xref: /netbsd-src/sys/netinet/tcp_output.c (revision c38e7cc395b1472a774ff828e46123de44c628e9)
1 /*	$NetBSD: tcp_output.c,v 1.206 2018/05/03 07:13:48 maxv 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 1.206 2018/05/03 07:13:48 maxv 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 #include <netinet/ip6.h>
170 #include <netinet6/in6_var.h>
171 #include <netinet6/ip6_var.h>
172 #include <netinet6/in6_pcb.h>
173 #include <netinet6/nd6.h>
174 #endif
175 
176 #ifdef IPSEC
177 #include <netipsec/ipsec.h>
178 #include <netipsec/key.h>
179 #ifdef INET6
180 #include <netipsec/ipsec6.h>
181 #endif
182 #endif
183 
184 #include <netinet/tcp.h>
185 #define	TCPOUTFLAGS
186 #include <netinet/tcp_fsm.h>
187 #include <netinet/tcp_seq.h>
188 #include <netinet/tcp_timer.h>
189 #include <netinet/tcp_var.h>
190 #include <netinet/tcp_private.h>
191 #include <netinet/tcp_congctl.h>
192 #include <netinet/tcp_debug.h>
193 #include <netinet/in_offload.h>
194 #include <netinet6/in6_offload.h>
195 
196 /*
197  * Knob to enable Congestion Window Monitoring, and control
198  * the burst size it allows.  Default burst is 4 packets, per
199  * the Internet draft.
200  */
201 int	tcp_cwm = 0;
202 int	tcp_cwm_burstsize = 4;
203 
204 int	tcp_do_autosndbuf = 1;
205 int	tcp_autosndbuf_inc = 8 * 1024;
206 int	tcp_autosndbuf_max = 256 * 1024;
207 
208 #ifdef TCP_OUTPUT_COUNTERS
209 #include <sys/device.h>
210 
211 extern struct evcnt tcp_output_bigheader;
212 extern struct evcnt tcp_output_predict_hit;
213 extern struct evcnt tcp_output_predict_miss;
214 extern struct evcnt tcp_output_copysmall;
215 extern struct evcnt tcp_output_copybig;
216 extern struct evcnt tcp_output_refbig;
217 
218 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
219 #else
220 
221 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
222 
223 #endif /* TCP_OUTPUT_COUNTERS */
224 
225 static int
226 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
227     bool *alwaysfragp)
228 {
229 	struct inpcb *inp = tp->t_inpcb;
230 #ifdef INET6
231 	struct in6pcb *in6p = tp->t_in6pcb;
232 #endif
233 	struct socket *so = NULL;
234 	struct rtentry *rt;
235 	struct ifnet *ifp;
236 	int size;
237 	int hdrlen;
238 	int optlen;
239 
240 	*alwaysfragp = false;
241 	size = tcp_mssdflt;
242 
243 	KASSERT(!(tp->t_inpcb && tp->t_in6pcb));
244 
245 	switch (tp->t_family) {
246 	case AF_INET:
247 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
248 		break;
249 #ifdef INET6
250 	case AF_INET6:
251 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
252 		break;
253 #endif
254 	default:
255 		goto out;
256 	}
257 
258 	rt = NULL;
259 	if (inp) {
260 		rt = in_pcbrtentry(inp);
261 		so = inp->inp_socket;
262 	}
263 #ifdef INET6
264 	if (in6p) {
265 		rt = in6_pcbrtentry(in6p);
266 		so = in6p->in6p_socket;
267 	}
268 #endif
269 	if (rt == NULL) {
270 		goto out;
271 	}
272 
273 	ifp = rt->rt_ifp;
274 
275 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
276 #ifdef INET6
277 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
278 			/*
279 			 * RFC2460 section 5, last paragraph: if path MTU is
280 			 * smaller than 1280, use 1280 as packet size and
281 			 * attach fragment header.
282 			 */
283 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
284 			*alwaysfragp = true;
285 		} else
286 			size = rt->rt_rmx.rmx_mtu - hdrlen;
287 #else
288 		size = rt->rt_rmx.rmx_mtu - hdrlen;
289 #endif
290 	} else if (ifp->if_flags & IFF_LOOPBACK)
291 		size = ifp->if_mtu - hdrlen;
292 	else if (inp && tp->t_mtudisc)
293 		size = ifp->if_mtu - hdrlen;
294 	else if (inp && in_localaddr(inp->inp_faddr))
295 		size = ifp->if_mtu - hdrlen;
296 #ifdef INET6
297 	else if (in6p) {
298 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
299 			/* mapped addr case */
300 			struct in_addr d;
301 			memcpy(&d, &in6p->in6p_faddr.s6_addr32[3], sizeof(d));
302 			if (tp->t_mtudisc || in_localaddr(d))
303 				size = ifp->if_mtu - hdrlen;
304 		} else {
305 			/*
306 			 * for IPv6, path MTU discovery is always turned on,
307 			 * or the node must use packet size <= 1280.
308 			 */
309 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
310 			size -= hdrlen;
311 		}
312 	}
313 #endif
314 	if (inp)
315 		in_pcbrtentry_unref(rt, inp);
316 #ifdef INET6
317 	if (in6p)
318 		in6_pcbrtentry_unref(rt, in6p);
319 #endif
320  out:
321 	/*
322 	 * Now we must make room for whatever extra TCP/IP options are in
323 	 * the packet.
324 	 */
325 	optlen = tcp_optlen(tp);
326 
327 	/*
328 	 * XXX tp->t_ourmss should have the right size, but without this code
329 	 * fragmentation will occur... need more investigation
330 	 */
331 
332 	if (inp) {
333 #if defined(IPSEC)
334 		if (ipsec_used &&
335 		    !ipsec_pcb_skip_ipsec(inp->inp_sp, IPSEC_DIR_OUTBOUND))
336 			optlen += ipsec4_hdrsiz_tcp(tp);
337 #endif
338 		optlen += ip_optlen(inp);
339 	}
340 
341 #ifdef INET6
342 	if (in6p && tp->t_family == AF_INET) {
343 #if defined(IPSEC)
344 		if (ipsec_used &&
345 		    !ipsec_pcb_skip_ipsec(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
346 			optlen += ipsec4_hdrsiz_tcp(tp);
347 #endif
348 		/* XXX size -= ip_optlen(in6p); */
349 	} else if (in6p && tp->t_family == AF_INET6) {
350 #if defined(IPSEC)
351 		if (ipsec_used &&
352 		    !ipsec_pcb_skip_ipsec(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
353 			optlen += ipsec6_hdrsiz_tcp(tp);
354 #endif
355 		optlen += ip6_optlen(in6p);
356 	}
357 #endif
358 	size -= optlen;
359 
360 	/*
361 	 * There may not be any room for data if mtu is too small. This
362 	 * includes zero-sized.
