xref: /netbsd-src/sys/netinet/tcp_output.c (revision c9496f6b604074a9451a67df576a5b423068e71e)
1 /*	$NetBSD: tcp_output.c,v 1.197 2017/08/03 06:32:51 ozaki-r 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  *
93  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
94  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
95  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
96  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
97  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
98  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
99  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
100  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
101  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
102  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
103  * POSSIBILITY OF SUCH DAMAGE.
104  */
105 
106 /*
107  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
108  *	The Regents of the University of California.  All rights reserved.
109  *
110  * Redistribution and use in source and binary forms, with or without
111  * modification, are permitted provided that the following conditions
112  * are met:
113  * 1. Redistributions of source code must retain the above copyright
114  *    notice, this list of conditions and the following disclaimer.
115  * 2. Redistributions in binary form must reproduce the above copyright
116  *    notice, this list of conditions and the following disclaimer in the
117  *    documentation and/or other materials provided with the distribution.
118  * 3. Neither the name of the University nor the names of its contributors
119  *    may be used to endorse or promote products derived from this software
120  *    without specific prior written permission.
121  *
122  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
123  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
124  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
125  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
126  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
127  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
128  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
129  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
130  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
131  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
132  * SUCH DAMAGE.
133  *
134  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
135  */
136 
137 #include <sys/cdefs.h>
138 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.197 2017/08/03 06:32:51 ozaki-r Exp $");
139 
140 #ifdef _KERNEL_OPT
141 #include "opt_inet.h"
142 #include "opt_ipsec.h"
143 #include "opt_tcp_debug.h"
144 #endif
145 
146 #include <sys/param.h>
147 #include <sys/systm.h>
148 #include <sys/mbuf.h>
149 #include <sys/protosw.h>
150 #include <sys/socket.h>
151 #include <sys/socketvar.h>
152 #include <sys/errno.h>
153 #include <sys/domain.h>
154 #include <sys/kernel.h>
155 #ifdef TCP_SIGNATURE
156 #include <sys/md5.h>
157 #endif
158 
159 #include <net/if.h>
160 #include <net/route.h>
161 
162 #include <netinet/in.h>
163 #include <netinet/in_systm.h>
164 #include <netinet/ip.h>
165 #include <netinet/in_pcb.h>
166 #include <netinet/ip_var.h>
167 
168 #ifdef INET6
169 #ifndef INET
170 #include <netinet/in.h>
171 #endif
172 #include <netinet/ip6.h>
173 #include <netinet6/in6_var.h>
174 #include <netinet6/ip6_var.h>
175 #include <netinet6/in6_pcb.h>
176 #include <netinet6/nd6.h>
177 #endif
178 
179 #ifdef IPSEC
180 #include <netipsec/ipsec.h>
181 #include <netipsec/key.h>
182 #ifdef INET6
183 #include <netipsec/ipsec6.h>
184 #endif
185 #endif	/* IPSEC*/
186 
187 #include <netinet/tcp.h>
188 #define	TCPOUTFLAGS
189 #include <netinet/tcp_fsm.h>
190 #include <netinet/tcp_seq.h>
191 #include <netinet/tcp_timer.h>
192 #include <netinet/tcp_var.h>
193 #include <netinet/tcp_private.h>
194 #include <netinet/tcp_congctl.h>
195 #include <netinet/tcpip.h>
196 #include <netinet/tcp_debug.h>
197 #include <netinet/in_offload.h>
198 #include <netinet6/in6_offload.h>
199 
200 #ifdef notyet
201 extern struct mbuf *m_copypack();
202 #endif
203 
204 /*
205  * Knob to enable Congestion Window Monitoring, and control
206  * the burst size it allows.  Default burst is 4 packets, per
207  * the Internet draft.
208  */
209 int	tcp_cwm = 0;
210 int	tcp_cwm_burstsize = 4;
211 
212 int	tcp_do_autosndbuf = 1;
213 int	tcp_autosndbuf_inc = 8 * 1024;
214 int	tcp_autosndbuf_max = 256 * 1024;
215 
216 #ifdef TCP_OUTPUT_COUNTERS
217 #include <sys/device.h>
218 
219 extern struct evcnt tcp_output_bigheader;
220 extern struct evcnt tcp_output_predict_hit;
221 extern struct evcnt tcp_output_predict_miss;
222 extern struct evcnt tcp_output_copysmall;
223 extern struct evcnt tcp_output_copybig;
224 extern struct evcnt tcp_output_refbig;
225 
226 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
227 #else
228 
229 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
230 
231 #endif /* TCP_OUTPUT_COUNTERS */
232 
233 static
234 #ifndef GPROF
235 inline
236 #endif
237 int
238 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
239     bool *alwaysfragp)
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 hdrlen;
252 	int optlen;
253 
254 	*alwaysfragp = false;
255 
256 #ifdef DIAGNOSTIC
257 	if (tp->t_inpcb && tp->t_in6pcb)
258 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
259 #endif
260 	switch (tp->t_family) {
261 #ifdef INET
262 	case AF_INET:
263 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
264 		break;
265 #endif
266 #ifdef INET6
267 	case AF_INET6:
268 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
269 		break;
270 #endif
271 	default:
272 		size = tcp_mssdflt;
273 		goto out;
274 	}
275 
276 	rt = NULL;
277 #ifdef INET
278 	if (inp) {
279 		rt = in_pcbrtentry(inp);
280 		so = inp->inp_socket;
281 	}
282 #endif
283 #ifdef INET6
284 	if (in6p) {
285 		rt = in6_pcbrtentry(in6p);
286 		so = in6p->in6p_socket;
287 	}
288 #endif
289 	if (rt == NULL) {
290 		size = tcp_mssdflt;
291 		goto out;
292 	}
293 
294 	ifp = rt->rt_ifp;
295 
296 	size = tcp_mssdflt;
297 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
298 #ifdef INET6
299 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
300 			/*
301 			 * RFC2460 section 5, last paragraph: if path MTU is
302 			 * smaller than 1280, use 1280 as packet size and
303 			 * attach fragment header.
304 			 */
305 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
306 			*alwaysfragp = true;
307 		} else
308 			size = rt->rt_rmx.rmx_mtu - hdrlen;
309 #else
310 		size = rt->rt_rmx.rmx_mtu - hdrlen;
311 #endif
312 	} else if (ifp->if_flags & IFF_LOOPBACK)
313 		size = ifp->if_mtu - hdrlen;
314 #ifdef INET
315 	else if (inp && tp->t_mtudisc)
316 		size = ifp->if_mtu - hdrlen;
317 	else if (inp && in_localaddr(inp->inp_faddr))
318 		size = ifp->if_mtu - hdrlen;
319 #endif
320 #ifdef INET6
321 	else if (in6p) {
322 #ifdef INET
323 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
324 			/* mapped addr case */
325 			struct in_addr d;
326 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
327 			if (tp->t_mtudisc || in_localaddr(d))
328 				size = ifp->if_mtu - hdrlen;
329 		} else
330 #endif
331 		{
332 			/*
333 			 * for IPv6, path MTU discovery is always turned on,
334 			 * or the node must use packet size <= 1280.
