408 		 * and the rest.  In any case, if the segment size
409 		 * has changed, chances are the path has, too, and
410 		 * our congestion window will be different.
411 		 */
412 		if (*txsegsizep < tp->t_segsz) {
413 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
414 					   * *txsegsizep, *txsegsizep);
415 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
416 						* *txsegsizep, *txsegsizep);
417 		}
418 		tp->t_segsz = *txsegsizep;
419 	}
420 
421 	return (0);
422 }
423 
424 static
425 #ifndef GPROF
426 __inline
427 #endif
428 int
429 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
430     long len, int hdrlen, struct mbuf **mp)
431 {
432 	struct mbuf *m, *m0;
433 
434 	if (tp->t_force && len == 1)
435 		tcpstat.tcps_sndprobe++;
436 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
437 		tcpstat.tcps_sndrexmitpack++;
438 		tcpstat.tcps_sndrexmitbyte += len;
439 	} else {
440 		tcpstat.tcps_sndpack++;
441 		tcpstat.tcps_sndbyte += len;
442 	}
443 #ifdef notyet
444 	if ((m = m_copypack(so->so_snd.sb_mb, off,
445 	    (int)len, max_linkhdr + hdrlen)) == 0)
446 		return (ENOBUFS);
447 	/*
448 	 * m_copypack left space for our hdr; use it.
449 	 */
450 	m->m_len += hdrlen;
451 	m->m_data -= hdrlen;
452 #else
453 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
454 	if (__predict_false(m == NULL))
455 		return (ENOBUFS);
456 	MCLAIM(m, &tcp_tx_mowner);
457 
458 	/*
459 	 * XXX Because other code assumes headers will fit in
460 	 * XXX one header mbuf.
461 	 *
462 	 * (This code should almost *never* be run.)
463 	 */
464 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
465 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
466 		MCLGET(m, M_DONTWAIT);
467 		if ((m->m_flags & M_EXT) == 0) {
468 			m_freem(m);
469 			return (ENOBUFS);
470 		}
471 	}
472 
473 	m->m_data += max_linkhdr;
474 	m->m_len = hdrlen;
475 
476 	/*
477 	 * To avoid traversing the whole sb_mb chain for correct
478 	 * data to send, remember last sent mbuf, its offset and
479 	 * the sent size.  When called the next time, see if the
480 	 * data to send is directly following the previous transfer.
481 	 * This is important for large TCP windows.
482 	 */
483 	if (off == 0 || tp->t_lastm == NULL ||
484 	    (tp->t_lastoff + tp->t_lastlen) != off) {
485 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
486 		/*
487 		 * Either a new packet or a retransmit.
488 		 * Start from the beginning.
489 		 */
490 		tp->t_lastm = so->so_snd.sb_mb;
491 		tp->t_inoff = off;
492 	} else {
493 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
494 		tp->t_inoff += tp->t_lastlen;
495 	}
496 
497 	/* Traverse forward to next packet */
498 	while (tp->t_inoff > 0) {
499 		if (tp->t_lastm == NULL)
500 			panic("tp->t_lastm == NULL");
501 		if (tp->t_inoff < tp->t_lastm->m_len)
502 			break;
503 		tp->t_inoff -= tp->t_lastm->m_len;
504 		tp->t_lastm = tp->t_lastm->m_next;
505 	}
506 
507 	tp->t_lastoff = off;
508 	tp->t_lastlen = len;
509 	m0 = tp->t_lastm;
510 	off = tp->t_inoff;
511 
512 	if (len <= M_TRAILINGSPACE(m)) {
513 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
514 		m->m_len += len;
515 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
516 	} else {
517 		m->m_next = m_copy(m0, off, (int) len);
518 		if (m->m_next == NULL) {
519 			m_freem(m);
520 			return (ENOBUFS);
521 		}
522 #ifdef TCP_OUTPUT_COUNTERS
523 		if (m->m_next->m_flags & M_EXT)
524 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
525 		else
526 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
527 #endif /* TCP_OUTPUT_COUNTERS */
528 	}
529 #endif
530 
531 	*mp = m;
532 	return (0);
533 }
534 
535 /*
536  * Tcp output routine: figure out what should be sent and send it.