363 	 */
364 	if (size <= 0) {
365 		return EMSGSIZE;
366 	}
367 
368 	/*
369 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
370 	 * assumes that path MTU is the same for both ways).  this is only
371 	 * for silly window avoidance, do not use the value for other purposes.
372 	 *
373 	 * ipseclen is subtracted from both sides, this may not be right.
374 	 * I'm not quite sure about this (could someone comment).
375 	 */
376 	*txsegsizep = min(tp->t_peermss - optlen, size);
377 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
378 
379 	/*
380 	 * Never send more than half a buffer full.  This insures that we can
381 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
382 	 * therefore acks will never be delayed unless we run out of data to
383 	 * transmit.
384 	 */
385 	if (so) {
386 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
387 	}
388 
389 	if (*txsegsizep != tp->t_segsz) {
390 		/*
391 		 * If the new segment size is larger, we don't want to
392 		 * mess up the congestion window, but if it is smaller
393 		 * we'll have to reduce the congestion window to ensure
394 		 * that we don't get into trouble with initial windows
395 		 * and the rest.  In any case, if the segment size
396 		 * has changed, chances are the path has, too, and
397 		 * our congestion window will be different.
398 		 */
399 		if (*txsegsizep < tp->t_segsz) {
400 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
401 			    * *txsegsizep, *txsegsizep);
402 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
403 			    * *txsegsizep, *txsegsizep);
404 		}
405 		tp->t_segsz = *txsegsizep;
406 	}
407 
408 	return 0;
409 }
410 
411 static int
412 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
413     long len, int hdrlen, struct mbuf **mp)
414 {
415 	struct mbuf *m, *m0;
416 	uint64_t *tcps;
417 
418 	tcps = TCP_STAT_GETREF();
419 	if (tp->t_force && len == 1)
420 		tcps[TCP_STAT_SNDPROBE]++;
421 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
422 		tp->t_sndrexmitpack++;
423 		tcps[TCP_STAT_SNDREXMITPACK]++;
424 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
425 	} else {
426 		tcps[TCP_STAT_SNDPACK]++;
427 		tcps[TCP_STAT_SNDBYTE] += len;
428 	}
429 	TCP_STAT_PUTREF();
430 
431 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
432 	if (__predict_false(m == NULL))
433 		return ENOBUFS;
434 	MCLAIM(m, &tcp_tx_mowner);
435 
436 	/*
437 	 * XXX Because other code assumes headers will fit in
438 	 * XXX one header mbuf.
439 	 *
440 	 * (This code should almost *never* be run.)
441 	 */
442 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
443 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
444 		MCLGET(m, M_DONTWAIT);
445 		if ((m->m_flags & M_EXT) == 0) {
446 			m_freem(m);
447 			return ENOBUFS;
448 		}
449 	}
450 
451 	m->m_data += max_linkhdr;
452 	m->m_len = hdrlen;
453 
454 	/*
455 	 * To avoid traversing the whole sb_mb chain for correct
456 	 * data to send, remember last sent mbuf, its offset and
457 	 * the sent size.  When called the next time, see if the
458 	 * data to send is directly following the previous transfer.
459 	 * This is important for large TCP windows.
460 	 */
461 	if (off == 0 || tp->t_lastm == NULL ||
462 	    (tp->t_lastoff + tp->t_lastlen) != off) {
463 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
464 		/*
465 		 * Either a new packet or a retransmit.
466 		 * Start from the beginning.
467 		 */
468 		tp->t_lastm = so->so_snd.sb_mb;
469 		tp->t_inoff = off;
470 	} else {
471 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
472 		tp->t_inoff += tp->t_lastlen;
473 	}
474 
475 	/* Traverse forward to next packet */
476 	while (tp->t_inoff > 0) {
477 		if (tp->t_lastm == NULL)
478 			panic("tp->t_lastm == NULL");
479 		if (tp->t_inoff < tp->t_lastm->m_len)
480 			break;
481 		tp->t_inoff -= tp->t_lastm->m_len;
482 		tp->t_lastm = tp->t_lastm->m_next;
483 	}
484 
485 	tp->t_lastoff = off;
486 	tp->t_lastlen = len;
487 	m0 = tp->t_lastm;
488 	off = tp->t_inoff;
489 
490 	if (len <= M_TRAILINGSPACE(m)) {
491 		m_copydata(m0, off, (int)len, mtod(m, char *) + hdrlen);
492 		m->m_len += len;
493 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
494 	} else {
495 		m->m_next = m_copym(m0, off, (int)len, M_DONTWAIT);
496 		if (m->m_next == NULL) {
497 			m_freem(m);
498 			return ENOBUFS;
499 		}
500 #ifdef TCP_OUTPUT_COUNTERS
501 		if (m->m_next->m_flags & M_EXT)
502 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
503 		else
504 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
505 #endif
506 	}
507 
508 	*mp = m;
509 	return 0;
510 }
511 
512 /*
513  * Tcp output routine: figure out what should be sent and send it.
