xref: /netbsd-src/sys/netinet/tcp_output.c (revision b5677b36047b601b9addaaa494a58ceae82c2a6c)
1 /*	$NetBSD: tcp_output.c,v 1.168 2009/03/18 16:00:22 cegger 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 1.168 2009/03/18 16:00:22 cegger Exp $");
139 
140 #include "opt_inet.h"
141 #include "opt_ipsec.h"
142 #include "opt_tcp_debug.h"
143 
144 #include <sys/param.h>
145 #include <sys/systm.h>
146 #include <sys/malloc.h>
147 #include <sys/mbuf.h>
148 #include <sys/protosw.h>
149 #include <sys/socket.h>
150 #include <sys/socketvar.h>
151 #include <sys/errno.h>
152 #include <sys/domain.h>
153 #include <sys/kernel.h>
154 #ifdef TCP_SIGNATURE
155 #include <sys/md5.h>
156 #endif
157 
158 #include <net/if.h>
159 #include <net/route.h>
160 
161 #include <netinet/in.h>
162 #include <netinet/in_systm.h>
163 #include <netinet/ip.h>
164 #include <netinet/in_pcb.h>
165 #include <netinet/ip_var.h>
166 
167 #ifdef INET6
168 #ifndef INET
169 #include <netinet/in.h>
170 #endif
171 #include <netinet/ip6.h>
172 #include <netinet6/in6_var.h>
173 #include <netinet6/ip6_var.h>
174 #include <netinet6/in6_pcb.h>
175 #include <netinet6/nd6.h>
176 #endif
177 
178 #ifdef FAST_IPSEC
179 #include <netipsec/ipsec.h>
180 #include <netipsec/key.h>
181 #ifdef INET6
182 #include <netipsec/ipsec6.h>
183 #endif
184 #endif	/* FAST_IPSEC*/
185 #ifdef IPSEC
186 #include <netinet6/ipsec.h>
187 #endif
188 
189 #include <netinet/tcp.h>
190 #define	TCPOUTFLAGS
191 #include <netinet/tcp_fsm.h>
192 #include <netinet/tcp_seq.h>
193 #include <netinet/tcp_timer.h>
194 #include <netinet/tcp_var.h>
195 #include <netinet/tcp_private.h>
196 #include <netinet/tcp_congctl.h>
197 #include <netinet/tcpip.h>
198 #include <netinet/tcp_debug.h>
199 #include <netinet/in_offload.h>
200 #include <netinet6/in6_offload.h>
201 
202 #ifdef IPSEC
203 #include <netkey/key.h>
204 #endif
205 
206 #ifdef notyet
207 extern struct mbuf *m_copypack();
208 #endif
209 
210 /*
211  * Knob to enable Congestion Window Monitoring, and control
212  * the burst size it allows.  Default burst is 4 packets, per
213  * the Internet draft.
214  */
215 int	tcp_cwm = 0;
216 int	tcp_cwm_burstsize = 4;
217 
218 int	tcp_do_autosndbuf = 0;
219 int	tcp_autosndbuf_inc = 8 * 1024;
220 int	tcp_autosndbuf_max = 256 * 1024;
221 
222 #ifdef TCP_OUTPUT_COUNTERS
223 #include <sys/device.h>
224 
225 extern struct evcnt tcp_output_bigheader;
226 extern struct evcnt tcp_output_predict_hit;
227 extern struct evcnt tcp_output_predict_miss;
228 extern struct evcnt tcp_output_copysmall;
229 extern struct evcnt tcp_output_copybig;
230 extern struct evcnt tcp_output_refbig;
231 
232 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
233 #else
234 
235 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
236 
237 #endif /* TCP_OUTPUT_COUNTERS */
238 
239 static
240 #ifndef GPROF
241 inline
242 #endif
243 int
244 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
245     bool *alwaysfragp)
246 {
247 #ifdef INET
248 	struct inpcb *inp = tp->t_inpcb;
249 #endif
250 #ifdef INET6
251 	struct in6pcb *in6p = tp->t_in6pcb;
252 #endif
253 	struct socket *so = NULL;
254 	struct rtentry *rt;
255 	struct ifnet *ifp;
256 	int size;
257 	int hdrlen;
258 	int optlen;
259 
260 	*alwaysfragp = false;
261 
262 #ifdef DIAGNOSTIC
263 	if (tp->t_inpcb && tp->t_in6pcb)
264 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
265 #endif
266 	switch (tp->t_family) {
267 #ifdef INET
268 	case AF_INET:
269 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
270 		break;
271 #endif
272 #ifdef INET6
273 	case AF_INET6:
274 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
275 		break;
276 #endif
277 	default:
278 		size = tcp_mssdflt;
279 		goto out;
280 	}
281 
282 	rt = NULL;
283 #ifdef INET
284 	if (inp) {
285 		rt = in_pcbrtentry(inp);
286 		so = inp->inp_socket;
287 	}
288 #endif
289 #ifdef INET6
290 	if (in6p) {
291 		rt = in6_pcbrtentry(in6p);
292 		so = in6p->in6p_socket;
293 	}
294 #endif
295 	if (rt == NULL) {
296 		size = tcp_mssdflt;
297 		goto out;
298 	}
299 
300 	ifp = rt->rt_ifp;
301 
302 	size = tcp_mssdflt;
303 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
304 #ifdef INET6
305 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
306 			/*
307 			 * RFC2460 section 5, last paragraph: if path MTU is
308 			 * smaller than 1280, use 1280 as packet size and
309 			 * attach fragment header.
310 			 */
311 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
312 			*alwaysfragp = true;
313 		} else
314 			size = rt->rt_rmx.rmx_mtu - hdrlen;
315 #else
316 		size = rt->rt_rmx.rmx_mtu - hdrlen;
317 #endif
318 	} else if (ifp->if_flags & IFF_LOOPBACK)
319 		size = ifp->if_mtu - hdrlen;
320 #ifdef INET
321 	else if (inp && tp->t_mtudisc)
322 		size = ifp->if_mtu - hdrlen;
323 	else if (inp && in_localaddr(inp->inp_faddr))
324 		size = ifp->if_mtu - hdrlen;
325 #endif
326 #ifdef INET6
327 	else if (in6p) {
328 #ifdef INET
329 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
330 			/* mapped addr case */
331 			struct in_addr d;
332 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
333 			if (tp->t_mtudisc || in_localaddr(d))
334 				size = ifp->if_mtu - hdrlen;
335 		} else
336 #endif
337 		{
338 			/*
339 			 * for IPv6, path MTU discovery is always turned on,
340 			 * or the node must use packet size <= 1280.
