xref: /netbsd-src/sys/netinet/tcp_output.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: tcp_output.c,v 1.162 2007/09/02 03:12:23 dyoung 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 1.162 2007/09/02 03:12:23 dyoung 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 #ifdef INET6
189 #include <netipsec/ipsec6.h>
190 #endif
191 #endif	/* FAST_IPSEC*/
192 #ifdef IPSEC
193 #include <netinet6/ipsec.h>
194 #endif
195 
196 #include <netinet/tcp.h>
197 #define	TCPOUTFLAGS
198 #include <netinet/tcp_fsm.h>
199 #include <netinet/tcp_seq.h>
200 #include <netinet/tcp_timer.h>
201 #include <netinet/tcp_var.h>
202 #include <netinet/tcp_congctl.h>
203 #include <netinet/tcpip.h>
204 #include <netinet/tcp_debug.h>
205 #include <netinet/in_offload.h>
206 #include <netinet6/in6_offload.h>
207 
208 #ifdef IPSEC
209 #include <netkey/key.h>
210 #endif
211 
212 #ifdef notyet
213 extern struct mbuf *m_copypack();
214 #endif
215 
216 /*
217  * Knob to enable Congestion Window Monitoring, and control
218  * the burst size it allows.  Default burst is 4 packets, per
219  * the Internet draft.
220  */
221 int	tcp_cwm = 0;
222 int	tcp_cwm_burstsize = 4;
223 
224 int	tcp_do_autosndbuf = 0;
225 int	tcp_autosndbuf_inc = 8 * 1024;
226 int	tcp_autosndbuf_max = 256 * 1024;
227 
228 #ifdef TCP_OUTPUT_COUNTERS
229 #include <sys/device.h>
230 
231 extern struct evcnt tcp_output_bigheader;
232 extern struct evcnt tcp_output_predict_hit;
233 extern struct evcnt tcp_output_predict_miss;
234 extern struct evcnt tcp_output_copysmall;
235 extern struct evcnt tcp_output_copybig;
236 extern struct evcnt tcp_output_refbig;
237 
238 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
239 #else
240 
241 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
242 
243 #endif /* TCP_OUTPUT_COUNTERS */
244 
245 static
246 #ifndef GPROF
247 inline
248 #endif
249 int
250 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
251     bool *alwaysfragp)
252 {
253 #ifdef INET
254 	struct inpcb *inp = tp->t_inpcb;
255 #endif
256 #ifdef INET6
257 	struct in6pcb *in6p = tp->t_in6pcb;
258 #endif
259 	struct socket *so = NULL;
260 	struct rtentry *rt;
261 	struct ifnet *ifp;
262 	int size;
263 	int hdrlen;
264 	int optlen;
265 
266 	*alwaysfragp = false;
267 
268 #ifdef DIAGNOSTIC
269 	if (tp->t_inpcb && tp->t_in6pcb)
270 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
271 #endif
272 	switch (tp->t_family) {
273 #ifdef INET
274 	case AF_INET:
275 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
276 		break;
277 #endif
278 #ifdef INET6
279 	case AF_INET6:
280 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
281 		break;
282 #endif
283 	default:
284 		size = tcp_mssdflt;
285 		goto out;
286 	}
287 
288 	rt = NULL;
289 #ifdef INET
290 	if (inp) {
291 		rt = in_pcbrtentry(inp);
292 		so = inp->inp_socket;
293 	}
294 #endif
295 #ifdef INET6
296 	if (in6p) {
297 		rt = in6_pcbrtentry(in6p);
298 		so = in6p->in6p_socket;
299 	}
300 #endif
301 	if (rt == NULL) {
302 		size = tcp_mssdflt;
303 		goto out;
304 	}
305 
306 	ifp = rt->rt_ifp;
307 
308 	size = tcp_mssdflt;
309 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
310 #ifdef INET6
311 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
312 			/*
313 			 * RFC2460 section 5, last paragraph: if path MTU is
314 			 * smaller than 1280, use 1280 as packet size and
315 			 * attach fragment header.
316 			 */
317 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
318 			*alwaysfragp = true;
319 		} else
320 			size = rt->rt_rmx.rmx_mtu - hdrlen;
321 #else
322 		size = rt->rt_rmx.rmx_mtu - hdrlen;
323 #endif
324 	} else if (ifp->if_flags & IFF_LOOPBACK)
325 		size = ifp->if_mtu - hdrlen;
326 #ifdef INET
327 	else if (inp && tp->t_mtudisc)
328 		size = ifp->if_mtu - hdrlen;
329 	else if (inp && in_localaddr(inp->inp_faddr))
330 		size = ifp->if_mtu - hdrlen;
331 #endif
332 #ifdef INET6
333 	else if (in6p) {
334 #ifdef INET
335 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
336 			/* mapped addr case */
337 			struct in_addr d;
338 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
339 			if (tp->t_mtudisc || in_localaddr(d))
340 				size = ifp->if_mtu - hdrlen;
341 		} else
342 #endif
343 		{
344 			/*
345 			 * for IPv6, path MTU discovery is always turned on,
346 			 * or the node must use packet size <= 1280.
347 			 */
348 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
349 			size -= hdrlen;
350 		}
351 	}
352 #endif
353  out:
354 	/*
355 	 * Now we must make room for whatever extra TCP/IP options are in
356 	 * the packet.
357 	 */
358 	optlen = tcp_optlen(tp);
359 
360 	/*
361 	 * XXX tp->t_ourmss should have the right size, but without this code
362 	 * fragmentation will occur... need more investigation
363 	 */
364 #ifdef INET
365 	if (inp) {
366 #if defined(IPSEC) || defined(FAST_IPSEC)
367 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
368 			optlen += ipsec4_hdrsiz_tcp(tp);
369 #endif
370 		optlen += ip_optlen(inp);
371 	}
372 #endif
373 #ifdef INET6
374 #ifdef INET
375 	if (in6p && tp->t_family == AF_INET) {
376 #if defined(IPSEC) || defined(FAST_IPSEC)
377 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
378 			optlen += ipsec4_hdrsiz_tcp(tp);
379 #endif
380 		/* XXX size -= ip_optlen(in6p); */
381 	} else
382 #endif
383 	if (in6p && tp->t_family == AF_INET6) {
384 #if defined(IPSEC) || defined(FAST_IPSEC)
385 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
386 			optlen += ipsec6_hdrsiz_tcp(tp);
387 #endif
388 		optlen += ip6_optlen(in6p);
389 	}
390 #endif
391 	size -= optlen;
392 
393 	/* there may not be any room for data if mtu is too small */
394 	if (size < 0)
395 		return (EMSGSIZE);
396 
397 	/*
398 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
399 	 * assumes that path MTU is the same for both ways).  this is only
400 	 * for silly window avoidance, do not use the value for other purposes.
