xref: /netbsd-src/sys/netinet/tcp_output.c (revision aad9773e38ed2370a628a6416e098f9008fc10a7)
1 /*	$NetBSD: tcp_output.c,v 1.179 2014/11/10 18:52:51 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 1.179 2014/11/10 18:52:51 maxv 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/mbuf.h>
147 #include <sys/protosw.h>
148 #include <sys/socket.h>
149 #include <sys/socketvar.h>
150 #include <sys/errno.h>
151 #include <sys/domain.h>
152 #include <sys/kernel.h>
153 #ifdef TCP_SIGNATURE
154 #include <sys/md5.h>
155 #endif
156 
157 #include <net/if.h>
158 #include <net/route.h>
159 
160 #include <netinet/in.h>
161 #include <netinet/in_systm.h>
162 #include <netinet/ip.h>
163 #include <netinet/in_pcb.h>
164 #include <netinet/ip_var.h>
165 
166 #ifdef INET6
167 #ifndef INET
168 #include <netinet/in.h>
169 #endif
170 #include <netinet/ip6.h>
171 #include <netinet6/in6_var.h>
172 #include <netinet6/ip6_var.h>
173 #include <netinet6/in6_pcb.h>
174 #include <netinet6/nd6.h>
175 #endif
176 
177 #ifdef IPSEC
178 #include <netipsec/ipsec.h>
179 #include <netipsec/key.h>
180 #ifdef INET6
181 #include <netipsec/ipsec6.h>
182 #endif
183 #endif	/* IPSEC*/
184 
185 #include <netinet/tcp.h>
186 #define	TCPOUTFLAGS
187 #include <netinet/tcp_fsm.h>
188 #include <netinet/tcp_seq.h>
189 #include <netinet/tcp_timer.h>
190 #include <netinet/tcp_var.h>
191 #include <netinet/tcp_private.h>
192 #include <netinet/tcp_congctl.h>
193 #include <netinet/tcpip.h>
194 #include <netinet/tcp_debug.h>
195 #include <netinet/in_offload.h>
196 #include <netinet6/in6_offload.h>
197 
198 #ifdef notyet
199 extern struct mbuf *m_copypack();
200 #endif
201 
202 /*
203  * Knob to enable Congestion Window Monitoring, and control
204  * the burst size it allows.  Default burst is 4 packets, per
205  * the Internet draft.
206  */
207 int	tcp_cwm = 0;
208 int	tcp_cwm_burstsize = 4;
209 
210 int	tcp_do_autosndbuf = 1;
211 int	tcp_autosndbuf_inc = 8 * 1024;
212 int	tcp_autosndbuf_max = 256 * 1024;
213 
214 #ifdef TCP_OUTPUT_COUNTERS
215 #include <sys/device.h>
216 
217 extern struct evcnt tcp_output_bigheader;
218 extern struct evcnt tcp_output_predict_hit;
219 extern struct evcnt tcp_output_predict_miss;
220 extern struct evcnt tcp_output_copysmall;
221 extern struct evcnt tcp_output_copybig;
222 extern struct evcnt tcp_output_refbig;
223 
224 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
225 #else
226 
227 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
228 
229 #endif /* TCP_OUTPUT_COUNTERS */
230 
231 static
232 #ifndef GPROF
233 inline
234 #endif
235 int
236 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
237     bool *alwaysfragp)
238 {
239 #ifdef INET
240 	struct inpcb *inp = tp->t_inpcb;
241 #endif
242 #ifdef INET6
243 	struct in6pcb *in6p = tp->t_in6pcb;
244 #endif
245 	struct socket *so = NULL;
246 	struct rtentry *rt;
247 	struct ifnet *ifp;
248 	int size;
249 	int hdrlen;
250 	int optlen;
251 
252 	*alwaysfragp = false;
253 
254 #ifdef DIAGNOSTIC
255 	if (tp->t_inpcb && tp->t_in6pcb)
256 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
257 #endif
258 	switch (tp->t_family) {
259 #ifdef INET
260 	case AF_INET:
261 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
262 		break;
263 #endif
264 #ifdef INET6
265 	case AF_INET6:
266 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
267 		break;
268 #endif
269 	default:
270 		size = tcp_mssdflt;
271 		goto out;
272 	}
273 
274 	rt = NULL;
275 #ifdef INET
276 	if (inp) {
277 		rt = in_pcbrtentry(inp);
278 		so = inp->inp_socket;
279 	}
280 #endif
281 #ifdef INET6
282 	if (in6p) {
283 		rt = in6_pcbrtentry(in6p);
284 		so = in6p->in6p_socket;
285 	}
286 #endif
287 	if (rt == NULL) {
288 		size = tcp_mssdflt;
289 		goto out;
290 	}
291 
292 	ifp = rt->rt_ifp;
293 
294 	size = tcp_mssdflt;
295 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
296 #ifdef INET6
297 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
298 			/*
299 			 * RFC2460 section 5, last paragraph: if path MTU is
300 			 * smaller than 1280, use 1280 as packet size and
301 			 * attach fragment header.
302 			 */
303 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
304 			*alwaysfragp = true;
305 		} else
306 			size = rt->rt_rmx.rmx_mtu - hdrlen;
307 #else
308 		size = rt->rt_rmx.rmx_mtu - hdrlen;
309 #endif
310 	} else if (ifp->if_flags & IFF_LOOPBACK)
311 		size = ifp->if_mtu - hdrlen;
312 #ifdef INET
313 	else if (inp && tp->t_mtudisc)
314 		size = ifp->if_mtu - hdrlen;
315 	else if (inp && in_localaddr(inp->inp_faddr))
316 		size = ifp->if_mtu - hdrlen;
317 #endif
318 #ifdef INET6
319 	else if (in6p) {
320 #ifdef INET
321 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
322 			/* mapped addr case */
323 			struct in_addr d;
324 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
325 			if (tp->t_mtudisc || in_localaddr(d))
326 				size = ifp->if_mtu - hdrlen;
327 		} else
328 #endif
329 		{
330 			/*
331 			 * for IPv6, path MTU discovery is always turned on,
332 			 * or the node must use packet size <= 1280.