341 			 */
342 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
343 			size -= hdrlen;
344 		}
345 	}
346 #endif
347  out:
348 	/*
349 	 * Now we must make room for whatever extra TCP/IP options are in
350 	 * the packet.
351 	 */
352 	optlen = tcp_optlen(tp);
353 
354 	/*
355 	 * XXX tp->t_ourmss should have the right size, but without this code
356 	 * fragmentation will occur... need more investigation
357 	 */
358 #ifdef INET
359 	if (inp) {
360 #if defined(IPSEC) || defined(FAST_IPSEC)
361 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
362 			optlen += ipsec4_hdrsiz_tcp(tp);
363 #endif
364 		optlen += ip_optlen(inp);
365 	}
366 #endif
367 #ifdef INET6
368 #ifdef INET
369 	if (in6p && tp->t_family == AF_INET) {
370 #if defined(IPSEC) || defined(FAST_IPSEC)
371 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
372 			optlen += ipsec4_hdrsiz_tcp(tp);
373 #endif
374 		/* XXX size -= ip_optlen(in6p); */
375 	} else
376 #endif
377 	if (in6p && tp->t_family == AF_INET6) {
378 #if defined(IPSEC) || defined(FAST_IPSEC)
379 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
380 			optlen += ipsec6_hdrsiz_tcp(tp);
381 #endif
382 		optlen += ip6_optlen(in6p);
383 	}
384 #endif
385 	size -= optlen;
386 
387 	/* there may not be any room for data if mtu is too small */
388 	if (size < 0)
389 		return (EMSGSIZE);
390 
391 	/*
392 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
393 	 * assumes that path MTU is the same for both ways).  this is only
394 	 * for silly window avoidance, do not use the value for other purposes.
395 	 *
396 	 * ipseclen is subtracted from both sides, this may not be right.
397 	 * I'm not quite sure about this (could someone comment).
398 	 */
399 	*txsegsizep = min(tp->t_peermss - optlen, size);
400 	/*
401 	 * Never send more than half a buffer full.  This insures that we can
402 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
403 	 * therefore acks will never be delayed unless we run out of data to
404 	 * transmit.
405 	 */
406 	if (so)
407 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
408 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
409 
410 	if (*txsegsizep != tp->t_segsz) {
411 		/*
412 		 * If the new segment size is larger, we don't want to
413 		 * mess up the congestion window, but if it is smaller
414 		 * we'll have to reduce the congestion window to ensure
415 		 * that we don't get into trouble with initial windows
416 		 * and the rest.  In any case, if the segment size
417 		 * has changed, chances are the path has, too, and
418 		 * our congestion window will be different.
419 		 */
420 		if (*txsegsizep < tp->t_segsz) {
421 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
422 					   * *txsegsizep, *txsegsizep);
423 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
424 						* *txsegsizep, *txsegsizep);
425 		}
426 		tp->t_segsz = *txsegsizep;
427 	}
428 
429 	return (0);
430 }
431 
432 static
433 #ifndef GPROF
434 inline
435 #endif
436 int
437 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
438     long len, int hdrlen, struct mbuf **mp)
439 {
440 	struct mbuf *m, *m0;
441 	uint64_t *tcps;
442 
443 	tcps = TCP_STAT_GETREF();
444 	if (tp->t_force && len == 1)
445 		tcps[TCP_STAT_SNDPROBE]++;
446 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
447 		tcps[TCP_STAT_SNDREXMITPACK]++;
448 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
449 	} else {
450 		tcps[TCP_STAT_SNDPACK]++;
451 		tcps[TCP_STAT_SNDBYTE] += len;
452 	}
453 	TCP_STAT_PUTREF();
454 #ifdef notyet
455 	if ((m = m_copypack(so->so_snd.sb_mb, off,
456 	    (int)len, max_linkhdr + hdrlen)) == 0)
457 		return (ENOBUFS);
458 	/*
459 	 * m_copypack left space for our hdr; use it.
460 	 */
461 	m->m_len += hdrlen;
462 	m->m_data -= hdrlen;
463 #else
464 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
465 	if (__predict_false(m == NULL))
466 		return (ENOBUFS);
467 	MCLAIM(m, &tcp_tx_mowner);
468 
469 	/*
470 	 * XXX Because other code assumes headers will fit in
471 	 * XXX one header mbuf.
472 	 *
473 	 * (This code should almost *never* be run.)
474 	 */
475 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
476 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
477 		MCLGET(m, M_DONTWAIT);
478 		if ((m->m_flags & M_EXT) == 0) {
479 			m_freem(m);
480 			return (ENOBUFS);
481 		}
482 	}
483 
484 	m->m_data += max_linkhdr;
485 	m->m_len = hdrlen;
486 
487 	/*
488 	 * To avoid traversing the whole sb_mb chain for correct
489 	 * data to send, remember last sent mbuf, its offset and
490 	 * the sent size.  When called the next time, see if the
491 	 * data to send is directly following the previous transfer.
492 	 * This is important for large TCP windows.
493 	 */
494 	if (off == 0 || tp->t_lastm == NULL ||
495 	    (tp->t_lastoff + tp->t_lastlen) != off) {
496 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
497 		/*
498 		 * Either a new packet or a retransmit.
499 		 * Start from the beginning.
500 		 */
501 		tp->t_lastm = so->so_snd.sb_mb;
502 		tp->t_inoff = off;
503 	} else {
504 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
505 		tp->t_inoff += tp->t_lastlen;
506 	}
507 
508 	/* Traverse forward to next packet */
509 	while (tp->t_inoff > 0) {
510 		if (tp->t_lastm == NULL)
511 			panic("tp->t_lastm == NULL");
512 		if (tp->t_inoff < tp->t_lastm->m_len)
513 			break;
514 		tp->t_inoff -= tp->t_lastm->m_len;
515 		tp->t_lastm = tp->t_lastm->m_next;
516 	}
517 
518 	tp->t_lastoff = off;
519 	tp->t_lastlen = len;
520 	m0 = tp->t_lastm;
521 	off = tp->t_inoff;
522 
523 	if (len <= M_TRAILINGSPACE(m)) {
524 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
525 		m->m_len += len;
526 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
527 	} else {
528 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
529 		if (m->m_next == NULL) {
530 			m_freem(m);
531 			return (ENOBUFS);
532 		}
533 #ifdef TCP_OUTPUT_COUNTERS
534 		if (m->m_next->m_flags & M_EXT)
535 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
536 		else
537 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
538 #endif /* TCP_OUTPUT_COUNTERS */
539 	}
540 #endif
541 
542 	*mp = m;
543 	return (0);
544 }
545 
546 /*
547  * Tcp output routine: figure out what should be sent and send it.