401 	 *
402 	 * ipseclen is subtracted from both sides, this may not be right.
403 	 * I'm not quite sure about this (could someone comment).
404 	 */
405 	*txsegsizep = min(tp->t_peermss - optlen, size);
406 	/*
407 	 * Never send more than half a buffer full.  This insures that we can
408 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
409 	 * therefore acks will never be delayed unless we run out of data to
410 	 * transmit.
411 	 */
412 	if (so)
413 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
414 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
415 
416 	if (*txsegsizep != tp->t_segsz) {
417 		/*
418 		 * If the new segment size is larger, we don't want to
419 		 * mess up the congestion window, but if it is smaller
420 		 * we'll have to reduce the congestion window to ensure
421 		 * that we don't get into trouble with initial windows
422 		 * and the rest.  In any case, if the segment size
423 		 * has changed, chances are the path has, too, and
424 		 * our congestion window will be different.
425 		 */
426 		if (*txsegsizep < tp->t_segsz) {
427 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
428 					   * *txsegsizep, *txsegsizep);
429 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
430 						* *txsegsizep, *txsegsizep);
431 		}
432 		tp->t_segsz = *txsegsizep;
433 	}
434 
435 	return (0);
436 }
437 
438 static
439 #ifndef GPROF
440 inline
441 #endif
442 int
443 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
444     long len, int hdrlen, struct mbuf **mp)
445 {
446 	struct mbuf *m, *m0;
447 
448 	if (tp->t_force && len == 1)
449 		tcpstat.tcps_sndprobe++;
450 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
451 		tcpstat.tcps_sndrexmitpack++;
452 		tcpstat.tcps_sndrexmitbyte += len;
453 	} else {
454 		tcpstat.tcps_sndpack++;
455 		tcpstat.tcps_sndbyte += len;
456 	}
457 #ifdef notyet
458 	if ((m = m_copypack(so->so_snd.sb_mb, off,
459 	    (int)len, max_linkhdr + hdrlen)) == 0)
460 		return (ENOBUFS);
461 	/*
462 	 * m_copypack left space for our hdr; use it.
463 	 */
464 	m->m_len += hdrlen;
465 	m->m_data -= hdrlen;
466 #else
467 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
468 	if (__predict_false(m == NULL))
469 		return (ENOBUFS);
470 	MCLAIM(m, &tcp_tx_mowner);
471 
472 	/*
473 	 * XXX Because other code assumes headers will fit in
474 	 * XXX one header mbuf.
475 	 *
476 	 * (This code should almost *never* be run.)
477 	 */
478 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
479 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
480 		MCLGET(m, M_DONTWAIT);
481 		if ((m->m_flags & M_EXT) == 0) {
482 			m_freem(m);
483 			return (ENOBUFS);
484 		}
485 	}
486 
487 	m->m_data += max_linkhdr;
488 	m->m_len = hdrlen;
489 
490 	/*
491 	 * To avoid traversing the whole sb_mb chain for correct
492 	 * data to send, remember last sent mbuf, its offset and
493 	 * the sent size.  When called the next time, see if the
494 	 * data to send is directly following the previous transfer.
495 	 * This is important for large TCP windows.
496 	 */
497 	if (off == 0 || tp->t_lastm == NULL ||
498 	    (tp->t_lastoff + tp->t_lastlen) != off) {
499 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
500 		/*
501 		 * Either a new packet or a retransmit.
502 		 * Start from the beginning.
503 		 */
504 		tp->t_lastm = so->so_snd.sb_mb;
505 		tp->t_inoff = off;
506 	} else {
507 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
508 		tp->t_inoff += tp->t_lastlen;
509 	}
510 
511 	/* Traverse forward to next packet */
512 	while (tp->t_inoff > 0) {
513 		if (tp->t_lastm == NULL)
514 			panic("tp->t_lastm == NULL");
515 		if (tp->t_inoff < tp->t_lastm->m_len)
516 			break;
517 		tp->t_inoff -= tp->t_lastm->m_len;
518 		tp->t_lastm = tp->t_lastm->m_next;
519 	}
520 
521 	tp->t_lastoff = off;
522 	tp->t_lastlen = len;
523 	m0 = tp->t_lastm;
524 	off = tp->t_inoff;
525 
526 	if (len <= M_TRAILINGSPACE(m)) {
527 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
528 		m->m_len += len;
529 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
530 	} else {
531 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
532 		if (m->m_next == NULL) {
533 			m_freem(m);
534 			return (ENOBUFS);
535 		}
536 #ifdef TCP_OUTPUT_COUNTERS
537 		if (m->m_next->m_flags & M_EXT)
538 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
539 		else
540 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
541 #endif /* TCP_OUTPUT_COUNTERS */
542 	}
543 #endif
544 
545 	*mp = m;
546 	return (0);
547 }
548 
549 /*
550  * Tcp output routine: figure out what should be sent and send it.