335 			 */
336 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
337 			size -= hdrlen;
338 		}
339 	}
340 #endif
341 #ifdef INET
342 	if (inp)
343 		in_pcbrtentry_unref(rt, inp);
344 #endif
345 #ifdef INET6
346 	if (in6p)
347 		in6_pcbrtentry_unref(rt, in6p);
348 #endif
349  out:
350 	/*
351 	 * Now we must make room for whatever extra TCP/IP options are in
352 	 * the packet.
353 	 */
354 	optlen = tcp_optlen(tp);
355 
356 	/*
357 	 * XXX tp->t_ourmss should have the right size, but without this code
358 	 * fragmentation will occur... need more investigation
359 	 */
360 #ifdef INET
361 	if (inp) {
362 #if defined(IPSEC)
363 		if (ipsec_used &&
364 		    !ipsec_pcb_skip_ipsec(inp->inp_sp, IPSEC_DIR_OUTBOUND))
365 			optlen += ipsec4_hdrsiz_tcp(tp);
366 #endif
367 		optlen += ip_optlen(inp);
368 	}
369 #endif
370 #ifdef INET6
371 #ifdef INET
372 	if (in6p && tp->t_family == AF_INET) {
373 #if defined(IPSEC)
374 		if (ipsec_used &&
375 		    !ipsec_pcb_skip_ipsec(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
376 			optlen += ipsec4_hdrsiz_tcp(tp);
377 #endif
378 		/* XXX size -= ip_optlen(in6p); */
379 	} else
380 #endif
381 	if (in6p && tp->t_family == AF_INET6) {
382 #if defined(IPSEC)
383 		if (ipsec_used &&
384 		    !ipsec_pcb_skip_ipsec(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
385 			optlen += ipsec6_hdrsiz_tcp(tp);
386 #endif
387 		optlen += ip6_optlen(in6p);
388 	}
389 #endif
390 	size -= optlen;
391 
392 	/* there may not be any room for data if mtu is too small */
393 	if (size < 0)
394 		return (EMSGSIZE);
395 
396 	/*
397 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
398 	 * assumes that path MTU is the same for both ways).  this is only
399 	 * for silly window avoidance, do not use the value for other purposes.
400 	 *
401 	 * ipseclen is subtracted from both sides, this may not be right.
402 	 * I'm not quite sure about this (could someone comment).
403 	 */
404 	*txsegsizep = min(tp->t_peermss - optlen, size);
405 	/*
406 	 * Never send more than half a buffer full.  This insures that we can
407 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
408 	 * therefore acks will never be delayed unless we run out of data to
409 	 * transmit.
410 	 */
411 	if (so)
412 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
413 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
414 
415 	if (*txsegsizep != tp->t_segsz) {
416 		/*
417 		 * If the new segment size is larger, we don't want to
418 		 * mess up the congestion window, but if it is smaller
419 		 * we'll have to reduce the congestion window to ensure
420 		 * that we don't get into trouble with initial windows
421 		 * and the rest.  In any case, if the segment size
422 		 * has changed, chances are the path has, too, and
423 		 * our congestion window will be different.
424 		 */
425 		if (*txsegsizep < tp->t_segsz) {
426 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
427 					   * *txsegsizep, *txsegsizep);
428 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
429 						* *txsegsizep, *txsegsizep);
430 		}
431 		tp->t_segsz = *txsegsizep;
432 	}
433 
434 	return (0);
435 }
436 
437 static
438 #ifndef GPROF
439 inline
440 #endif
441 int
442 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
443     long len, int hdrlen, struct mbuf **mp)
444 {
445 	struct mbuf *m, *m0;
446 	uint64_t *tcps;
447 
448 	tcps = TCP_STAT_GETREF();
449 	if (tp->t_force && len == 1)
450 		tcps[TCP_STAT_SNDPROBE]++;
451 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
452 		tp->t_sndrexmitpack++;
453 		tcps[TCP_STAT_SNDREXMITPACK]++;
454 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
455 	} else {
456 		tcps[TCP_STAT_SNDPACK]++;
457 		tcps[TCP_STAT_SNDBYTE] += len;
458 	}
459 	TCP_STAT_PUTREF();
460 #ifdef notyet
461 	if ((m = m_copypack(so->so_snd.sb_mb, off,
462 	    (int)len, max_linkhdr + hdrlen)) == 0)
463 		return (ENOBUFS);
464 	/*
465 	 * m_copypack left space for our hdr; use it.
466 	 */
467 	m->m_len += hdrlen;
468 	m->m_data -= hdrlen;
469 #else
470 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
471 	if (__predict_false(m == NULL))
472 		return (ENOBUFS);
473 	MCLAIM(m, &tcp_tx_mowner);
474 
475 	/*
476 	 * XXX Because other code assumes headers will fit in
477 	 * XXX one header mbuf.
478 	 *
479 	 * (This code should almost *never* be run.)
480 	 */
481 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
482 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
483 		MCLGET(m, M_DONTWAIT);
484 		if ((m->m_flags & M_EXT) == 0) {
485 			m_freem(m);
486 			return (ENOBUFS);
487 		}
488 	}
489 
490 	m->m_data += max_linkhdr;
491 	m->m_len = hdrlen;
492 
493 	/*
494 	 * To avoid traversing the whole sb_mb chain for correct
495 	 * data to send, remember last sent mbuf, its offset and
496 	 * the sent size.  When called the next time, see if the
497 	 * data to send is directly following the previous transfer.
498 	 * This is important for large TCP windows.
499 	 */
500 	if (off == 0 || tp->t_lastm == NULL ||
501 	    (tp->t_lastoff + tp->t_lastlen) != off) {
502 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
503 		/*
504 		 * Either a new packet or a retransmit.
505 		 * Start from the beginning.
506 		 */
507 		tp->t_lastm = so->so_snd.sb_mb;
508 		tp->t_inoff = off;
509 	} else {
510 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
511 		tp->t_inoff += tp->t_lastlen;
512 	}
513 
514 	/* Traverse forward to next packet */
515 	while (tp->t_inoff > 0) {
516 		if (tp->t_lastm == NULL)
517 			panic("tp->t_lastm == NULL");
518 		if (tp->t_inoff < tp->t_lastm->m_len)
519 			break;
520 		tp->t_inoff -= tp->t_lastm->m_len;
521 		tp->t_lastm = tp->t_lastm->m_next;
522 	}
523 
524 	tp->t_lastoff = off;
525 	tp->t_lastlen = len;
526 	m0 = tp->t_lastm;
527 	off = tp->t_inoff;
528 
529 	if (len <= M_TRAILINGSPACE(m)) {
530 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
531 		m->m_len += len;
532 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
533 	} else {
534 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
535 		if (m->m_next == NULL) {
536 			m_freem(m);
537 			return (ENOBUFS);
538 		}
539 #ifdef TCP_OUTPUT_COUNTERS
540 		if (m->m_next->m_flags & M_EXT)
541 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
542 		else
543 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
544 #endif /* TCP_OUTPUT_COUNTERS */
545 	}
546 #endif
547 
548 	*mp = m;
549 	return (0);
550 }
551 
552 /*
553  * Tcp output routine: figure out what should be sent and send it.