514  */
515 int
516 tcp_output(struct tcpcb *tp)
517 {
518 	struct rtentry *rt = NULL;
519 	struct socket *so;
520 	struct route *ro;
521 	long len, win;
522 	int off, flags, error;
523 	struct mbuf *m;
524 	struct ip *ip;
525 #ifdef INET6
526 	struct ip6_hdr *ip6;
527 #endif
528 	struct tcphdr *th;
529 	u_char opt[MAX_TCPOPTLEN], *optp;
530 #define OPT_FITS(more)	((optlen + (more)) <= sizeof(opt))
531 	unsigned optlen, hdrlen, packetlen;
532 	unsigned int sack_numblks;
533 	int idle, sendalot, txsegsize, rxsegsize;
534 	int txsegsize_nosack;
535 	int maxburst = TCP_MAXBURST;
536 	int af;		/* address family on the wire */
537 	int iphdrlen;
538 	int has_tso4, has_tso6;
539 	int has_tso, use_tso;
540 	bool alwaysfrag;
541 	int sack_rxmit;
542 	int sack_bytes_rxmt;
543 	int ecn_tos;
544 	struct sackhole *p;
545 #ifdef TCP_SIGNATURE
546 	int sigoff = 0;
547 #endif
548 	uint64_t *tcps;
549 
550 	KASSERT(!(tp->t_inpcb && tp->t_in6pcb));
551 
552 	so = NULL;
553 	ro = NULL;
554 	if (tp->t_inpcb) {
555 		so = tp->t_inpcb->inp_socket;
556 		ro = &tp->t_inpcb->inp_route;
557 	}
558 #ifdef INET6
559 	else if (tp->t_in6pcb) {
560 		so = tp->t_in6pcb->in6p_socket;
561 		ro = &tp->t_in6pcb->in6p_route;
562 	}
563 #endif
564 
565 	switch (af = tp->t_family) {
566 	case AF_INET:
567 		if (tp->t_inpcb)
568 			break;
569 #ifdef INET6
570 		/* mapped addr case */
571 		if (tp->t_in6pcb)
572 			break;
573 #endif
574 		return EINVAL;
575 #ifdef INET6
576 	case AF_INET6:
577 		if (tp->t_in6pcb)
578 			break;
579 		return EINVAL;
580 #endif
581 	default:
582 		return EAFNOSUPPORT;
583 	}
584 
585 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
586 		return EMSGSIZE;
587 
588 	idle = (tp->snd_max == tp->snd_una);
589 
590 	/*
591 	 * Determine if we can use TCP segmentation offload:
592 	 * - If we're using IPv4
593 	 * - If there is not an IPsec policy that prevents it
594 	 * - If the interface can do it
595 	 */
596 	has_tso4 = has_tso6 = false;
597 
598 	has_tso4 = tp->t_inpcb != NULL &&
599 #if defined(IPSEC)
600 	    (!ipsec_used || ipsec_pcb_skip_ipsec(tp->t_inpcb->inp_sp,
601 	    IPSEC_DIR_OUTBOUND)) &&
602 #endif
603 	    (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
604 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
605 	if (rt != NULL) {
606 		rtcache_unref(rt, &tp->t_inpcb->inp_route);
607 		rt = NULL;
608 	}
609 
610 #if defined(INET6)
611 	has_tso6 = tp->t_in6pcb != NULL &&
612 #if defined(IPSEC)
613 	    (!ipsec_used || ipsec_pcb_skip_ipsec(tp->t_in6pcb->in6p_sp,
614 	    IPSEC_DIR_OUTBOUND)) &&
615 #endif
616 	    (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
617 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
618 	if (rt != NULL)
619 		rtcache_unref(rt, &tp->t_in6pcb->in6p_route);
620 #endif /* defined(INET6) */
621 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
622 
623 	/*
624 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
625 	 *
626 	 *	Optionally, a TCP MAY set the restart window to the
627 	 *	minimum of the value used for the initial window and
628 	 *	the current value of cwnd (in other words, using a
629 	 *	larger value for the restart window should never increase
630 	 *	the size of cwnd).
631 	 */
632 	if (tcp_cwm) {
633 		/*
634 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
635 		 * Count the number of packets currently pending
636 		 * acknowledgement, and limit our congestion window
637 		 * to a pre-determined allowed burst size plus that count.
638 		 * This prevents bursting once all pending packets have
639 		 * been acknowledged (i.e. transmission is idle).
640 		 *
641 		 * XXX Link this to Initial Window?
642 		 */
643 		tp->snd_cwnd = min(tp->snd_cwnd,
644 		    (tcp_cwm_burstsize * txsegsize) +
645 		    (tp->snd_nxt - tp->snd_una));
646 	} else {
647 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
648 			/*
649 			 * We have been idle for "a while" and no acks are
650 			 * expected to clock out any data we send --
651 			 * slow start to get ack "clock" running again.
652 			 */
653 			int ss = tcp_init_win;
654 			if (tp->t_inpcb &&
655 			    in_localaddr(tp->t_inpcb->inp_faddr))
656 				ss = tcp_init_win_local;
657 #ifdef INET6
658 			if (tp->t_in6pcb &&
659 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
660 				ss = tcp_init_win_local;
661 #endif
662 			tp->snd_cwnd = min(tp->snd_cwnd,
663 			    TCP_INITIAL_WINDOW(ss, txsegsize));
664 		}
665 	}
666 
667 	txsegsize_nosack = txsegsize;
668 again:
669 	ecn_tos = 0;
670 	use_tso = has_tso;
671 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
672 		/* don't duplicate CWR/ECE. */
673 		use_tso = 0;
674 	}
675 	TCP_REASS_LOCK(tp);
676 	sack_numblks = tcp_sack_numblks(tp);
677 	if (sack_numblks) {
678 		int sackoptlen;
679 
680 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
681 		if (sackoptlen > txsegsize_nosack) {
682 			sack_numblks = 0; /* give up SACK */
683 			txsegsize = txsegsize_nosack;
684 		} else {
685 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
686 				/* don't duplicate D-SACK. */
687 				use_tso = 0;
688 			}
689 			txsegsize = txsegsize_nosack - sackoptlen;
690 		}
691 	} else {
692 		txsegsize = txsegsize_nosack;
693 	}
694 
695 	/*
696 	 * Determine length of data that should be transmitted, and
697 	 * flags that should be used.  If there is some data or critical
698 	 * controls (SYN, RST) to send, then transmit; otherwise,
699 	 * investigate further.
700 	 *
701 	 * Readjust SACK information to avoid resending duplicate data.
702 	 */
703 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
704 		tcp_sack_adjust(tp);
705 	sendalot = 0;
706 	off = tp->snd_nxt - tp->snd_una;
707 	win = min(tp->snd_wnd, tp->snd_cwnd);
708 
709 	flags = tcp_outflags[tp->t_state];
710 
711 	/*
712 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
713 	 * to send out new data (when sendalot is 1), bypass this function.
714 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
715 	 * we're replacing a (future) new transmission with a retransmission
716 	 * now, and we previously incremented snd_cwnd in tcp_input().
717 	 */
718 	/*
719 	 * Still in sack recovery, reset rxmit flag to zero.
720 	 */
721 	sack_rxmit = 0;
722 	sack_bytes_rxmt = 0;
723 	len = 0;
724 	p = NULL;
725 	do {
726 		long cwin;
727 		if (!TCP_SACK_ENABLED(tp))
728 			break;
729 		if (tp->t_partialacks < 0)
730 			break;
731 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
732 		if (p == NULL)
733 			break;
734 
735 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
736 		if (cwin < 0)
737 			cwin = 0;
738 		/* Do not retransmit SACK segments beyond snd_recover */
739 		if (SEQ_GT(p->end, tp->snd_recover)) {
740 			/*
741 			 * (At least) part of sack hole extends beyond
742 			 * snd_recover. Check to see if we can rexmit data
743 			 * for this hole.
744 			 */
745 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
746 				/*
747 				 * Can't rexmit any more data for this hole.
748 				 * That data will be rexmitted in the next
749 				 * sack recovery episode, when snd_recover
750 				 * moves past p->rxmit.
751 				 */
752 				p = NULL;
753 				break;
754 			}
755 			/* Can rexmit part of the current hole */
756 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
757 		} else
758 			len = ((long)ulmin(cwin, p->end - p->rxmit));
759 		off = p->rxmit - tp->snd_una;
760 		if (off + len > so->so_snd.sb_cc) {
761 			/* 1 for TH_FIN */
762 			KASSERT(off + len == so->so_snd.sb_cc + 1);
763 			KASSERT(p->rxmit + len == tp->snd_max);
764 			len = so->so_snd.sb_cc - off;
765 		}
766 		if (len > 0) {
767 			sack_rxmit = 1;
768 			sendalot = 1;
769 		}
770 	} while (/*CONSTCOND*/0);
771 
772 	/*
773 	 * If in persist timeout with window of 0, send 1 byte.