551  */
552 int
553 tcp_output(struct tcpcb *tp)
554 {
555 	struct socket *so;
556 	struct route *ro;
557 	long len, win;
558 	int off, flags, error;
559 	struct mbuf *m;
560 	struct ip *ip;
561 #ifdef INET6
562 	struct ip6_hdr *ip6;
563 #endif
564 	struct tcphdr *th;
565 	u_char opt[MAX_TCPOPTLEN];
566 	unsigned optlen, hdrlen, packetlen;
567 	unsigned int sack_numblks;
568 	int idle, sendalot, txsegsize, rxsegsize;
569 	int txsegsize_nosack;
570 	int maxburst = TCP_MAXBURST;
571 	int af;		/* address family on the wire */
572 	int iphdrlen;
573 	int has_tso4, has_tso6;
574 	int has_tso, use_tso;
575 	bool alwaysfrag;
576 	int sack_rxmit;
577 	int sack_bytes_rxmt;
578 	struct sackhole *p;
579 #ifdef TCP_SIGNATURE
580 	int sigoff = 0;
581 #endif
582 
583 #ifdef DIAGNOSTIC
584 	if (tp->t_inpcb && tp->t_in6pcb)
585 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
586 #endif
587 	so = NULL;
588 	ro = NULL;
589 	if (tp->t_inpcb) {
590 		so = tp->t_inpcb->inp_socket;
591 		ro = &tp->t_inpcb->inp_route;
592 	}
593 #ifdef INET6
594 	else if (tp->t_in6pcb) {
595 		so = tp->t_in6pcb->in6p_socket;
596 		ro = &tp->t_in6pcb->in6p_route;
597 	}
598 #endif
599 
600 	switch (af = tp->t_family) {
601 #ifdef INET
602 	case AF_INET:
603 		if (tp->t_inpcb)
604 			break;
605 #ifdef INET6
606 		/* mapped addr case */
607 		if (tp->t_in6pcb)
608 			break;
609 #endif
610 		return (EINVAL);
611 #endif
612 #ifdef INET6
613 	case AF_INET6:
614 		if (tp->t_in6pcb)
615 			break;
616 		return (EINVAL);
617 #endif
618 	default:
619 		return (EAFNOSUPPORT);
620 	}
621 
622 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
623 		return (EMSGSIZE);
624 
625 	idle = (tp->snd_max == tp->snd_una);
626 
627 	/*
628 	 * Determine if we can use TCP segmentation offload:
629 	 * - If we're using IPv4
630 	 * - If there is not an IPsec policy that prevents it
631 	 * - If the interface can do it
632 	 */
633 	has_tso4 = has_tso6 = false;
634 #if defined(INET)
635 	has_tso4 = tp->t_inpcb != NULL &&
636 #if defined(IPSEC) || defined(FAST_IPSEC)
637 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
638 		  		       IPSEC_DIR_OUTBOUND) &&
639 #endif
640 		  tp->t_inpcb->inp_route.ro_rt != NULL &&
641 		  (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable &
642 		   IFCAP_TSOv4) != 0;
643 #endif /* defined(INET) */
644 #if defined(INET6)
645 	has_tso6 = tp->t_in6pcb != NULL &&
646 #if defined(IPSEC) || defined(FAST_IPSEC)
647 		  IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
648 		  		       IPSEC_DIR_OUTBOUND) &&
649 #endif
650 		  tp->t_in6pcb->in6p_route.ro_rt != NULL &&
651 		  (tp->t_in6pcb->in6p_route.ro_rt->rt_ifp->if_capenable &
652 		   IFCAP_TSOv6) != 0;
653 #endif /* defined(INET6) */
654 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
655 
656 	/*
657 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
658 	 *
659 	 *	Optionally, a TCP MAY set the restart window to the
660 	 *	minimum of the value used for the initial window and
661 	 *	the current value of cwnd (in other words, using a
662 	 *	larger value for the restart window should never increase
663 	 *	the size of cwnd).
664 	 */
665 	if (tcp_cwm) {
666 		/*
667 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
668 		 * Count the number of packets currently pending
669 		 * acknowledgement, and limit our congestion window
670 		 * to a pre-determined allowed burst size plus that count.
671 		 * This prevents bursting once all pending packets have
672 		 * been acknowledged (i.e. transmission is idle).
673 		 *
674 		 * XXX Link this to Initial Window?
675 		 */
676 		tp->snd_cwnd = min(tp->snd_cwnd,
677 		    (tcp_cwm_burstsize * txsegsize) +
678 		    (tp->snd_nxt - tp->snd_una));
679 	} else {
680 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
681 			/*
682 			 * We have been idle for "a while" and no acks are
683 			 * expected to clock out any data we send --
684 			 * slow start to get ack "clock" running again.
685 			 */
686 			int ss = tcp_init_win;
687 #ifdef INET
688 			if (tp->t_inpcb &&
689 			    in_localaddr(tp->t_inpcb->inp_faddr))
690 				ss = tcp_init_win_local;
691 #endif
692 #ifdef INET6
693 			if (tp->t_in6pcb &&
694 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
695 				ss = tcp_init_win_local;
696 #endif
697 			tp->snd_cwnd = min(tp->snd_cwnd,
698 			    TCP_INITIAL_WINDOW(ss, txsegsize));
699 		}
700 	}
701 
702 	txsegsize_nosack = txsegsize;
703 again:
704 	use_tso = has_tso;
705 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
706 		/* don't duplicate CWR/ECE. */
707 		use_tso = 0;
708 	}
709 	TCP_REASS_LOCK(tp);
710 	sack_numblks = tcp_sack_numblks(tp);
711 	if (sack_numblks) {
712 		int sackoptlen;
713 
714 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
715 		if (sackoptlen > txsegsize_nosack) {
716 			sack_numblks = 0; /* give up SACK */
717 			txsegsize = txsegsize_nosack;
718 		} else {
719 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
720 				/* don't duplicate D-SACK. */
721 				use_tso = 0;
722 			}
723 			txsegsize = txsegsize_nosack - sackoptlen;
724 		}
725 	} else {
726 		txsegsize = txsegsize_nosack;
727 	}
728 
729 	/*
730 	 * Determine length of data that should be transmitted, and
731 	 * flags that should be used.  If there is some data or critical
732 	 * controls (SYN, RST) to send, then transmit; otherwise,
733 	 * investigate further.
734 	 *
735 	 * Readjust SACK information to avoid resending duplicate data.
736 	 */
737 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
738 		tcp_sack_adjust(tp);
739 	sendalot = 0;
740 	off = tp->snd_nxt - tp->snd_una;
741 	win = min(tp->snd_wnd, tp->snd_cwnd);
742 
743 	flags = tcp_outflags[tp->t_state];
744 
745 	/*
746 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
747 	 * to send out new data (when sendalot is 1), bypass this function.
748 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
749 	 * we're replacing a (future) new transmission with a retransmission
750 	 * now, and we previously incremented snd_cwnd in tcp_input().
751 	 */
752 	/*
753 	 * Still in sack recovery , reset rxmit flag to zero.
754 	 */
755 	sack_rxmit = 0;
756 	sack_bytes_rxmt = 0;
757 	len = 0;
758 	p = NULL;
759 	do {
760 		long cwin;
761 		if (!TCP_SACK_ENABLED(tp))
762 			break;
763 		if (tp->t_partialacks < 0)
764 			break;
765 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
766 		if (p == NULL)
767 			break;
768 
769 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
770 		if (cwin < 0)
771 			cwin = 0;
772 		/* Do not retransmit SACK segments beyond snd_recover */
773 		if (SEQ_GT(p->end, tp->snd_recover)) {
774 			/*
775 			 * (At least) part of sack hole extends beyond
776 			 * snd_recover. Check to see if we can rexmit data
777 			 * for this hole.
