xref: /netbsd-src/sys/netinet/tcp_output.c (revision aaf4ece63a859a04e37cf3a7229b5fab0157cc06)
1 /*	$NetBSD: tcp_output.c,v 1.141 2005/12/24 20:45:09 perry 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 1.141 2005/12/24 20:45:09 perry Exp $");
144 
145 #include "opt_inet.h"
146 #include "opt_ipsec.h"
147 #include "opt_tcp_debug.h"
148 
149 #include <sys/param.h>
150 #include <sys/systm.h>
151 #include <sys/malloc.h>
152 #include <sys/mbuf.h>
153 #include <sys/protosw.h>
154 #include <sys/socket.h>
155 #include <sys/socketvar.h>
156 #include <sys/errno.h>
157 #include <sys/domain.h>
158 #include <sys/kernel.h>
159 #ifdef TCP_SIGNATURE
160 #include <sys/md5.h>
161 #endif
162 
163 #include <net/if.h>
164 #include <net/route.h>
165 
166 #include <netinet/in.h>
167 #include <netinet/in_systm.h>
168 #include <netinet/ip.h>
169 #include <netinet/in_pcb.h>
170 #include <netinet/ip_var.h>
171 
172 #ifdef INET6
173 #ifndef INET
174 #include <netinet/in.h>
175 #endif
176 #include <netinet/ip6.h>
177 #include <netinet6/in6_var.h>
178 #include <netinet6/ip6_var.h>
179 #include <netinet6/in6_pcb.h>
180 #include <netinet6/nd6.h>
181 #endif
182 
183 #ifdef FAST_IPSEC
184 #include <netipsec/ipsec.h>
185 #include <netipsec/key.h>
186 #endif	/* FAST_IPSEC*/
187 #ifdef IPSEC
188 #include <netinet6/ipsec.h>
189 #endif
190 
191 #include <netinet/tcp.h>
192 #define	TCPOUTFLAGS
193 #include <netinet/tcp_fsm.h>
194 #include <netinet/tcp_seq.h>
195 #include <netinet/tcp_timer.h>
196 #include <netinet/tcp_var.h>
197 #include <netinet/tcpip.h>
198 #include <netinet/tcp_debug.h>
199 #include <netinet/in_offload.h>
200 
201 #ifdef IPSEC
202 #include <netkey/key.h>
203 #endif
204 
205 #ifdef notyet
206 extern struct mbuf *m_copypack();
207 #endif
208 
209 /*
210  * Knob to enable Congestion Window Monitoring, and control
211  * the burst size it allows.  Default burst is 4 packets, per
212  * the Internet draft.
213  */
214 int	tcp_cwm = 0;
215 int	tcp_cwm_burstsize = 4;
216 
217 #ifdef TCP_OUTPUT_COUNTERS
218 #include <sys/device.h>
219 
220 extern struct evcnt tcp_output_bigheader;
221 extern struct evcnt tcp_output_predict_hit;
222 extern struct evcnt tcp_output_predict_miss;
223 extern struct evcnt tcp_output_copysmall;
224 extern struct evcnt tcp_output_copybig;
225 extern struct evcnt tcp_output_refbig;
226 
227 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
228 #else
229 
230 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
231 
232 #endif /* TCP_OUTPUT_COUNTERS */
233 
234 static
235 #ifndef GPROF
236 inline
237 #endif
238 int
239 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
240 {
241 #ifdef INET
242 	struct inpcb *inp = tp->t_inpcb;
243 #endif
244 #ifdef INET6
245 	struct in6pcb *in6p = tp->t_in6pcb;
246 #endif
247 	struct socket *so = NULL;
248 	struct rtentry *rt;
249 	struct ifnet *ifp;
250 	int size;
251 	int iphlen;
252 	int optlen;
253 
254 #ifdef DIAGNOSTIC
255 	if (tp->t_inpcb && tp->t_in6pcb)
256 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
257 #endif
258 	switch (tp->t_family) {
259 #ifdef INET
260 	case AF_INET:
261 		iphlen = sizeof(struct ip);
262 		break;
263 #endif
264 #ifdef INET6
265 	case AF_INET6:
266 		iphlen = sizeof(struct ip6_hdr);
267 		break;
268 #endif
269 	default:
270 		size = tcp_mssdflt;
271 		goto out;
272 	}
273 
274 	rt = NULL;
275 #ifdef INET
276 	if (inp) {
277 		rt = in_pcbrtentry(inp);
278 		so = inp->inp_socket;
279 	}
280 #endif
281 #ifdef INET6
282 	if (in6p) {
283 		rt = in6_pcbrtentry(in6p);
284 		so = in6p->in6p_socket;
285 	}
286 #endif
287 	if (rt == NULL) {
288 		size = tcp_mssdflt;
289 		goto out;
290 	}
291 
292 	ifp = rt->rt_ifp;
293 
294 	size = tcp_mssdflt;
295 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
296 #ifdef INET6
297 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
298 			/*
299 			 * RFC2460 section 5, last paragraph: if path MTU is
300 			 * smaller than 1280, use 1280 as packet size and
301 			 * attach fragment header.
302 			 */
303 			size = IPV6_MMTU - iphlen - sizeof(struct ip6_frag) -
304 			    sizeof(struct tcphdr);
305 		} else
306 			size = rt->rt_rmx.rmx_mtu - iphlen -
307 			    sizeof(struct tcphdr);
308 #else
309 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
310 #endif
311 	} else if (ifp->if_flags & IFF_LOOPBACK)
312 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
313 #ifdef INET
314 	else if (inp && tp->t_mtudisc)
315 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
316 	else if (inp && in_localaddr(inp->inp_faddr))
317 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
318 #endif
319 #ifdef INET6
320 	else if (in6p) {
321 #ifdef INET
322 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
323 			/* mapped addr case */
324 			struct in_addr d;
325 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
326 			if (tp->t_mtudisc || in_localaddr(d))
327 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
328 		} else
329 #endif
330 		{
331 			/*
332 			 * for IPv6, path MTU discovery is always turned on,
333 			 * or the node must use packet size <= 1280.
334 			 */
335 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
336 			size -= (iphlen + sizeof(struct tcphdr));
337 		}
338 	}
339 #endif
340  out:
341 	/*
342 	 * Now we must make room for whatever extra TCP/IP options are in
343 	 * the packet.