774 	 * Otherwise, if window is small but nonzero
775 	 * and timer expired, we will send what we can
776 	 * and go to transmit state.
777 	 */
778 	if (tp->t_force) {
779 		if (win == 0) {
780 			/*
781 			 * If we still have some data to send, then
782 			 * clear the FIN bit.  Usually this would
783 			 * happen below when it realizes that we
784 			 * aren't sending all the data.  However,
785 			 * if we have exactly 1 byte of unset data,
786 			 * then it won't clear the FIN bit below,
787 			 * and if we are in persist state, we wind
788 			 * up sending the packet without recording
789 			 * that we sent the FIN bit.
790 			 *
791 			 * We can't just blindly clear the FIN bit,
792 			 * because if we don't have any more data
793 			 * to send then the probe will be the FIN
794 			 * itself.
795 			 */
796 			if (off < so->so_snd.sb_cc)
797 				flags &= ~TH_FIN;
798 			win = 1;
799 		} else {
800 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
801 			tp->t_rxtshift = 0;
802 		}
803 	}
804 
805 	if (sack_rxmit == 0) {
806 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
807 			long cwin;
808 
809 			/*
810 			 * We are inside of a SACK recovery episode and are
811 			 * sending new data, having retransmitted all the
812 			 * data possible in the scoreboard.
813 			 */
814 			if (tp->snd_wnd < so->so_snd.sb_cc) {
815 				len = tp->snd_wnd - off;
816 				flags &= ~TH_FIN;
817 			} else {
818 				len = so->so_snd.sb_cc - off;
819 			}
820 
821 			/*
822 			 * From FreeBSD:
823 			 *  Don't remove this (len > 0) check !
824 			 *  We explicitly check for len > 0 here (although it
825 			 *  isn't really necessary), to work around a gcc
826 			 *  optimization issue - to force gcc to compute
827 			 *  len above. Without this check, the computation
828 			 *  of len is bungled by the optimizer.
829 			 */
830 			if (len > 0) {
831 				cwin = tp->snd_cwnd -
832 				    (tp->snd_nxt - tp->sack_newdata) -
833 				    sack_bytes_rxmt;
834 				if (cwin < 0)
835 					cwin = 0;
836 				if (cwin < len) {
837 					len = cwin;
838 					flags &= ~TH_FIN;
839 				}
840 			}
841 		} else if (win < so->so_snd.sb_cc) {
842 			len = win - off;
843 			flags &= ~TH_FIN;
844 		} else {
845 			len = so->so_snd.sb_cc - off;
846 		}
847 	}
848 
849 	if (len < 0) {
850 		/*
851 		 * If FIN has been sent but not acked,
852 		 * but we haven't been called to retransmit,
853 		 * len will be -1.  Otherwise, window shrank
854 		 * after we sent into it.  If window shrank to 0,
855 		 * cancel pending retransmit, pull snd_nxt back
856 		 * to (closed) window, and set the persist timer
857 		 * if it isn't already going.  If the window didn't
858 		 * close completely, just wait for an ACK.
859 		 *
860 		 * If we have a pending FIN, either it has already been
861 		 * transmitted or it is outside the window, so drop it.
862 		 * If the FIN has been transmitted, but this is not a
863 		 * retransmission, then len must be -1.  Therefore we also
864 		 * prevent here the sending of `gratuitous FINs'.  This
865 		 * eliminates the need to check for that case below (e.g.
866 		 * to back up snd_nxt before the FIN so that the sequence
867 		 * number is correct).
868 		 */
869 		len = 0;
870 		flags &= ~TH_FIN;
871 		if (win == 0) {
872 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
873 			tp->t_rxtshift = 0;
874 			tp->snd_nxt = tp->snd_una;
875 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
876 				tcp_setpersist(tp);
877 		}
878 	}
879 
880 	/*
881 	 * Automatic sizing enables the performance of large buffers
882 	 * and most of the efficiency of small ones by only allocating
883 	 * space when it is needed.
884 	 *
885 	 * The criteria to step up the send buffer one notch are:
886 	 *  1. receive window of remote host is larger than send buffer
887 	 *     (with a fudge factor of 5/4th);
888 	 *  2. send buffer is filled to 7/8th with data (so we actually
889 	 *     have data to make use of it);
890 	 *  3. send buffer fill has not hit maximal automatic size;
891 	 *  4. our send window (slow start and cogestion controlled) is
892 	 *     larger than sent but unacknowledged data in send buffer.
893 	 *
894 	 * The remote host receive window scaling factor may limit the
895 	 * growing of the send buffer before it reaches its allowed
896 	 * maximum.
897 	 *
898 	 * It scales directly with slow start or congestion window
899 	 * and does at most one step per received ACK.  This fast
900 	 * scaling has the drawback of growing the send buffer beyond
901 	 * what is strictly necessary to make full use of a given
902 	 * delay*bandwith product.  However testing has shown this not
903 	 * to be much of an problem.  At worst we are trading wasting
904 	 * of available bandwith (the non-use of it) for wasting some
905 	 * socket buffer memory.
906 	 *
907 	 * TODO: Shrink send buffer during idle periods together
908 	 * with congestion window.  Requires another timer.
909 	 */
910 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
911 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
912 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
913 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
914 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
915 			if (!sbreserve(&so->so_snd,
916 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
917 			     tcp_autosndbuf_max), so))
918 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
919 		}
920 	}
921 
922 	if (len > txsegsize) {
923 		if (use_tso) {
924 			/*
925 			 * Truncate TSO transfers to IP_MAXPACKET, and make
926 			 * sure that we send equal size transfers down the
927 			 * stack (rather than big-small-big-small-...).
928 			 */
929 #ifdef INET6
930 			CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET);
931 #endif
932 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
933 			if (len <= txsegsize) {
934 				use_tso = 0;
935 			}
936 		} else
937 			len = txsegsize;
938 		flags &= ~TH_FIN;
939 		sendalot = 1;
940 	} else
941 		use_tso = 0;
942 	if (sack_rxmit) {
943 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
944 			flags &= ~TH_FIN;
945 	}
946 
947 	win = sbspace(&so->so_rcv);
948 
949 	/*
950 	 * Sender silly window avoidance.  If connection is idle
951 	 * and can send all data, a maximum segment,
952 	 * at least a maximum default-size segment do it,
953 	 * or are forced, do it; otherwise don't bother.
954 	 * If peer's buffer is tiny, then send
955 	 * when window is at least half open.
956 	 * If retransmitting (possibly after persist timer forced us
957 	 * to send into a small window), then must resend.