333 			 */
334 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
335 			size -= hdrlen;
336 		}
337 	}
338 #endif
339  out:
340 	/*
341 	 * Now we must make room for whatever extra TCP/IP options are in
342 	 * the packet.
343 	 */
344 	optlen = tcp_optlen(tp);
345 
346 	/*
347 	 * XXX tp->t_ourmss should have the right size, but without this code
348 	 * fragmentation will occur... need more investigation
349 	 */
350 #ifdef INET
351 	if (inp) {
352 #if defined(IPSEC)
353 		if (ipsec_used &&
354 		    !IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
355 			optlen += ipsec4_hdrsiz_tcp(tp);
356 #endif
357 		optlen += ip_optlen(inp);
358 	}
359 #endif
360 #ifdef INET6
361 #ifdef INET
362 	if (in6p && tp->t_family == AF_INET) {
363 #if defined(IPSEC)
364 		if (ipsec_used &&
365 		    !IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
366 			optlen += ipsec4_hdrsiz_tcp(tp);
367 #endif
368 		/* XXX size -= ip_optlen(in6p); */
369 	} else
370 #endif
371 	if (in6p && tp->t_family == AF_INET6) {
372 #if defined(IPSEC)
373 		if (ipsec_used &&
374 		    !IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
375 			optlen += ipsec6_hdrsiz_tcp(tp);
376 #endif
377 		optlen += ip6_optlen(in6p);
378 	}
379 #endif
380 	size -= optlen;
381 
382 	/* there may not be any room for data if mtu is too small */
383 	if (size < 0)
384 		return (EMSGSIZE);
385 
386 	/*
387 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
388 	 * assumes that path MTU is the same for both ways).  this is only
389 	 * for silly window avoidance, do not use the value for other purposes.
390 	 *
391 	 * ipseclen is subtracted from both sides, this may not be right.
392 	 * I'm not quite sure about this (could someone comment).
393 	 */
394 	*txsegsizep = min(tp->t_peermss - optlen, size);
395 	/*
396 	 * Never send more than half a buffer full.  This insures that we can
397 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
398 	 * therefore acks will never be delayed unless we run out of data to
399 	 * transmit.
400 	 */
401 	if (so)
402 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
403 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
404 
405 	if (*txsegsizep != tp->t_segsz) {
406 		/*
407 		 * If the new segment size is larger, we don't want to
408 		 * mess up the congestion window, but if it is smaller
409 		 * we'll have to reduce the congestion window to ensure
410 		 * that we don't get into trouble with initial windows
411 		 * and the rest.  In any case, if the segment size
412 		 * has changed, chances are the path has, too, and
413 		 * our congestion window will be different.
414 		 */
415 		if (*txsegsizep < tp->t_segsz) {
416 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
417 					   * *txsegsizep, *txsegsizep);
418 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
419 						* *txsegsizep, *txsegsizep);
420 		}
421 		tp->t_segsz = *txsegsizep;
422 	}
423 
424 	return (0);
425 }
426 
427 static
428 #ifndef GPROF
429 inline
430 #endif
431 int
432 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
433     long len, int hdrlen, struct mbuf **mp)
434 {
435 	struct mbuf *m, *m0;
436 	uint64_t *tcps;
437 
438 	tcps = TCP_STAT_GETREF();
439 	if (tp->t_force && len == 1)
440 		tcps[TCP_STAT_SNDPROBE]++;
441 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
442 		tcps[TCP_STAT_SNDREXMITPACK]++;
443 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
444 	} else {
445 		tcps[TCP_STAT_SNDPACK]++;
446 		tcps[TCP_STAT_SNDBYTE] += len;
447 	}
448 	TCP_STAT_PUTREF();
449 #ifdef notyet
450 	if ((m = m_copypack(so->so_snd.sb_mb, off,
451 	    (int)len, max_linkhdr + hdrlen)) == 0)
452 		return (ENOBUFS);
453 	/*
454 	 * m_copypack left space for our hdr; use it.
455 	 */
456 	m->m_len += hdrlen;
457 	m->m_data -= hdrlen;
458 #else
459 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
460 	if (__predict_false(m == NULL))
461 		return (ENOBUFS);
462 	MCLAIM(m, &tcp_tx_mowner);
463 
464 	/*
465 	 * XXX Because other code assumes headers will fit in
466 	 * XXX one header mbuf.
467 	 *
468 	 * (This code should almost *never* be run.)
469 	 */
470 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
471 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
472 		MCLGET(m, M_DONTWAIT);
473 		if ((m->m_flags & M_EXT) == 0) {
474 			m_freem(m);
475 			return (ENOBUFS);
476 		}
477 	}
478 
479 	m->m_data += max_linkhdr;
480 	m->m_len = hdrlen;
481 
482 	/*
483 	 * To avoid traversing the whole sb_mb chain for correct
484 	 * data to send, remember last sent mbuf, its offset and
485 	 * the sent size.  When called the next time, see if the
486 	 * data to send is directly following the previous transfer.
487 	 * This is important for large TCP windows.
488 	 */
489 	if (off == 0 || tp->t_lastm == NULL ||
490 	    (tp->t_lastoff + tp->t_lastlen) != off) {
491 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
492 		/*
493 		 * Either a new packet or a retransmit.
494 		 * Start from the beginning.