554  */
555 int
556 tcp_output(struct tcpcb *tp)
557 {
558 	struct rtentry *rt = NULL;
559 	struct socket *so;
560 	struct route *ro;
561 	long len, win;
562 	int off, flags, error;
563 	struct mbuf *m;
564 	struct ip *ip;
565 #ifdef INET6
566 	struct ip6_hdr *ip6;
567 #endif
568 	struct tcphdr *th;
569 	u_char opt[MAX_TCPOPTLEN], *optp;
570 #define OPT_FITS(more)	((optlen + (more)) <= sizeof(opt))
571 	unsigned optlen, hdrlen, packetlen;
572 	unsigned int sack_numblks;
573 	int idle, sendalot, txsegsize, rxsegsize;
574 	int txsegsize_nosack;
575 	int maxburst = TCP_MAXBURST;
576 	int af;		/* address family on the wire */
577 	int iphdrlen;
578 	int has_tso4, has_tso6;
579 	int has_tso, use_tso;
580 	bool alwaysfrag;
581 	int sack_rxmit;
582 	int sack_bytes_rxmt;
583 	int ecn_tos;
584 	struct sackhole *p;
585 #ifdef TCP_SIGNATURE
586 	int sigoff = 0;
587 #endif
588 	uint64_t *tcps;
589 
590 #ifdef DIAGNOSTIC
591 	if (tp->t_inpcb && tp->t_in6pcb)
592 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
593 #endif
594 	so = NULL;
595 	ro = NULL;
596 	if (tp->t_inpcb) {
597 		so = tp->t_inpcb->inp_socket;
598 		ro = &tp->t_inpcb->inp_route;
599 	}
600 #ifdef INET6
601 	else if (tp->t_in6pcb) {
602 		so = tp->t_in6pcb->in6p_socket;
603 		ro = &tp->t_in6pcb->in6p_route;
604 	}
605 #endif
606 
607 	switch (af = tp->t_family) {
608 #ifdef INET
609 	case AF_INET:
610 		if (tp->t_inpcb)
611 			break;
612 #ifdef INET6
613 		/* mapped addr case */
614 		if (tp->t_in6pcb)
615 			break;
616 #endif
617 		return (EINVAL);
618 #endif
619 #ifdef INET6
620 	case AF_INET6:
621 		if (tp->t_in6pcb)
622 			break;
623 		return (EINVAL);
624 #endif
625 	default:
626 		return (EAFNOSUPPORT);
627 	}
628 
629 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
630 		return (EMSGSIZE);
631 
632 	idle = (tp->snd_max == tp->snd_una);
633 
634 	/*
635 	 * Determine if we can use TCP segmentation offload:
636 	 * - If we're using IPv4
637 	 * - If there is not an IPsec policy that prevents it
638 	 * - If the interface can do it
639 	 */
640 	has_tso4 = has_tso6 = false;
641 #if defined(INET)
642 	has_tso4 = tp->t_inpcb != NULL &&
643 #if defined(IPSEC)
644 	    (!ipsec_used || ipsec_pcb_skip_ipsec(tp->t_inpcb->inp_sp,
645 	    IPSEC_DIR_OUTBOUND)) &&
646 #endif
647 	    (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
648 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
649 	if (rt != NULL) {
650 		rtcache_unref(rt, &tp->t_inpcb->inp_route);
651 		rt = NULL;
652 	}
653 #endif /* defined(INET) */
654 #if defined(INET6)
655 	has_tso6 = tp->t_in6pcb != NULL &&
656 #if defined(IPSEC)
657 	    (!ipsec_used || ipsec_pcb_skip_ipsec(tp->t_in6pcb->in6p_sp,
658 	    IPSEC_DIR_OUTBOUND)) &&
659 #endif
660 	    (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
661 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
662 	if (rt != NULL)
663 		rtcache_unref(rt, &tp->t_in6pcb->in6p_route);
664 #endif /* defined(INET6) */
665 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
666 
667 	/*
668 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
669 	 *
670 	 *	Optionally, a TCP MAY set the restart window to the
671 	 *	minimum of the value used for the initial window and
672 	 *	the current value of cwnd (in other words, using a
673 	 *	larger value for the restart window should never increase
674 	 *	the size of cwnd).
675 	 */
676 	if (tcp_cwm) {
677 		/*
678 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
679 		 * Count the number of packets currently pending
680 		 * acknowledgement, and limit our congestion window
681 		 * to a pre-determined allowed burst size plus that count.
682 		 * This prevents bursting once all pending packets have
683 		 * been acknowledged (i.e. transmission is idle).
684 		 *
685 		 * XXX Link this to Initial Window?
686 		 */
687 		tp->snd_cwnd = min(tp->snd_cwnd,
688 		    (tcp_cwm_burstsize * txsegsize) +
689 		    (tp->snd_nxt - tp->snd_una));
690 	} else {
691 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
692 			/*
693 			 * We have been idle for "a while" and no acks are
694 			 * expected to clock out any data we send --
695 			 * slow start to get ack "clock" running again.
696 			 */
697 			int ss = tcp_init_win;
698 #ifdef INET
699 			if (tp->t_inpcb &&
700 			    in_localaddr(tp->t_inpcb->inp_faddr))
701 				ss = tcp_init_win_local;
702 #endif
703 #ifdef INET6
704 			if (tp->t_in6pcb &&
705 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
706 				ss = tcp_init_win_local;
707 #endif
708 			tp->snd_cwnd = min(tp->snd_cwnd,
709 			    TCP_INITIAL_WINDOW(ss, txsegsize));
710 		}
711 	}
712 
713 	txsegsize_nosack = txsegsize;
714 again:
715 	ecn_tos = 0;
716 	use_tso = has_tso;
717 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
718 		/* don't duplicate CWR/ECE. */
719 		use_tso = 0;
720 	}
721 	TCP_REASS_LOCK(tp);
722 	sack_numblks = tcp_sack_numblks(tp);
723 	if (sack_numblks) {
724 		int sackoptlen;
725 
726 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
727 		if (sackoptlen > txsegsize_nosack) {
728 			sack_numblks = 0; /* give up SACK */
729 			txsegsize = txsegsize_nosack;
730 		} else {
731 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
732 				/* don't duplicate D-SACK. */
733 				use_tso = 0;
734 			}
735 			txsegsize = txsegsize_nosack - sackoptlen;
736 		}
737 	} else {
738 		txsegsize = txsegsize_nosack;
739 	}
740 
741 	/*
742 	 * Determine length of data that should be transmitted, and
743 	 * flags that should be used.  If there is some data or critical
744 	 * controls (SYN, RST) to send, then transmit; otherwise,
745 	 * investigate further.
746 	 *
747 	 * Readjust SACK information to avoid resending duplicate data.
748 	 */
749 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
750 		tcp_sack_adjust(tp);
751 	sendalot = 0;
752 	off = tp->snd_nxt - tp->snd_una;
753 	win = min(tp->snd_wnd, tp->snd_cwnd);
754 
755 	flags = tcp_outflags[tp->t_state];
756 
757 	/*
758 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
759 	 * to send out new data (when sendalot is 1), bypass this function.
760 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
761 	 * we're replacing a (future) new transmission with a retransmission
762 	 * now, and we previously incremented snd_cwnd in tcp_input().
763 	 */
764 	/*
765 	 * Still in sack recovery , reset rxmit flag to zero.