958 	 */
959 	if (len) {
960 		if (len >= txsegsize)
961 			goto send;
962 		if ((so->so_state & SS_MORETOCOME) == 0 &&
963 		    ((idle || tp->t_flags & TF_NODELAY) &&
964 		     len + off >= so->so_snd.sb_cc))
965 			goto send;
966 		if (tp->t_force)
967 			goto send;
968 		if (len >= tp->max_sndwnd / 2)
969 			goto send;
970 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
971 			goto send;
972 		if (sack_rxmit)
973 			goto send;
974 	}
975 
976 	/*
977 	 * Compare available window to amount of window known to peer
978 	 * (as advertised window less next expected input).  If the
979 	 * difference is at least twice the size of the largest segment
980 	 * we expect to receive (i.e. two segments) or at least 50% of
981 	 * the maximum possible window, then want to send a window update
982 	 * to peer.
983 	 */
984 	if (win > 0) {
985 		/*
986 		 * "adv" is the amount we can increase the window,
987 		 * taking into account that we are limited by
988 		 * TCP_MAXWIN << tp->rcv_scale.
989 		 */
990 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
991 			(tp->rcv_adv - tp->rcv_nxt);
992 
993 		/*
994 		 * If the new window size ends up being the same as the old
995 		 * size when it is scaled, then don't force a window update.
996 		 */
997 		if ((tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale ==
998 		    (adv + tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale)
999 			goto dontupdate;
1000 		if (adv >= (long) (2 * rxsegsize))
1001 			goto send;
1002 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1003 			goto send;
1004 	}
1005 dontupdate:
1006 
1007 	/*
1008 	 * Send if we owe peer an ACK.
1009 	 */
1010 	if (tp->t_flags & TF_ACKNOW)
1011 		goto send;
1012 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1013 		goto send;
1014 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1015 		goto send;
1016 	/*
1017 	 * In SACK, it is possible for tcp_output to fail to send a segment
1018 	 * after the retransmission timer has been turned off.  Make sure
1019 	 * that the retransmission timer is set.
1020 	 */
1021 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1022 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1023 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1024 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1025 		goto just_return;
1026 	}
1027 
1028 	/*
1029 	 * TCP window updates are not reliable, rather a polling protocol
1030 	 * using ``persist'' packets is used to insure receipt of window
1031 	 * updates.  The three ``states'' for the output side are:
1032 	 *	idle			not doing retransmits or persists
1033 	 *	persisting		to move a small or zero window
1034 	 *	(re)transmitting	and thereby not persisting
1035 	 *
1036 	 * tp->t_timer[TCPT_PERSIST]
1037 	 *	is set when we are in persist state.
1038 	 * tp->t_force
1039 	 *	is set when we are called to send a persist packet.
1040 	 * tp->t_timer[TCPT_REXMT]
1041 	 *	is set when we are retransmitting
1042 	 * The output side is idle when both timers are zero.
1043 	 *
1044 	 * If send window is too small, there is data to transmit, and no
1045 	 * retransmit or persist is pending, then go to persist state.
1046 	 * If nothing happens soon, send when timer expires:
1047 	 * if window is nonzero, transmit what we can,
1048 	 * otherwise force out a byte.
1049 	 */
1050 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1051 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1052 		tp->t_rxtshift = 0;
1053 		tcp_setpersist(tp);
1054 	}
1055 
1056 	/*
1057 	 * No reason to send a segment, just return.
1058 	 */
1059 just_return:
1060 	TCP_REASS_UNLOCK(tp);
1061 	return 0;
1062 
1063 send:
1064 	/*
1065 	 * Before ESTABLISHED, force sending of initial options
1066 	 * unless TCP set not to do any options.
1067 	 * NOTE: we assume that the IP/TCP header plus TCP options
1068 	 * always fit in a single mbuf, leaving room for a maximum
1069 	 * link header, i.e.
1070 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1071 	 */
1072 	optlen = 0;
1073 	optp = opt;
1074 	switch (af) {
1075 	case AF_INET:
1076 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1077 		break;
1078 #ifdef INET6
1079 	case AF_INET6:
1080 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1081 		break;
1082 #endif
1083 	default:	/*pacify gcc*/
1084 		iphdrlen = 0;
1085 		break;
1086 	}
1087 	hdrlen = iphdrlen;
1088 	if (flags & TH_SYN) {
1089 		struct rtentry *synrt;
1090 
1091 		synrt = NULL;
1092 		if (tp->t_inpcb)
1093 			synrt = in_pcbrtentry(tp->t_inpcb);
1094 #ifdef INET6
1095 		if (tp->t_in6pcb)
1096 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1097 #endif
1098 
1099 		tp->snd_nxt = tp->iss;
1100 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1101 						    synrt->rt_ifp : NULL, af);
1102 		if (tp->t_inpcb)
1103 			in_pcbrtentry_unref(synrt, tp->t_inpcb);
1104 #ifdef INET6
1105 		if (tp->t_in6pcb)
1106 			in6_pcbrtentry_unref(synrt, tp->t_in6pcb);
1107 #endif
1108 		if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(TCPOLEN_MAXSEG)) {
1109 			*optp++ = TCPOPT_MAXSEG;
1110 			*optp++ = TCPOLEN_MAXSEG;
1111 			*optp++ = (tp->t_ourmss >> 8) & 0xff;
1112 			*optp++ = tp->t_ourmss & 0xff;
1113 			optlen += TCPOLEN_MAXSEG;
1114 
1115 			if ((tp->t_flags & TF_REQ_SCALE) &&
1116 			    ((flags & TH_ACK) == 0 ||
1117 			    (tp->t_flags & TF_RCVD_SCALE)) &&
1118 			    OPT_FITS(TCPOLEN_WINDOW + TCPOLEN_NOP)) {
1119 				*((uint32_t *)optp) = htonl(
1120 					TCPOPT_NOP << 24 |
1121 					TCPOPT_WINDOW << 16 |
1122 					TCPOLEN_WINDOW << 8 |
1123 					tp->request_r_scale);
1124 				optp += TCPOLEN_WINDOW + TCPOLEN_NOP;
1125 				optlen += TCPOLEN_WINDOW + TCPOLEN_NOP;
1126 			}
1127 			if (tcp_do_sack && OPT_FITS(TCPOLEN_SACK_PERMITTED)) {
1128 				*optp++ = TCPOPT_SACK_PERMITTED;
1129 				*optp++ = TCPOLEN_SACK_PERMITTED;
1130 				optlen += TCPOLEN_SACK_PERMITTED;
1131 			}
1132 		}
1133 	}
1134 
1135 	/*
1136 	 * Send a timestamp and echo-reply if this is a SYN and our side
1137 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1138 	 * and our peer have sent timestamps in our SYN's.