495 		 */
496 		tp->t_lastm = so->so_snd.sb_mb;
497 		tp->t_inoff = off;
498 	} else {
499 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
500 		tp->t_inoff += tp->t_lastlen;
501 	}
502 
503 	/* Traverse forward to next packet */
504 	while (tp->t_inoff > 0) {
505 		if (tp->t_lastm == NULL)
506 			panic("tp->t_lastm == NULL");
507 		if (tp->t_inoff < tp->t_lastm->m_len)
508 			break;
509 		tp->t_inoff -= tp->t_lastm->m_len;
510 		tp->t_lastm = tp->t_lastm->m_next;
511 	}
512 
513 	tp->t_lastoff = off;
514 	tp->t_lastlen = len;
515 	m0 = tp->t_lastm;
516 	off = tp->t_inoff;
517 
518 	if (len <= M_TRAILINGSPACE(m)) {
519 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
520 		m->m_len += len;
521 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
522 	} else {
523 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
524 		if (m->m_next == NULL) {
525 			m_freem(m);
526 			return (ENOBUFS);
527 		}
528 #ifdef TCP_OUTPUT_COUNTERS
529 		if (m->m_next->m_flags & M_EXT)
530 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
531 		else
532 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
533 #endif /* TCP_OUTPUT_COUNTERS */
534 	}
535 #endif
536 
537 	*mp = m;
538 	return (0);
539 }
540 
541 /*
542  * Tcp output routine: figure out what should be sent and send it.
543  */
544 int
545 tcp_output(struct tcpcb *tp)
546 {
547 	struct rtentry *rt;
548 	struct socket *so;
549 	struct route *ro;
550 	long len, win;
551 	int off, flags, error;
552 	struct mbuf *m;
553 	struct ip *ip;
554 #ifdef INET6
555 	struct ip6_hdr *ip6;
556 #endif
557 	struct tcphdr *th;
558 	u_char opt[MAX_TCPOPTLEN];
559 #define OPT_FITS(more)	((optlen + (more)) < sizeof(opt))
560 	unsigned optlen, hdrlen, packetlen;
561 	unsigned int sack_numblks;
562 	int idle, sendalot, txsegsize, rxsegsize;
563 	int txsegsize_nosack;
564 	int maxburst = TCP_MAXBURST;
565 	int af;		/* address family on the wire */
566 	int iphdrlen;
567 	int has_tso4, has_tso6;
568 	int has_tso, use_tso;
569 	bool alwaysfrag;
570 	int sack_rxmit;
571 	int sack_bytes_rxmt;
572 	int ecn_tos;
573 	struct sackhole *p;
574 #ifdef TCP_SIGNATURE
575 	int sigoff = 0;
576 #endif
577 	uint64_t *tcps;
578 
579 #ifdef DIAGNOSTIC
580 	if (tp->t_inpcb && tp->t_in6pcb)
581 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
582 #endif
583 	so = NULL;
584 	ro = NULL;
585 	if (tp->t_inpcb) {
586 		so = tp->t_inpcb->inp_socket;
587 		ro = &tp->t_inpcb->inp_route;
588 	}
589 #ifdef INET6
590 	else if (tp->t_in6pcb) {
591 		so = tp->t_in6pcb->in6p_socket;
592 		ro = &tp->t_in6pcb->in6p_route;
593 	}
594 #endif
595 
596 	switch (af = tp->t_family) {
597 #ifdef INET
598 	case AF_INET:
599 		if (tp->t_inpcb)
600 			break;
601 #ifdef INET6
602 		/* mapped addr case */
603 		if (tp->t_in6pcb)
604 			break;
605 #endif
606 		return (EINVAL);
607 #endif
608 #ifdef INET6
609 	case AF_INET6:
610 		if (tp->t_in6pcb)
611 			break;
612 		return (EINVAL);
613 #endif
614 	default:
615 		return (EAFNOSUPPORT);
616 	}
617 
618 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
619 		return (EMSGSIZE);
620 
621 	idle = (tp->snd_max == tp->snd_una);
622 
623 	/*
624 	 * Determine if we can use TCP segmentation offload:
625 	 * - If we're using IPv4
626 	 * - If there is not an IPsec policy that prevents it
627 	 * - If the interface can do it
628 	 */
629 	has_tso4 = has_tso6 = false;
630 #if defined(INET)
631 	has_tso4 = tp->t_inpcb != NULL &&
632 #if defined(IPSEC)
633 	    (!ipsec_used || IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
634 	    IPSEC_DIR_OUTBOUND)) &&
635 #endif
636 	    (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
637 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
638 #endif /* defined(INET) */
639 #if defined(INET6)
640 	has_tso6 = tp->t_in6pcb != NULL &&
641 #if defined(IPSEC)
642 	    (!ipsec_used || IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
643 	    IPSEC_DIR_OUTBOUND)) &&
644 #endif
645 	    (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
646 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
647 #endif /* defined(INET6) */
648 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
649 
650 	/*
651 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
652 	 *
653 	 *	Optionally, a TCP MAY set the restart window to the
654 	 *	minimum of the value used for the initial window and
655 	 *	the current value of cwnd (in other words, using a
656 	 *	larger value for the restart window should never increase
657 	 *	the size of cwnd).
658 	 */
659 	if (tcp_cwm) {
660 		/*
661 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
662 		 * Count the number of packets currently pending
663 		 * acknowledgement, and limit our congestion window
664 		 * to a pre-determined allowed burst size plus that count.
665 		 * This prevents bursting once all pending packets have
666 		 * been acknowledged (i.e. transmission is idle).
667 		 *
668 		 * XXX Link this to Initial Window?
669 		 */
670 		tp->snd_cwnd = min(tp->snd_cwnd,
671 		    (tcp_cwm_burstsize * txsegsize) +
672 		    (tp->snd_nxt - tp->snd_una));
673 	} else {
674 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
675 			/*
676 			 * We have been idle for "a while" and no acks are
677 			 * expected to clock out any data we send --
678 			 * slow start to get ack "clock" running again.