548  */
549 int
550 tcp_output(struct tcpcb *tp)
551 {
552 	struct rtentry *rt;
553 	struct socket *so;
554 	struct route *ro;
555 	long len, win;
556 	int off, flags, error;
557 	struct mbuf *m;
558 	struct ip *ip;
559 #ifdef INET6
560 	struct ip6_hdr *ip6;
561 #endif
562 	struct tcphdr *th;
563 	u_char opt[MAX_TCPOPTLEN];
564 	unsigned optlen, hdrlen, packetlen;
565 	unsigned int sack_numblks;
566 	int idle, sendalot, txsegsize, rxsegsize;
567 	int txsegsize_nosack;
568 	int maxburst = TCP_MAXBURST;
569 	int af;		/* address family on the wire */
570 	int iphdrlen;
571 	int has_tso4, has_tso6;
572 	int has_tso, use_tso;
573 	bool alwaysfrag;
574 	int sack_rxmit;
575 	int sack_bytes_rxmt;
576 	struct sackhole *p;
577 #ifdef TCP_SIGNATURE
578 	int sigoff = 0;
579 #endif
580 	uint64_t *tcps;
581 
582 #ifdef DIAGNOSTIC
583 	if (tp->t_inpcb && tp->t_in6pcb)
584 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
585 #endif
586 	so = NULL;
587 	ro = NULL;
588 	if (tp->t_inpcb) {
589 		so = tp->t_inpcb->inp_socket;
590 		ro = &tp->t_inpcb->inp_route;
591 	}
592 #ifdef INET6
593 	else if (tp->t_in6pcb) {
594 		so = tp->t_in6pcb->in6p_socket;
595 		ro = &tp->t_in6pcb->in6p_route;
596 	}
597 #endif
598 
599 	switch (af = tp->t_family) {
600 #ifdef INET
601 	case AF_INET:
602 		if (tp->t_inpcb)
603 			break;
604 #ifdef INET6
605 		/* mapped addr case */
606 		if (tp->t_in6pcb)
607 			break;
608 #endif
609 		return (EINVAL);
610 #endif
611 #ifdef INET6
612 	case AF_INET6:
613 		if (tp->t_in6pcb)
614 			break;
615 		return (EINVAL);
616 #endif
617 	default:
618 		return (EAFNOSUPPORT);
619 	}
620 
621 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
622 		return (EMSGSIZE);
623 
624 	idle = (tp->snd_max == tp->snd_una);
625 
626 	/*
627 	 * Determine if we can use TCP segmentation offload:
628 	 * - If we're using IPv4
629 	 * - If there is not an IPsec policy that prevents it
630 	 * - If the interface can do it
631 	 */
632 	has_tso4 = has_tso6 = false;
633 #if defined(INET)
634 	has_tso4 = tp->t_inpcb != NULL &&
635 #if defined(IPSEC) || defined(FAST_IPSEC)
636 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
637 		  		       IPSEC_DIR_OUTBOUND) &&
638 #endif
639 		  (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
640 		  (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
641 #endif /* defined(INET) */
642 #if defined(INET6)
643 	has_tso6 = tp->t_in6pcb != NULL &&
644 #if defined(IPSEC) || defined(FAST_IPSEC)
645 		  IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
646 		  		       IPSEC_DIR_OUTBOUND) &&
647 #endif
648 		  (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
649 		  (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
650 #endif /* defined(INET6) */
651 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
652 
653 	/*
654 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
655 	 *
656 	 *	Optionally, a TCP MAY set the restart window to the
657 	 *	minimum of the value used for the initial window and
658 	 *	the current value of cwnd (in other words, using a
659 	 *	larger value for the restart window should never increase
660 	 *	the size of cwnd).
661 	 */
662 	if (tcp_cwm) {
663 		/*
664 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
665 		 * Count the number of packets currently pending
666 		 * acknowledgement, and limit our congestion window
667 		 * to a pre-determined allowed burst size plus that count.
668 		 * This prevents bursting once all pending packets have
669 		 * been acknowledged (i.e. transmission is idle).
670 		 *
671 		 * XXX Link this to Initial Window?
672 		 */
673 		tp->snd_cwnd = min(tp->snd_cwnd,
674 		    (tcp_cwm_burstsize * txsegsize) +
675 		    (tp->snd_nxt - tp->snd_una));
676 	} else {
677 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
678 			/*
679 			 * We have been idle for "a while" and no acks are
680 			 * expected to clock out any data we send --
681 			 * slow start to get ack "clock" running again.
682 			 */
683 			int ss = tcp_init_win;
684 #ifdef INET
685 			if (tp->t_inpcb &&
686 			    in_localaddr(tp->t_inpcb->inp_faddr))
687 				ss = tcp_init_win_local;
688 #endif
689 #ifdef INET6
690 			if (tp->t_in6pcb &&
691 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
692 				ss = tcp_init_win_local;
693 #endif
694 			tp->snd_cwnd = min(tp->snd_cwnd,
695 			    TCP_INITIAL_WINDOW(ss, txsegsize));
696 		}
697 	}
698 
699 	txsegsize_nosack = txsegsize;
700 again:
701 	use_tso = has_tso;
702 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
703 		/* don't duplicate CWR/ECE. */
704 		use_tso = 0;
705 	}
706 	TCP_REASS_LOCK(tp);
707 	sack_numblks = tcp_sack_numblks(tp);
708 	if (sack_numblks) {
709 		int sackoptlen;
710 
711 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
712 		if (sackoptlen > txsegsize_nosack) {
713 			sack_numblks = 0; /* give up SACK */
714 			txsegsize = txsegsize_nosack;
715 		} else {
716 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
717 				/* don't duplicate D-SACK. */
718 				use_tso = 0;
719 			}
720 			txsegsize = txsegsize_nosack - sackoptlen;
721 		}
722 	} else {
723 		txsegsize = txsegsize_nosack;
724 	}
725 
726 	/*
727 	 * Determine length of data that should be transmitted, and
728 	 * flags that should be used.  If there is some data or critical
729 	 * controls (SYN, RST) to send, then transmit; otherwise,
730 	 * investigate further.