778 			 */
779 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
780 				/*
781 				 * Can't rexmit any more data for this hole.
782 				 * That data will be rexmitted in the next
783 				 * sack recovery episode, when snd_recover
784 				 * moves past p->rxmit.
785 				 */
786 				p = NULL;
787 				break;
788 			}
789 			/* Can rexmit part of the current hole */
790 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
791 		} else
792 			len = ((long)ulmin(cwin, p->end - p->rxmit));
793 		off = p->rxmit - tp->snd_una;
794 		if (off + len > so->so_snd.sb_cc) {
795 			/* 1 for TH_FIN */
796 			KASSERT(off + len == so->so_snd.sb_cc + 1);
797 			KASSERT(p->rxmit + len == tp->snd_max);
798 			len = so->so_snd.sb_cc - off;
799 		}
800 		if (len > 0) {
801 			sack_rxmit = 1;
802 			sendalot = 1;
803 		}
804 	} while (/*CONSTCOND*/0);
805 
806 	/*
807 	 * If in persist timeout with window of 0, send 1 byte.
808 	 * Otherwise, if window is small but nonzero
809 	 * and timer expired, we will send what we can
810 	 * and go to transmit state.
811 	 */
812 	if (tp->t_force) {
813 		if (win == 0) {
814 			/*
815 			 * If we still have some data to send, then
816 			 * clear the FIN bit.  Usually this would
817 			 * happen below when it realizes that we
818 			 * aren't sending all the data.  However,
819 			 * if we have exactly 1 byte of unset data,
820 			 * then it won't clear the FIN bit below,
821 			 * and if we are in persist state, we wind
822 			 * up sending the packet without recording
823 			 * that we sent the FIN bit.
824 			 *
825 			 * We can't just blindly clear the FIN bit,
826 			 * because if we don't have any more data
827 			 * to send then the probe will be the FIN
828 			 * itself.
829 			 */
830 			if (off < so->so_snd.sb_cc)
831 				flags &= ~TH_FIN;
832 			win = 1;
833 		} else {
834 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
835 			tp->t_rxtshift = 0;
836 		}
837 	}
838 
839 	if (sack_rxmit == 0) {
840 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
841 			long cwin;
842 
843 			/*
844 			 * We are inside of a SACK recovery episode and are
845 			 * sending new data, having retransmitted all the
846 			 * data possible in the scoreboard.
847 			 */
848 			if (tp->snd_wnd < so->so_snd.sb_cc) {
849 				len = tp->snd_wnd - off;
850 				flags &= ~TH_FIN;
851 			} else {
852 				len = so->so_snd.sb_cc - off;
853 			}
854 
855 			/*
856 			 * From FreeBSD:
857 			 *  Don't remove this (len > 0) check !
858 			 *  We explicitly check for len > 0 here (although it
859 			 *  isn't really necessary), to work around a gcc
860 			 *  optimization issue - to force gcc to compute
861 			 *  len above. Without this check, the computation
862 			 *  of len is bungled by the optimizer.
863 			 */
864 			if (len > 0) {
865 				cwin = tp->snd_cwnd -
866 				    (tp->snd_nxt - tp->sack_newdata) -
867 				    sack_bytes_rxmt;
868 				if (cwin < 0)
869 					cwin = 0;
870 				if (cwin < len) {
871 					len = cwin;
872 					flags &= ~TH_FIN;
873 				}
874 			}
875 		} else if (win < so->so_snd.sb_cc) {
876 			len = win - off;
877 			flags &= ~TH_FIN;
878 		} else {
879 			len = so->so_snd.sb_cc - off;
880 		}
881 	}
882 
883 	if (len < 0) {
884 		/*
885 		 * If FIN has been sent but not acked,
886 		 * but we haven't been called to retransmit,
887 		 * len will be -1.  Otherwise, window shrank
888 		 * after we sent into it.  If window shrank to 0,
889 		 * cancel pending retransmit, pull snd_nxt back
890 		 * to (closed) window, and set the persist timer
891 		 * if it isn't already going.  If the window didn't
892 		 * close completely, just wait for an ACK.
893 		 *
894 		 * If we have a pending FIN, either it has already been
895 		 * transmitted or it is outside the window, so drop it.
896 		 * If the FIN has been transmitted, but this is not a
897 		 * retransmission, then len must be -1.  Therefore we also
898 		 * prevent here the sending of `gratuitous FINs'.  This
899 		 * eliminates the need to check for that case below (e.g.
900 		 * to back up snd_nxt before the FIN so that the sequence
901 		 * number is correct).
902 		 */
903 		len = 0;
904 		flags &= ~TH_FIN;
905 		if (win == 0) {
906 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
907 			tp->t_rxtshift = 0;
908 			tp->snd_nxt = tp->snd_una;
909 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
910 				tcp_setpersist(tp);
911 		}
912 	}
913 
914 	/*
915 	 * Automatic sizing enables the performance of large buffers
916 	 * and most of the efficiency of small ones by only allocating
917 	 * space when it is needed.
918 	 *
919 	 * The criteria to step up the send buffer one notch are:
920 	 *  1. receive window of remote host is larger than send buffer
921 	 *     (with a fudge factor of 5/4th);
922 	 *  2. send buffer is filled to 7/8th with data (so we actually
923 	 *     have data to make use of it);
924 	 *  3. send buffer fill has not hit maximal automatic size;
925 	 *  4. our send window (slow start and cogestion controlled) is
926 	 *     larger than sent but unacknowledged data in send buffer.
927 	 *
928 	 * The remote host receive window scaling factor may limit the
929 	 * growing of the send buffer before it reaches its allowed
930 	 * maximum.
931 	 *
932 	 * It scales directly with slow start or congestion window
933 	 * and does at most one step per received ACK.  This fast
934 	 * scaling has the drawback of growing the send buffer beyond
935 	 * what is strictly necessary to make full use of a given
936 	 * delay*bandwith product.  However testing has shown this not
937 	 * to be much of an problem.  At worst we are trading wasting
938 	 * of available bandwith (the non-use of it) for wasting some
939 	 * socket buffer memory.
940 	 *
941 	 * TODO: Shrink send buffer during idle periods together
942 	 * with congestion window.  Requires another timer.