679 			 */
680 			int ss = tcp_init_win;
681 #ifdef INET
682 			if (tp->t_inpcb &&
683 			    in_localaddr(tp->t_inpcb->inp_faddr))
684 				ss = tcp_init_win_local;
685 #endif
686 #ifdef INET6
687 			if (tp->t_in6pcb &&
688 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
689 				ss = tcp_init_win_local;
690 #endif
691 			tp->snd_cwnd = min(tp->snd_cwnd,
692 			    TCP_INITIAL_WINDOW(ss, txsegsize));
693 		}
694 	}
695 
696 	txsegsize_nosack = txsegsize;
697 again:
698 	ecn_tos = 0;
699 	use_tso = has_tso;
700 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
701 		/* don't duplicate CWR/ECE. */
702 		use_tso = 0;
703 	}
704 	TCP_REASS_LOCK(tp);
705 	sack_numblks = tcp_sack_numblks(tp);
706 	if (sack_numblks) {
707 		int sackoptlen;
708 
709 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
710 		if (sackoptlen > txsegsize_nosack) {
711 			sack_numblks = 0; /* give up SACK */
712 			txsegsize = txsegsize_nosack;
713 		} else {
714 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
715 				/* don't duplicate D-SACK. */
716 				use_tso = 0;
717 			}
718 			txsegsize = txsegsize_nosack - sackoptlen;
719 		}
720 	} else {
721 		txsegsize = txsegsize_nosack;
722 	}
723 
724 	/*
725 	 * Determine length of data that should be transmitted, and
726 	 * flags that should be used.  If there is some data or critical
727 	 * controls (SYN, RST) to send, then transmit; otherwise,
728 	 * investigate further.
729 	 *
730 	 * Readjust SACK information to avoid resending duplicate data.
731 	 */
732 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
733 		tcp_sack_adjust(tp);
734 	sendalot = 0;
735 	off = tp->snd_nxt - tp->snd_una;
736 	win = min(tp->snd_wnd, tp->snd_cwnd);
737 
738 	flags = tcp_outflags[tp->t_state];
739 
740 	/*
741 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
742 	 * to send out new data (when sendalot is 1), bypass this function.
743 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
744 	 * we're replacing a (future) new transmission with a retransmission
745 	 * now, and we previously incremented snd_cwnd in tcp_input().
746 	 */
747 	/*
748 	 * Still in sack recovery , reset rxmit flag to zero.
749 	 */
750 	sack_rxmit = 0;
751 	sack_bytes_rxmt = 0;
752 	len = 0;
753 	p = NULL;
754 	do {
755 		long cwin;
756 		if (!TCP_SACK_ENABLED(tp))
757 			break;
758 		if (tp->t_partialacks < 0)
759 			break;
760 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
761 		if (p == NULL)
762 			break;
763 
764 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
765 		if (cwin < 0)
766 			cwin = 0;
767 		/* Do not retransmit SACK segments beyond snd_recover */
768 		if (SEQ_GT(p->end, tp->snd_recover)) {
769 			/*
770 			 * (At least) part of sack hole extends beyond
771 			 * snd_recover. Check to see if we can rexmit data
772 			 * for this hole.
773 			 */
774 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
775 				/*
776 				 * Can't rexmit any more data for this hole.
777 				 * That data will be rexmitted in the next
778 				 * sack recovery episode, when snd_recover
779 				 * moves past p->rxmit.
780 				 */
781 				p = NULL;
782 				break;
783 			}
784 			/* Can rexmit part of the current hole */
785 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
786 		} else
787 			len = ((long)ulmin(cwin, p->end - p->rxmit));
788 		off = p->rxmit - tp->snd_una;
789 		if (off + len > so->so_snd.sb_cc) {
790 			/* 1 for TH_FIN */
791 			KASSERT(off + len == so->so_snd.sb_cc + 1);
792 			KASSERT(p->rxmit + len == tp->snd_max);
793 			len = so->so_snd.sb_cc - off;
794 		}
795 		if (len > 0) {
796 			sack_rxmit = 1;
797 			sendalot = 1;
798 		}
799 	} while (/*CONSTCOND*/0);
800 
801 	/*
802 	 * If in persist timeout with window of 0, send 1 byte.
803 	 * Otherwise, if window is small but nonzero
804 	 * and timer expired, we will send what we can
805 	 * and go to transmit state.
806 	 */
807 	if (tp->t_force) {
808 		if (win == 0) {
809 			/*
810 			 * If we still have some data to send, then
811 			 * clear the FIN bit.  Usually this would
812 			 * happen below when it realizes that we
813 			 * aren't sending all the data.  However,
814 			 * if we have exactly 1 byte of unset data,
815 			 * then it won't clear the FIN bit below,
816 			 * and if we are in persist state, we wind
817 			 * up sending the packet without recording
818 			 * that we sent the FIN bit.
819 			 *
820 			 * We can't just blindly clear the FIN bit,
821 			 * because if we don't have any more data
822 			 * to send then the probe will be the FIN
823 			 * itself.
824 			 */
825 			if (off < so->so_snd.sb_cc)
826 				flags &= ~TH_FIN;
827 			win = 1;
828 		} else {
829 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
830 			tp->t_rxtshift = 0;
831 		}
832 	}
833 
834 	if (sack_rxmit == 0) {
835 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
836 			long cwin;
837 
838 			/*
839 			 * We are inside of a SACK recovery episode and are
840 			 * sending new data, having retransmitted all the
841 			 * data possible in the scoreboard.
842 			 */
843 			if (tp->snd_wnd < so->so_snd.sb_cc) {
844 				len = tp->snd_wnd - off;
845 				flags &= ~TH_FIN;
846 			} else {
847 				len = so->so_snd.sb_cc - off;
848 			}
849 
850 			/*
851 			 * From FreeBSD:
852 			 *  Don't remove this (len > 0) check !
853 			 *  We explicitly check for len > 0 here (although it
854 			 *  isn't really necessary), to work around a gcc
855 			 *  optimization issue - to force gcc to compute
856 			 *  len above. Without this check, the computation
857 			 *  of len is bungled by the optimizer.