537  */
538 int
539 tcp_output(struct tcpcb *tp)
540 {
541 	struct socket *so;
542 	struct route *ro;
543 	long len, win;
544 	int off, flags, error;
545 	struct mbuf *m;
546 	struct ip *ip;
547 #ifdef INET6
548 	struct ip6_hdr *ip6;
549 #endif
550 	struct tcphdr *th;
551 	u_char opt[MAX_TCPOPTLEN];
552 	unsigned optlen, hdrlen;
553 	unsigned int sack_numblks;
554 	int idle, sendalot, txsegsize, rxsegsize;
555 	int txsegsize_nosack;
556 	int maxburst = TCP_MAXBURST;
557 	int af;		/* address family on the wire */
558 	int iphdrlen;
559 	int has_tso, use_tso;
560 	int sack_rxmit;
561 	int sack_bytes_rxmt;
562 	struct sackhole *p;
563 #ifdef TCP_SIGNATURE
564 	int sigoff = 0;
565 #endif
566 
567 #ifdef DIAGNOSTIC
568 	if (tp->t_inpcb && tp->t_in6pcb)
569 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
570 #endif
571 	so = NULL;
572 	ro = NULL;
573 	if (tp->t_inpcb) {
574 		so = tp->t_inpcb->inp_socket;
575 		ro = &tp->t_inpcb->inp_route;
576 	}
577 #ifdef INET6
578 	else if (tp->t_in6pcb) {
579 		so = tp->t_in6pcb->in6p_socket;
580 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
581 	}
582 #endif
583 
584 	switch (af = tp->t_family) {
585 #ifdef INET
586 	case AF_INET:
587 		if (tp->t_inpcb)
588 			break;
589 #ifdef INET6
590 		/* mapped addr case */
591 		if (tp->t_in6pcb)
592 			break;
593 #endif
594 		return (EINVAL);
595 #endif
596 #ifdef INET6
597 	case AF_INET6:
598 		if (tp->t_in6pcb)
599 			break;
600 		return (EINVAL);
601 #endif
602 	default:
603 		return (EAFNOSUPPORT);
604 	}
605 
606 	if (tcp_segsize(tp, &txsegsize, &rxsegsize))
607 		return (EMSGSIZE);
608 
609 	idle = (tp->snd_max == tp->snd_una);
610 
611 	/*
612 	 * Determine if we can use TCP segmentation offload:
613 	 * - If we're using IPv4
614 	 * - If there is not an IPsec policy that prevents it
615 	 * - If the interface can do it
616 	 */
617 	has_tso = tp->t_inpcb != NULL &&
618 #if defined(IPSEC) || defined(FAST_IPSEC)
619 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
620 		  		       IPSEC_DIR_OUTBOUND) &&
621 #endif
622 		  tp->t_inpcb->inp_route.ro_rt != NULL &&
623 		  (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable &
624 		   IFCAP_TSOv4) != 0;
625 
626 	/*
627 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
628 	 *
629 	 *	Optionally, a TCP MAY set the restart window to the
630 	 *	minimum of the value used for the initial window and
631 	 *	the current value of cwnd (in other words, using a
632 	 *	larger value for the restart window should never increase
633 	 *	the size of cwnd).
634 	 */
635 	if (tcp_cwm) {
636 		/*
637 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
638 		 * Count the number of packets currently pending
639 		 * acknowledgement, and limit our congestion window
640 		 * to a pre-determined allowed burst size plus that count.
641 		 * This prevents bursting once all pending packets have
642 		 * been acknowledged (i.e. transmission is idle).
643 		 *
644 		 * XXX Link this to Initial Window?
645 		 */
646 		tp->snd_cwnd = min(tp->snd_cwnd,
647 		    (tcp_cwm_burstsize * txsegsize) +
648 		    (tp->snd_nxt - tp->snd_una));
649 	} else {
650 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
651 			/*
652 			 * We have been idle for "a while" and no acks are
653 			 * expected to clock out any data we send --
654 			 * slow start to get ack "clock" running again.
655 			 */
656 			int ss = tcp_init_win;
657 #ifdef INET
658 			if (tp->t_inpcb &&
659 			    in_localaddr(tp->t_inpcb->inp_faddr))
660 				ss = tcp_init_win_local;
661 #endif
662 #ifdef INET6
663 			if (tp->t_in6pcb &&
664 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
665 				ss = tcp_init_win_local;
666 #endif
667 			tp->snd_cwnd = min(tp->snd_cwnd,
668 			    TCP_INITIAL_WINDOW(ss, txsegsize));
669 		}
670 	}
671 
672 	txsegsize_nosack = txsegsize;
673 again:
674 	use_tso = has_tso;
675 	TCP_REASS_LOCK(tp);
676 	sack_numblks = tcp_sack_numblks(tp);
677 	if (sack_numblks) {
678 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
679 			/* don't duplicate D-SACK. */
680 			use_tso = 0;
681 		}
682 		txsegsize = txsegsize_nosack - TCP_SACK_OPTLEN(sack_numblks);
683 	} else {
684 		txsegsize = txsegsize_nosack;
685 	}
686 
687 	/*
688 	 * Determine length of data that should be transmitted, and
689 	 * flags that should be used.  If there is some data or critical
690 	 * controls (SYN, RST) to send, then transmit; otherwise,
691 	 * investigate further.
692 	 *
693 	 * Readjust SACK information to avoid resending duplicate data.
694 	 */
695 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
696 		tcp_sack_adjust(tp);
697 	sendalot = 0;
698 	off = tp->snd_nxt - tp->snd_una;
699 	win = min(tp->snd_wnd, tp->snd_cwnd);
700 
701 	flags = tcp_outflags[tp->t_state];
702 
703 	/*
704 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
705 	 * to send out new data (when sendalot is 1), bypass this function.
706 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
707 	 * we're replacing a (future) new transmission with a retransmission
708 	 * now, and we previously incremented snd_cwnd in tcp_input().