943 	 */
944 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
945 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
946 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
947 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
948 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
949 			if (!sbreserve(&so->so_snd,
950 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
951 			     tcp_autosndbuf_max), so))
952 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
953 		}
954 	}
955 
956 	if (len > txsegsize) {
957 		if (use_tso) {
958 			/*
959 			 * Truncate TSO transfers to IP_MAXPACKET, and make
960 			 * sure that we send equal size transfers down the
961 			 * stack (rather than big-small-big-small-...).
962 			 */
963 #ifdef INET6
964 #if IPV6_MAXPACKET != IP_MAXPACKET
965 #error IPV6_MAXPACKET != IP_MAXPACKET
966 #endif
967 #endif
968 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
969 			if (len <= txsegsize) {
970 				use_tso = 0;
971 			}
972 		} else
973 			len = txsegsize;
974 		flags &= ~TH_FIN;
975 		sendalot = 1;
976 	} else
977 		use_tso = 0;
978 	if (sack_rxmit) {
979 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
980 			flags &= ~TH_FIN;
981 	}
982 
983 	win = sbspace(&so->so_rcv);
984 
985 	/*
986 	 * Sender silly window avoidance.  If connection is idle
987 	 * and can send all data, a maximum segment,
988 	 * at least a maximum default-size segment do it,
989 	 * or are forced, do it; otherwise don't bother.
990 	 * If peer's buffer is tiny, then send
991 	 * when window is at least half open.
992 	 * If retransmitting (possibly after persist timer forced us
993 	 * to send into a small window), then must resend.
994 	 */
995 	if (len) {
996 		if (len >= txsegsize)
997 			goto send;
998 		if ((so->so_state & SS_MORETOCOME) == 0 &&
999 		    ((idle || tp->t_flags & TF_NODELAY) &&
1000 		     len + off >= so->so_snd.sb_cc))
1001 			goto send;
1002 		if (tp->t_force)
1003 			goto send;
1004 		if (len >= tp->max_sndwnd / 2)
1005 			goto send;
1006 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1007 			goto send;
1008 		if (sack_rxmit)
1009 			goto send;
1010 	}
1011 
1012 	/*
1013 	 * Compare available window to amount of window known to peer
1014 	 * (as advertised window less next expected input).  If the
1015 	 * difference is at least twice the size of the largest segment
1016 	 * we expect to receive (i.e. two segments) or at least 50% of
1017 	 * the maximum possible window, then want to send a window update
1018 	 * to peer.
1019 	 */
1020 	if (win > 0) {
1021 		/*
1022 		 * "adv" is the amount we can increase the window,
1023 		 * taking into account that we are limited by
1024 		 * TCP_MAXWIN << tp->rcv_scale.
1025 		 */
1026 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1027 			(tp->rcv_adv - tp->rcv_nxt);
1028 
1029 		if (adv >= (long) (2 * rxsegsize))
1030 			goto send;
1031 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1032 			goto send;
1033 	}
1034 
1035 	/*
1036 	 * Send if we owe peer an ACK.
1037 	 */
1038 	if (tp->t_flags & TF_ACKNOW)
1039 		goto send;
1040 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1041 		goto send;
1042 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1043 		goto send;
1044 	/*
1045 	 * In SACK, it is possible for tcp_output to fail to send a segment
1046 	 * after the retransmission timer has been turned off.  Make sure
1047 	 * that the retransmission timer is set.
1048 	 */
1049 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1050 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1051 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1052 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1053 		goto just_return;
1054 	}
1055 
1056 	/*
1057 	 * TCP window updates are not reliable, rather a polling protocol
1058 	 * using ``persist'' packets is used to insure receipt of window
1059 	 * updates.  The three ``states'' for the output side are:
1060 	 *	idle			not doing retransmits or persists
1061 	 *	persisting		to move a small or zero window
1062 	 *	(re)transmitting	and thereby not persisting
1063 	 *
1064 	 * tp->t_timer[TCPT_PERSIST]
1065 	 *	is set when we are in persist state.
1066 	 * tp->t_force
1067 	 *	is set when we are called to send a persist packet.
1068 	 * tp->t_timer[TCPT_REXMT]
1069 	 *	is set when we are retransmitting
1070 	 * The output side is idle when both timers are zero.
1071 	 *
1072 	 * If send window is too small, there is data to transmit, and no
1073 	 * retransmit or persist is pending, then go to persist state.
1074 	 * If nothing happens soon, send when timer expires:
1075 	 * if window is nonzero, transmit what we can,
1076 	 * otherwise force out a byte.
1077 	 */
1078 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1079 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1080 		tp->t_rxtshift = 0;
1081 		tcp_setpersist(tp);
1082 	}
1083 
1084 	/*
1085 	 * No reason to send a segment, just return.
1086 	 */
1087 just_return:
1088 	TCP_REASS_UNLOCK(tp);
1089 	return (0);
1090 
1091 send:
1092 	/*
1093 	 * Before ESTABLISHED, force sending of initial options
1094 	 * unless TCP set not to do any options.
1095 	 * NOTE: we assume that the IP/TCP header plus TCP options
1096 	 * always fit in a single mbuf, leaving room for a maximum
1097 	 * link header, i.e.
1098 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1099 	 */
1100 	optlen = 0;
1101 	switch (af) {
1102 #ifdef INET
1103 	case AF_INET:
1104 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1105 		break;
1106 #endif
1107 #ifdef INET6
1108 	case AF_INET6:
1109 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1110 		break;
1111 #endif
1112 	default:	/*pacify gcc*/
1113 		iphdrlen = 0;
1114 		break;
1115 	}
1116 	hdrlen = iphdrlen;
1117 	if (flags & TH_SYN) {
1118 		struct rtentry *rt;
1119 
1120 		rt = NULL;
1121 #ifdef INET
1122 		if (tp->t_inpcb)
1123 			rt = in_pcbrtentry(tp->t_inpcb);
1124 #endif
1125 #ifdef INET6
1126 		if (tp->t_in6pcb)
1127 			rt = in6_pcbrtentry(tp->t_in6pcb);
1128 #endif
1129 
1130 		tp->snd_nxt = tp->iss;
1131 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
1132 						    rt->rt_ifp : NULL, af);
1133 		if ((tp->t_flags & TF_NOOPT) == 0) {
1134 			opt[0] = TCPOPT_MAXSEG;
1135 			opt[1] = 4;
1136 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1137 			opt[3] = tp->t_ourmss & 0xff;
1138 			optlen = 4;
1139 
1140 			if ((tp->t_flags & TF_REQ_SCALE) &&
1141 			    ((flags & TH_ACK) == 0 ||
1142 			    (tp->t_flags & TF_RCVD_SCALE))) {
1143 				*((u_int32_t *) (opt + optlen)) = htonl(
1144 					TCPOPT_NOP << 24 |
1145 					TCPOPT_WINDOW << 16 |
1146 					TCPOLEN_WINDOW << 8 |
1147 					tp->request_r_scale);
1148 				optlen += 4;
1149 			}
1150 			if (tcp_do_sack) {
1151 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1152 
1153 				cp[0] = TCPOPT_SACK_PERMITTED;
1154 				cp[1] = 2;
1155 				cp[2] = TCPOPT_NOP;
1156 				cp[3] = TCPOPT_NOP;
1157 				optlen += 4;
1158 			}
1159 		}
1160 	}
1161 
1162 	/*
1163 	 * Send a timestamp and echo-reply if this is a SYN and our side
1164 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1165 	 * and our peer have sent timestamps in our SYN's.