858 			 */
859 			if (len > 0) {
860 				cwin = tp->snd_cwnd -
861 				    (tp->snd_nxt - tp->sack_newdata) -
862 				    sack_bytes_rxmt;
863 				if (cwin < 0)
864 					cwin = 0;
865 				if (cwin < len) {
866 					len = cwin;
867 					flags &= ~TH_FIN;
868 				}
869 			}
870 		} else if (win < so->so_snd.sb_cc) {
871 			len = win - off;
872 			flags &= ~TH_FIN;
873 		} else {
874 			len = so->so_snd.sb_cc - off;
875 		}
876 	}
877 
878 	if (len < 0) {
879 		/*
880 		 * If FIN has been sent but not acked,
881 		 * but we haven't been called to retransmit,
882 		 * len will be -1.  Otherwise, window shrank
883 		 * after we sent into it.  If window shrank to 0,
884 		 * cancel pending retransmit, pull snd_nxt back
885 		 * to (closed) window, and set the persist timer
886 		 * if it isn't already going.  If the window didn't
887 		 * close completely, just wait for an ACK.
888 		 *
889 		 * If we have a pending FIN, either it has already been
890 		 * transmitted or it is outside the window, so drop it.
891 		 * If the FIN has been transmitted, but this is not a
892 		 * retransmission, then len must be -1.  Therefore we also
893 		 * prevent here the sending of `gratuitous FINs'.  This
894 		 * eliminates the need to check for that case below (e.g.
895 		 * to back up snd_nxt before the FIN so that the sequence
896 		 * number is correct).
897 		 */
898 		len = 0;
899 		flags &= ~TH_FIN;
900 		if (win == 0) {
901 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
902 			tp->t_rxtshift = 0;
903 			tp->snd_nxt = tp->snd_una;
904 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
905 				tcp_setpersist(tp);
906 		}
907 	}
908 
909 	/*
910 	 * Automatic sizing enables the performance of large buffers
911 	 * and most of the efficiency of small ones by only allocating
912 	 * space when it is needed.
913 	 *
914 	 * The criteria to step up the send buffer one notch are:
915 	 *  1. receive window of remote host is larger than send buffer
916 	 *     (with a fudge factor of 5/4th);
917 	 *  2. send buffer is filled to 7/8th with data (so we actually
918 	 *     have data to make use of it);
919 	 *  3. send buffer fill has not hit maximal automatic size;
920 	 *  4. our send window (slow start and cogestion controlled) is
921 	 *     larger than sent but unacknowledged data in send buffer.
922 	 *
923 	 * The remote host receive window scaling factor may limit the
924 	 * growing of the send buffer before it reaches its allowed
925 	 * maximum.
926 	 *
927 	 * It scales directly with slow start or congestion window
928 	 * and does at most one step per received ACK.  This fast
929 	 * scaling has the drawback of growing the send buffer beyond
930 	 * what is strictly necessary to make full use of a given
931 	 * delay*bandwith product.  However testing has shown this not
932 	 * to be much of an problem.  At worst we are trading wasting
933 	 * of available bandwith (the non-use of it) for wasting some
934 	 * socket buffer memory.
935 	 *
936 	 * TODO: Shrink send buffer during idle periods together
937 	 * with congestion window.  Requires another timer.
938 	 */
939 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
940 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
941 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
942 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
943 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
944 			if (!sbreserve(&so->so_snd,
945 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
946 			     tcp_autosndbuf_max), so))
947 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
948 		}
949 	}
950 
951 	if (len > txsegsize) {
952 		if (use_tso) {
953 			/*
954 			 * Truncate TSO transfers to IP_MAXPACKET, and make
955 			 * sure that we send equal size transfers down the
956 			 * stack (rather than big-small-big-small-...).
957 			 */
958 #ifdef INET6
959 			CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET);
960 #endif
961 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
962 			if (len <= txsegsize) {
963 				use_tso = 0;
964 			}
965 		} else
966 			len = txsegsize;
967 		flags &= ~TH_FIN;
968 		sendalot = 1;
969 	} else
970 		use_tso = 0;
971 	if (sack_rxmit) {
972 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
973 			flags &= ~TH_FIN;
974 	}
975 
976 	win = sbspace(&so->so_rcv);
977 
978 	/*
979 	 * Sender silly window avoidance.  If connection is idle
980 	 * and can send all data, a maximum segment,
981 	 * at least a maximum default-size segment do it,
982 	 * or are forced, do it; otherwise don't bother.
983 	 * If peer's buffer is tiny, then send
984 	 * when window is at least half open.
985 	 * If retransmitting (possibly after persist timer forced us
986 	 * to send into a small window), then must resend.
987 	 */
988 	if (len) {
989 		if (len >= txsegsize)
990 			goto send;
991 		if ((so->so_state & SS_MORETOCOME) == 0 &&
992 		    ((idle || tp->t_flags & TF_NODELAY) &&
993 		     len + off >= so->so_snd.sb_cc))
994 			goto send;
995 		if (tp->t_force)
996 			goto send;
997 		if (len >= tp->max_sndwnd / 2)
998 			goto send;
999 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1000 			goto send;
1001 		if (sack_rxmit)
1002 			goto send;
1003 	}
1004 
1005 	/*
1006 	 * Compare available window to amount of window known to peer
1007 	 * (as advertised window less next expected input).  If the
1008 	 * difference is at least twice the size of the largest segment
1009 	 * we expect to receive (i.e. two segments) or at least 50% of
1010 	 * the maximum possible window, then want to send a window update
1011 	 * to peer.
1012 	 */
1013 	if (win > 0) {
1014 		/*
1015 		 * "adv" is the amount we can increase the window,
1016 		 * taking into account that we are limited by
1017 		 * TCP_MAXWIN << tp->rcv_scale.
1018 		 */
1019 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1020 			(tp->rcv_adv - tp->rcv_nxt);
1021 
1022 		if (adv >= (long) (2 * rxsegsize))
1023 			goto send;
1024 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1025 			goto send;
1026 	}
1027 
1028 	/*
1029 	 * Send if we owe peer an ACK.
1030 	 */
1031 	if (tp->t_flags & TF_ACKNOW)
1032 		goto send;
1033 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1034 		goto send;
1035 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1036 		goto send;
1037 	/*
1038 	 * In SACK, it is possible for tcp_output to fail to send a segment
1039 	 * after the retransmission timer has been turned off.  Make sure
1040 	 * that the retransmission timer is set.