731 	 *
732 	 * Readjust SACK information to avoid resending duplicate data.
733 	 */
734 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
735 		tcp_sack_adjust(tp);
736 	sendalot = 0;
737 	off = tp->snd_nxt - tp->snd_una;
738 	win = min(tp->snd_wnd, tp->snd_cwnd);
739 
740 	flags = tcp_outflags[tp->t_state];
741 
742 	/*
743 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
744 	 * to send out new data (when sendalot is 1), bypass this function.
745 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
746 	 * we're replacing a (future) new transmission with a retransmission
747 	 * now, and we previously incremented snd_cwnd in tcp_input().
748 	 */
749 	/*
750 	 * Still in sack recovery , reset rxmit flag to zero.
751 	 */
752 	sack_rxmit = 0;
753 	sack_bytes_rxmt = 0;
754 	len = 0;
755 	p = NULL;
756 	do {
757 		long cwin;
758 		if (!TCP_SACK_ENABLED(tp))
759 			break;
760 		if (tp->t_partialacks < 0)
761 			break;
762 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
763 		if (p == NULL)
764 			break;
765 
766 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
767 		if (cwin < 0)
768 			cwin = 0;
769 		/* Do not retransmit SACK segments beyond snd_recover */
770 		if (SEQ_GT(p->end, tp->snd_recover)) {
771 			/*
772 			 * (At least) part of sack hole extends beyond
773 			 * snd_recover. Check to see if we can rexmit data
774 			 * for this hole.
775 			 */
776 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
777 				/*
778 				 * Can't rexmit any more data for this hole.
779 				 * That data will be rexmitted in the next
780 				 * sack recovery episode, when snd_recover
781 				 * moves past p->rxmit.
782 				 */
783 				p = NULL;
784 				break;
785 			}
786 			/* Can rexmit part of the current hole */
787 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
788 		} else
789 			len = ((long)ulmin(cwin, p->end - p->rxmit));
790 		off = p->rxmit - tp->snd_una;
791 		if (off + len > so->so_snd.sb_cc) {
792 			/* 1 for TH_FIN */
793 			KASSERT(off + len == so->so_snd.sb_cc + 1);
794 			KASSERT(p->rxmit + len == tp->snd_max);
795 			len = so->so_snd.sb_cc - off;
796 		}
797 		if (len > 0) {
798 			sack_rxmit = 1;
799 			sendalot = 1;
800 		}
801 	} while (/*CONSTCOND*/0);
802 
803 	/*
804 	 * If in persist timeout with window of 0, send 1 byte.
805 	 * Otherwise, if window is small but nonzero
806 	 * and timer expired, we will send what we can
807 	 * and go to transmit state.
808 	 */
809 	if (tp->t_force) {
810 		if (win == 0) {
811 			/*
812 			 * If we still have some data to send, then
813 			 * clear the FIN bit.  Usually this would
814 			 * happen below when it realizes that we
815 			 * aren't sending all the data.  However,
816 			 * if we have exactly 1 byte of unset data,
817 			 * then it won't clear the FIN bit below,
818 			 * and if we are in persist state, we wind
819 			 * up sending the packet without recording
820 			 * that we sent the FIN bit.
821 			 *
822 			 * We can't just blindly clear the FIN bit,
823 			 * because if we don't have any more data
824 			 * to send then the probe will be the FIN
825 			 * itself.
826 			 */
827 			if (off < so->so_snd.sb_cc)
828 				flags &= ~TH_FIN;
829 			win = 1;
830 		} else {
831 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
832 			tp->t_rxtshift = 0;
833 		}
834 	}
835 
836 	if (sack_rxmit == 0) {
837 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
838 			long cwin;
839 
840 			/*
841 			 * We are inside of a SACK recovery episode and are
842 			 * sending new data, having retransmitted all the
843 			 * data possible in the scoreboard.
844 			 */
845 			if (tp->snd_wnd < so->so_snd.sb_cc) {
846 				len = tp->snd_wnd - off;
847 				flags &= ~TH_FIN;
848 			} else {
849 				len = so->so_snd.sb_cc - off;
850 			}
851 
852 			/*
853 			 * From FreeBSD:
854 			 *  Don't remove this (len > 0) check !
855 			 *  We explicitly check for len > 0 here (although it
856 			 *  isn't really necessary), to work around a gcc
857 			 *  optimization issue - to force gcc to compute
858 			 *  len above. Without this check, the computation
859 			 *  of len is bungled by the optimizer.
860 			 */
861 			if (len > 0) {
862 				cwin = tp->snd_cwnd -
863 				    (tp->snd_nxt - tp->sack_newdata) -
864 				    sack_bytes_rxmt;
865 				if (cwin < 0)
866 					cwin = 0;
867 				if (cwin < len) {
868 					len = cwin;
869 					flags &= ~TH_FIN;
870 				}
871 			}
872 		} else if (win < so->so_snd.sb_cc) {
873 			len = win - off;
874 			flags &= ~TH_FIN;
875 		} else {
876 			len = so->so_snd.sb_cc - off;
877 		}
878 	}
879 
880 	if (len < 0) {
881 		/*
882 		 * If FIN has been sent but not acked,
883 		 * but we haven't been called to retransmit,
884 		 * len will be -1.  Otherwise, window shrank
885 		 * after we sent into it.  If window shrank to 0,
886 		 * cancel pending retransmit, pull snd_nxt back
887 		 * to (closed) window, and set the persist timer
888 		 * if it isn't already going.  If the window didn't
889 		 * close completely, just wait for an ACK.
890 		 *
891 		 * If we have a pending FIN, either it has already been
892 		 * transmitted or it is outside the window, so drop it.
893 		 * If the FIN has been transmitted, but this is not a
894 		 * retransmission, then len must be -1.  Therefore we also
895 		 * prevent here the sending of `gratuitous FINs'.  This
896 		 * eliminates the need to check for that case below (e.g.
897 		 * to back up snd_nxt before the FIN so that the sequence
898 		 * number is correct).
899 		 */
900 		len = 0;
901 		flags &= ~TH_FIN;
902 		if (win == 0) {
903 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
904 			tp->t_rxtshift = 0;
905 			tp->snd_nxt = tp->snd_una;
906 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
907 				tcp_setpersist(tp);
908 		}
909 	}
910 
911 	/*
912 	 * Automatic sizing enables the performance of large buffers
913 	 * and most of the efficiency of small ones by only allocating
914 	 * space when it is needed.