344 	 */
345 	optlen = tcp_optlen(tp);
346 
347 	/*
348 	 * XXX tp->t_ourmss should have the right size, but without this code
349 	 * fragmentation will occur... need more investigation
350 	 */
351 #ifdef INET
352 	if (inp) {
353 #if defined(IPSEC) || defined(FAST_IPSEC)
354 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
355 			optlen += ipsec4_hdrsiz_tcp(tp);
356 #endif
357 		optlen += ip_optlen(inp);
358 	}
359 #endif
360 #ifdef INET6
361 #ifdef INET
362 	if (in6p && tp->t_family == AF_INET) {
363 #if defined(IPSEC) || defined(FAST_IPSEC)
364 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
365 			optlen += ipsec4_hdrsiz_tcp(tp);
366 #endif
367 		/* XXX size -= ip_optlen(in6p); */
368 	} else
369 #endif
370 	if (in6p && tp->t_family == AF_INET6) {
371 #ifdef IPSEC
372 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
373 			optlen += ipsec6_hdrsiz_tcp(tp);
374 #endif
375 		optlen += ip6_optlen(in6p);
376 	}
377 #endif
378 	size -= optlen;
379 
380 	/* there may not be any room for data if mtu is too small */
381 	if (size < 0)
382 		return (EMSGSIZE);
383 
384 	/*
385 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
386 	 * assumes that path MTU is the same for both ways).  this is only
387 	 * for silly window avoidance, do not use the value for other purposes.
388 	 *
389 	 * ipseclen is subtracted from both sides, this may not be right.
390 	 * I'm not quite sure about this (could someone comment).
391 	 */
392 	*txsegsizep = min(tp->t_peermss - optlen, size);
393 	/*
394 	 * Never send more than half a buffer full.  This insures that we can
395 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
396 	 * therefore acks will never be delayed unless we run out of data to
397 	 * transmit.
398 	 */
399 	if (so)
400 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
401 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
402 
403 	if (*txsegsizep != tp->t_segsz) {
404 		/*
405 		 * If the new segment size is larger, we don't want to
406 		 * mess up the congestion window, but if it is smaller
407 		 * we'll have to reduce the congestion window to ensure
408 		 * that we don't get into trouble with initial windows
409 		 * and the rest.  In any case, if the segment size
410 		 * has changed, chances are the path has, too, and
411 		 * our congestion window will be different.
412 		 */
413 		if (*txsegsizep < tp->t_segsz) {
414 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
415 					   * *txsegsizep, *txsegsizep);
416 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
417 						* *txsegsizep, *txsegsizep);
418 		}
419 		tp->t_segsz = *txsegsizep;
420 	}
421 
422 	return (0);
423 }
424 
425 static
426 #ifndef GPROF
427 inline
428 #endif
429 int
430 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
431     long len, int hdrlen, struct mbuf **mp)
432 {
433 	struct mbuf *m, *m0;
434 
435 	if (tp->t_force && len == 1)
436 		tcpstat.tcps_sndprobe++;
437 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
438 		tcpstat.tcps_sndrexmitpack++;
439 		tcpstat.tcps_sndrexmitbyte += len;
440 	} else {
441 		tcpstat.tcps_sndpack++;
442 		tcpstat.tcps_sndbyte += len;
443 	}
444 #ifdef notyet
445 	if ((m = m_copypack(so->so_snd.sb_mb, off,
446 	    (int)len, max_linkhdr + hdrlen)) == 0)
447 		return (ENOBUFS);
448 	/*
449 	 * m_copypack left space for our hdr; use it.
450 	 */
451 	m->m_len += hdrlen;
452 	m->m_data -= hdrlen;
453 #else
454 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
455 	if (__predict_false(m == NULL))
456 		return (ENOBUFS);
457 	MCLAIM(m, &tcp_tx_mowner);
458 
459 	/*
460 	 * XXX Because other code assumes headers will fit in
461 	 * XXX one header mbuf.
462 	 *
463 	 * (This code should almost *never* be run.)
464 	 */
465 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
466 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
467 		MCLGET(m, M_DONTWAIT);
468 		if ((m->m_flags & M_EXT) == 0) {
469 			m_freem(m);
470 			return (ENOBUFS);
471 		}
472 	}
473 
474 	m->m_data += max_linkhdr;
475 	m->m_len = hdrlen;
476 
477 	/*
478 	 * To avoid traversing the whole sb_mb chain for correct
479 	 * data to send, remember last sent mbuf, its offset and
480 	 * the sent size.  When called the next time, see if the
481 	 * data to send is directly following the previous transfer.
482 	 * This is important for large TCP windows.
483 	 */
484 	if (off == 0 || tp->t_lastm == NULL ||
485 	    (tp->t_lastoff + tp->t_lastlen) != off) {
486 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
487 		/*
488 		 * Either a new packet or a retransmit.
489 		 * Start from the beginning.
490 		 */
491 		tp->t_lastm = so->so_snd.sb_mb;
492 		tp->t_inoff = off;
493 	} else {
494 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
495 		tp->t_inoff += tp->t_lastlen;
496 	}
497 
498 	/* Traverse forward to next packet */
499 	while (tp->t_inoff > 0) {
500 		if (tp->t_lastm == NULL)
501 			panic("tp->t_lastm == NULL");
502 		if (tp->t_inoff < tp->t_lastm->m_len)
503 			break;
504 		tp->t_inoff -= tp->t_lastm->m_len;
505 		tp->t_lastm = tp->t_lastm->m_next;
506 	}
507 
508 	tp->t_lastoff = off;
509 	tp->t_lastlen = len;
510 	m0 = tp->t_lastm;
511 	off = tp->t_inoff;
512 
513 	if (len <= M_TRAILINGSPACE(m)) {
514 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
515 		m->m_len += len;
516 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
517 	} else {
518 		m->m_next = m_copy(m0, off, (int) len);
519 		if (m->m_next == NULL) {
520 			m_freem(m);
521 			return (ENOBUFS);
522 		}
523 #ifdef TCP_OUTPUT_COUNTERS
524 		if (m->m_next->m_flags & M_EXT)
525 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
526 		else
527 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
528 #endif /* TCP_OUTPUT_COUNTERS */
529 	}
530 #endif
531 
532 	*mp = m;
533 	return (0);
534 }
535 
536 /*
537  * Tcp output routine: figure out what should be sent and send it.