709 	 */
710 	/*
711 	 * Still in sack recovery , reset rxmit flag to zero.
712 	 */
713 	sack_rxmit = 0;
714 	sack_bytes_rxmt = 0;
715 	len = 0;
716 	p = NULL;
717 	do {
718 		long cwin;
719 		if (!TCP_SACK_ENABLED(tp))
720 			break;
721 		if (tp->t_partialacks < 0)
722 			break;
723 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
724 		if (p == NULL)
725 			break;
726 
727 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
728 		if (cwin < 0)
729 			cwin = 0;
730 		/* Do not retransmit SACK segments beyond snd_recover */
731 		if (SEQ_GT(p->end, tp->snd_recover)) {
732 			/*
733 			 * (At least) part of sack hole extends beyond
734 			 * snd_recover. Check to see if we can rexmit data
735 			 * for this hole.
736 			 */
737 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
738 				/*
739 				 * Can't rexmit any more data for this hole.
740 				 * That data will be rexmitted in the next
741 				 * sack recovery episode, when snd_recover
742 				 * moves past p->rxmit.
743 				 */
744 				p = NULL;
745 				break;
746 			}
747 			/* Can rexmit part of the current hole */
748 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
749 		} else
750 			len = ((long)ulmin(cwin, p->end - p->rxmit));
751 		off = p->rxmit - tp->snd_una;
752 		if (len > 0) {
753 			sack_rxmit = 1;
754 			sendalot = 1;
755 		}
756 	} while (/*CONSTCOND*/0);
757 
758 	/*
759 	 * If in persist timeout with window of 0, send 1 byte.
760 	 * Otherwise, if window is small but nonzero
761 	 * and timer expired, we will send what we can
762 	 * and go to transmit state.
763 	 */
764 	if (tp->t_force) {
765 		if (win == 0) {
766 			/*
767 			 * If we still have some data to send, then
768 			 * clear the FIN bit.  Usually this would
769 			 * happen below when it realizes that we
770 			 * aren't sending all the data.  However,
771 			 * if we have exactly 1 byte of unset data,
772 			 * then it won't clear the FIN bit below,
773 			 * and if we are in persist state, we wind
774 			 * up sending the packet without recording
775 			 * that we sent the FIN bit.
776 			 *
777 			 * We can't just blindly clear the FIN bit,
778 			 * because if we don't have any more data
779 			 * to send then the probe will be the FIN
780 			 * itself.
781 			 */
782 			if (off < so->so_snd.sb_cc)
783 				flags &= ~TH_FIN;
784 			win = 1;
785 		} else {
786 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
787 			tp->t_rxtshift = 0;
788 		}
789 	}
790 
791 	if (!TCP_SACK_ENABLED(tp)) {
792 		if (win < so->so_snd.sb_cc) {
793 			len = win - off;
794 			flags &= ~TH_FIN;
795 		} else
796 			len = so->so_snd.sb_cc - off;
797 	} else if (sack_rxmit == 0) {
798 		if (sack_bytes_rxmt != 0) {
799 			long cwin;
800 
801 			/*
802 			 * We are inside of a SACK recovery episode and are
803 			 * sending new data, having retransmitted all the
804 			 * data possible in the scoreboard.
805 			 */
806 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
807 				       - off);
808 			/*
809 			 * From FreeBSD:
810 			 *  Don't remove this (len > 0) check !
811 			 *  We explicitly check for len > 0 here (although it
812 			 *  isn't really necessary), to work around a gcc
813 			 *  optimization issue - to force gcc to compute
814 			 *  len above. Without this check, the computation
815 			 *  of len is bungled by the optimizer.
816 			 */
817 			if (len > 0) {
818 				cwin = tp->snd_cwnd -
819 						(tp->snd_nxt - tp->sack_newdata) -
820 						sack_bytes_rxmt;
821 				if (cwin < 0)
822 					cwin = 0;
823 				len = lmin(len, cwin);
824 			}
825 		} else if (win < so->so_snd.sb_cc) {
826 			len = win - off;
827 			flags &= ~TH_FIN;
828 		} else
829 			len = so->so_snd.sb_cc - off;
830 	}
831 
832 	if (len < 0) {
833 		/*
834 		 * If FIN has been sent but not acked,
835 		 * but we haven't been called to retransmit,
836 		 * len will be -1.  Otherwise, window shrank
837 		 * after we sent into it.  If window shrank to 0,
838 		 * cancel pending retransmit, pull snd_nxt back
839 		 * to (closed) window, and set the persist timer
840 		 * if it isn't already going.  If the window didn't
841 		 * close completely, just wait for an ACK.
842 		 *
843 		 * If we have a pending FIN, either it has already been
844 		 * transmitted or it is outside the window, so drop it.
845 		 * If the FIN has been transmitted, but this is not a
846 		 * retransmission, then len must be -1.  Therefore we also
847 		 * prevent here the sending of `gratuitous FINs'.  This
848 		 * eliminates the need to check for that case below (e.g.