1166 	 */
1167 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1168 	     (flags & TH_RST) == 0 &&
1169 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1170 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1171 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1172 
1173 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1174 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1175 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1176 		*lp   = htonl(tp->ts_recent);
1177 		optlen += TCPOLEN_TSTAMP_APPA;
1178 
1179 		/* Set receive buffer autosizing timestamp. */
1180 		if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1181 			tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1182 	}
1183 
1184 	/*
1185 	 * Tack on the SACK block if it is necessary.
1186 	 */
1187 	if (sack_numblks) {
1188 		int sack_len;
1189 		u_char *bp = (u_char *)(opt + optlen);
1190 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1191 		struct ipqent *tiqe;
1192 
1193 		sack_len = sack_numblks * 8 + 2;
1194 		bp[0] = TCPOPT_NOP;
1195 		bp[1] = TCPOPT_NOP;
1196 		bp[2] = TCPOPT_SACK;
1197 		bp[3] = sack_len;
1198 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1199 			sack_numblks--;
1200 			*lp++ = htonl(tp->rcv_dsack_block.left);
1201 			*lp++ = htonl(tp->rcv_dsack_block.right);
1202 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1203 		}
1204 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1205 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1206 			KASSERT(tiqe != NULL);
1207 			sack_numblks--;
1208 			*lp++ = htonl(tiqe->ipqe_seq);
1209 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1210 			    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1211 		}
1212 		optlen += sack_len + 2;
1213 	}
1214 	TCP_REASS_UNLOCK(tp);
1215 
1216 #ifdef TCP_SIGNATURE
1217 	if (tp->t_flags & TF_SIGNATURE) {
1218 		u_char *bp;
1219 		/*
1220 		 * Initialize TCP-MD5 option (RFC2385)
1221 		 */
1222 		bp = (u_char *)opt + optlen;
1223 		*bp++ = TCPOPT_SIGNATURE;
1224 		*bp++ = TCPOLEN_SIGNATURE;
1225 		sigoff = optlen + 2;
1226 		bzero(bp, TCP_SIGLEN);
1227 		bp += TCP_SIGLEN;
1228 		optlen += TCPOLEN_SIGNATURE;
1229 		/*
1230 		 * Terminate options list and maintain 32-bit alignment.
1231  		 */
1232 		*bp++ = TCPOPT_NOP;
1233 		*bp++ = TCPOPT_EOL;
1234  		optlen += 2;
1235  	}
1236 #endif /* TCP_SIGNATURE */
1237 
1238 	hdrlen += optlen;
1239 
1240 #ifdef DIAGNOSTIC
1241 	if (!use_tso && len > txsegsize)
1242 		panic("tcp data to be sent is larger than segment");
1243 	else if (use_tso && len > IP_MAXPACKET)
1244 		panic("tcp data to be sent is larger than max TSO size");
1245 	if (max_linkhdr + hdrlen > MCLBYTES)
1246 		panic("tcphdr too big");
1247 #endif
1248 
1249 	/*
1250 	 * Grab a header mbuf, attaching a copy of data to
1251 	 * be transmitted, and initialize the header from
1252 	 * the template for sends on this connection.
1253 	 */
1254 	if (len) {
1255 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1256 		if (error)
1257 			goto out;
1258 		/*
1259 		 * If we're sending everything we've got, set PUSH.
1260 		 * (This will keep happy those implementations which only
1261 		 * give data to the user when a buffer fills or
1262 		 * a PUSH comes in.)
1263 		 */
1264 		if (off + len == so->so_snd.sb_cc)
1265 			flags |= TH_PUSH;
1266 	} else {
1267 		if (tp->t_flags & TF_ACKNOW)
1268 			tcpstat.tcps_sndacks++;
1269 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1270 			tcpstat.tcps_sndctrl++;
1271 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1272 			tcpstat.tcps_sndurg++;
1273 		else
1274 			tcpstat.tcps_sndwinup++;
1275 
1276 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1277 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1278 			MCLGET(m, M_DONTWAIT);
1279 			if ((m->m_flags & M_EXT) == 0) {
1280 				m_freem(m);
1281 				m = NULL;
1282 			}
1283 		}
1284 		if (m == NULL) {
1285 			error = ENOBUFS;
1286 			goto out;
1287 		}
1288 		MCLAIM(m, &tcp_tx_mowner);
1289 		m->m_data += max_linkhdr;
1290 		m->m_len = hdrlen;
1291 	}
1292 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1293 	switch (af) {
1294 #ifdef INET
1295 	case AF_INET:
1296 		ip = mtod(m, struct ip *);
1297 #ifdef INET6
1298 		ip6 = NULL;
1299 #endif
1300 		th = (struct tcphdr *)(ip + 1);
1301 		break;
1302 #endif
1303 #ifdef INET6
1304 	case AF_INET6:
1305 		ip = NULL;
1306 		ip6 = mtod(m, struct ip6_hdr *);
1307 		th = (struct tcphdr *)(ip6 + 1);
1308 		break;
1309 #endif
1310 	default:	/*pacify gcc*/
1311 		ip = NULL;
1312 #ifdef INET6
1313 		ip6 = NULL;
1314 #endif
1315 		th = NULL;
1316 		break;
1317 	}
1318 	if (tp->t_template == 0)
1319 		panic("tcp_output");
1320 	if (tp->t_template->m_len < iphdrlen)
1321 		panic("tcp_output");
1322 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1323 
1324 	/*
1325 	 * If we are starting a connection, send ECN setup
1326 	 * SYN packet. If we are on a retransmit, we may
1327 	 * resend those bits a number of times as per
1328 	 * RFC 3168.