538  */
539 int
540 tcp_output(struct tcpcb *tp)
541 {
542 	struct socket *so;
543 	struct route *ro;
544 	long len, win;
545 	int off, flags, error;
546 	struct mbuf *m;
547 	struct ip *ip;
548 #ifdef INET6
549 	struct ip6_hdr *ip6;
550 #endif
551 	struct tcphdr *th;
552 	u_char opt[MAX_TCPOPTLEN];
553 	unsigned optlen, hdrlen, packetlen;
554 	unsigned int sack_numblks;
555 	int idle, sendalot, txsegsize, rxsegsize;
556 	int txsegsize_nosack;
557 	int maxburst = TCP_MAXBURST;
558 	int af;		/* address family on the wire */
559 	int iphdrlen;
560 	int has_tso, use_tso;
561 	int sack_rxmit;
562 	int sack_bytes_rxmt;
563 	struct sackhole *p;
564 #ifdef TCP_SIGNATURE
565 	int sigoff = 0;
566 #endif
567 
568 #ifdef DIAGNOSTIC
569 	if (tp->t_inpcb && tp->t_in6pcb)
570 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
571 #endif
572 	so = NULL;
573 	ro = NULL;
574 	if (tp->t_inpcb) {
575 		so = tp->t_inpcb->inp_socket;
576 		ro = &tp->t_inpcb->inp_route;
577 	}
578 #ifdef INET6
579 	else if (tp->t_in6pcb) {
580 		so = tp->t_in6pcb->in6p_socket;
581 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
582 	}
583 #endif
584 
585 	switch (af = tp->t_family) {
586 #ifdef INET
587 	case AF_INET:
588 		if (tp->t_inpcb)
589 			break;
590 #ifdef INET6
591 		/* mapped addr case */
592 		if (tp->t_in6pcb)
593 			break;
594 #endif
595 		return (EINVAL);
596 #endif
597 #ifdef INET6
598 	case AF_INET6:
599 		if (tp->t_in6pcb)
600 			break;
601 		return (EINVAL);
602 #endif
603 	default:
604 		return (EAFNOSUPPORT);
605 	}
606 
607 	if (tcp_segsize(tp, &txsegsize, &rxsegsize))
608 		return (EMSGSIZE);
609 
610 	idle = (tp->snd_max == tp->snd_una);
611 
612 	/*
613 	 * Determine if we can use TCP segmentation offload:
614 	 * - If we're using IPv4
615 	 * - If there is not an IPsec policy that prevents it
616 	 * - If the interface can do it
617 	 */
618 	has_tso = tp->t_inpcb != NULL &&
619 #if defined(IPSEC) || defined(FAST_IPSEC)
620 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
621 		  		       IPSEC_DIR_OUTBOUND) &&
622 #endif
623 		  tp->t_inpcb->inp_route.ro_rt != NULL &&
624 		  (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable &
625 		   IFCAP_TSOv4) != 0;
626 
627 	/*
628 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
629 	 *
630 	 *	Optionally, a TCP MAY set the restart window to the
631 	 *	minimum of the value used for the initial window and
632 	 *	the current value of cwnd (in other words, using a
633 	 *	larger value for the restart window should never increase
634 	 *	the size of cwnd).
635 	 */
636 	if (tcp_cwm) {
637 		/*
638 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
639 		 * Count the number of packets currently pending
640 		 * acknowledgement, and limit our congestion window
641 		 * to a pre-determined allowed burst size plus that count.
642 		 * This prevents bursting once all pending packets have
643 		 * been acknowledged (i.e. transmission is idle).
644 		 *
645 		 * XXX Link this to Initial Window?
646 		 */
647 		tp->snd_cwnd = min(tp->snd_cwnd,
648 		    (tcp_cwm_burstsize * txsegsize) +
649 		    (tp->snd_nxt - tp->snd_una));
650 	} else {
651 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
652 			/*
653 			 * We have been idle for "a while" and no acks are
654 			 * expected to clock out any data we send --
655 			 * slow start to get ack "clock" running again.
656 			 */
657 			int ss = tcp_init_win;
658 #ifdef INET
659 			if (tp->t_inpcb &&
660 			    in_localaddr(tp->t_inpcb->inp_faddr))
661 				ss = tcp_init_win_local;
662 #endif
663 #ifdef INET6
664 			if (tp->t_in6pcb &&
665 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
666 				ss = tcp_init_win_local;
667 #endif
668 			tp->snd_cwnd = min(tp->snd_cwnd,
669 			    TCP_INITIAL_WINDOW(ss, txsegsize));
670 		}
671 	}
672 
673 	txsegsize_nosack = txsegsize;
674 again:
675 	use_tso = has_tso;
676 	TCP_REASS_LOCK(tp);
677 	sack_numblks = tcp_sack_numblks(tp);
678 	if (sack_numblks) {
679 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
680 			/* don't duplicate D-SACK. */
681 			use_tso = 0;
682 		}
683 		txsegsize = txsegsize_nosack - TCP_SACK_OPTLEN(sack_numblks);
684 	} else {
685 		txsegsize = txsegsize_nosack;
686 	}
687 
688 	/*
689 	 * Determine length of data that should be transmitted, and
690 	 * flags that should be used.  If there is some data or critical
691 	 * controls (SYN, RST) to send, then transmit; otherwise,
692 	 * investigate further.
693 	 *
694 	 * Readjust SACK information to avoid resending duplicate data.
695 	 */
696 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
697 		tcp_sack_adjust(tp);
698 	sendalot = 0;
699 	off = tp->snd_nxt - tp->snd_una;
700 	win = min(tp->snd_wnd, tp->snd_cwnd);
701 
702 	flags = tcp_outflags[tp->t_state];
703 
704 	/*
705 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
706 	 * to send out new data (when sendalot is 1), bypass this function.
707 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
708 	 * we're replacing a (future) new transmission with a retransmission
709 	 * now, and we previously incremented snd_cwnd in tcp_input().
710 	 */
711 	/*
712 	 * Still in sack recovery , reset rxmit flag to zero.
713 	 */
714 	sack_rxmit = 0;
715 	sack_bytes_rxmt = 0;
716 	len = 0;
717 	p = NULL;
718 	do {
719 		long cwin;
720 		if (!TCP_SACK_ENABLED(tp))
721 			break;
722 		if (tp->t_partialacks < 0)
723 			break;
724 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
725 		if (p == NULL)
726 			break;
727 
728 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
729 		if (cwin < 0)
730 			cwin = 0;
731 		/* Do not retransmit SACK segments beyond snd_recover */
732 		if (SEQ_GT(p->end, tp->snd_recover)) {
733 			/*
734 			 * (At least) part of sack hole extends beyond
735 			 * snd_recover. Check to see if we can rexmit data
736 			 * for this hole.
737 			 */
738 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
739 				/*
740 				 * Can't rexmit any more data for this hole.
741 				 * That data will be rexmitted in the next
742 				 * sack recovery episode, when snd_recover
743 				 * moves past p->rxmit.