1041 	 */
1042 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1043 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1044 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1045 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1046 		goto just_return;
1047 	}
1048 
1049 	/*
1050 	 * TCP window updates are not reliable, rather a polling protocol
1051 	 * using ``persist'' packets is used to insure receipt of window
1052 	 * updates.  The three ``states'' for the output side are:
1053 	 *	idle			not doing retransmits or persists
1054 	 *	persisting		to move a small or zero window
1055 	 *	(re)transmitting	and thereby not persisting
1056 	 *
1057 	 * tp->t_timer[TCPT_PERSIST]
1058 	 *	is set when we are in persist state.
1059 	 * tp->t_force
1060 	 *	is set when we are called to send a persist packet.
1061 	 * tp->t_timer[TCPT_REXMT]
1062 	 *	is set when we are retransmitting
1063 	 * The output side is idle when both timers are zero.
1064 	 *
1065 	 * If send window is too small, there is data to transmit, and no
1066 	 * retransmit or persist is pending, then go to persist state.
1067 	 * If nothing happens soon, send when timer expires:
1068 	 * if window is nonzero, transmit what we can,
1069 	 * otherwise force out a byte.
1070 	 */
1071 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1072 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1073 		tp->t_rxtshift = 0;
1074 		tcp_setpersist(tp);
1075 	}
1076 
1077 	/*
1078 	 * No reason to send a segment, just return.
1079 	 */
1080 just_return:
1081 	TCP_REASS_UNLOCK(tp);
1082 	return (0);
1083 
1084 send:
1085 	/*
1086 	 * Before ESTABLISHED, force sending of initial options
1087 	 * unless TCP set not to do any options.
1088 	 * NOTE: we assume that the IP/TCP header plus TCP options
1089 	 * always fit in a single mbuf, leaving room for a maximum
1090 	 * link header, i.e.
1091 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1092 	 */
1093 	optlen = 0;
1094 	switch (af) {
1095 #ifdef INET
1096 	case AF_INET:
1097 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1098 		break;
1099 #endif
1100 #ifdef INET6
1101 	case AF_INET6:
1102 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1103 		break;
1104 #endif
1105 	default:	/*pacify gcc*/
1106 		iphdrlen = 0;
1107 		break;
1108 	}
1109 	hdrlen = iphdrlen;
1110 	if (flags & TH_SYN) {
1111 		struct rtentry *synrt;
1112 
1113 		synrt = NULL;
1114 #ifdef INET
1115 		if (tp->t_inpcb)
1116 			synrt = in_pcbrtentry(tp->t_inpcb);
1117 #endif
1118 #ifdef INET6
1119 		if (tp->t_in6pcb)
1120 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1121 #endif
1122 
1123 		tp->snd_nxt = tp->iss;
1124 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1125 						    synrt->rt_ifp : NULL, af);
1126 		if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(4)) {
1127 			opt[0] = TCPOPT_MAXSEG;
1128 			opt[1] = 4;
1129 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1130 			opt[3] = tp->t_ourmss & 0xff;
1131 			optlen = 4;
1132 
1133 			if ((tp->t_flags & TF_REQ_SCALE) &&
1134 			    ((flags & TH_ACK) == 0 ||
1135 			    (tp->t_flags & TF_RCVD_SCALE)) &&
1136 			    OPT_FITS(4)) {
1137 				*((u_int32_t *) (opt + optlen)) = htonl(
1138 					TCPOPT_NOP << 24 |
1139 					TCPOPT_WINDOW << 16 |
1140 					TCPOLEN_WINDOW << 8 |
1141 					tp->request_r_scale);
1142 				optlen += 4;
1143 			}
1144 			if (tcp_do_sack && OPT_FITS(4)) {
1145 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1146 
1147 				cp[0] = TCPOPT_SACK_PERMITTED;
1148 				cp[1] = 2;
1149 				cp[2] = TCPOPT_NOP;
1150 				cp[3] = TCPOPT_NOP;
1151 				optlen += 4;
1152 			}
1153 		}
1154 	}
1155 
1156 	/*
1157 	 * Send a timestamp and echo-reply if this is a SYN and our side
1158 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1159 	 * and our peer have sent timestamps in our SYN's.
1160 	 */
1161 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1162 	     (flags & TH_RST) == 0 &&
1163 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1164 	     (tp->t_flags & TF_RCVD_TSTMP)) && OPT_FITS(TCPOLEN_TSTAMP_APPA)) {
1165 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1166 
1167 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1168 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1169 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1170 		*lp   = htonl(tp->ts_recent);
1171 		optlen += TCPOLEN_TSTAMP_APPA;
1172 
1173 		/* Set receive buffer autosizing timestamp. */
1174 		if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1175 			tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1176 	}
1177 
1178 	/*
1179 	 * Tack on the SACK block if it is necessary.
1180 	 */
1181 	if (sack_numblks) {
1182 		int sack_len;
1183 		u_char *bp = (u_char *)(opt + optlen);
1184 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1185 		struct ipqent *tiqe;
1186 
1187 		sack_len = sack_numblks * 8 + 2;
1188 		if (OPT_FITS(sack_len + 2)) {
1189 			bp[0] = TCPOPT_NOP;
1190 			bp[1] = TCPOPT_NOP;
1191 			bp[2] = TCPOPT_SACK;
1192 			bp[3] = sack_len;
1193 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1194 				sack_numblks--;
1195 				*lp++ = htonl(tp->rcv_dsack_block.left);
1196 				*lp++ = htonl(tp->rcv_dsack_block.right);
1197 				tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1198 			}
1199 			for (tiqe = TAILQ_FIRST(&tp->timeq);
1200 			    sack_numblks > 0;
1201 			    tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1202 				KASSERT(tiqe != NULL);
1203 				sack_numblks--;
1204 				*lp++ = htonl(tiqe->ipqe_seq);
1205 				*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1206 				    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1207 			}
1208 			optlen += sack_len + 2;
1209 		}
1210 	}
1211 	TCP_REASS_UNLOCK(tp);
1212 
1213 #ifdef TCP_SIGNATURE
1214 	if ((tp->t_flags & TF_SIGNATURE) && OPT_FITS(TCPOLEN_SIGNATURE + 2)) {
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 = NULL;
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 			ecn_tos = IPTOS_ECN_ECT0;
1348 			TCP_STATINC(TCP_STAT_ECN_ECT);
1349 		}
1350 
1351 		/*
1352 		 * Reply with proper ECN notifications.