766 	 */
767 	sack_rxmit = 0;
768 	sack_bytes_rxmt = 0;
769 	len = 0;
770 	p = NULL;
771 	do {
772 		long cwin;
773 		if (!TCP_SACK_ENABLED(tp))
774 			break;
775 		if (tp->t_partialacks < 0)
776 			break;
777 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
778 		if (p == NULL)
779 			break;
780 
781 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
782 		if (cwin < 0)
783 			cwin = 0;
784 		/* Do not retransmit SACK segments beyond snd_recover */
785 		if (SEQ_GT(p->end, tp->snd_recover)) {
786 			/*
787 			 * (At least) part of sack hole extends beyond
788 			 * snd_recover. Check to see if we can rexmit data
789 			 * for this hole.
790 			 */
791 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
792 				/*
793 				 * Can't rexmit any more data for this hole.
794 				 * That data will be rexmitted in the next
795 				 * sack recovery episode, when snd_recover
796 				 * moves past p->rxmit.
797 				 */
798 				p = NULL;
799 				break;
800 			}
801 			/* Can rexmit part of the current hole */
802 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
803 		} else
804 			len = ((long)ulmin(cwin, p->end - p->rxmit));
805 		off = p->rxmit - tp->snd_una;
806 		if (off + len > so->so_snd.sb_cc) {
807 			/* 1 for TH_FIN */
808 			KASSERT(off + len == so->so_snd.sb_cc + 1);
809 			KASSERT(p->rxmit + len == tp->snd_max);
810 			len = so->so_snd.sb_cc - off;
811 		}
812 		if (len > 0) {
813 			sack_rxmit = 1;
814 			sendalot = 1;
815 		}
816 	} while (/*CONSTCOND*/0);
817 
818 	/*
819 	 * If in persist timeout with window of 0, send 1 byte.
820 	 * Otherwise, if window is small but nonzero
821 	 * and timer expired, we will send what we can
822 	 * and go to transmit state.
823 	 */
824 	if (tp->t_force) {
825 		if (win == 0) {
826 			/*
827 			 * If we still have some data to send, then
828 			 * clear the FIN bit.  Usually this would
829 			 * happen below when it realizes that we
830 			 * aren't sending all the data.  However,
831 			 * if we have exactly 1 byte of unset data,
832 			 * then it won't clear the FIN bit below,
833 			 * and if we are in persist state, we wind
834 			 * up sending the packet without recording
835 			 * that we sent the FIN bit.
836 			 *
837 			 * We can't just blindly clear the FIN bit,
838 			 * because if we don't have any more data
839 			 * to send then the probe will be the FIN
840 			 * itself.
841 			 */
842 			if (off < so->so_snd.sb_cc)
843 				flags &= ~TH_FIN;
844 			win = 1;
845 		} else {
846 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
847 			tp->t_rxtshift = 0;
848 		}
849 	}
850 
851 	if (sack_rxmit == 0) {
852 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
853 			long cwin;
854 
855 			/*
856 			 * We are inside of a SACK recovery episode and are
857 			 * sending new data, having retransmitted all the
858 			 * data possible in the scoreboard.
859 			 */
860 			if (tp->snd_wnd < so->so_snd.sb_cc) {
861 				len = tp->snd_wnd - off;
862 				flags &= ~TH_FIN;
863 			} else {
864 				len = so->so_snd.sb_cc - off;
865 			}
866 
867 			/*
868 			 * From FreeBSD:
869 			 *  Don't remove this (len > 0) check !
870 			 *  We explicitly check for len > 0 here (although it
871 			 *  isn't really necessary), to work around a gcc
872 			 *  optimization issue - to force gcc to compute
873 			 *  len above. Without this check, the computation
874 			 *  of len is bungled by the optimizer.
875 			 */
876 			if (len > 0) {
877 				cwin = tp->snd_cwnd -
878 				    (tp->snd_nxt - tp->sack_newdata) -
879 				    sack_bytes_rxmt;
880 				if (cwin < 0)
881 					cwin = 0;
882 				if (cwin < len) {
883 					len = cwin;
884 					flags &= ~TH_FIN;
885 				}
886 			}
887 		} else if (win < so->so_snd.sb_cc) {
888 			len = win - off;
889 			flags &= ~TH_FIN;
890 		} else {
891 			len = so->so_snd.sb_cc - off;
892 		}
893 	}
894 
895 	if (len < 0) {
896 		/*
897 		 * If FIN has been sent but not acked,
898 		 * but we haven't been called to retransmit,
899 		 * len will be -1.  Otherwise, window shrank
900 		 * after we sent into it.  If window shrank to 0,
901 		 * cancel pending retransmit, pull snd_nxt back
902 		 * to (closed) window, and set the persist timer
903 		 * if it isn't already going.  If the window didn't
904 		 * close completely, just wait for an ACK.
905 		 *
906 		 * If we have a pending FIN, either it has already been
907 		 * transmitted or it is outside the window, so drop it.
908 		 * If the FIN has been transmitted, but this is not a
909 		 * retransmission, then len must be -1.  Therefore we also
910 		 * prevent here the sending of `gratuitous FINs'.  This
911 		 * eliminates the need to check for that case below (e.g.
912 		 * to back up snd_nxt before the FIN so that the sequence
913 		 * number is correct).
914 		 */
915 		len = 0;
916 		flags &= ~TH_FIN;
917 		if (win == 0) {
918 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
919 			tp->t_rxtshift = 0;
920 			tp->snd_nxt = tp->snd_una;
921 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
922 				tcp_setpersist(tp);
923 		}
924 	}
925 
926 	/*
927 	 * Automatic sizing enables the performance of large buffers
928 	 * and most of the efficiency of small ones by only allocating
929 	 * space when it is needed.
930 	 *
931 	 * The criteria to step up the send buffer one notch are:
932 	 *  1. receive window of remote host is larger than send buffer
933 	 *     (with a fudge factor of 5/4th);
934 	 *  2. send buffer is filled to 7/8th with data (so we actually
935 	 *     have data to make use of it);
936 	 *  3. send buffer fill has not hit maximal automatic size;
937 	 *  4. our send window (slow start and cogestion controlled) is
938 	 *     larger than sent but unacknowledged data in send buffer.
939 	 *
940 	 * The remote host receive window scaling factor may limit the
941 	 * growing of the send buffer before it reaches its allowed
942 	 * maximum.
943 	 *
944 	 * It scales directly with slow start or congestion window
945 	 * and does at most one step per received ACK.  This fast
946 	 * scaling has the drawback of growing the send buffer beyond
947 	 * what is strictly necessary to make full use of a given
948 	 * delay*bandwith product.  However testing has shown this not
949 	 * to be much of an problem.  At worst we are trading wasting
950 	 * of available bandwith (the non-use of it) for wasting some
951 	 * socket buffer memory.
952 	 *
953 	 * TODO: Shrink send buffer during idle periods together
954 	 * with congestion window.  Requires another timer.
955 	 */
956 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
957 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
958 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
959 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
960 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
961 			if (!sbreserve(&so->so_snd,
962 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
963 			     tcp_autosndbuf_max), so))
964 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
965 		}
966 	}
967 
968 	if (len > txsegsize) {
969 		if (use_tso) {
970 			/*
971 			 * Truncate TSO transfers to IP_MAXPACKET, and make
972 			 * sure that we send equal size transfers down the
973 			 * stack (rather than big-small-big-small-...).