849 		 * to back up snd_nxt before the FIN so that the sequence
850 		 * number is correct).
851 		 */
852 		len = 0;
853 		flags &= ~TH_FIN;
854 		if (win == 0) {
855 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
856 			tp->t_rxtshift = 0;
857 			tp->snd_nxt = tp->snd_una;
858 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
859 				tcp_setpersist(tp);
860 		}
861 	}
862 	if (len > txsegsize) {
863 		if (use_tso) {
864 			/*
865 			 * Truncate TSO transfers to IP_MAXPACKET, and make
866 			 * sure that we send equal size transfers down the
867 			 * stack (rather than big-small-big-small-...).
868 			 */
869 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
870 			if (len <= txsegsize) {
871 				use_tso = 0;
872 			}
873 		} else
874 			len = txsegsize;
875 		flags &= ~TH_FIN;
876 		sendalot = 1;
877 	} else
878 		use_tso = 0;
879 	if (sack_rxmit) {
880 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
881 			flags &= ~TH_FIN;
882 	}
883 
884 	win = sbspace(&so->so_rcv);
885 
886 	/*
887 	 * Sender silly window avoidance.  If connection is idle
888 	 * and can send all data, a maximum segment,
889 	 * at least a maximum default-size segment do it,
890 	 * or are forced, do it; otherwise don't bother.
891 	 * If peer's buffer is tiny, then send
892 	 * when window is at least half open.
893 	 * If retransmitting (possibly after persist timer forced us
894 	 * to send into a small window), then must resend.
895 	 */
896 	if (len) {
897 		if (len >= txsegsize)
898 			goto send;
899 		if ((so->so_state & SS_MORETOCOME) == 0 &&
900 		    ((idle || tp->t_flags & TF_NODELAY) &&
901 		     len + off >= so->so_snd.sb_cc))
902 			goto send;
903 		if (tp->t_force)
904 			goto send;
905 		if (len >= tp->max_sndwnd / 2)
906 			goto send;
907 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
908 			goto send;
909 		if (sack_rxmit)
910 			goto send;
911 	}
912 
913 	/*
914 	 * Compare available window to amount of window known to peer
915 	 * (as advertised window less next expected input).  If the
916 	 * difference is at least twice the size of the largest segment
917 	 * we expect to receive (i.e. two segments) or at least 50% of
918 	 * the maximum possible window, then want to send a window update
919 	 * to peer.
920 	 */
921 	if (win > 0) {
922 		/*
923 		 * "adv" is the amount we can increase the window,
924 		 * taking into account that we are limited by
925 		 * TCP_MAXWIN << tp->rcv_scale.
926 		 */
927 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
928 			(tp->rcv_adv - tp->rcv_nxt);
929 
930 		if (adv >= (long) (2 * rxsegsize))
931 			goto send;
932 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
933 			goto send;
934 	}
935 
936 	/*
937 	 * Send if we owe peer an ACK.
938 	 */
939 	if (tp->t_flags & TF_ACKNOW)
940 		goto send;
941 	if (flags & (TH_SYN|TH_FIN|TH_RST))
942 		goto send;
943 	if (SEQ_GT(tp->snd_up, tp->snd_una))
944 		goto send;
945 	/*
946 	 * In SACK, it is possible for tcp_output to fail to send a segment
947 	 * after the retransmission timer has been turned off.  Make sure
948 	 * that the retransmission timer is set.
949 	 */
950 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
951 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
952 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
953 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
954 		goto just_return;
955 	}
956 
957 	/*
958 	 * TCP window updates are not reliable, rather a polling protocol
959 	 * using ``persist'' packets is used to insure receipt of window
960 	 * updates.  The three ``states'' for the output side are:
961 	 *	idle			not doing retransmits or persists
962 	 *	persisting		to move a small or zero window
963 	 *	(re)transmitting	and thereby not persisting
964 	 *
965 	 * tp->t_timer[TCPT_PERSIST]
966 	 *	is set when we are in persist state.
967 	 * tp->t_force
968 	 *	is set when we are called to send a persist packet.
969 	 * tp->t_timer[TCPT_REXMT]
970 	 *	is set when we are retransmitting
971 	 * The output side is idle when both timers are zero.
972 	 *
973 	 * If send window is too small, there is data to transmit, and no
974 	 * retransmit or persist is pending, then go to persist state.
975 	 * If nothing happens soon, send when timer expires:
976 	 * if window is nonzero, transmit what we can,
977 	 * otherwise force out a byte.
978 	 */
979 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
980 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
981 		tp->t_rxtshift = 0;
982 		tcp_setpersist(tp);
983 	}
984 
985 	/*
986 	 * No reason to send a segment, just return.
987 	 */
988 just_return:
989 	TCP_REASS_UNLOCK(tp);
990 	return (0);
991 
992 send:
993 	/*
994 	 * Before ESTABLISHED, force sending of initial options
995 	 * unless TCP set not to do any options.
996 	 * NOTE: we assume that the IP/TCP header plus TCP options
997 	 * always fit in a single mbuf, leaving room for a maximum
998 	 * link header, i.e.