1139 	 */
1140 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1141 	     (flags & TH_RST) == 0 &&
1142 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1143 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1144 		int alen = 0;
1145 		while (optlen % 4 != 2) {
1146 			optlen += TCPOLEN_NOP;
1147 			*optp++ = TCPOPT_NOP;
1148 			alen++;
1149 		}
1150 		if (OPT_FITS(TCPOLEN_TIMESTAMP)) {
1151 			*optp++ = TCPOPT_TIMESTAMP;
1152 			*optp++ = TCPOLEN_TIMESTAMP;
1153 			uint32_t *lp = (uint32_t *)optp;
1154 			/* Form timestamp option (appendix A of RFC 1323) */
1155 			*lp++ = htonl(TCP_TIMESTAMP(tp));
1156 			*lp   = htonl(tp->ts_recent);
1157 			optp += TCPOLEN_TIMESTAMP - 2;
1158 			optlen += TCPOLEN_TIMESTAMP;
1159 
1160 			/* Set receive buffer autosizing timestamp. */
1161 			if (tp->rfbuf_ts == 0 &&
1162 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
1163 				tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1164 		} else {
1165 			optp -= alen;
1166 			optlen -= alen;
1167 		}
1168 	}
1169 
1170 #ifdef TCP_SIGNATURE
1171 	if (tp->t_flags & TF_SIGNATURE) {
1172 		/*
1173 		 * Initialize TCP-MD5 option (RFC2385)
1174 		 */
1175 		if (!OPT_FITS(TCPOLEN_SIGNATURE))
1176 			goto reset;
1177 
1178 		*optp++ = TCPOPT_SIGNATURE;
1179 		*optp++ = TCPOLEN_SIGNATURE;
1180 		sigoff = optlen + 2;
1181 		memset(optp, 0, TCP_SIGLEN);
1182 		optlen += TCPOLEN_SIGNATURE;
1183 		optp += TCP_SIGLEN;
1184 	}
1185 #endif
1186 
1187 	/*
1188 	 * Tack on the SACK block if it is necessary.
1189 	 */
1190 	if (sack_numblks) {
1191 		int alen = 0;
1192 		int sack_len = sack_numblks * 8;
1193 		while (optlen % 4 != 2) {
1194 			optlen += TCPOLEN_NOP;
1195 			*optp++ = TCPOPT_NOP;
1196 			alen++;
1197 		}
1198 		if (OPT_FITS(sack_len + 2)) {
1199 			struct ipqent *tiqe;
1200 			*optp++ = TCPOPT_SACK;
1201 			*optp++ = sack_len + 2;
1202 			uint32_t *lp = (uint32_t *)optp;
1203 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1204 				sack_numblks--;
1205 				*lp++ = htonl(tp->rcv_dsack_block.left);
1206 				*lp++ = htonl(tp->rcv_dsack_block.right);
1207 				tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1208 			}
1209 			for (tiqe = TAILQ_FIRST(&tp->timeq);
1210 			    sack_numblks > 0;
1211 			    tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1212 				KASSERT(tiqe != NULL);
1213 				sack_numblks--;
1214 				*lp++ = htonl(tiqe->ipqe_seq);
1215 				*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1216 				    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1217 			}
1218 			optlen += sack_len + 2;
1219 			optp += sack_len;
1220 		} else {
1221 			optp -= alen;
1222 			optlen -= alen;
1223 		}
1224 	}
1225 
1226 	/* Terminate and pad TCP options to a 4 byte boundary. */
1227 	if (optlen % 4) {
1228 		if (!OPT_FITS(TCPOLEN_EOL)) {
1229 reset:			TCP_REASS_UNLOCK(tp);
1230 			error = ECONNABORTED;
1231 			goto out;
1232 		}
1233 		optlen += TCPOLEN_EOL;
1234 		*optp++ = TCPOPT_EOL;
1235 	}
1236 	/*
1237 	 * According to RFC 793 (STD0007):
1238 	 *   "The content of the header beyond the End-of-Option option
1239 	 *    must be header padding (i.e., zero)."
1240 	 *   and later: "The padding is composed of zeros."
1241 	 */
1242 	while (optlen % 4) {
1243 		if (!OPT_FITS(TCPOLEN_PAD))
1244 			goto reset;
1245 		optlen += TCPOLEN_PAD;
1246 		*optp++ = TCPOPT_PAD;
1247 	}
1248 
1249 	TCP_REASS_UNLOCK(tp);
1250 
1251 	hdrlen += optlen;
1252 
1253 #ifdef DIAGNOSTIC
1254 	if (!use_tso && len > txsegsize)
1255 		panic("tcp data to be sent is larger than segment");
1256 	else if (use_tso && len > IP_MAXPACKET)
1257 		panic("tcp data to be sent is larger than max TSO size");
1258 	if (max_linkhdr + hdrlen > MCLBYTES)
1259 		panic("tcphdr too big");
1260 #endif
1261 
1262 	/*
1263 	 * Grab a header mbuf, attaching a copy of data to
1264 	 * be transmitted, and initialize the header from
1265 	 * the template for sends on this connection.
1266 	 */
1267 	if (len) {
1268 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1269 		if (error)
1270 			goto out;
1271 		/*
1272 		 * If we're sending everything we've got, set PUSH.
1273 		 * (This will keep happy those implementations which only
1274 		 * give data to the user when a buffer fills or
1275 		 * a PUSH comes in.)
1276 		 */
1277 		if (off + len == so->so_snd.sb_cc)
1278 			flags |= TH_PUSH;
1279 	} else {
1280 		tcps = TCP_STAT_GETREF();
1281 		if (tp->t_flags & TF_ACKNOW)
1282 			tcps[TCP_STAT_SNDACKS]++;
1283 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1284 			tcps[TCP_STAT_SNDCTRL]++;
1285 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1286 			tcps[TCP_STAT_SNDURG]++;
1287 		else
1288 			tcps[TCP_STAT_SNDWINUP]++;
1289 		TCP_STAT_PUTREF();
1290 
1291 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1292 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1293 			MCLGET(m, M_DONTWAIT);
1294 			if ((m->m_flags & M_EXT) == 0) {
1295 				m_freem(m);
1296 				m = NULL;
1297 			}
1298 		}
1299 		if (m == NULL) {
1300 			error = ENOBUFS;
1301 			goto out;
1302 		}
1303 		MCLAIM(m, &tcp_tx_mowner);
1304 		m->m_data += max_linkhdr;
1305 		m->m_len = hdrlen;
1306 	}
1307 	m_reset_rcvif(m);
1308 	switch (af) {
1309 	case AF_INET:
1310 		ip = mtod(m, struct ip *);
1311 #ifdef INET6
1312 		ip6 = NULL;
1313 #endif
1314 		th = (struct tcphdr *)(ip + 1);
1315 		break;
1316 #ifdef INET6
1317 	case AF_INET6:
1318 		ip = NULL;
1319 		ip6 = mtod(m, struct ip6_hdr *);
1320 		th = (struct tcphdr *)(ip6 + 1);
1321 		break;
1322 #endif
1323 	default:	/*pacify gcc*/
1324 		ip = NULL;
1325 #ifdef INET6
1326 		ip6 = NULL;
1327 #endif
1328 		th = NULL;
1329 		break;
1330 	}
1331 	if (tp->t_template == NULL)
1332 		panic("tcp_output");
1333 	if (tp->t_template->m_len < iphdrlen)
1334 		panic("tcp_output");
1335 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1336 
1337 	/*
1338 	 * If we are starting a connection, send ECN setup
1339 	 * SYN packet. If we are on a retransmit, we may
1340 	 * resend those bits a number of times as per
1341 	 * RFC 3168.