974 			 */
975 #ifdef INET6
976 			CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET);
977 #endif
978 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
979 			if (len <= txsegsize) {
980 				use_tso = 0;
981 			}
982 		} else
983 			len = txsegsize;
984 		flags &= ~TH_FIN;
985 		sendalot = 1;
986 	} else
987 		use_tso = 0;
988 	if (sack_rxmit) {
989 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
990 			flags &= ~TH_FIN;
991 	}
992 
993 	win = sbspace(&so->so_rcv);
994 
995 	/*
996 	 * Sender silly window avoidance.  If connection is idle
997 	 * and can send all data, a maximum segment,
998 	 * at least a maximum default-size segment do it,
999 	 * or are forced, do it; otherwise don't bother.
1000 	 * If peer's buffer is tiny, then send
1001 	 * when window is at least half open.
1002 	 * If retransmitting (possibly after persist timer forced us
1003 	 * to send into a small window), then must resend.
1004 	 */
1005 	if (len) {
1006 		if (len >= txsegsize)
1007 			goto send;
1008 		if ((so->so_state & SS_MORETOCOME) == 0 &&
1009 		    ((idle || tp->t_flags & TF_NODELAY) &&
1010 		     len + off >= so->so_snd.sb_cc))
1011 			goto send;
1012 		if (tp->t_force)
1013 			goto send;
1014 		if (len >= tp->max_sndwnd / 2)
1015 			goto send;
1016 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1017 			goto send;
1018 		if (sack_rxmit)
1019 			goto send;
1020 	}
1021 
1022 	/*
1023 	 * Compare available window to amount of window known to peer
1024 	 * (as advertised window less next expected input).  If the
1025 	 * difference is at least twice the size of the largest segment
1026 	 * we expect to receive (i.e. two segments) or at least 50% of
1027 	 * the maximum possible window, then want to send a window update
1028 	 * to peer.
1029 	 */
1030 	if (win > 0) {
1031 		/*
1032 		 * "adv" is the amount we can increase the window,
1033 		 * taking into account that we are limited by
1034 		 * TCP_MAXWIN << tp->rcv_scale.
1035 		 */
1036 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1037 			(tp->rcv_adv - tp->rcv_nxt);
1038 
1039 		/*
1040 		 * If the new window size ends up being the same as the old
1041 		 * size when it is scaled, then don't force a window update.
1042 		 */
1043 		if ((tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale ==
1044 		    (adv + tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale)
1045 			goto dontupdate;
1046 		if (adv >= (long) (2 * rxsegsize))
1047 			goto send;
1048 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1049 			goto send;
1050 	}
1051 dontupdate:
1052 
1053 	/*
1054 	 * Send if we owe peer an ACK.
1055 	 */
1056 	if (tp->t_flags & TF_ACKNOW)
1057 		goto send;
1058 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1059 		goto send;
1060 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1061 		goto send;
1062 	/*
1063 	 * In SACK, it is possible for tcp_output to fail to send a segment
1064 	 * after the retransmission timer has been turned off.  Make sure
1065 	 * that the retransmission timer is set.
1066 	 */
1067 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1068 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1069 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1070 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1071 		goto just_return;
1072 	}
1073 
1074 	/*
1075 	 * TCP window updates are not reliable, rather a polling protocol
1076 	 * using ``persist'' packets is used to insure receipt of window
1077 	 * updates.  The three ``states'' for the output side are:
1078 	 *	idle			not doing retransmits or persists
1079 	 *	persisting		to move a small or zero window
1080 	 *	(re)transmitting	and thereby not persisting
1081 	 *
1082 	 * tp->t_timer[TCPT_PERSIST]
1083 	 *	is set when we are in persist state.
1084 	 * tp->t_force
1085 	 *	is set when we are called to send a persist packet.
1086 	 * tp->t_timer[TCPT_REXMT]
1087 	 *	is set when we are retransmitting
1088 	 * The output side is idle when both timers are zero.
1089 	 *
1090 	 * If send window is too small, there is data to transmit, and no
1091 	 * retransmit or persist is pending, then go to persist state.
1092 	 * If nothing happens soon, send when timer expires:
1093 	 * if window is nonzero, transmit what we can,
1094 	 * otherwise force out a byte.
1095 	 */
1096 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1097 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1098 		tp->t_rxtshift = 0;
1099 		tcp_setpersist(tp);
1100 	}
1101 
1102 	/*
1103 	 * No reason to send a segment, just return.
1104 	 */
1105 just_return:
1106 	TCP_REASS_UNLOCK(tp);
1107 	return (0);
1108 
1109 send:
1110 	/*
1111 	 * Before ESTABLISHED, force sending of initial options
1112 	 * unless TCP set not to do any options.
1113 	 * NOTE: we assume that the IP/TCP header plus TCP options
1114 	 * always fit in a single mbuf, leaving room for a maximum
1115 	 * link header, i.e.
1116 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1117 	 */
1118 	optlen = 0;
1119 	optp = opt;
1120 	switch (af) {
1121 #ifdef INET
1122 	case AF_INET:
1123 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1124 		break;
1125 #endif
1126 #ifdef INET6
1127 	case AF_INET6:
1128 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1129 		break;
1130 #endif
1131 	default:	/*pacify gcc*/
1132 		iphdrlen = 0;
1133 		break;
1134 	}
1135 	hdrlen = iphdrlen;
1136 	if (flags & TH_SYN) {
1137 		struct rtentry *synrt;
1138 
1139 		synrt = NULL;
1140 #ifdef INET
1141 		if (tp->t_inpcb)
1142 			synrt = in_pcbrtentry(tp->t_inpcb);
1143 #endif
1144 #ifdef INET6
1145 		if (tp->t_in6pcb)
1146 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1147 #endif
1148 
1149 		tp->snd_nxt = tp->iss;
1150 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1151 						    synrt->rt_ifp : NULL, af);
1152 #ifdef INET
1153 		if (tp->t_inpcb)
1154 			in_pcbrtentry_unref(synrt, tp->t_inpcb);
1155 #endif
1156 #ifdef INET6
1157 		if (tp->t_in6pcb)
1158 			in6_pcbrtentry_unref(synrt, tp->t_in6pcb);
1159 #endif
1160 		if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(TCPOLEN_MAXSEG)) {
1161 			*optp++ = TCPOPT_MAXSEG;
1162 			*optp++ = TCPOLEN_MAXSEG;
1163 			*optp++ = (tp->t_ourmss >> 8) & 0xff;
1164 			*optp++ = tp->t_ourmss & 0xff;
1165 			optlen += TCPOLEN_MAXSEG;
1166 
1167 			if ((tp->t_flags & TF_REQ_SCALE) &&
1168 			    ((flags & TH_ACK) == 0 ||
1169 			    (tp->t_flags & TF_RCVD_SCALE)) &&
1170 			    OPT_FITS(TCPOLEN_WINDOW + TCPOLEN_NOP)) {
1171 				*((uint32_t *)optp) = htonl(
1172 					TCPOPT_NOP << 24 |
1173 					TCPOPT_WINDOW << 16 |
1174 					TCPOLEN_WINDOW << 8 |
1175 					tp->request_r_scale);
1176 				optp += TCPOLEN_WINDOW + TCPOLEN_NOP;
1177 				optlen += TCPOLEN_WINDOW + TCPOLEN_NOP;
1178 			}
1179 			if (tcp_do_sack && OPT_FITS(TCPOLEN_SACK_PERMITTED)) {
1180 				*optp++ = TCPOPT_SACK_PERMITTED;
1181 				*optp++ = TCPOLEN_SACK_PERMITTED;
1182 				optlen += TCPOLEN_SACK_PERMITTED;
1183 			}
1184 		}
1185 	}
1186 
1187 	/*
1188 	 * Send a timestamp and echo-reply if this is a SYN and our side
1189 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1190 	 * and our peer have sent timestamps in our SYN's.