915 	 *
916 	 * The criteria to step up the send buffer one notch are:
917 	 *  1. receive window of remote host is larger than send buffer
918 	 *     (with a fudge factor of 5/4th);
919 	 *  2. send buffer is filled to 7/8th with data (so we actually
920 	 *     have data to make use of it);
921 	 *  3. send buffer fill has not hit maximal automatic size;
922 	 *  4. our send window (slow start and cogestion controlled) is
923 	 *     larger than sent but unacknowledged data in send buffer.
924 	 *
925 	 * The remote host receive window scaling factor may limit the
926 	 * growing of the send buffer before it reaches its allowed
927 	 * maximum.
928 	 *
929 	 * It scales directly with slow start or congestion window
930 	 * and does at most one step per received ACK.  This fast
931 	 * scaling has the drawback of growing the send buffer beyond
932 	 * what is strictly necessary to make full use of a given
933 	 * delay*bandwith product.  However testing has shown this not
934 	 * to be much of an problem.  At worst we are trading wasting
935 	 * of available bandwith (the non-use of it) for wasting some
936 	 * socket buffer memory.
937 	 *
938 	 * TODO: Shrink send buffer during idle periods together
939 	 * with congestion window.  Requires another timer.
940 	 */
941 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
942 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
943 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
944 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
945 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
946 			if (!sbreserve(&so->so_snd,
947 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
948 			     tcp_autosndbuf_max), so))
949 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
950 		}
951 	}
952 
953 	if (len > txsegsize) {
954 		if (use_tso) {
955 			/*
956 			 * Truncate TSO transfers to IP_MAXPACKET, and make
957 			 * sure that we send equal size transfers down the
958 			 * stack (rather than big-small-big-small-...).
959 			 */
960 #ifdef INET6
961 #if IPV6_MAXPACKET != IP_MAXPACKET
962 #error IPV6_MAXPACKET != IP_MAXPACKET
963 #endif
964 #endif
965 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
966 			if (len <= txsegsize) {
967 				use_tso = 0;
968 			}
969 		} else
970 			len = txsegsize;
971 		flags &= ~TH_FIN;
972 		sendalot = 1;
973 	} else
974 		use_tso = 0;
975 	if (sack_rxmit) {
976 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
977 			flags &= ~TH_FIN;
978 	}
979 
980 	win = sbspace(&so->so_rcv);
981 
982 	/*
983 	 * Sender silly window avoidance.  If connection is idle
984 	 * and can send all data, a maximum segment,
985 	 * at least a maximum default-size segment do it,
986 	 * or are forced, do it; otherwise don't bother.
987 	 * If peer's buffer is tiny, then send
988 	 * when window is at least half open.
989 	 * If retransmitting (possibly after persist timer forced us
990 	 * to send into a small window), then must resend.
991 	 */
992 	if (len) {
993 		if (len >= txsegsize)
994 			goto send;
995 		if ((so->so_state & SS_MORETOCOME) == 0 &&
996 		    ((idle || tp->t_flags & TF_NODELAY) &&
997 		     len + off >= so->so_snd.sb_cc))
998 			goto send;
999 		if (tp->t_force)
1000 			goto send;
1001 		if (len >= tp->max_sndwnd / 2)
1002 			goto send;
1003 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1004 			goto send;
1005 		if (sack_rxmit)
1006 			goto send;
1007 	}
1008 
1009 	/*
1010 	 * Compare available window to amount of window known to peer
1011 	 * (as advertised window less next expected input).  If the
1012 	 * difference is at least twice the size of the largest segment
1013 	 * we expect to receive (i.e. two segments) or at least 50% of
1014 	 * the maximum possible window, then want to send a window update
1015 	 * to peer.
1016 	 */
1017 	if (win > 0) {
1018 		/*
1019 		 * "adv" is the amount we can increase the window,
1020 		 * taking into account that we are limited by
1021 		 * TCP_MAXWIN << tp->rcv_scale.
1022 		 */
1023 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1024 			(tp->rcv_adv - tp->rcv_nxt);
1025 
1026 		if (adv >= (long) (2 * rxsegsize))
1027 			goto send;
1028 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1029 			goto send;
1030 	}
1031 
1032 	/*
1033 	 * Send if we owe peer an ACK.
1034 	 */
1035 	if (tp->t_flags & TF_ACKNOW)
1036 		goto send;
1037 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1038 		goto send;
1039 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1040 		goto send;
1041 	/*
1042 	 * In SACK, it is possible for tcp_output to fail to send a segment
1043 	 * after the retransmission timer has been turned off.  Make sure
1044 	 * that the retransmission timer is set.
1045 	 */
1046 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1047 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1048 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1049 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1050 		goto just_return;
1051 	}
1052 
1053 	/*
1054 	 * TCP window updates are not reliable, rather a polling protocol
1055 	 * using ``persist'' packets is used to insure receipt of window
1056 	 * updates.  The three ``states'' for the output side are:
1057 	 *	idle			not doing retransmits or persists
1058 	 *	persisting		to move a small or zero window
1059 	 *	(re)transmitting	and thereby not persisting
1060 	 *
1061 	 * tp->t_timer[TCPT_PERSIST]
1062 	 *	is set when we are in persist state.
1063 	 * tp->t_force
1064 	 *	is set when we are called to send a persist packet.
1065 	 * tp->t_timer[TCPT_REXMT]
1066 	 *	is set when we are retransmitting
1067 	 * The output side is idle when both timers are zero.
1068 	 *
1069 	 * If send window is too small, there is data to transmit, and no
1070 	 * retransmit or persist is pending, then go to persist state.
1071 	 * If nothing happens soon, send when timer expires:
1072 	 * if window is nonzero, transmit what we can,
1073 	 * otherwise force out a byte.
1074 	 */
1075 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1076 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1077 		tp->t_rxtshift = 0;
1078 		tcp_setpersist(tp);
1079 	}
1080 
1081 	/*
1082 	 * No reason to send a segment, just return.
1083 	 */
1084 just_return:
1085 	TCP_REASS_UNLOCK(tp);
1086 	return (0);
1087 
1088 send:
1089 	/*
1090 	 * Before ESTABLISHED, force sending of initial options
1091 	 * unless TCP set not to do any options.