744 				 */
745 				p = NULL;
746 				break;
747 			}
748 			/* Can rexmit part of the current hole */
749 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
750 		} else
751 			len = ((long)ulmin(cwin, p->end - p->rxmit));
752 		off = p->rxmit - tp->snd_una;
753 		if (off + len > so->so_snd.sb_cc) {
754 			/* 1 for TH_FIN */
755 			KASSERT(off + len == so->so_snd.sb_cc + 1);
756 			KASSERT(p->rxmit + len == tp->snd_max);
757 			len = so->so_snd.sb_cc - off;
758 		}
759 		if (len > 0) {
760 			sack_rxmit = 1;
761 			sendalot = 1;
762 		}
763 	} while (/*CONSTCOND*/0);
764 
765 	/*
766 	 * If in persist timeout with window of 0, send 1 byte.
767 	 * Otherwise, if window is small but nonzero
768 	 * and timer expired, we will send what we can
769 	 * and go to transmit state.
770 	 */
771 	if (tp->t_force) {
772 		if (win == 0) {
773 			/*
774 			 * If we still have some data to send, then
775 			 * clear the FIN bit.  Usually this would
776 			 * happen below when it realizes that we
777 			 * aren't sending all the data.  However,
778 			 * if we have exactly 1 byte of unset data,
779 			 * then it won't clear the FIN bit below,
780 			 * and if we are in persist state, we wind
781 			 * up sending the packet without recording
782 			 * that we sent the FIN bit.
783 			 *
784 			 * We can't just blindly clear the FIN bit,
785 			 * because if we don't have any more data
786 			 * to send then the probe will be the FIN
787 			 * itself.
788 			 */
789 			if (off < so->so_snd.sb_cc)
790 				flags &= ~TH_FIN;
791 			win = 1;
792 		} else {
793 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
794 			tp->t_rxtshift = 0;
795 		}
796 	}
797 
798 	if (!TCP_SACK_ENABLED(tp)) {
799 		if (win < so->so_snd.sb_cc) {
800 			len = win - off;
801 			flags &= ~TH_FIN;
802 		} else
803 			len = so->so_snd.sb_cc - off;
804 	} else if (sack_rxmit == 0) {
805 		if (sack_bytes_rxmt != 0) {
806 			long cwin;
807 
808 			/*
809 			 * We are inside of a SACK recovery episode and are
810 			 * sending new data, having retransmitted all the
811 			 * data possible in the scoreboard.
812 			 */
813 			if (tp->snd_wnd < so->so_snd.sb_cc) {
814 				len = tp->snd_wnd - off;
815 				flags &= ~TH_FIN;
816 			} else {
817 				len = so->so_snd.sb_cc - off;
818 			}
819 
820 			/*
821 			 * From FreeBSD:
822 			 *  Don't remove this (len > 0) check !
823 			 *  We explicitly check for len > 0 here (although it
824 			 *  isn't really necessary), to work around a gcc
825 			 *  optimization issue - to force gcc to compute
826 			 *  len above. Without this check, the computation
827 			 *  of len is bungled by the optimizer.
828 			 */
829 			if (len > 0) {
830 				cwin = tp->snd_cwnd -
831 						(tp->snd_nxt - tp->sack_newdata) -
832 						sack_bytes_rxmt;
833 				if (cwin < 0)
834 					cwin = 0;
835 				if (cwin < len) {
836 					len = cwin;
837 					flags &= ~TH_FIN;
838 				}
839 			}
840 		} else if (win < so->so_snd.sb_cc) {
841 			len = win - off;
842 			flags &= ~TH_FIN;
843 		} else
844 			len = so->so_snd.sb_cc - off;
845 	}
846 
847 	if (len < 0) {
848 		/*
849 		 * If FIN has been sent but not acked,
850 		 * but we haven't been called to retransmit,
851 		 * len will be -1.  Otherwise, window shrank
852 		 * after we sent into it.  If window shrank to 0,
853 		 * cancel pending retransmit, pull snd_nxt back
854 		 * to (closed) window, and set the persist timer
855 		 * if it isn't already going.  If the window didn't
856 		 * close completely, just wait for an ACK.
857 		 *
858 		 * If we have a pending FIN, either it has already been
859 		 * transmitted or it is outside the window, so drop it.
860 		 * If the FIN has been transmitted, but this is not a
861 		 * retransmission, then len must be -1.  Therefore we also
862 		 * prevent here the sending of `gratuitous FINs'.  This
863 		 * eliminates the need to check for that case below (e.g.
864 		 * to back up snd_nxt before the FIN so that the sequence
865 		 * number is correct).
866 		 */
867 		len = 0;
868 		flags &= ~TH_FIN;
869 		if (win == 0) {
870 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
871 			tp->t_rxtshift = 0;
872 			tp->snd_nxt = tp->snd_una;
873 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
874 				tcp_setpersist(tp);
875 		}
876 	}
877 	if (len > txsegsize) {
878 		if (use_tso) {
879 			/*
880 			 * Truncate TSO transfers to IP_MAXPACKET, and make
881 			 * sure that we send equal size transfers down the
882 			 * stack (rather than big-small-big-small-...).
883 			 */
884 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
885 			if (len <= txsegsize) {
886 				use_tso = 0;
887 			}
888 		} else
889 			len = txsegsize;
890 		flags &= ~TH_FIN;
891 		sendalot = 1;
892 	} else
893 		use_tso = 0;
894 	if (sack_rxmit) {
895 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
896 			flags &= ~TH_FIN;
897 	}
898 
899 	win = sbspace(&so->so_rcv);
900 
901 	/*
902 	 * Sender silly window avoidance.  If connection is idle
903 	 * and can send all data, a maximum segment,
904 	 * at least a maximum default-size segment do it,
905 	 * or are forced, do it; otherwise don't bother.
906 	 * If peer's buffer is tiny, then send
907 	 * when window is at least half open.
908 	 * If retransmitting (possibly after persist timer forced us
909 	 * to send into a small window), then must resend.
910 	 */
911 	if (len) {
912 		if (len >= txsegsize)
913 			goto send;
914 		if ((so->so_state & SS_MORETOCOME) == 0 &&
915 		    ((idle || tp->t_flags & TF_NODELAY) &&
916 		     len + off >= so->so_snd.sb_cc))
917 			goto send;
918 		if (tp->t_force)
919 			goto send;
920 		if (len >= tp->max_sndwnd / 2)
921 			goto send;
922 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
923 			goto send;
924 		if (sack_rxmit)
925 			goto send;
926 	}
927 
928 	/*
929 	 * Compare available window to amount of window known to peer
930 	 * (as advertised window less next expected input).  If the
931 	 * difference is at least twice the size of the largest segment
932 	 * we expect to receive (i.e. two segments) or at least 50% of
933 	 * the maximum possible window, then want to send a window update
934 	 * to peer.