1191 	 */
1192 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1193 	     (flags & TH_RST) == 0 &&
1194 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1195 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1196 		int alen = 0;
1197 		while (optlen % 4 != 2) {
1198 			optlen += TCPOLEN_NOP;
1199 			*optp++ = TCPOPT_NOP;
1200 			alen++;
1201 		}
1202 		if (OPT_FITS(TCPOLEN_TIMESTAMP)) {
1203 			*optp++ = TCPOPT_TIMESTAMP;
1204 			*optp++ = TCPOLEN_TIMESTAMP;
1205 			uint32_t *lp = (uint32_t *)optp;
1206 			/* Form timestamp option (appendix A of RFC 1323) */
1207 			*lp++ = htonl(TCP_TIMESTAMP(tp));
1208 			*lp   = htonl(tp->ts_recent);
1209 			optp += TCPOLEN_TIMESTAMP - 2;
1210 			optlen += TCPOLEN_TIMESTAMP;
1211 
1212 			/* Set receive buffer autosizing timestamp. */
1213 			if (tp->rfbuf_ts == 0 &&
1214 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
1215 				tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1216 		} else {
1217 			optp -= alen;
1218 			optlen -= alen;
1219 		}
1220 	}
1221 
1222 #ifdef TCP_SIGNATURE
1223 	if (tp->t_flags & TF_SIGNATURE) {
1224 		/*
1225 		 * Initialize TCP-MD5 option (RFC2385)
1226 		 */
1227 		if (!OPT_FITS(TCPOLEN_SIGNATURE))
1228 			goto reset;
1229 
1230 		*optp++ = TCPOPT_SIGNATURE;
1231 		*optp++ = TCPOLEN_SIGNATURE;
1232 		sigoff = optlen + 2;
1233 		memset(optp, 0, TCP_SIGLEN);
1234 		optlen += TCPOLEN_SIGNATURE;
1235 		optp += TCP_SIGLEN;
1236 	}
1237 #endif /* TCP_SIGNATURE */
1238 
1239 	/*
1240 	 * Tack on the SACK block if it is necessary.
1241 	 */
1242 	if (sack_numblks) {
1243 		int alen = 0;
1244 		int sack_len = sack_numblks * 8;
1245 		while (optlen % 4 != 2) {
1246 			optlen += TCPOLEN_NOP;
1247 			*optp++ = TCPOPT_NOP;
1248 			alen++;
1249 		}
1250 		if (OPT_FITS(sack_len + 2)) {
1251 			struct ipqent *tiqe;
1252 			*optp++ = TCPOPT_SACK;
1253 			*optp++ = sack_len + 2;
1254 			uint32_t *lp = (uint32_t *)optp;
1255 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1256 				sack_numblks--;
1257 				*lp++ = htonl(tp->rcv_dsack_block.left);
1258 				*lp++ = htonl(tp->rcv_dsack_block.right);
1259 				tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1260 			}
1261 			for (tiqe = TAILQ_FIRST(&tp->timeq);
1262 			    sack_numblks > 0;
1263 			    tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1264 				KASSERT(tiqe != NULL);
1265 				sack_numblks--;
1266 				*lp++ = htonl(tiqe->ipqe_seq);
1267 				*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1268 				    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1269 			}
1270 			optlen += sack_len + 2;
1271 			optp += sack_len;
1272 		} else {
1273 			optp -= alen;
1274 			optlen -= alen;
1275 		}
1276 	}
1277 
1278 	/* Terminate and pad TCP options to a 4 byte boundary. */
1279 	if (optlen % 4) {
1280 		if (!OPT_FITS(TCPOLEN_EOL)) {
1281 reset:			TCP_REASS_UNLOCK(tp);
1282 			error = ECONNABORTED;
1283 			goto out;
1284 		}
1285 		optlen += TCPOLEN_EOL;
1286 		*optp++ = TCPOPT_EOL;
1287 	}
1288 	/*
1289 	 * According to RFC 793 (STD0007):
1290 	 *   "The content of the header beyond the End-of-Option option
1291 	 *    must be header padding (i.e., zero)."
1292 	 *   and later: "The padding is composed of zeros."
1293 	 */
1294 	while (optlen % 4) {
1295 		if (!OPT_FITS(TCPOLEN_PAD))
1296 			goto reset;
1297 		optlen += TCPOLEN_PAD;
1298 		*optp++ = TCPOPT_PAD;
1299 	}
1300 
1301 	TCP_REASS_UNLOCK(tp);
1302 
1303 	hdrlen += optlen;
1304 
1305 #ifdef DIAGNOSTIC
1306 	if (!use_tso && len > txsegsize)
1307 		panic("tcp data to be sent is larger than segment");
1308 	else if (use_tso && len > IP_MAXPACKET)
1309 		panic("tcp data to be sent is larger than max TSO size");
1310 	if (max_linkhdr + hdrlen > MCLBYTES)
1311 		panic("tcphdr too big");
1312 #endif
1313 
1314 	/*
1315 	 * Grab a header mbuf, attaching a copy of data to
1316 	 * be transmitted, and initialize the header from
1317 	 * the template for sends on this connection.
1318 	 */
1319 	if (len) {
1320 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1321 		if (error)
1322 			goto out;
1323 		/*
1324 		 * If we're sending everything we've got, set PUSH.
1325 		 * (This will keep happy those implementations which only
1326 		 * give data to the user when a buffer fills or
1327 		 * a PUSH comes in.)