1092 	 * NOTE: we assume that the IP/TCP header plus TCP options
1093 	 * always fit in a single mbuf, leaving room for a maximum
1094 	 * link header, i.e.
1095 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1096 	 */
1097 	optlen = 0;
1098 	switch (af) {
1099 #ifdef INET
1100 	case AF_INET:
1101 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1102 		break;
1103 #endif
1104 #ifdef INET6
1105 	case AF_INET6:
1106 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1107 		break;
1108 #endif
1109 	default:	/*pacify gcc*/
1110 		iphdrlen = 0;
1111 		break;
1112 	}
1113 	hdrlen = iphdrlen;
1114 	if (flags & TH_SYN) {
1115 		struct rtentry *synrt;
1116 
1117 		synrt = NULL;
1118 #ifdef INET
1119 		if (tp->t_inpcb)
1120 			synrt = in_pcbrtentry(tp->t_inpcb);
1121 #endif
1122 #ifdef INET6
1123 		if (tp->t_in6pcb)
1124 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1125 #endif
1126 
1127 		tp->snd_nxt = tp->iss;
1128 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1129 						    synrt->rt_ifp : NULL, af);
1130 		if ((tp->t_flags & TF_NOOPT) == 0) {
1131 			opt[0] = TCPOPT_MAXSEG;
1132 			opt[1] = 4;
1133 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1134 			opt[3] = tp->t_ourmss & 0xff;
1135 			optlen = 4;
1136 
1137 			if ((tp->t_flags & TF_REQ_SCALE) &&
1138 			    ((flags & TH_ACK) == 0 ||
1139 			    (tp->t_flags & TF_RCVD_SCALE))) {
1140 				*((u_int32_t *) (opt + optlen)) = htonl(
1141 					TCPOPT_NOP << 24 |
1142 					TCPOPT_WINDOW << 16 |
1143 					TCPOLEN_WINDOW << 8 |
1144 					tp->request_r_scale);
1145 				optlen += 4;
1146 			}
1147 			if (tcp_do_sack) {
1148 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1149 
1150 				cp[0] = TCPOPT_SACK_PERMITTED;
1151 				cp[1] = 2;
1152 				cp[2] = TCPOPT_NOP;
1153 				cp[3] = TCPOPT_NOP;
1154 				optlen += 4;
1155 			}
1156 		}
1157 	}
1158 
1159 	/*
1160 	 * Send a timestamp and echo-reply if this is a SYN and our side
1161 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1162 	 * and our peer have sent timestamps in our SYN's.
1163 	 */
1164 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1165 	     (flags & TH_RST) == 0 &&
1166 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1167 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1168 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1169 
1170 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1171 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1172 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1173 		*lp   = htonl(tp->ts_recent);
1174 		optlen += TCPOLEN_TSTAMP_APPA;
1175 
1176 		/* Set receive buffer autosizing timestamp. */
1177 		if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1178 			tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1179 	}
1180 
1181 	/*
1182 	 * Tack on the SACK block if it is necessary.
1183 	 */
1184 	if (sack_numblks) {
1185 		int sack_len;
1186 		u_char *bp = (u_char *)(opt + optlen);
1187 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1188 		struct ipqent *tiqe;
1189 
1190 		sack_len = sack_numblks * 8 + 2;
1191 		bp[0] = TCPOPT_NOP;
1192 		bp[1] = TCPOPT_NOP;
1193 		bp[2] = TCPOPT_SACK;
1194 		bp[3] = sack_len;
1195 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1196 			sack_numblks--;
1197 			*lp++ = htonl(tp->rcv_dsack_block.left);
1198 			*lp++ = htonl(tp->rcv_dsack_block.right);
1199 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1200 		}
1201 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1202 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1203 			KASSERT(tiqe != NULL);
1204 			sack_numblks--;
1205 			*lp++ = htonl(tiqe->ipqe_seq);
1206 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1207 			    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1208 		}
1209 		optlen += sack_len + 2;
1210 	}
1211 	TCP_REASS_UNLOCK(tp);
1212 
1213 #ifdef TCP_SIGNATURE
1214 	if (tp->t_flags & TF_SIGNATURE) {
1215 		u_char *bp;
1216 		/*
1217 		 * Initialize TCP-MD5 option (RFC2385)
1218 		 */
1219 		bp = (u_char *)opt + optlen;
1220 		*bp++ = TCPOPT_SIGNATURE;
1221 		*bp++ = TCPOLEN_SIGNATURE;
1222 		sigoff = optlen + 2;
1223 		memset(bp, 0, TCP_SIGLEN);
1224 		bp += TCP_SIGLEN;
1225 		optlen += TCPOLEN_SIGNATURE;
1226 		/*
1227 		 * Terminate options list and maintain 32-bit alignment.
1228  		 */
1229 		*bp++ = TCPOPT_NOP;
1230 		*bp++ = TCPOPT_EOL;
1231  		optlen += 2;
1232  	}
1233 #endif /* TCP_SIGNATURE */
1234 
1235 	hdrlen += optlen;
1236 
1237 #ifdef DIAGNOSTIC
1238 	if (!use_tso && len > txsegsize)
1239 		panic("tcp data to be sent is larger than segment");
1240 	else if (use_tso && len > IP_MAXPACKET)
1241 		panic("tcp data to be sent is larger than max TSO size");
1242 	if (max_linkhdr + hdrlen > MCLBYTES)
1243 		panic("tcphdr too big");
1244 #endif
1245 
1246 	/*
1247 	 * Grab a header mbuf, attaching a copy of data to
1248 	 * be transmitted, and initialize the header from
1249 	 * the template for sends on this connection.
1250 	 */
1251 	if (len) {
1252 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1253 		if (error)
1254 			goto out;
1255 		/*
1256 		 * If we're sending everything we've got, set PUSH.
1257 		 * (This will keep happy those implementations which only
1258 		 * give data to the user when a buffer fills or
1259 		 * a PUSH comes in.)