935 	 */
936 	if (win > 0) {
937 		/*
938 		 * "adv" is the amount we can increase the window,
939 		 * taking into account that we are limited by
940 		 * TCP_MAXWIN << tp->rcv_scale.
941 		 */
942 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
943 			(tp->rcv_adv - tp->rcv_nxt);
944 
945 		if (adv >= (long) (2 * rxsegsize))
946 			goto send;
947 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
948 			goto send;
949 	}
950 
951 	/*
952 	 * Send if we owe peer an ACK.
953 	 */
954 	if (tp->t_flags & TF_ACKNOW)
955 		goto send;
956 	if (flags & (TH_SYN|TH_FIN|TH_RST))
957 		goto send;
958 	if (SEQ_GT(tp->snd_up, tp->snd_una))
959 		goto send;
960 	/*
961 	 * In SACK, it is possible for tcp_output to fail to send a segment
962 	 * after the retransmission timer has been turned off.  Make sure
963 	 * that the retransmission timer is set.
964 	 */
965 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
966 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
967 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
968 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
969 		goto just_return;
970 	}
971 
972 	/*
973 	 * TCP window updates are not reliable, rather a polling protocol
974 	 * using ``persist'' packets is used to insure receipt of window
975 	 * updates.  The three ``states'' for the output side are:
976 	 *	idle			not doing retransmits or persists
977 	 *	persisting		to move a small or zero window
978 	 *	(re)transmitting	and thereby not persisting
979 	 *
980 	 * tp->t_timer[TCPT_PERSIST]
981 	 *	is set when we are in persist state.
982 	 * tp->t_force
983 	 *	is set when we are called to send a persist packet.
984 	 * tp->t_timer[TCPT_REXMT]
985 	 *	is set when we are retransmitting
986 	 * The output side is idle when both timers are zero.
987 	 *
988 	 * If send window is too small, there is data to transmit, and no
989 	 * retransmit or persist is pending, then go to persist state.
990 	 * If nothing happens soon, send when timer expires:
991 	 * if window is nonzero, transmit what we can,
992 	 * otherwise force out a byte.
993 	 */
994 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
995 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
996 		tp->t_rxtshift = 0;
997 		tcp_setpersist(tp);
998 	}
999 
1000 	/*
1001 	 * No reason to send a segment, just return.
1002 	 */
1003 just_return:
1004 	TCP_REASS_UNLOCK(tp);
1005 	return (0);
1006 
1007 send:
1008 	/*
1009 	 * Before ESTABLISHED, force sending of initial options
1010 	 * unless TCP set not to do any options.
1011 	 * NOTE: we assume that the IP/TCP header plus TCP options
1012 	 * always fit in a single mbuf, leaving room for a maximum
1013 	 * link header, i.e.
1014 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1015 	 */
1016 	optlen = 0;
1017 	switch (af) {
1018 #ifdef INET
1019 	case AF_INET:
1020 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1021 		break;
1022 #endif
1023 #ifdef INET6
1024 	case AF_INET6:
1025 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1026 		break;
1027 #endif
1028 	default:	/*pacify gcc*/
1029 		iphdrlen = 0;
1030 		break;
1031 	}
1032 	hdrlen = iphdrlen;
1033 	if (flags & TH_SYN) {
1034 		struct rtentry *rt;
1035 
1036 		rt = NULL;
1037 #ifdef INET
1038 		if (tp->t_inpcb)
1039 			rt = in_pcbrtentry(tp->t_inpcb);
1040 #endif
1041 #ifdef INET6
1042 		if (tp->t_in6pcb)
1043 			rt = in6_pcbrtentry(tp->t_in6pcb);
1044 #endif
1045 
1046 		tp->snd_nxt = tp->iss;
1047 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
1048 						    rt->rt_ifp : NULL, af);
1049 		if ((tp->t_flags & TF_NOOPT) == 0) {
1050 			opt[0] = TCPOPT_MAXSEG;
1051 			opt[1] = 4;
1052 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1053 			opt[3] = tp->t_ourmss & 0xff;
1054 			optlen = 4;
1055 
1056 			if ((tp->t_flags & TF_REQ_SCALE) &&
1057 			    ((flags & TH_ACK) == 0 ||
1058 			    (tp->t_flags & TF_RCVD_SCALE))) {
1059 				*((u_int32_t *) (opt + optlen)) = htonl(
1060 					TCPOPT_NOP << 24 |
1061 					TCPOPT_WINDOW << 16 |
1062 					TCPOLEN_WINDOW << 8 |
1063 					tp->request_r_scale);
1064 				optlen += 4;
1065 			}
1066 			if (tcp_do_sack) {
1067 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1068 
1069 				cp[0] = TCPOPT_SACK_PERMITTED;
1070 				cp[1] = 2;
1071 				cp[2] = TCPOPT_NOP;
1072 				cp[3] = TCPOPT_NOP;
1073 				optlen += 4;
1074 			}
1075 		}
1076 	}
1077 
1078 	/*
1079 	 * Send a timestamp and echo-reply if this is a SYN and our side
1080 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1081 	 * and our peer have sent timestamps in our SYN's.
1082 	 */
1083 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1084 	     (flags & TH_RST) == 0 &&
1085 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1086 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1087 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1088 
1089 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1090 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1091 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1092 		*lp   = htonl(tp->ts_recent);
1093 		optlen += TCPOLEN_TSTAMP_APPA;
1094 	}
1095 
1096 	/*
1097 	 * Tack on the SACK block if it is necessary.