1353 		 */
1354 		if (tp->t_flags & TF_ECN_SND_CWR) {
1355 			flags |= TH_CWR;
1356 			tp->t_flags &= ~TF_ECN_SND_CWR;
1357 		}
1358 		if (tp->t_flags & TF_ECN_SND_ECE) {
1359 			flags |= TH_ECE;
1360 		}
1361 	}
1362 
1363 
1364 	/*
1365 	 * If we are doing retransmissions, then snd_nxt will
1366 	 * not reflect the first unsent octet.  For ACK only
1367 	 * packets, we do not want the sequence number of the
1368 	 * retransmitted packet, we want the sequence number
1369 	 * of the next unsent octet.  So, if there is no data
1370 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1371 	 * when filling in ti_seq.  But if we are in persist
1372 	 * state, snd_max might reflect one byte beyond the
1373 	 * right edge of the window, so use snd_nxt in that
1374 	 * case, since we know we aren't doing a retransmission.
1375 	 * (retransmit and persist are mutually exclusive...)
1376 	 */
1377 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1378 		th->th_seq = htonl(p->rxmit);
1379 		p->rxmit += len;
1380 	} else {
1381 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1382 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1383 			th->th_seq = htonl(tp->snd_nxt);
1384 		else
1385 			th->th_seq = htonl(tp->snd_max);
1386 	}
1387 	th->th_ack = htonl(tp->rcv_nxt);
1388 	if (optlen) {
1389 		bcopy((void *)opt, (void *)(th + 1), optlen);
1390 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1391 	}
1392 	th->th_flags = flags;
1393 	/*
1394 	 * Calculate receive window.  Don't shrink window,
1395 	 * but avoid silly window syndrome.
1396 	 */
1397 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1398 		win = 0;
1399 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1400 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1401 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1402 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1403 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1404 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1405 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1406 		if (urp > IP_MAXPACKET)
1407 			urp = IP_MAXPACKET;
1408 		th->th_urp = htons((u_int16_t)urp);
1409 		th->th_flags |= TH_URG;
1410 	} else
1411 		/*
1412 		 * If no urgent pointer to send, then we pull
1413 		 * the urgent pointer to the left edge of the send window
1414 		 * so that it doesn't drift into the send window on sequence
1415 		 * number wraparound.
1416 		 */
1417 		tp->snd_up = tp->snd_una;		/* drag it along */
1418 
1419 #ifdef TCP_SIGNATURE
1420 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1421 		struct secasvar *sav;
1422 		u_int8_t *sigp;
1423 
1424 		sav = tcp_signature_getsav(m, th);
1425 
1426 		if (sav == NULL) {
1427 			if (m)
1428 				m_freem(m);
1429 			return (EPERM);
1430 		}
1431 
1432 		m->m_pkthdr.len = hdrlen + len;
1433 		sigp = (char *)th + sizeof(*th) + sigoff;
1434 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1435 
1436 		key_sa_recordxfer(sav, m);
1437 		KEY_FREESAV(&sav);
1438 	}
1439 #endif
1440 
1441 	/*
1442 	 * Set ourselves up to be checksummed just before the packet
1443 	 * hits the wire.
1444 	 */
1445 	switch (af) {
1446 #ifdef INET
1447 	case AF_INET:
1448 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1449 		if (use_tso) {
1450 			m->m_pkthdr.segsz = txsegsize;
1451 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1452 		} else {
1453 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1454 			if (len + optlen) {
1455 				/* Fixup the pseudo-header checksum. */
1456 				/* XXXJRT Not IP Jumbogram safe. */
1457 				th->th_sum = in_cksum_addword(th->th_sum,
1458 				    htons((u_int16_t) (len + optlen)));
1459 			}
1460 		}
1461 		break;
1462 #endif
1463 #ifdef INET6
1464 	case AF_INET6:
1465 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1466 		if (use_tso) {
1467 			m->m_pkthdr.segsz = txsegsize;
1468 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1469 		} else {
1470 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1471 			if (len + optlen) {
1472 				/* Fixup the pseudo-header checksum. */
1473 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1474 				th->th_sum = in_cksum_addword(th->th_sum,
1475 				    htons((u_int16_t) (len + optlen)));
1476 			}
1477 		}
1478 		break;
1479 #endif
1480 	}
1481 
1482 	/*
1483 	 * In transmit state, time the transmission and arrange for
1484 	 * the retransmit.  In persist state, just set snd_max.
1485 	 */
1486 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1487 		tcp_seq startseq = tp->snd_nxt;
1488 
1489 		/*
1490 		 * Advance snd_nxt over sequence space of this segment.
1491 		 * There are no states in which we send both a SYN and a FIN,
1492 		 * so we collapse the tests for these flags.
1493 		 */
1494 		if (flags & (TH_SYN|TH_FIN))
1495 			tp->snd_nxt++;
1496 		if (sack_rxmit)
1497 			goto timer;
1498 		tp->snd_nxt += len;
1499 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1500 			tp->snd_max = tp->snd_nxt;
1501 			/*
1502 			 * Time this transmission if not a retransmission and
1503 			 * not currently timing anything.
1504 			 */
1505 			if (tp->t_rtttime == 0) {
1506 				tp->t_rtttime = tcp_now;
1507 				tp->t_rtseq = startseq;
1508 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1509 			}
1510 		}
1511 
1512 		/*
1513 		 * Set retransmit timer if not currently set,
1514 		 * and not doing an ack or a keep-alive probe.