1342 	 */
1343 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1344 		if (tp->t_flags & TF_SYN_REXMT) {
1345 			if (tp->t_ecn_retries--)
1346 				flags |= TH_ECE|TH_CWR;
1347 		} else {
1348 			flags |= TH_ECE|TH_CWR;
1349 			tp->t_ecn_retries = tcp_ecn_maxretries;
1350 		}
1351 	}
1352 
1353 	if (TCP_ECN_ALLOWED(tp)) {
1354 		/*
1355 		 * If the peer has ECN, mark data packets
1356 		 * ECN capable. Ignore pure ack packets, retransmissions
1357 		 * and window probes.
1358 		 */
1359 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1360 		    !(tp->t_force && len == 1)) {
1361 			ecn_tos = IPTOS_ECN_ECT0;
1362 			TCP_STATINC(TCP_STAT_ECN_ECT);
1363 		}
1364 
1365 		/*
1366 		 * Reply with proper ECN notifications.
1367 		 */
1368 		if (tp->t_flags & TF_ECN_SND_CWR) {
1369 			flags |= TH_CWR;
1370 			tp->t_flags &= ~TF_ECN_SND_CWR;
1371 		}
1372 		if (tp->t_flags & TF_ECN_SND_ECE) {
1373 			flags |= TH_ECE;
1374 		}
1375 	}
1376 
1377 	/*
1378 	 * If we are doing retransmissions, then snd_nxt will
1379 	 * not reflect the first unsent octet.  For ACK only
1380 	 * packets, we do not want the sequence number of the
1381 	 * retransmitted packet, we want the sequence number
1382 	 * of the next unsent octet.  So, if there is no data
1383 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1384 	 * when filling in ti_seq.  But if we are in persist
1385 	 * state, snd_max might reflect one byte beyond the
1386 	 * right edge of the window, so use snd_nxt in that
1387 	 * case, since we know we aren't doing a retransmission.
1388 	 * (retransmit and persist are mutually exclusive...)
1389 	 */
1390 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1391 		th->th_seq = htonl(p->rxmit);
1392 		p->rxmit += len;
1393 	} else {
1394 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1395 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1396 			th->th_seq = htonl(tp->snd_nxt);
1397 		else
1398 			th->th_seq = htonl(tp->snd_max);
1399 	}
1400 	th->th_ack = htonl(tp->rcv_nxt);
1401 	if (optlen) {
1402 		memcpy(th + 1, opt, optlen);
1403 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1404 	}
1405 	th->th_flags = flags;
1406 	/*
1407 	 * Calculate receive window.  Don't shrink window,
1408 	 * but avoid silly window syndrome.
1409 	 */
1410 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1411 		win = 0;
1412 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1413 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1414 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1415 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1416 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1417 	if (th->th_win == 0) {
1418 		tp->t_sndzerowin++;
1419 	}
1420 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1421 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1422 		if (urp > IP_MAXPACKET)
1423 			urp = IP_MAXPACKET;
1424 		th->th_urp = htons((u_int16_t)urp);
1425 		th->th_flags |= TH_URG;
1426 	} else
1427 		/*
1428 		 * If no urgent pointer to send, then we pull
1429 		 * the urgent pointer to the left edge of the send window
1430 		 * so that it doesn't drift into the send window on sequence
1431 		 * number wraparound.
1432 		 */
1433 		tp->snd_up = tp->snd_una;		/* drag it along */
1434 
1435 #ifdef TCP_SIGNATURE
1436 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1437 		struct secasvar *sav;
1438 		u_int8_t *sigp;
1439 
1440 		sav = tcp_signature_getsav(m);
1441 		if (sav == NULL) {
1442 			if (m)
1443 				m_freem(m);
1444 			return EPERM;
1445 		}
1446 
1447 		m->m_pkthdr.len = hdrlen + len;
1448 		sigp = (char *)th + sizeof(*th) + sigoff;
1449 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1450 
1451 		key_sa_recordxfer(sav, m);
1452 		KEY_SA_UNREF(&sav);
1453 	}
1454 #endif
1455 
1456 	/*
1457 	 * Set ourselves up to be checksummed just before the packet
1458 	 * hits the wire.
1459 	 */
1460 	switch (af) {
1461 	case AF_INET:
1462 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1463 		if (use_tso) {
1464 			m->m_pkthdr.segsz = txsegsize;
1465 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1466 		} else {
1467 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1468 			if (len + optlen) {
1469 				/* Fixup the pseudo-header checksum. */
1470 				/* XXXJRT Not IP Jumbogram safe. */
1471 				th->th_sum = in_cksum_addword(th->th_sum,
1472 				    htons((u_int16_t) (len + optlen)));
1473 			}
1474 		}
1475 		break;
1476 #ifdef INET6
1477 	case AF_INET6:
1478 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1479 		if (use_tso) {
1480 			m->m_pkthdr.segsz = txsegsize;
1481 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1482 		} else {
1483 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1484 			if (len + optlen) {
1485 				/* Fixup the pseudo-header checksum. */
1486 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1487 				th->th_sum = in_cksum_addword(th->th_sum,
1488 				    htons((u_int16_t) (len + optlen)));
1489 			}
1490 		}
1491 		break;
1492 #endif
1493 	}
1494 
1495 	/*
1496 	 * In transmit state, time the transmission and arrange for
1497 	 * the retransmit.  In persist state, just set snd_max.
1498 	 */
1499 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1500 		tcp_seq startseq = tp->snd_nxt;
1501 
1502 		/*
1503 		 * Advance snd_nxt over sequence space of this segment.
1504 		 * There are no states in which we send both a SYN and a FIN,
1505 		 * so we collapse the tests for these flags.
1506 		 */
1507 		if (flags & (TH_SYN|TH_FIN))
1508 			tp->snd_nxt++;
1509 		if (sack_rxmit)
1510 			goto timer;
1511 		tp->snd_nxt += len;
1512 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1513 			tp->snd_max = tp->snd_nxt;
1514 			/*
1515 			 * Time this transmission if not a retransmission and
1516 			 * not currently timing anything.