1098 	 */
1099 	if (sack_numblks) {
1100 		int sack_len;
1101 		u_char *bp = (u_char *)(opt + optlen);
1102 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1103 		struct ipqent *tiqe;
1104 
1105 		sack_len = sack_numblks * 8 + 2;
1106 		bp[0] = TCPOPT_NOP;
1107 		bp[1] = TCPOPT_NOP;
1108 		bp[2] = TCPOPT_SACK;
1109 		bp[3] = sack_len;
1110 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1111 			sack_numblks--;
1112 			*lp++ = htonl(tp->rcv_dsack_block.left);
1113 			*lp++ = htonl(tp->rcv_dsack_block.right);
1114 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1115 		}
1116 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1117 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1118 			KASSERT(tiqe != NULL);
1119 			sack_numblks--;
1120 			*lp++ = htonl(tiqe->ipqe_seq);
1121 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1122 			    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1123 		}
1124 		optlen += sack_len + 2;
1125 	}
1126 	TCP_REASS_UNLOCK(tp);
1127 
1128 #ifdef TCP_SIGNATURE
1129 #if defined(INET6) && defined(FAST_IPSEC)
1130 	if (tp->t_family == AF_INET)
1131 #endif
1132 	if (tp->t_flags & TF_SIGNATURE) {
1133 		u_char *bp;
1134 		/*
1135 		 * Initialize TCP-MD5 option (RFC2385)
1136 		 */
1137 		bp = (u_char *)opt + optlen;
1138 		*bp++ = TCPOPT_SIGNATURE;
1139 		*bp++ = TCPOLEN_SIGNATURE;
1140 		sigoff = optlen + 2;
1141 		bzero(bp, TCP_SIGLEN);
1142 		bp += TCP_SIGLEN;
1143 		optlen += TCPOLEN_SIGNATURE;
1144 		/*
1145 		 * Terminate options list and maintain 32-bit alignment.
1146  		 */
1147 		*bp++ = TCPOPT_NOP;
1148 		*bp++ = TCPOPT_EOL;
1149  		optlen += 2;
1150  	}
1151 #endif /* TCP_SIGNATURE */
1152 
1153 	hdrlen += optlen;
1154 
1155 #ifdef DIAGNOSTIC
1156 	if (!use_tso && len > txsegsize)
1157 		panic("tcp data to be sent is larger than segment");
1158 	else if (use_tso && len > IP_MAXPACKET)
1159 		panic("tcp data to be sent is larger than max TSO size");
1160 	if (max_linkhdr + hdrlen > MCLBYTES)
1161 		panic("tcphdr too big");
1162 #endif
1163 
1164 	/*
1165 	 * Grab a header mbuf, attaching a copy of data to
1166 	 * be transmitted, and initialize the header from
1167 	 * the template for sends on this connection.
1168 	 */
1169 	if (len) {
1170 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1171 		if (error)
1172 			goto out;
1173 		/*
1174 		 * If we're sending everything we've got, set PUSH.
1175 		 * (This will keep happy those implementations which only
1176 		 * give data to the user when a buffer fills or
1177 		 * a PUSH comes in.)
1178 		 */
1179 		if (off + len == so->so_snd.sb_cc)
1180 			flags |= TH_PUSH;
1181 	} else {
1182 		if (tp->t_flags & TF_ACKNOW)
1183 			tcpstat.tcps_sndacks++;
1184 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1185 			tcpstat.tcps_sndctrl++;
1186 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1187 			tcpstat.tcps_sndurg++;
1188 		else
1189 			tcpstat.tcps_sndwinup++;
1190 
1191 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1192 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1193 			MCLGET(m, M_DONTWAIT);
1194 			if ((m->m_flags & M_EXT) == 0) {
1195 				m_freem(m);
1196 				m = NULL;
1197 			}
1198 		}
1199 		if (m == NULL) {
1200 			error = ENOBUFS;
1201 			goto out;
1202 		}
1203 		MCLAIM(m, &tcp_tx_mowner);
1204 		m->m_data += max_linkhdr;
1205 		m->m_len = hdrlen;
1206 	}
1207 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1208 	switch (af) {
1209 #ifdef INET
1210 	case AF_INET:
1211 		ip = mtod(m, struct ip *);
1212 #ifdef INET6
1213 		ip6 = NULL;
1214 #endif
1215 		th = (struct tcphdr *)(ip + 1);
1216 		break;
1217 #endif
1218 #ifdef INET6
1219 	case AF_INET6:
1220 		ip = NULL;
1221 		ip6 = mtod(m, struct ip6_hdr *);
1222 		th = (struct tcphdr *)(ip6 + 1);
1223 		break;
1224 #endif
1225 	default:	/*pacify gcc*/
1226 		ip = NULL;
1227 #ifdef INET6
1228 		ip6 = NULL;
1229 #endif
1230 		th = NULL;
1231 		break;
1232 	}
1233 	if (tp->t_template == 0)
1234 		panic("tcp_output");
1235 	if (tp->t_template->m_len < iphdrlen)
1236 		panic("tcp_output");
1237 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1238 
1239 	/*
1240 	 * If we are doing retransmissions, then snd_nxt will
1241 	 * not reflect the first unsent octet.  For ACK only
1242 	 * packets, we do not want the sequence number of the
1243 	 * retransmitted packet, we want the sequence number
1244 	 * of the next unsent octet.  So, if there is no data
1245 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1246 	 * when filling in ti_seq.  But if we are in persist
1247 	 * state, snd_max might reflect one byte beyond the
1248 	 * right edge of the window, so use snd_nxt in that
1249 	 * case, since we know we aren't doing a retransmission.
1250 	 * (retransmit and persist are mutually exclusive...)
1251 	 */
1252 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1253 		th->th_seq = htonl(p->rxmit);
1254 		p->rxmit += len;
1255 	} else {
1256 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1257 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1258 			th->th_seq = htonl(tp->snd_nxt);
1259 		else
1260 			th->th_seq = htonl(tp->snd_max);
1261 	}
1262 	th->th_ack = htonl(tp->rcv_nxt);
1263 	if (optlen) {
1264 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1265 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1266 	}
1267 	th->th_flags = flags;
1268 	/*
1269 	 * Calculate receive window.  Don't shrink window,
1270 	 * but avoid silly window syndrome.
1271 	 */
1272 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1273 		win = 0;
1274 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1275 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1276 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1277 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1278 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1279 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1280 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1281 		if (urp > IP_MAXPACKET)
1282 			urp = IP_MAXPACKET;
1283 		th->th_urp = htons((u_int16_t)urp);
1284 		th->th_flags |= TH_URG;
1285 	} else
1286 		/*
1287 		 * If no urgent pointer to send, then we pull
1288 		 * the urgent pointer to the left edge of the send window
1289 		 * so that it doesn't drift into the send window on sequence
1290 		 * number wraparound.
1291 		 */
1292 		tp->snd_up = tp->snd_una;		/* drag it along */
1293 
1294 #ifdef TCP_SIGNATURE
1295 #if defined(INET6) && defined(FAST_IPSEC)
1296 	if (tp->t_family == AF_INET) /* XXX */
1297 #endif
1298 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1299 		struct secasvar *sav;
1300 		u_int8_t *sigp;
1301 
1302 		sav = tcp_signature_getsav(m, th);
1303 
1304 		if (sav == NULL) {
1305 			if (m)
1306 				m_freem(m);
1307 			return (EPERM);
1308 		}
1309 
1310 		m->m_pkthdr.len = hdrlen + len;
1311 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1312 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1313 
1314 		key_sa_recordxfer(sav, m);
1315 #ifdef FAST_IPSEC
1316 		KEY_FREESAV(&sav);
1317 #else
1318 		key_freesav(sav);
1319 #endif
1320 	}
1321 #endif
1322 
1323 	/*
1324 	 * Set ourselves up to be checksummed just before the packet
1325 	 * hits the wire.