999 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1000 	 */
1001 	optlen = 0;
1002 	switch (af) {
1003 #ifdef INET
1004 	case AF_INET:
1005 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1006 		break;
1007 #endif
1008 #ifdef INET6
1009 	case AF_INET6:
1010 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1011 		break;
1012 #endif
1013 	default:	/*pacify gcc*/
1014 		iphdrlen = 0;
1015 		break;
1016 	}
1017 	hdrlen = iphdrlen;
1018 	if (flags & TH_SYN) {
1019 		struct rtentry *rt;
1020 
1021 		rt = NULL;
1022 #ifdef INET
1023 		if (tp->t_inpcb)
1024 			rt = in_pcbrtentry(tp->t_inpcb);
1025 #endif
1026 #ifdef INET6
1027 		if (tp->t_in6pcb)
1028 			rt = in6_pcbrtentry(tp->t_in6pcb);
1029 #endif
1030 
1031 		tp->snd_nxt = tp->iss;
1032 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
1033 						    rt->rt_ifp : NULL, af);
1034 		if ((tp->t_flags & TF_NOOPT) == 0) {
1035 			opt[0] = TCPOPT_MAXSEG;
1036 			opt[1] = 4;
1037 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1038 			opt[3] = tp->t_ourmss & 0xff;
1039 			optlen = 4;
1040 
1041 			if ((tp->t_flags & TF_REQ_SCALE) &&
1042 			    ((flags & TH_ACK) == 0 ||
1043 			    (tp->t_flags & TF_RCVD_SCALE))) {
1044 				*((u_int32_t *) (opt + optlen)) = htonl(
1045 					TCPOPT_NOP << 24 |
1046 					TCPOPT_WINDOW << 16 |
1047 					TCPOLEN_WINDOW << 8 |
1048 					tp->request_r_scale);
1049 				optlen += 4;
1050 			}
1051 			if (tcp_do_sack) {
1052 				u_int8_t *p = (u_int8_t *)(opt + optlen);
1053 
1054 				p[0] = TCPOPT_SACK_PERMITTED;
1055 				p[1] = 2;
1056 				p[2] = TCPOPT_NOP;
1057 				p[3] = TCPOPT_NOP;
1058 				optlen += 4;
1059 			}
1060 		}
1061 	}
1062 
1063 	/*
1064 	 * Send a timestamp and echo-reply if this is a SYN and our side
1065 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1066 	 * and our peer have sent timestamps in our SYN's.
1067 	 */
1068 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1069 	     (flags & TH_RST) == 0 &&
1070 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1071 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1072 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1073 
1074 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1075 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1076 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1077 		*lp   = htonl(tp->ts_recent);
1078 		optlen += TCPOLEN_TSTAMP_APPA;
1079 	}
1080 
1081 	/*
1082 	 * Tack on the SACK block if it is necessary.
1083 	 */
1084 	if (sack_numblks) {
1085 		int sack_len;
1086 		u_char *bp = (u_char *)(opt + optlen);
1087 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1088 		struct ipqent *tiqe;
1089 
1090 		sack_len = sack_numblks * 8 + 2;
1091 		bp[0] = TCPOPT_NOP;
1092 		bp[1] = TCPOPT_NOP;
1093 		bp[2] = TCPOPT_SACK;
1094 		bp[3] = sack_len;
1095 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1096 			sack_numblks--;
1097 			*lp++ = htonl(tp->rcv_dsack_block.left);
1098 			*lp++ = htonl(tp->rcv_dsack_block.right);
1099 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1100 		}
1101 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1102 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1103 			KASSERT(tiqe != NULL);
1104 			sack_numblks--;
1105 			*lp++ = htonl(tiqe->ipqe_seq);
1106 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len);
1107 		}
1108 		optlen += sack_len + 2;
1109 	}
1110 	TCP_REASS_UNLOCK(tp);
1111 
1112 #ifdef TCP_SIGNATURE
1113 #if defined(INET6) && defined(FAST_IPSEC)
1114 	if (tp->t_family == AF_INET)
1115 #endif
1116 	if (tp->t_flags & TF_SIGNATURE) {
1117 		u_char *bp;
1118 		/*
1119 		 * Initialize TCP-MD5 option (RFC2385)
1120 		 */
1121 		bp = (u_char *)opt + optlen;
1122 		*bp++ = TCPOPT_SIGNATURE;
1123 		*bp++ = TCPOLEN_SIGNATURE;
1124 		sigoff = optlen + 2;
1125 		bzero(bp, TCP_SIGLEN);
1126 		bp += TCP_SIGLEN;
1127 		optlen += TCPOLEN_SIGNATURE;
1128 		/*
1129 		 * Terminate options list and maintain 32-bit alignment.