1329 	 */
1330 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1331 		if (tp->t_flags & TF_SYN_REXMT) {
1332 			if (tp->t_ecn_retries--)
1333 				flags |= TH_ECE|TH_CWR;
1334 		} else {
1335 			flags |= TH_ECE|TH_CWR;
1336 			tp->t_ecn_retries = tcp_ecn_maxretries;
1337 		}
1338 	}
1339 
1340 	if (TCP_ECN_ALLOWED(tp)) {
1341 		/*
1342 		 * If the peer has ECN, mark data packets
1343 		 * ECN capable. Ignore pure ack packets, retransmissions
1344 		 * and window probes.
1345 		 */
1346 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1347 		    !(tp->t_force && len == 1)) {
1348 			switch (af) {
1349 #ifdef INET
1350 			case AF_INET:
1351 				tp->t_inpcb->inp_ip.ip_tos |= IPTOS_ECN_ECT0;
1352 				break;
1353 #endif
1354 #ifdef INET6
1355 			case AF_INET6:
1356 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1357 				break;
1358 #endif
1359 			}
1360 			tcpstat.tcps_ecn_ect++;
1361 		}
1362 
1363 		/*
1364 		 * Reply with proper ECN notifications.
1365 		 */
1366 		if (tp->t_flags & TF_ECN_SND_CWR) {
1367 			flags |= TH_CWR;
1368 			tp->t_flags &= ~TF_ECN_SND_CWR;
1369 		}
1370 		if (tp->t_flags & TF_ECN_SND_ECE) {
1371 			flags |= TH_ECE;
1372 		}
1373 	}
1374 
1375 
1376 	/*
1377 	 * If we are doing retransmissions, then snd_nxt will
1378 	 * not reflect the first unsent octet.  For ACK only
1379 	 * packets, we do not want the sequence number of the
1380 	 * retransmitted packet, we want the sequence number
1381 	 * of the next unsent octet.  So, if there is no data
1382 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1383 	 * when filling in ti_seq.  But if we are in persist
1384 	 * state, snd_max might reflect one byte beyond the
1385 	 * right edge of the window, so use snd_nxt in that
1386 	 * case, since we know we aren't doing a retransmission.
1387 	 * (retransmit and persist are mutually exclusive...)
1388 	 */
1389 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1390 		th->th_seq = htonl(p->rxmit);
1391 		p->rxmit += len;
1392 	} else {
1393 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1394 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1395 			th->th_seq = htonl(tp->snd_nxt);
1396 		else
1397 			th->th_seq = htonl(tp->snd_max);
1398 	}
1399 	th->th_ack = htonl(tp->rcv_nxt);
1400 	if (optlen) {
1401 		bcopy((void *)opt, (void *)(th + 1), optlen);
1402 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1403 	}
1404 	th->th_flags = flags;
1405 	/*
1406 	 * Calculate receive window.  Don't shrink window,
1407 	 * but avoid silly window syndrome.
1408 	 */
1409 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1410 		win = 0;
1411 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1412 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1413 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1414 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1415 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1416 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1417 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1418 		if (urp > IP_MAXPACKET)
1419 			urp = IP_MAXPACKET;
1420 		th->th_urp = htons((u_int16_t)urp);
1421 		th->th_flags |= TH_URG;
1422 	} else
1423 		/*
1424 		 * If no urgent pointer to send, then we pull
1425 		 * the urgent pointer to the left edge of the send window
1426 		 * so that it doesn't drift into the send window on sequence
1427 		 * number wraparound.
1428 		 */
1429 		tp->snd_up = tp->snd_una;		/* drag it along */
1430 
1431 #ifdef TCP_SIGNATURE
1432 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1433 		struct secasvar *sav;
1434 		u_int8_t *sigp;
1435 
1436 		sav = tcp_signature_getsav(m, th);
1437 
1438 		if (sav == NULL) {
1439 			if (m)
1440 				m_freem(m);
1441 			return (EPERM);
1442 		}
1443 
1444 		m->m_pkthdr.len = hdrlen + len;
1445 		sigp = (char *)th + sizeof(*th) + sigoff;
1446 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1447 
1448 		key_sa_recordxfer(sav, m);
1449 #ifdef FAST_IPSEC
1450 		KEY_FREESAV(&sav);
1451 #else
1452 		key_freesav(sav);
1453 #endif
1454 	}
1455 #endif
1456 
1457 	/*
1458 	 * Set ourselves up to be checksummed just before the packet
1459 	 * hits the wire.
1460 	 */
1461 	switch (af) {
1462 #ifdef INET
1463 	case AF_INET:
1464 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1465 		if (use_tso) {
1466 			m->m_pkthdr.segsz = txsegsize;
1467 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1468 		} else {
1469 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1470 			if (len + optlen) {
1471 				/* Fixup the pseudo-header checksum. */
1472 				/* XXXJRT Not IP Jumbogram safe. */
1473 				th->th_sum = in_cksum_addword(th->th_sum,
1474 				    htons((u_int16_t) (len + optlen)));
1475 			}
1476 		}
1477 		break;
1478 #endif
1479 #ifdef INET6
1480 	case AF_INET6:
1481 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1482 		if (use_tso) {
1483 			m->m_pkthdr.segsz = txsegsize;
1484 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1485 		} else {
1486 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1487 			if (len + optlen) {
1488 				/* Fixup the pseudo-header checksum. */
1489 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1490 				th->th_sum = in_cksum_addword(th->th_sum,
1491 				    htons((u_int16_t) (len + optlen)));
1492 			}
1493 		}
1494 		break;
1495 #endif
1496 	}
1497 
1498 	/*
1499 	 * In transmit state, time the transmission and arrange for
1500 	 * the retransmit.  In persist state, just set snd_max.
1501 	 */
1502 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1503 		tcp_seq startseq = tp->snd_nxt;
1504 
1505 		/*
1506 		 * Advance snd_nxt over sequence space of this segment.
1507 		 * There are no states in which we send both a SYN and a FIN,
1508 		 * so we collapse the tests for these flags.