1326 	 */
1327 	switch (af) {
1328 #ifdef INET
1329 	case AF_INET:
1330 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1331 		if (use_tso) {
1332 			m->m_pkthdr.segsz = txsegsize;
1333 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1334 		} else {
1335 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1336 			if (len + optlen) {
1337 				/* Fixup the pseudo-header checksum. */
1338 				/* XXXJRT Not IP Jumbogram safe. */
1339 				th->th_sum = in_cksum_addword(th->th_sum,
1340 				    htons((u_int16_t) (len + optlen)));
1341 			}
1342 		}
1343 		break;
1344 #endif
1345 #ifdef INET6
1346 	case AF_INET6:
1347 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1348 		m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1349 		if (len + optlen) {
1350 			/* Fixup the pseudo-header checksum. */
1351 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1352 			th->th_sum = in_cksum_addword(th->th_sum,
1353 			    htons((u_int16_t) (len + optlen)));
1354 		}
1355 		break;
1356 #endif
1357 	}
1358 
1359 	/*
1360 	 * In transmit state, time the transmission and arrange for
1361 	 * the retransmit.  In persist state, just set snd_max.
1362 	 */
1363 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1364 		tcp_seq startseq = tp->snd_nxt;
1365 
1366 		/*
1367 		 * Advance snd_nxt over sequence space of this segment.
1368 		 * There are no states in which we send both a SYN and a FIN,
1369 		 * so we collapse the tests for these flags.
1370 		 */
1371 		if (flags & (TH_SYN|TH_FIN))
1372 			tp->snd_nxt++;
1373 		if (sack_rxmit)
1374 			goto timer;
1375 		tp->snd_nxt += len;
1376 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1377 			tp->snd_max = tp->snd_nxt;
1378 			/*
1379 			 * Time this transmission if not a retransmission and
1380 			 * not currently timing anything.
1381 			 */
1382 			if (tp->t_rtttime == 0) {
1383 				tp->t_rtttime = tcp_now;
1384 				tp->t_rtseq = startseq;
1385 				tcpstat.tcps_segstimed++;
1386 			}
1387 		}
1388 
1389 		/*
1390 		 * Set retransmit timer if not currently set,
1391 		 * and not doing an ack or a keep-alive probe.
1392 		 * Initial value for retransmit timer is smoothed
1393 		 * round-trip time + 2 * round-trip time variance.
1394 		 * Initialize shift counter which is used for backoff
1395 		 * of retransmit time.
1396 		 */
1397 timer:
1398 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1399 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1400 		    tp->snd_nxt != tp->snd_una)) {
1401 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1402 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1403 				tp->t_rxtshift = 0;
1404 			}
1405 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1406 		}
1407 	} else
1408 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1409 			tp->snd_max = tp->snd_nxt + len;
1410 
1411 #ifdef TCP_DEBUG
1412 	/*
1413 	 * Trace.
1414 	 */
1415 	if (so->so_options & SO_DEBUG)
1416 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1417 #endif
1418 
1419 	/*
1420 	 * Fill in IP length and desired time to live and
1421 	 * send to IP level.  There should be a better way
1422 	 * to handle ttl and tos; we could keep them in
1423 	 * the template, but need a way to checksum without them.
1424 	 */
1425 	m->m_pkthdr.len = hdrlen + len;
1426 
1427 	switch (af) {
1428 #ifdef INET
1429 	case AF_INET:
1430 		ip->ip_len = htons(m->m_pkthdr.len);
1431 		packetlen = m->m_pkthdr.len;
1432 		if (tp->t_inpcb) {
1433 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1434 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1435 		}
1436 #ifdef INET6
1437 		else if (tp->t_in6pcb) {
1438 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1439 			ip->ip_tos = 0;	/*XXX*/
1440 		}
1441 #endif
1442 		break;
1443 #endif
1444 #ifdef INET6
1445 	case AF_INET6:
1446 		packetlen = m->m_pkthdr.len;
1447 		ip6->ip6_nxt = IPPROTO_TCP;
1448 		if (tp->t_in6pcb) {
1449 			/*
1450 			 * we separately set hoplimit for every segment, since
1451 			 * the user might want to change the value via
1452 			 * setsockopt. Also, desired default hop limit might
1453 			 * be changed via Neighbor Discovery.
1454 			 */
1455 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1456 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1457 		}
1458 		/* ip6->ip6_flow = ??? */
1459 		/* ip6_plen will be filled in ip6_output(). */
1460 		break;
1461 #endif
1462 	default:	/*pacify gcc*/
1463 		packetlen = 0;
1464 		break;
1465 	}
1466 
1467 	switch (af) {
1468 #ifdef INET
1469 	case AF_INET:
1470 	    {
1471 		struct mbuf *opts;
1472 
1473 		if (tp->t_inpcb)
1474 			opts = tp->t_inpcb->inp_options;
1475 		else
1476 			opts = NULL;
1477 		error = ip_output(m, opts, ro,
1478 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1479 			(so->so_options & SO_DONTROUTE),
1480 			(struct ip_moptions *)0, so);
1481 		break;
1482 	    }
1483 #endif
1484 #ifdef INET6
1485 	case AF_INET6:
1486 	    {
1487 		struct ip6_pktopts *opts;
1488 
1489 		if (tp->t_in6pcb)
1490 			opts = tp->t_in6pcb->in6p_outputopts;
1491 		else
1492 			opts = NULL;
1493 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1494 			so->so_options & SO_DONTROUTE,
1495 			(struct ip6_moptions *)0, so, NULL);
1496 		break;
1497 	    }
1498 #endif
1499 	default:
1500 		error = EAFNOSUPPORT;
1501 		break;
1502 	}
1503 	if (error) {
1504 out:
1505 		if (error == ENOBUFS) {
1506 			tcpstat.tcps_selfquench++;
1507 #ifdef INET
1508 			if (tp->t_inpcb)
1509 				tcp_quench(tp->t_inpcb, 0);
1510 #endif
1511 #ifdef INET6
1512 			if (tp->t_in6pcb)
1513 				tcp6_quench(tp->t_in6pcb, 0);
1514 #endif
1515 			error = 0;
1516 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1517 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1518 			tp->t_softerror = error;
1519 			error = 0;
1520 		}
1521 
1522 		/* Back out the seqence number advance. */
1523 		if (sack_rxmit)
1524 			p->rxmit -= len;
1525 
1526 		/* Restart the delayed ACK timer, if necessary. */
1527 		if (tp->t_flags & TF_DELACK)
1528 			TCP_RESTART_DELACK(tp);
1529 
1530 		return (error);
1531 	}
1532 
1533 	if (packetlen > tp->t_pmtud_mtu_sent)
1534 		tp->t_pmtud_mtu_sent = packetlen;
1535 
1536 	tcpstat.tcps_sndtotal++;
1537 	if (tp->t_flags & TF_DELACK)
1538 		tcpstat.tcps_delack++;
1539 
1540 	/*
1541 	 * Data sent (as far as we can tell).