1130  		 */
1131 		*bp++ = TCPOPT_NOP;
1132 		*bp++ = TCPOPT_EOL;
1133  		optlen += 2;
1134  	}
1135 #endif /* TCP_SIGNATURE */
1136 
1137 	hdrlen += optlen;
1138 
1139 #ifdef DIAGNOSTIC
1140 	if (!use_tso && len > txsegsize)
1141 		panic("tcp data to be sent is larger than segment");
1142 	else if (use_tso && len > IP_MAXPACKET)
1143 		panic("tcp data to be sent is larger than max TSO size");
1144 	if (max_linkhdr + hdrlen > MCLBYTES)
1145 		panic("tcphdr too big");
1146 #endif
1147 
1148 	/*
1149 	 * Grab a header mbuf, attaching a copy of data to
1150 	 * be transmitted, and initialize the header from
1151 	 * the template for sends on this connection.
1152 	 */
1153 	if (len) {
1154 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1155 		if (error)
1156 			goto out;
1157 		/*
1158 		 * If we're sending everything we've got, set PUSH.
1159 		 * (This will keep happy those implementations which only
1160 		 * give data to the user when a buffer fills or
1161 		 * a PUSH comes in.)
1162 		 */
1163 		if (off + len == so->so_snd.sb_cc)
1164 			flags |= TH_PUSH;
1165 	} else {
1166 		if (tp->t_flags & TF_ACKNOW)
1167 			tcpstat.tcps_sndacks++;
1168 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1169 			tcpstat.tcps_sndctrl++;
1170 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1171 			tcpstat.tcps_sndurg++;
1172 		else
1173 			tcpstat.tcps_sndwinup++;
1174 
1175 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1176 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1177 			MCLGET(m, M_DONTWAIT);
1178 			if ((m->m_flags & M_EXT) == 0) {
1179 				m_freem(m);
1180 				m = NULL;
1181 			}
1182 		}
1183 		if (m == NULL) {
1184 			error = ENOBUFS;
1185 			goto out;
1186 		}
1187 		MCLAIM(m, &tcp_tx_mowner);
1188 		m->m_data += max_linkhdr;
1189 		m->m_len = hdrlen;
1190 	}
1191 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1192 	switch (af) {
1193 #ifdef INET
1194 	case AF_INET:
1195 		ip = mtod(m, struct ip *);
1196 #ifdef INET6
1197 		ip6 = NULL;
1198 #endif
1199 		th = (struct tcphdr *)(ip + 1);
1200 		break;
1201 #endif
1202 #ifdef INET6
1203 	case AF_INET6:
1204 		ip = NULL;
1205 		ip6 = mtod(m, struct ip6_hdr *);
1206 		th = (struct tcphdr *)(ip6 + 1);
1207 		break;
1208 #endif
1209 	default:	/*pacify gcc*/
1210 		ip = NULL;
1211 #ifdef INET6
1212 		ip6 = NULL;
1213 #endif
1214 		th = NULL;
1215 		break;
1216 	}
1217 	if (tp->t_template == 0)
1218 		panic("tcp_output");
1219 	if (tp->t_template->m_len < iphdrlen)
1220 		panic("tcp_output");
1221 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1222 
1223 	/*
1224 	 * If we are doing retransmissions, then snd_nxt will
1225 	 * not reflect the first unsent octet.  For ACK only
1226 	 * packets, we do not want the sequence number of the
1227 	 * retransmitted packet, we want the sequence number
1228 	 * of the next unsent octet.  So, if there is no data
1229 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1230 	 * when filling in ti_seq.  But if we are in persist
1231 	 * state, snd_max might reflect one byte beyond the
1232 	 * right edge of the window, so use snd_nxt in that
1233 	 * case, since we know we aren't doing a retransmission.
1234 	 * (retransmit and persist are mutually exclusive...)
1235 	 */
1236 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1237 		th->th_seq = htonl(p->rxmit);
1238 		p->rxmit += len;
1239 	} else {
1240 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1241 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1242 			th->th_seq = htonl(tp->snd_nxt);
1243 		else
1244 			th->th_seq = htonl(tp->snd_max);
1245 	}
1246 	th->th_ack = htonl(tp->rcv_nxt);
1247 	if (optlen) {
1248 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1249 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1250 	}
1251 	th->th_flags = flags;
1252 	/*
1253 	 * Calculate receive window.  Don't shrink window,
1254 	 * but avoid silly window syndrome.
1255 	 */
1256 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1257 		win = 0;
1258 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1259 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1260 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1261 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1262 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1263 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1264 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1265 		if (urp > IP_MAXPACKET)
1266 			urp = IP_MAXPACKET;
1267 		th->th_urp = htons((u_int16_t)urp);
1268 		th->th_flags |= TH_URG;
1269 	} else
1270 		/*
1271 		 * If no urgent pointer to send, then we pull
1272 		 * the urgent pointer to the left edge of the send window
1273 		 * so that it doesn't drift into the send window on sequence
1274 		 * number wraparound.