1328 		 */
1329 		if (off + len == so->so_snd.sb_cc)
1330 			flags |= TH_PUSH;
1331 	} else {
1332 		tcps = TCP_STAT_GETREF();
1333 		if (tp->t_flags & TF_ACKNOW)
1334 			tcps[TCP_STAT_SNDACKS]++;
1335 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1336 			tcps[TCP_STAT_SNDCTRL]++;
1337 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1338 			tcps[TCP_STAT_SNDURG]++;
1339 		else
1340 			tcps[TCP_STAT_SNDWINUP]++;
1341 		TCP_STAT_PUTREF();
1342 
1343 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1344 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1345 			MCLGET(m, M_DONTWAIT);
1346 			if ((m->m_flags & M_EXT) == 0) {
1347 				m_freem(m);
1348 				m = NULL;
1349 			}
1350 		}
1351 		if (m == NULL) {
1352 			error = ENOBUFS;
1353 			goto out;
1354 		}
1355 		MCLAIM(m, &tcp_tx_mowner);
1356 		m->m_data += max_linkhdr;
1357 		m->m_len = hdrlen;
1358 	}
1359 	m_reset_rcvif(m);
1360 	switch (af) {
1361 #ifdef INET
1362 	case AF_INET:
1363 		ip = mtod(m, struct ip *);
1364 #ifdef INET6
1365 		ip6 = NULL;
1366 #endif
1367 		th = (struct tcphdr *)(ip + 1);
1368 		break;
1369 #endif
1370 #ifdef INET6
1371 	case AF_INET6:
1372 		ip = NULL;
1373 		ip6 = mtod(m, struct ip6_hdr *);
1374 		th = (struct tcphdr *)(ip6 + 1);
1375 		break;
1376 #endif
1377 	default:	/*pacify gcc*/
1378 		ip = NULL;
1379 #ifdef INET6
1380 		ip6 = NULL;
1381 #endif
1382 		th = NULL;
1383 		break;
1384 	}
1385 	if (tp->t_template == 0)
1386 		panic("tcp_output");
1387 	if (tp->t_template->m_len < iphdrlen)
1388 		panic("tcp_output");
1389 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1390 
1391 	/*
1392 	 * If we are starting a connection, send ECN setup
1393 	 * SYN packet. If we are on a retransmit, we may
1394 	 * resend those bits a number of times as per
1395 	 * RFC 3168.
1396 	 */
1397 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1398 		if (tp->t_flags & TF_SYN_REXMT) {
1399 			if (tp->t_ecn_retries--)
1400 				flags |= TH_ECE|TH_CWR;
1401 		} else {
1402 			flags |= TH_ECE|TH_CWR;
1403 			tp->t_ecn_retries = tcp_ecn_maxretries;
1404 		}
1405 	}
1406 
1407 	if (TCP_ECN_ALLOWED(tp)) {
1408 		/*
1409 		 * If the peer has ECN, mark data packets
1410 		 * ECN capable. Ignore pure ack packets, retransmissions
1411 		 * and window probes.
1412 		 */
1413 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1414 		    !(tp->t_force && len == 1)) {
1415 			ecn_tos = IPTOS_ECN_ECT0;
1416 			TCP_STATINC(TCP_STAT_ECN_ECT);
1417 		}
1418 
1419 		/*
1420 		 * Reply with proper ECN notifications.
1421 		 */
1422 		if (tp->t_flags & TF_ECN_SND_CWR) {
1423 			flags |= TH_CWR;
1424 			tp->t_flags &= ~TF_ECN_SND_CWR;
1425 		}
1426 		if (tp->t_flags & TF_ECN_SND_ECE) {
1427 			flags |= TH_ECE;
1428 		}
1429 	}
1430 
1431 
1432 	/*
1433 	 * If we are doing retransmissions, then snd_nxt will
1434 	 * not reflect the first unsent octet.  For ACK only
1435 	 * packets, we do not want the sequence number of the
1436 	 * retransmitted packet, we want the sequence number
1437 	 * of the next unsent octet.  So, if there is no data
1438 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1439 	 * when filling in ti_seq.  But if we are in persist
1440 	 * state, snd_max might reflect one byte beyond the
1441 	 * right edge of the window, so use snd_nxt in that
1442 	 * case, since we know we aren't doing a retransmission.
1443 	 * (retransmit and persist are mutually exclusive...)
1444 	 */
1445 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1446 		th->th_seq = htonl(p->rxmit);
1447 		p->rxmit += len;
1448 	} else {
1449 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1450 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1451 			th->th_seq = htonl(tp->snd_nxt);
1452 		else
1453 			th->th_seq = htonl(tp->snd_max);
1454 	}
1455 	th->th_ack = htonl(tp->rcv_nxt);
1456 	if (optlen) {
1457 		bcopy((void *)opt, (void *)(th + 1), optlen);
1458 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1459 	}
1460 	th->th_flags = flags;
1461 	/*
1462 	 * Calculate receive window.  Don't shrink window,
1463 	 * but avoid silly window syndrome.
1464 	 */
1465 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1466 		win = 0;
1467 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1468 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1469 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1470 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1471 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1472 	if (th->th_win == 0) {
1473 		tp->t_sndzerowin++;
1474 	}
1475 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1476 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1477 		if (urp > IP_MAXPACKET)
1478 			urp = IP_MAXPACKET;
1479 		th->th_urp = htons((u_int16_t)urp);
1480 		th->th_flags |= TH_URG;
1481 	} else
1482 		/*
1483 		 * If no urgent pointer to send, then we pull
1484 		 * the urgent pointer to the left edge of the send window
1485 		 * so that it doesn't drift into the send window on sequence
1486 		 * number wraparound.
1487 		 */
1488 		tp->snd_up = tp->snd_una;		/* drag it along */
1489 
1490 #ifdef TCP_SIGNATURE
1491 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1492 		struct secasvar *sav;
1493 		u_int8_t *sigp;
1494 
1495 		sav = tcp_signature_getsav(m, th);
1496 
1497 		if (sav == NULL) {
1498 			if (m)
1499 				m_freem(m);
1500 			return (EPERM);
1501 		}
1502 
1503 		m->m_pkthdr.len = hdrlen + len;
1504 		sigp = (char *)th + sizeof(*th) + sigoff;
1505 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1506 
1507 		key_sa_recordxfer(sav, m);
1508 		KEY_SA_UNREF(&sav);
1509 	}
1510 #endif
1511 
1512 	/*
1513 	 * Set ourselves up to be checksummed just before the packet
1514 	 * hits the wire.
1515 	 */
1516 	switch (af) {
1517 #ifdef INET
1518 	case AF_INET:
1519 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1520 		if (use_tso) {
1521 			m->m_pkthdr.segsz = txsegsize;
1522 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1523 		} else {
1524 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1525 			if (len + optlen) {
1526 				/* Fixup the pseudo-header checksum. */
1527 				/* XXXJRT Not IP Jumbogram safe. */
1528 				th->th_sum = in_cksum_addword(th->th_sum,
1529 				    htons((u_int16_t) (len + optlen)));
1530 			}
1531 		}
1532 		break;
1533 #endif
1534 #ifdef INET6
1535 	case AF_INET6:
1536 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1537 		if (use_tso) {
1538 			m->m_pkthdr.segsz = txsegsize;
1539 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1540 		} else {
1541 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1542 			if (len + optlen) {
1543 				/* Fixup the pseudo-header checksum. */
1544 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1545 				th->th_sum = in_cksum_addword(th->th_sum,
1546 				    htons((u_int16_t) (len + optlen)));
1547 			}
1548 		}
1549 		break;
1550 #endif
1551 	}
1552 
1553 	/*
1554 	 * In transmit state, time the transmission and arrange for
1555 	 * the retransmit.  In persist state, just set snd_max.
1556 	 */
1557 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1558 		tcp_seq startseq = tp->snd_nxt;
1559 
1560 		/*
1561 		 * Advance snd_nxt over sequence space of this segment.
1562 		 * There are no states in which we send both a SYN and a FIN,
1563 		 * so we collapse the tests for these flags.