1542 	 * If this advertises a larger window than any other segment,
1543 	 * then remember the size of the advertised window.
1544 	 * Any pending ACK has now been sent.
1545 	 */
1546 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1547 		tp->rcv_adv = tp->rcv_nxt + win;
1548 	tp->last_ack_sent = tp->rcv_nxt;
1549 	tp->t_flags &= ~TF_ACKNOW;
1550 	TCP_CLEAR_DELACK(tp);
1551 #ifdef DIAGNOSTIC
1552 	if (maxburst < 0)
1553 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1554 #endif
1555 	if (sendalot && (!tcp_do_newreno || --maxburst))
1556 		goto again;
1557 	return (0);
1558 }
1559 
1560 void
1561 tcp_setpersist(struct tcpcb *tp)
1562 {
1563 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1564 	int nticks;
1565 
1566 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1567 		panic("tcp_output REXMT");
1568 	/*
1569 	 * Start/restart persistance timer.
1570 	 */
1571 	if (t < tp->t_rttmin)
1572 		t = tp->t_rttmin;
1573 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1574 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1575 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1576 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1577 		tp->t_rxtshift++;
1578 }
1579 
1580 #if defined(INET)
1581 /*
1582  * tcp4_segment: handle M_CSUM_TSOv4 by software.
1583  *
1584  * => always consume m.
1585  * => call output_func with output_arg for each segments.
1586  */
1587 
1588 int
1589 tcp4_segment(struct mbuf *m, int (*output_func)(void *, struct mbuf *),
1590     void *output_arg)
1591 {
1592 	int mss;
1593 	int iphlen;
1594 	int thlen;
1595 	int hlen;
1596 	int len;
1597 	struct ip *iph;
1598 	struct tcphdr *th;
1599 	uint16_t ipid;
1600 	uint32_t tcpseq;
1601 	struct mbuf *hdr = NULL;
1602 	struct mbuf *t;
1603 	int error = 0;
1604 
1605 	KASSERT((m->m_flags & M_PKTHDR) != 0);
1606 	KASSERT((m->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0);
1607 
1608 	m->m_pkthdr.csum_flags = 0;
1609 
1610 	len = m->m_pkthdr.len;
1611 	KASSERT(len >= sizeof(*iph) + sizeof(*th));
1612 
1613 	if (m->m_len < sizeof(*iph)) {
1614 		m = m_pullup(m, sizeof(*iph));
1615 		if (m == NULL) {
1616 			error = ENOMEM;
1617 			goto quit;
1618 		}
1619 	}
1620 	iph = mtod(m, struct ip *);
1621 	iphlen = iph->ip_hl * 4;
1622 	KASSERT(iph->ip_v == IPVERSION);
1623 	KASSERT(iphlen >= sizeof(*iph));
1624 	KASSERT(iph->ip_p == IPPROTO_TCP);
1625 	ipid = ntohs(iph->ip_id);
1626 
1627 	hlen = iphlen + sizeof(*th);
1628 	if (m->m_len < hlen) {
1629 		m = m_pullup(m, hlen);
1630 		if (m == NULL) {
1631 			error = ENOMEM;
1632 			goto quit;
1633 		}
1634 	}
1635 	th = (void *)(mtod(m, char *) + iphlen);
1636 	tcpseq = ntohl(th->th_seq);
1637 	thlen = th->th_off * 4;
1638 	hlen = iphlen + thlen;
1639 
1640 	mss = m->m_pkthdr.segsz;
1641 	KASSERT(mss != 0);
1642 	KASSERT(len > hlen);
1643 
1644 	t = m_split(m, hlen, M_NOWAIT);
1645 	if (t == NULL) {
1646 		error = ENOMEM;
1647 		goto quit;
1648 	}
1649 	hdr = m;
1650 	m = t;
1651 	len -= hlen;
1652 	KASSERT(len % mss == 0);
1653 	while (len > 0) {
1654 		struct mbuf *n;
1655 
1656 		n = m_dup(hdr, 0, hlen, M_NOWAIT);
1657 		if (n == NULL) {
1658 			error = ENOMEM;
1659 			goto quit;
1660 		}
1661 		KASSERT(n->m_len == hlen); /* XXX */
1662 
1663 		t = m_split(m, mss, M_NOWAIT);
1664 		if (t == NULL) {
1665 			m_freem(n);
1666 			error = ENOMEM;
1667 			goto quit;
1668 		}
1669 		m_cat(n, m);
1670 		m = t;
1671 
1672 		KASSERT(n->m_len >= hlen); /* XXX */
1673 
1674 		n->m_pkthdr.len = hlen + mss;
1675 		iph = mtod(n, struct ip *);
1676 		KASSERT(iph->ip_v == IPVERSION);
1677 		iph->ip_len = htons(n->m_pkthdr.len);
1678 		iph->ip_id = htons(ipid);
1679 		th = (void *)(mtod(n, char *) + iphlen);
1680 		th->th_seq = htonl(tcpseq);
1681 		iph->ip_sum = 0;
1682 		iph->ip_sum = in_cksum(n, iphlen);
1683 		th->th_sum = 0;
1684 		th->th_sum = in4_cksum(n, IPPROTO_TCP, iphlen, thlen + mss);
1685 
1686 		error = (*output_func)(output_arg, n);
1687 		if (error) {
1688 			goto quit;
1689 		}
1690 
1691 		tcpseq += mss;
1692 		ipid++;
1693 		len -= mss;
1694 	}
1695 
1696 quit:
1697 	if (hdr != NULL) {
1698 		m_freem(hdr);
1699 	}
1700 	if (m != NULL) {
1701 		m_freem(m);
1702 	}
1703 
1704 	return error;
1705 }
1706 #endif /* defined(INET) */
1707