1275 		 */
1276 		tp->snd_up = tp->snd_una;		/* drag it along */
1277 
1278 #ifdef TCP_SIGNATURE
1279 #if defined(INET6) && defined(FAST_IPSEC)
1280 	if (tp->t_family == AF_INET) /* XXX */
1281 #endif
1282 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1283 		struct secasvar *sav;
1284 		u_int8_t *sigp;
1285 
1286 		sav = tcp_signature_getsav(m, th);
1287 
1288 		if (sav == NULL) {
1289 			if (m)
1290 				m_freem(m);
1291 			return (EPERM);
1292 		}
1293 
1294 		m->m_pkthdr.len = hdrlen + len;
1295 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1296 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1297 
1298 		key_sa_recordxfer(sav, m);
1299 #ifdef FAST_IPSEC
1300 		KEY_FREESAV(&sav);
1301 #else
1302 		key_freesav(sav);
1303 #endif
1304 	}
1305 #endif
1306 
1307 	/*
1308 	 * Set ourselves up to be checksummed just before the packet
1309 	 * hits the wire.  Maybe skip checksums on loopback interfaces.
1310 	 */
1311 	switch (af) {
1312 #ifdef INET
1313 	case AF_INET:
1314 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1315 		if (use_tso) {
1316 			m->m_pkthdr.segsz = txsegsize;
1317 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1318 		} else {
1319 			if (__predict_true(ro->ro_rt == NULL ||
1320 					   !(ro->ro_rt->rt_ifp->if_flags &
1321 					     IFF_LOOPBACK) ||
1322 					   tcp_do_loopback_cksum))
1323 				m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1324 			else
1325 				m->m_pkthdr.csum_flags = 0;
1326 			if (len + optlen) {
1327 				/* Fixup the pseudo-header checksum. */
1328 				/* XXXJRT Not IP Jumbogram safe. */
1329 				th->th_sum = in_cksum_addword(th->th_sum,
1330 				    htons((u_int16_t) (len + optlen)));
1331 			}
1332 		}
1333 		break;
1334 #endif
1335 #ifdef INET6
1336 	case AF_INET6:
1337 		/*
1338 		 * XXX Actually delaying the checksum is Hard
1339 		 * XXX (well, maybe not for Itojun, but it is
1340 		 * XXX for me), but we can still take advantage
1341 		 * XXX of the cached pseudo-header checksum.
1342 		 */
1343 		/* equals to hdrlen + len */
1344 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
1345 			+ sizeof(struct tcphdr) + optlen + len;
1346 #ifdef notyet
1347 		if (__predict_true(ro->ro_rt == NULL ||
1348 				   !(ro->ro_rt->rt_ifp->if_flags &
1349 				     IFF_LOOPBACK) ||
1350 				   tcp_do_loopback_cksum))
1351 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1352 		else
1353 			m->m_pkthdr.csum_flags = 0;
1354 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1355 #endif
1356 		if (len + optlen) {
1357 			/* Fixup the pseudo-header checksum. */
1358 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1359 			th->th_sum = in_cksum_addword(th->th_sum,
1360 			    htons((u_int16_t) (len + optlen)));
1361 		}
1362 #ifndef notyet
1363 		if (__predict_true(ro->ro_rt == NULL ||
1364 				   !(ro->ro_rt->rt_ifp->if_flags &
1365 				     IFF_LOOPBACK) ||
1366 				   tcp_do_loopback_cksum))
1367 			th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
1368 			    sizeof(struct tcphdr) + optlen + len);
1369 #endif
1370 		break;
1371 #endif
1372 	}
1373 
1374 	/*
1375 	 * In transmit state, time the transmission and arrange for
1376 	 * the retransmit.  In persist state, just set snd_max.
1377 	 */
1378 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1379 		tcp_seq startseq = tp->snd_nxt;
1380 
1381 		/*
1382 		 * Advance snd_nxt over sequence space of this segment.
1383 		 * There are no states in which we send both a SYN and a FIN,
1384 		 * so we collapse the tests for these flags.
1385 		 */
1386 		if (flags & (TH_SYN|TH_FIN))
1387 			tp->snd_nxt++;
1388 		if (sack_rxmit)
1389 			goto timer;
1390 		tp->snd_nxt += len;
1391 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1392 			tp->snd_max = tp->snd_nxt;
1393 			/*
1394 			 * Time this transmission if not a retransmission and
1395 			 * not currently timing anything.
1396 			 */
1397 			if (tp->t_rtttime == 0) {
1398 				tp->t_rtttime = tcp_now;
1399 				tp->t_rtseq = startseq;
1400 				tcpstat.tcps_segstimed++;
1401 			}
1402 		}
1403 
1404 		/*
1405 		 * Set retransmit timer if not currently set,
1406 		 * and not doing an ack or a keep-alive probe.
1407 		 * Initial value for retransmit timer is smoothed
1408 		 * round-trip time + 2 * round-trip time variance.
1409 		 * Initialize shift counter which is used for backoff
1410 		 * of retransmit time.
1411 		 */
1412 timer:
1413 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1414 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1415 		    tp->snd_nxt != tp->snd_una)) {
1416 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1417 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1418 				tp->t_rxtshift = 0;
1419 			}
1420 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1421 		}
1422 	} else
1423 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1424 			tp->snd_max = tp->snd_nxt + len;
1425 
1426 #ifdef TCP_DEBUG
1427 	/*
1428 	 * Trace.