1260 		 */
1261 		if (off + len == so->so_snd.sb_cc)
1262 			flags |= TH_PUSH;
1263 	} else {
1264 		tcps = TCP_STAT_GETREF();
1265 		if (tp->t_flags & TF_ACKNOW)
1266 			tcps[TCP_STAT_SNDACKS]++;
1267 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1268 			tcps[TCP_STAT_SNDCTRL]++;
1269 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1270 			tcps[TCP_STAT_SNDURG]++;
1271 		else
1272 			tcps[TCP_STAT_SNDWINUP]++;
1273 		TCP_STAT_PUTREF();
1274 
1275 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1276 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1277 			MCLGET(m, M_DONTWAIT);
1278 			if ((m->m_flags & M_EXT) == 0) {
1279 				m_freem(m);
1280 				m = NULL;
1281 			}
1282 		}
1283 		if (m == NULL) {
1284 			error = ENOBUFS;
1285 			goto out;
1286 		}
1287 		MCLAIM(m, &tcp_tx_mowner);
1288 		m->m_data += max_linkhdr;
1289 		m->m_len = hdrlen;
1290 	}
1291 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1292 	switch (af) {
1293 #ifdef INET
1294 	case AF_INET:
1295 		ip = mtod(m, struct ip *);
1296 #ifdef INET6
1297 		ip6 = NULL;
1298 #endif
1299 		th = (struct tcphdr *)(ip + 1);
1300 		break;
1301 #endif
1302 #ifdef INET6
1303 	case AF_INET6:
1304 		ip = NULL;
1305 		ip6 = mtod(m, struct ip6_hdr *);
1306 		th = (struct tcphdr *)(ip6 + 1);
1307 		break;
1308 #endif
1309 	default:	/*pacify gcc*/
1310 		ip = NULL;
1311 #ifdef INET6
1312 		ip6 = NULL;
1313 #endif
1314 		th = NULL;
1315 		break;
1316 	}
1317 	if (tp->t_template == 0)
1318 		panic("tcp_output");
1319 	if (tp->t_template->m_len < iphdrlen)
1320 		panic("tcp_output");
1321 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1322 
1323 	/*
1324 	 * If we are starting a connection, send ECN setup
1325 	 * SYN packet. If we are on a retransmit, we may
1326 	 * resend those bits a number of times as per
1327 	 * RFC 3168.
1328 	 */
1329 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1330 		if (tp->t_flags & TF_SYN_REXMT) {
1331 			if (tp->t_ecn_retries--)
1332 				flags |= TH_ECE|TH_CWR;
1333 		} else {
1334 			flags |= TH_ECE|TH_CWR;
1335 			tp->t_ecn_retries = tcp_ecn_maxretries;
1336 		}
1337 	}
1338 
1339 	if (TCP_ECN_ALLOWED(tp)) {
1340 		/*
1341 		 * If the peer has ECN, mark data packets
1342 		 * ECN capable. Ignore pure ack packets, retransmissions
1343 		 * and window probes.
1344 		 */
1345 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1346 		    !(tp->t_force && len == 1)) {
1347 			switch (af) {
1348 #ifdef INET
1349 			case AF_INET:
1350 				tp->t_inpcb->inp_ip.ip_tos |= IPTOS_ECN_ECT0;
1351 				break;
1352 #endif
1353 #ifdef INET6
1354 			case AF_INET6:
1355 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1356 				break;
1357 #endif
1358 			}
1359 			TCP_STATINC(TCP_STAT_ECN_ECT);
1360 		}
1361 
1362 		/*
1363 		 * Reply with proper ECN notifications.
1364 		 */
1365 		if (tp->t_flags & TF_ECN_SND_CWR) {
1366 			flags |= TH_CWR;
1367 			tp->t_flags &= ~TF_ECN_SND_CWR;
1368 		}
1369 		if (tp->t_flags & TF_ECN_SND_ECE) {
1370 			flags |= TH_ECE;
1371 		}
1372 	}
1373 
1374 
1375 	/*
1376 	 * If we are doing retransmissions, then snd_nxt will
1377 	 * not reflect the first unsent octet.  For ACK only
1378 	 * packets, we do not want the sequence number of the
1379 	 * retransmitted packet, we want the sequence number
1380 	 * of the next unsent octet.  So, if there is no data
1381 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1382 	 * when filling in ti_seq.  But if we are in persist
1383 	 * state, snd_max might reflect one byte beyond the
1384 	 * right edge of the window, so use snd_nxt in that
1385 	 * case, since we know we aren't doing a retransmission.
1386 	 * (retransmit and persist are mutually exclusive...)
1387 	 */
1388 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1389 		th->th_seq = htonl(p->rxmit);
1390 		p->rxmit += len;
1391 	} else {
1392 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1393 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1394 			th->th_seq = htonl(tp->snd_nxt);
1395 		else
1396 			th->th_seq = htonl(tp->snd_max);
1397 	}
1398 	th->th_ack = htonl(tp->rcv_nxt);
1399 	if (optlen) {
1400 		bcopy((void *)opt, (void *)(th + 1), optlen);
1401 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1402 	}
1403 	th->th_flags = flags;
1404 	/*
1405 	 * Calculate receive window.  Don't shrink window,
1406 	 * but avoid silly window syndrome.
1407 	 */
1408 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1409 		win = 0;
1410 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1411 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1412 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1413 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1414 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1415 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1416 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1417 		if (urp > IP_MAXPACKET)
1418 			urp = IP_MAXPACKET;
1419 		th->th_urp = htons((u_int16_t)urp);
1420 		th->th_flags |= TH_URG;
1421 	} else
1422 		/*
1423 		 * If no urgent pointer to send, then we pull
1424 		 * the urgent pointer to the left edge of the send window
1425 		 * so that it doesn't drift into the send window on sequence
1426 		 * number wraparound.
1427 		 */
1428 		tp->snd_up = tp->snd_una;		/* drag it along */
1429 
1430 #ifdef TCP_SIGNATURE
1431 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1432 		struct secasvar *sav;
1433 		u_int8_t *sigp;
1434 
1435 		sav = tcp_signature_getsav(m, th);
1436 
1437 		if (sav == NULL) {
1438 			if (m)
1439 				m_freem(m);
1440 			return (EPERM);
1441 		}
1442 
1443 		m->m_pkthdr.len = hdrlen + len;
1444 		sigp = (char *)th + sizeof(*th) + sigoff;
1445 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1446 
1447 		key_sa_recordxfer(sav, m);
1448 #ifdef FAST_IPSEC
1449 		KEY_FREESAV(&sav);
1450 #else
1451 		key_freesav(sav);
1452 #endif
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 #ifdef INET
1462 	case AF_INET:
1463 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1464 		if (use_tso) {
1465 			m->m_pkthdr.segsz = txsegsize;
1466 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1467 		} else {
1468 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1469 			if (len + optlen) {
1470 				/* Fixup the pseudo-header checksum. */
1471 				/* XXXJRT Not IP Jumbogram safe. */
1472 				th->th_sum = in_cksum_addword(th->th_sum,
1473 				    htons((u_int16_t) (len + optlen)));
1474 			}
1475 		}
1476 		break;
1477 #endif
1478 #ifdef INET6
1479 	case AF_INET6:
1480 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1481 		if (use_tso) {
1482 			m->m_pkthdr.segsz = txsegsize;
1483 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1484 		} else {
1485 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1486 			if (len + optlen) {
1487 				/* Fixup the pseudo-header checksum. */
1488 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1489 				th->th_sum = in_cksum_addword(th->th_sum,
1490 				    htons((u_int16_t) (len + optlen)));
1491 			}
1492 		}
1493 		break;
1494 #endif
1495 	}
1496 
1497 	/*
1498 	 * In transmit state, time the transmission and arrange for
1499 	 * the retransmit.  In persist state, just set snd_max.