1517 			 */
1518 			if (tp->t_rtttime == 0) {
1519 				tp->t_rtttime = tcp_now;
1520 				tp->t_rtseq = startseq;
1521 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1522 			}
1523 		}
1524 
1525 		/*
1526 		 * Set retransmit timer if not currently set,
1527 		 * and not doing an ack or a keep-alive probe.
1528 		 * Initial value for retransmit timer is smoothed
1529 		 * round-trip time + 2 * round-trip time variance.
1530 		 * Initialize shift counter which is used for backoff
1531 		 * of retransmit time.
1532 		 */
1533 timer:
1534 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0) {
1535 			if ((sack_rxmit && tp->snd_nxt != tp->snd_max)
1536 			    || tp->snd_nxt != tp->snd_una) {
1537 				if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1538 					TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1539 					tp->t_rxtshift = 0;
1540 				}
1541 				TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1542 			} else if (len == 0 && so->so_snd.sb_cc > 0
1543 			    && TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1544 				/*
1545 				 * If we are sending a window probe and there's
1546 				 * unacked data in the socket, make sure at
1547 				 * least the persist timer is running.
1548 				 */
1549 				tp->t_rxtshift = 0;
1550 				tcp_setpersist(tp);
1551 			}
1552 		}
1553 	} else
1554 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1555 			tp->snd_max = tp->snd_nxt + len;
1556 
1557 #ifdef TCP_DEBUG
1558 	/*
1559 	 * Trace.
1560 	 */
1561 	if (so->so_options & SO_DEBUG)
1562 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1563 #endif
1564 
1565 	/*
1566 	 * Fill in IP length and desired time to live and
1567 	 * send to IP level.  There should be a better way
1568 	 * to handle ttl and tos; we could keep them in
1569 	 * the template, but need a way to checksum without them.
1570 	 */
1571 	m->m_pkthdr.len = hdrlen + len;
1572 
1573 	switch (af) {
1574 	case AF_INET:
1575 		ip->ip_len = htons(m->m_pkthdr.len);
1576 		packetlen = m->m_pkthdr.len;
1577 		if (tp->t_inpcb) {
1578 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1579 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos;
1580 		}
1581 #ifdef INET6
1582 		else if (tp->t_in6pcb) {
1583 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1584 			ip->ip_tos = ecn_tos;	/*XXX*/
1585 		}
1586 #endif
1587 		break;
1588 #ifdef INET6
1589 	case AF_INET6:
1590 		packetlen = m->m_pkthdr.len;
1591 		ip6->ip6_nxt = IPPROTO_TCP;
1592 		if (tp->t_in6pcb) {
1593 			/*
1594 			 * we separately set hoplimit for every segment, since
1595 			 * the user might want to change the value via
1596 			 * setsockopt. Also, desired default hop limit might
1597 			 * be changed via Neighbor Discovery.
1598 			 */
1599 			ip6->ip6_hlim = in6_selecthlim_rt(tp->t_in6pcb);
1600 		}
1601 		ip6->ip6_flow |= htonl(ecn_tos << 20);
1602 		/* ip6->ip6_flow = ??? (from template) */
1603 		/* ip6_plen will be filled in ip6_output(). */
1604 		break;
1605 #endif
1606 	default:	/*pacify gcc*/
1607 		packetlen = 0;
1608 		break;
1609 	}
1610 
1611 	switch (af) {
1612 	case AF_INET:
1613 	    {
1614 		struct mbuf *opts;
1615 
1616 		if (tp->t_inpcb)
1617 			opts = tp->t_inpcb->inp_options;
1618 		else
1619 			opts = NULL;
1620 		error = ip_output(m, opts, ro,
1621 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1622 			(so->so_options & SO_DONTROUTE), NULL, tp->t_inpcb);
1623 		break;
1624 	    }
1625 #ifdef INET6
1626 	case AF_INET6:
1627 	    {
1628 		struct ip6_pktopts *opts;
1629 
1630 		if (tp->t_in6pcb)
1631 			opts = tp->t_in6pcb->in6p_outputopts;
1632 		else
1633 			opts = NULL;
1634 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1635 			NULL, tp->t_in6pcb, NULL);
1636 		break;
1637 	    }
1638 #endif
1639 	default:
1640 		error = EAFNOSUPPORT;
1641 		break;
1642 	}
1643 	if (error) {
1644 out:
1645 		if (error == ENOBUFS) {
1646 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1647 			if (tp->t_inpcb)
1648 				tcp_quench(tp->t_inpcb, 0);
1649 #ifdef INET6
1650 			if (tp->t_in6pcb)
1651 				tcp6_quench(tp->t_in6pcb, 0);
1652 #endif
1653 			error = 0;
1654 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1655 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1656 			tp->t_softerror = error;
1657 			error = 0;
1658 		}
1659 
1660 		/* Back out the sequence number advance. */
1661 		if (sack_rxmit)
1662 			p->rxmit -= len;
1663 
1664 		/* Restart the delayed ACK timer, if necessary. */
1665 		if (tp->t_flags & TF_DELACK)
1666 			TCP_RESTART_DELACK(tp);
1667 
1668 		return error;
1669 	}
1670 
1671 	if (packetlen > tp->t_pmtud_mtu_sent)
1672 		tp->t_pmtud_mtu_sent = packetlen;
1673 
1674 	tcps = TCP_STAT_GETREF();
1675 	tcps[TCP_STAT_SNDTOTAL]++;
1676 	if (tp->t_flags & TF_DELACK)
1677 		tcps[TCP_STAT_DELACK]++;
1678 	TCP_STAT_PUTREF();
1679 
1680 	/*
1681 	 * Data sent (as far as we can tell).
1682 	 * If this advertises a larger window than any other segment,
1683 	 * then remember the size of the advertised window.
1684 	 * Any pending ACK has now been sent.
1685 	 */
1686 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1687 		tp->rcv_adv = tp->rcv_nxt + win;
1688 	tp->last_ack_sent = tp->rcv_nxt;
1689 	tp->t_flags &= ~TF_ACKNOW;
1690 	TCP_CLEAR_DELACK(tp);
1691 #ifdef DIAGNOSTIC
1692 	if (maxburst < 0)
1693 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1694 #endif
1695 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1696 		goto again;
1697 	return 0;
1698 }
1699 
1700 void
1701 tcp_setpersist(struct tcpcb *tp)
1702 {
1703 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1704 	int nticks;
1705 
1706 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1707 		panic("tcp_output REXMT");
1708 	/*
1709 	 * Start/restart persistance timer.
1710 	 */
1711 	if (t < tp->t_rttmin)
1712 		t = tp->t_rttmin;
1713 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1714 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1715 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1716 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1717 		tp->t_rxtshift++;
1718 }
1719