1509 		 */
1510 		if (flags & (TH_SYN|TH_FIN))
1511 			tp->snd_nxt++;
1512 		if (sack_rxmit)
1513 			goto timer;
1514 		tp->snd_nxt += len;
1515 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1516 			tp->snd_max = tp->snd_nxt;
1517 			/*
1518 			 * Time this transmission if not a retransmission and
1519 			 * not currently timing anything.
1520 			 */
1521 			if (tp->t_rtttime == 0) {
1522 				tp->t_rtttime = tcp_now;
1523 				tp->t_rtseq = startseq;
1524 				tcpstat.tcps_segstimed++;
1525 			}
1526 		}
1527 
1528 		/*
1529 		 * Set retransmit timer if not currently set,
1530 		 * and not doing an ack or a keep-alive probe.
1531 		 * Initial value for retransmit timer is smoothed
1532 		 * round-trip time + 2 * round-trip time variance.
1533 		 * Initialize shift counter which is used for backoff
1534 		 * of retransmit time.
1535 		 */
1536 timer:
1537 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1538 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1539 		    tp->snd_nxt != tp->snd_una)) {
1540 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1541 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1542 				tp->t_rxtshift = 0;
1543 			}
1544 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1545 		}
1546 	} else
1547 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1548 			tp->snd_max = tp->snd_nxt + len;
1549 
1550 #ifdef TCP_DEBUG
1551 	/*
1552 	 * Trace.
1553 	 */
1554 	if (so->so_options & SO_DEBUG)
1555 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1556 #endif
1557 
1558 	/*
1559 	 * Fill in IP length and desired time to live and
1560 	 * send to IP level.  There should be a better way
1561 	 * to handle ttl and tos; we could keep them in
1562 	 * the template, but need a way to checksum without them.
1563 	 */
1564 	m->m_pkthdr.len = hdrlen + len;
1565 
1566 	switch (af) {
1567 #ifdef INET
1568 	case AF_INET:
1569 		ip->ip_len = htons(m->m_pkthdr.len);
1570 		packetlen = m->m_pkthdr.len;
1571 		if (tp->t_inpcb) {
1572 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1573 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1574 		}
1575 #ifdef INET6
1576 		else if (tp->t_in6pcb) {
1577 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1578 			ip->ip_tos = 0;	/*XXX*/
1579 		}
1580 #endif
1581 		break;
1582 #endif
1583 #ifdef INET6
1584 	case AF_INET6:
1585 		packetlen = m->m_pkthdr.len;
1586 		ip6->ip6_nxt = IPPROTO_TCP;
1587 		if (tp->t_in6pcb) {
1588 			/*
1589 			 * we separately set hoplimit for every segment, since
1590 			 * the user might want to change the value via
1591 			 * setsockopt. Also, desired default hop limit might
1592 			 * be changed via Neighbor Discovery.
1593 			 */
1594 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1595 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1596 		}
1597 		/* ip6->ip6_flow = ??? */
1598 		/* ip6_plen will be filled in ip6_output(). */
1599 		break;
1600 #endif
1601 	default:	/*pacify gcc*/
1602 		packetlen = 0;
1603 		break;
1604 	}
1605 
1606 	switch (af) {
1607 #ifdef INET
1608 	case AF_INET:
1609 	    {
1610 		struct mbuf *opts;
1611 
1612 		if (tp->t_inpcb)
1613 			opts = tp->t_inpcb->inp_options;
1614 		else
1615 			opts = NULL;
1616 		error = ip_output(m, opts, ro,
1617 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1618 			(so->so_options & SO_DONTROUTE),
1619 			(struct ip_moptions *)0, so);
1620 		break;
1621 	    }
1622 #endif
1623 #ifdef INET6
1624 	case AF_INET6:
1625 	    {
1626 		struct ip6_pktopts *opts;
1627 
1628 		if (tp->t_in6pcb)
1629 			opts = tp->t_in6pcb->in6p_outputopts;
1630 		else
1631 			opts = NULL;
1632 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1633 			NULL, so, NULL);
1634 		break;
1635 	    }
1636 #endif
1637 	default:
1638 		error = EAFNOSUPPORT;
1639 		break;
1640 	}
1641 	if (error) {
1642 out:
1643 		if (error == ENOBUFS) {
1644 			tcpstat.tcps_selfquench++;
1645 #ifdef INET
1646 			if (tp->t_inpcb)
1647 				tcp_quench(tp->t_inpcb, 0);
1648 #endif
1649 #ifdef INET6
1650 			if (tp->t_in6pcb)
1651 				tcp6_quench(tp->t_in6pcb, 0);
1652 #endif
1653 			error = 0;
1654 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1655 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1656 			tp->t_softerror = error;
1657 			error = 0;
1658 		}
1659 
1660 		/* Back out the seqence number advance. */
1661 		if (sack_rxmit)
1662 			p->rxmit -= len;
1663 
1664 		/* Restart the delayed ACK timer, if necessary. */
1665 		if (tp->t_flags & TF_DELACK)
1666 			TCP_RESTART_DELACK(tp);
1667 
1668 		return (error);
1669 	}
1670 
1671 	if (packetlen > tp->t_pmtud_mtu_sent)
1672 		tp->t_pmtud_mtu_sent = packetlen;
1673 
1674 	tcpstat.tcps_sndtotal++;
1675 	if (tp->t_flags & TF_DELACK)
1676 		tcpstat.tcps_delack++;
1677 
1678 	/*
1679 	 * Data sent (as far as we can tell).
1680 	 * If this advertises a larger window than any other segment,
1681 	 * then remember the size of the advertised window.
1682 	 * Any pending ACK has now been sent.
1683 	 */
1684 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1685 		tp->rcv_adv = tp->rcv_nxt + win;
1686 	tp->last_ack_sent = tp->rcv_nxt;
1687 	tp->t_flags &= ~TF_ACKNOW;
1688 	TCP_CLEAR_DELACK(tp);
1689 #ifdef DIAGNOSTIC
1690 	if (maxburst < 0)
1691 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1692 #endif
1693 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1694 		goto again;
1695 	return (0);
1696 }
1697 
1698 void
1699 tcp_setpersist(struct tcpcb *tp)
1700 {
1701 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1702 	int nticks;
1703 
1704 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1705 		panic("tcp_output REXMT");
1706 	/*
1707 	 * Start/restart persistance timer.
1708 	 */
1709 	if (t < tp->t_rttmin)
1710 		t = tp->t_rttmin;
1711 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1712 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1713 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1714 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1715 		tp->t_rxtshift++;
1716 }
1717