1500 	 */
1501 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1502 		tcp_seq startseq = tp->snd_nxt;
1503 
1504 		/*
1505 		 * Advance snd_nxt over sequence space of this segment.
1506 		 * There are no states in which we send both a SYN and a FIN,
1507 		 * so we collapse the tests for these flags.
1508 		 */
1509 		if (flags & (TH_SYN|TH_FIN))
1510 			tp->snd_nxt++;
1511 		if (sack_rxmit)
1512 			goto timer;
1513 		tp->snd_nxt += len;
1514 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1515 			tp->snd_max = tp->snd_nxt;
1516 			/*
1517 			 * Time this transmission if not a retransmission and
1518 			 * not currently timing anything.
1519 			 */
1520 			if (tp->t_rtttime == 0) {
1521 				tp->t_rtttime = tcp_now;
1522 				tp->t_rtseq = startseq;
1523 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1524 			}
1525 		}
1526 
1527 		/*
1528 		 * Set retransmit timer if not currently set,
1529 		 * and not doing an ack or a keep-alive probe.
1530 		 * Initial value for retransmit timer is smoothed
1531 		 * round-trip time + 2 * round-trip time variance.
1532 		 * Initialize shift counter which is used for backoff
1533 		 * of retransmit time.
1534 		 */
1535 timer:
1536 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1537 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1538 		    tp->snd_nxt != tp->snd_una)) {
1539 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1540 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1541 				tp->t_rxtshift = 0;
1542 			}
1543 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1544 		}
1545 	} else
1546 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1547 			tp->snd_max = tp->snd_nxt + len;
1548 
1549 #ifdef TCP_DEBUG
1550 	/*
1551 	 * Trace.
1552 	 */
1553 	if (so->so_options & SO_DEBUG)
1554 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1555 #endif
1556 
1557 	/*
1558 	 * Fill in IP length and desired time to live and
1559 	 * send to IP level.  There should be a better way
1560 	 * to handle ttl and tos; we could keep them in
1561 	 * the template, but need a way to checksum without them.
1562 	 */
1563 	m->m_pkthdr.len = hdrlen + len;
1564 
1565 	switch (af) {
1566 #ifdef INET
1567 	case AF_INET:
1568 		ip->ip_len = htons(m->m_pkthdr.len);
1569 		packetlen = m->m_pkthdr.len;
1570 		if (tp->t_inpcb) {
1571 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1572 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1573 		}
1574 #ifdef INET6
1575 		else if (tp->t_in6pcb) {
1576 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1577 			ip->ip_tos = 0;	/*XXX*/
1578 		}
1579 #endif
1580 		break;
1581 #endif
1582 #ifdef INET6
1583 	case AF_INET6:
1584 		packetlen = m->m_pkthdr.len;
1585 		ip6->ip6_nxt = IPPROTO_TCP;
1586 		if (tp->t_in6pcb) {
1587 			/*
1588 			 * we separately set hoplimit for every segment, since
1589 			 * the user might want to change the value via
1590 			 * setsockopt. Also, desired default hop limit might
1591 			 * be changed via Neighbor Discovery.
1592 			 */
1593 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1594 				(rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp
1595 				                                    : NULL);
1596 		}
1597 		/* ip6->ip6_flow = ??? */
1598 		/* ip6_plen will be filled in ip6_output(). */
1599 		break;
1600 #endif
1601 	default:	/*pacify gcc*/
1602 		packetlen = 0;
1603 		break;
1604 	}
1605 
1606 	switch (af) {
1607 #ifdef INET
1608 	case AF_INET:
1609 	    {
1610 		struct mbuf *opts;
1611 
1612 		if (tp->t_inpcb)
1613 			opts = tp->t_inpcb->inp_options;
1614 		else
1615 			opts = NULL;
1616 		error = ip_output(m, opts, ro,
1617 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1618 			(so->so_options & SO_DONTROUTE),
1619 			(struct ip_moptions *)0, so);
1620 		break;
1621 	    }
1622 #endif
1623 #ifdef INET6
1624 	case AF_INET6:
1625 	    {
1626 		struct ip6_pktopts *opts;
1627 
1628 		if (tp->t_in6pcb)
1629 			opts = tp->t_in6pcb->in6p_outputopts;
1630 		else
1631 			opts = NULL;
1632 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1633 			NULL, so, NULL);
1634 		break;
1635 	    }
1636 #endif
1637 	default:
1638 		error = EAFNOSUPPORT;
1639 		break;
1640 	}
1641 	if (error) {
1642 out:
1643 		if (error == ENOBUFS) {
1644 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1645 #ifdef INET
1646 			if (tp->t_inpcb)
1647 				tcp_quench(tp->t_inpcb, 0);
1648 #endif
1649 #ifdef INET6
1650 			if (tp->t_in6pcb)
1651 				tcp6_quench(tp->t_in6pcb, 0);
1652 #endif
1653 			error = 0;
1654 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1655 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1656 			tp->t_softerror = error;
1657 			error = 0;
1658 		}
1659 
1660 		/* Back out the seqence number advance. */
1661 		if (sack_rxmit)
1662 			p->rxmit -= len;
1663 
1664 		/* Restart the delayed ACK timer, if necessary. */
1665 		if (tp->t_flags & TF_DELACK)
1666 			TCP_RESTART_DELACK(tp);
1667 
1668 		return (error);
1669 	}
1670 
1671 	if (packetlen > tp->t_pmtud_mtu_sent)
1672 		tp->t_pmtud_mtu_sent = packetlen;
1673 
1674 	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