1429 	 */
1430 	if (so->so_options & SO_DEBUG)
1431 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1432 #endif
1433 
1434 	/*
1435 	 * Fill in IP length and desired time to live and
1436 	 * send to IP level.  There should be a better way
1437 	 * to handle ttl and tos; we could keep them in
1438 	 * the template, but need a way to checksum without them.
1439 	 */
1440 	m->m_pkthdr.len = hdrlen + len;
1441 
1442 	switch (af) {
1443 #ifdef INET
1444 	case AF_INET:
1445 		ip->ip_len = htons(m->m_pkthdr.len);
1446 		if (tp->t_inpcb) {
1447 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1448 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1449 		}
1450 #ifdef INET6
1451 		else if (tp->t_in6pcb) {
1452 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1453 			ip->ip_tos = 0;	/*XXX*/
1454 		}
1455 #endif
1456 		break;
1457 #endif
1458 #ifdef INET6
1459 	case AF_INET6:
1460 		ip6->ip6_nxt = IPPROTO_TCP;
1461 		if (tp->t_in6pcb) {
1462 			/*
1463 			 * we separately set hoplimit for every segment, since
1464 			 * the user might want to change the value via
1465 			 * setsockopt. Also, desired default hop limit might
1466 			 * be changed via Neighbor Discovery.
1467 			 */
1468 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1469 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1470 		}
1471 		/* ip6->ip6_flow = ??? */
1472 		/* ip6_plen will be filled in ip6_output(). */
1473 		break;
1474 #endif
1475 	}
1476 
1477 	switch (af) {
1478 #ifdef INET
1479 	case AF_INET:
1480 	    {
1481 		struct mbuf *opts;
1482 
1483 		if (tp->t_inpcb)
1484 			opts = tp->t_inpcb->inp_options;
1485 		else
1486 			opts = NULL;
1487 		error = ip_output(m, opts, ro,
1488 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1489 			(so->so_options & SO_DONTROUTE),
1490 			(struct ip_moptions *)0, so);
1491 		break;
1492 	    }
1493 #endif
1494 #ifdef INET6
1495 	case AF_INET6:
1496 	    {
1497 		struct ip6_pktopts *opts;
1498 
1499 		if (tp->t_in6pcb)
1500 			opts = tp->t_in6pcb->in6p_outputopts;
1501 		else
1502 			opts = NULL;
1503 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1504 			so->so_options & SO_DONTROUTE,
1505 			(struct ip6_moptions *)0, so, NULL);
1506 		break;
1507 	    }
1508 #endif
1509 	default:
1510 		error = EAFNOSUPPORT;
1511 		break;
1512 	}
1513 	if (error) {
1514 out:
1515 		if (error == ENOBUFS) {
1516 			tcpstat.tcps_selfquench++;
1517 #ifdef INET
1518 			if (tp->t_inpcb)
1519 				tcp_quench(tp->t_inpcb, 0);
1520 #endif
1521 #ifdef INET6
1522 			if (tp->t_in6pcb)
1523 				tcp6_quench(tp->t_in6pcb, 0);
1524 #endif
1525 			error = 0;
1526 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1527 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1528 			tp->t_softerror = error;
1529 			error = 0;
1530 		}
1531 
1532 		/* Back out the seqence number advance. */
1533 		if (sack_rxmit)
1534 			p->rxmit -= len;
1535 
1536 		/* Restart the delayed ACK timer, if necessary. */
1537 		if (tp->t_flags & TF_DELACK)
1538 			TCP_RESTART_DELACK(tp);
1539 
1540 		return (error);
1541 	}
1542 	tcpstat.tcps_sndtotal++;
1543 	if (tp->t_flags & TF_DELACK)
1544 		tcpstat.tcps_delack++;
1545 
1546 	/*
1547 	 * Data sent (as far as we can tell).
1548 	 * If this advertises a larger window than any other segment,
1549 	 * then remember the size of the advertised window.
1550 	 * Any pending ACK has now been sent.
1551 	 */
1552 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1553 		tp->rcv_adv = tp->rcv_nxt + win;
1554 	tp->last_ack_sent = tp->rcv_nxt;
1555 	tp->t_flags &= ~TF_ACKNOW;
1556 	TCP_CLEAR_DELACK(tp);
1557 #ifdef DIAGNOSTIC
1558 	if (maxburst < 0)
1559 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1560 #endif
1561 	if (sendalot && (!tcp_do_newreno || --maxburst))
1562 		goto again;
1563 	return (0);
1564 }
1565 
1566 void
1567 tcp_setpersist(struct tcpcb *tp)
1568 {
1569 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1570 	int nticks;
1571 
1572 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1573 		panic("tcp_output REXMT");
1574 	/*
1575 	 * Start/restart persistance timer.
1576 	 */
1577 	if (t < tp->t_rttmin)
1578 		t = tp->t_rttmin;
1579 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1580 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1581 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1582 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1583 		tp->t_rxtshift++;
1584 }
1585