1564 		 */
1565 		if (flags & (TH_SYN|TH_FIN))
1566 			tp->snd_nxt++;
1567 		if (sack_rxmit)
1568 			goto timer;
1569 		tp->snd_nxt += len;
1570 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1571 			tp->snd_max = tp->snd_nxt;
1572 			/*
1573 			 * Time this transmission if not a retransmission and
1574 			 * not currently timing anything.
1575 			 */
1576 			if (tp->t_rtttime == 0) {
1577 				tp->t_rtttime = tcp_now;
1578 				tp->t_rtseq = startseq;
1579 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1580 			}
1581 		}
1582 
1583 		/*
1584 		 * Set retransmit timer if not currently set,
1585 		 * and not doing an ack or a keep-alive probe.
1586 		 * Initial value for retransmit timer is smoothed
1587 		 * round-trip time + 2 * round-trip time variance.
1588 		 * Initialize shift counter which is used for backoff
1589 		 * of retransmit time.
1590 		 */
1591 timer:
1592 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0) {
1593 			if ((sack_rxmit && tp->snd_nxt != tp->snd_max)
1594 			    || tp->snd_nxt != tp->snd_una) {
1595 				if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1596 					TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1597 					tp->t_rxtshift = 0;
1598 				}
1599 				TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1600 			} else if (len == 0 && so->so_snd.sb_cc > 0
1601 			    && TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1602 				/*
1603 				 * If we are sending a window probe and there's
1604 				 * unacked data in the socket, make sure at
1605 				 * least the persist timer is running.
1606 				 */
1607 				tp->t_rxtshift = 0;
1608 				tcp_setpersist(tp);
1609 			}
1610 		}
1611 	} else
1612 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1613 			tp->snd_max = tp->snd_nxt + len;
1614 
1615 #ifdef TCP_DEBUG
1616 	/*
1617 	 * Trace.
1618 	 */
1619 	if (so->so_options & SO_DEBUG)
1620 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1621 #endif
1622 
1623 	/*
1624 	 * Fill in IP length and desired time to live and
1625 	 * send to IP level.  There should be a better way
1626 	 * to handle ttl and tos; we could keep them in
1627 	 * the template, but need a way to checksum without them.
1628 	 */
1629 	m->m_pkthdr.len = hdrlen + len;
1630 
1631 	switch (af) {
1632 #ifdef INET
1633 	case AF_INET:
1634 		ip->ip_len = htons(m->m_pkthdr.len);
1635 		packetlen = m->m_pkthdr.len;
1636 		if (tp->t_inpcb) {
1637 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1638 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos;
1639 		}
1640 #ifdef INET6
1641 		else if (tp->t_in6pcb) {
1642 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1643 			ip->ip_tos = ecn_tos;	/*XXX*/
1644 		}
1645 #endif
1646 		break;
1647 #endif
1648 #ifdef INET6
1649 	case AF_INET6:
1650 		packetlen = m->m_pkthdr.len;
1651 		ip6->ip6_nxt = IPPROTO_TCP;
1652 		if (tp->t_in6pcb) {
1653 			/*
1654 			 * we separately set hoplimit for every segment, since
1655 			 * the user might want to change the value via
1656 			 * setsockopt. Also, desired default hop limit might
1657 			 * be changed via Neighbor Discovery.
1658 			 */
1659 			ip6->ip6_hlim = in6_selecthlim_rt(tp->t_in6pcb);
1660 		}
1661 		ip6->ip6_flow |= htonl(ecn_tos << 20);
1662 		/* ip6->ip6_flow = ??? (from template) */
1663 		/* ip6_plen will be filled in ip6_output(). */
1664 		break;
1665 #endif
1666 	default:	/*pacify gcc*/
1667 		packetlen = 0;
1668 		break;
1669 	}
1670 
1671 	switch (af) {
1672 #ifdef INET
1673 	case AF_INET:
1674 	    {
1675 		struct mbuf *opts;
1676 
1677 		if (tp->t_inpcb)
1678 			opts = tp->t_inpcb->inp_options;
1679 		else
1680 			opts = NULL;
1681 		error = ip_output(m, opts, ro,
1682 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1683 			(so->so_options & SO_DONTROUTE), NULL, tp->t_inpcb);
1684 		break;
1685 	    }
1686 #endif
1687 #ifdef INET6
1688 	case AF_INET6:
1689 	    {
1690 		struct ip6_pktopts *opts;
1691 
1692 		if (tp->t_in6pcb)
1693 			opts = tp->t_in6pcb->in6p_outputopts;
1694 		else
1695 			opts = NULL;
1696 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1697 			NULL, tp->t_in6pcb, NULL);
1698 		break;
1699 	    }
1700 #endif
1701 	default:
1702 		error = EAFNOSUPPORT;
1703 		break;
1704 	}
1705 	if (error) {
1706 out:
1707 		if (error == ENOBUFS) {
1708 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1709 #ifdef INET
1710 			if (tp->t_inpcb)
1711 				tcp_quench(tp->t_inpcb, 0);
1712 #endif
1713 #ifdef INET6
1714 			if (tp->t_in6pcb)
1715 				tcp6_quench(tp->t_in6pcb, 0);
1716 #endif
1717 			error = 0;
1718 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1719 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1720 			tp->t_softerror = error;
1721 			error = 0;
1722 		}
1723 
1724 		/* Back out the seqence number advance. */
1725 		if (sack_rxmit)
1726 			p->rxmit -= len;
1727 
1728 		/* Restart the delayed ACK timer, if necessary. */
1729 		if (tp->t_flags & TF_DELACK)
1730 			TCP_RESTART_DELACK(tp);
1731 
1732 		return (error);
1733 	}
1734 
1735 	if (packetlen > tp->t_pmtud_mtu_sent)
1736 		tp->t_pmtud_mtu_sent = packetlen;
1737 
1738 	tcps = TCP_STAT_GETREF();
1739 	tcps[TCP_STAT_SNDTOTAL]++;
1740 	if (tp->t_flags & TF_DELACK)
1741 		tcps[TCP_STAT_DELACK]++;
1742 	TCP_STAT_PUTREF();
1743 
1744 	/*
1745 	 * Data sent (as far as we can tell).
1746 	 * If this advertises a larger window than any other segment,
1747 	 * then remember the size of the advertised window.
1748 	 * Any pending ACK has now been sent.
1749 	 */
1750 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1751 		tp->rcv_adv = tp->rcv_nxt + win;
1752 	tp->last_ack_sent = tp->rcv_nxt;
1753 	tp->t_flags &= ~TF_ACKNOW;
1754 	TCP_CLEAR_DELACK(tp);
1755 #ifdef DIAGNOSTIC
1756 	if (maxburst < 0)
1757 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1758 #endif
1759 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1760 		goto again;
1761 	return (0);
1762 }
1763 
1764 void
1765 tcp_setpersist(struct tcpcb *tp)
1766 {
1767 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1768 	int nticks;
1769 
1770 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1771 		panic("tcp_output REXMT");
1772 	/*
1773 	 * Start/restart persistance timer.
1774 	 */
1775 	if (t < tp->t_rttmin)
1776 		t = tp->t_rttmin;
1777 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1778 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1779 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1780 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1781 		tp->t_rxtshift++;
1782 }
1783