1515 		 * Initial value for retransmit timer is smoothed
1516 		 * round-trip time + 2 * round-trip time variance.
1517 		 * Initialize shift counter which is used for backoff
1518 		 * of retransmit time.
1519 		 */
1520 timer:
1521 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1522 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1523 		    tp->snd_nxt != tp->snd_una)) {
1524 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1525 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1526 				tp->t_rxtshift = 0;
1527 			}
1528 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1529 		}
1530 	} else
1531 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1532 			tp->snd_max = tp->snd_nxt + len;
1533 
1534 #ifdef TCP_DEBUG
1535 	/*
1536 	 * Trace.
1537 	 */
1538 	if (so->so_options & SO_DEBUG)
1539 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1540 #endif
1541 
1542 	/*
1543 	 * Fill in IP length and desired time to live and
1544 	 * send to IP level.  There should be a better way
1545 	 * to handle ttl and tos; we could keep them in
1546 	 * the template, but need a way to checksum without them.
1547 	 */
1548 	m->m_pkthdr.len = hdrlen + len;
1549 
1550 	switch (af) {
1551 #ifdef INET
1552 	case AF_INET:
1553 		ip->ip_len = htons(m->m_pkthdr.len);
1554 		packetlen = m->m_pkthdr.len;
1555 		if (tp->t_inpcb) {
1556 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1557 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos;
1558 		}
1559 #ifdef INET6
1560 		else if (tp->t_in6pcb) {
1561 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1562 			ip->ip_tos = ecn_tos;	/*XXX*/
1563 		}
1564 #endif
1565 		break;
1566 #endif
1567 #ifdef INET6
1568 	case AF_INET6:
1569 		packetlen = m->m_pkthdr.len;
1570 		ip6->ip6_nxt = IPPROTO_TCP;
1571 		if (tp->t_in6pcb) {
1572 			/*
1573 			 * we separately set hoplimit for every segment, since
1574 			 * the user might want to change the value via
1575 			 * setsockopt. Also, desired default hop limit might
1576 			 * be changed via Neighbor Discovery.
1577 			 */
1578 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1579 				(rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp
1580 				                                    : NULL);
1581 		}
1582 		ip6->ip6_flow |= htonl(ecn_tos << 20);
1583 		/* ip6->ip6_flow = ??? (from template) */
1584 		/* ip6_plen will be filled in ip6_output(). */
1585 		break;
1586 #endif
1587 	default:	/*pacify gcc*/
1588 		packetlen = 0;
1589 		break;
1590 	}
1591 
1592 	switch (af) {
1593 #ifdef INET
1594 	case AF_INET:
1595 	    {
1596 		struct mbuf *opts;
1597 
1598 		if (tp->t_inpcb)
1599 			opts = tp->t_inpcb->inp_options;
1600 		else
1601 			opts = NULL;
1602 		error = ip_output(m, opts, ro,
1603 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1604 			(so->so_options & SO_DONTROUTE), NULL, so);
1605 		break;
1606 	    }
1607 #endif
1608 #ifdef INET6
1609 	case AF_INET6:
1610 	    {
1611 		struct ip6_pktopts *opts;
1612 
1613 		if (tp->t_in6pcb)
1614 			opts = tp->t_in6pcb->in6p_outputopts;
1615 		else
1616 			opts = NULL;
1617 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1618 			NULL, so, NULL);
1619 		break;
1620 	    }
1621 #endif
1622 	default:
1623 		error = EAFNOSUPPORT;
1624 		break;
1625 	}
1626 	if (error) {
1627 out:
1628 		if (error == ENOBUFS) {
1629 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1630 #ifdef INET
1631 			if (tp->t_inpcb)
1632 				tcp_quench(tp->t_inpcb, 0);
1633 #endif
1634 #ifdef INET6
1635 			if (tp->t_in6pcb)
1636 				tcp6_quench(tp->t_in6pcb, 0);
1637 #endif
1638 			error = 0;
1639 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1640 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1641 			tp->t_softerror = error;
1642 			error = 0;
1643 		}
1644 
1645 		/* Back out the seqence number advance. */
1646 		if (sack_rxmit)
1647 			p->rxmit -= len;
1648 
1649 		/* Restart the delayed ACK timer, if necessary. */
1650 		if (tp->t_flags & TF_DELACK)
1651 			TCP_RESTART_DELACK(tp);
1652 
1653 		return (error);
1654 	}
1655 
1656 	if (packetlen > tp->t_pmtud_mtu_sent)
1657 		tp->t_pmtud_mtu_sent = packetlen;
1658 
1659 	tcps = TCP_STAT_GETREF();
1660 	tcps[TCP_STAT_SNDTOTAL]++;
1661 	if (tp->t_flags & TF_DELACK)
1662 		tcps[TCP_STAT_DELACK]++;
1663 	TCP_STAT_PUTREF();
1664 
1665 	/*
1666 	 * Data sent (as far as we can tell).
1667 	 * If this advertises a larger window than any other segment,
1668 	 * then remember the size of the advertised window.
1669 	 * Any pending ACK has now been sent.
1670 	 */
1671 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1672 		tp->rcv_adv = tp->rcv_nxt + win;
1673 	tp->last_ack_sent = tp->rcv_nxt;
1674 	tp->t_flags &= ~TF_ACKNOW;
1675 	TCP_CLEAR_DELACK(tp);
1676 #ifdef DIAGNOSTIC
1677 	if (maxburst < 0)
1678 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1679 #endif
1680 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1681 		goto again;
1682 	return (0);
1683 }
1684 
1685 void
1686 tcp_setpersist(struct tcpcb *tp)
1687 {
1688 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1689 	int nticks;
1690 
1691 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1692 		panic("tcp_output REXMT");
1693 	/*
1694 	 * Start/restart persistance timer.
1695 	 */
1696 	if (t < tp->t_rttmin)
1697 		t = tp->t_rttmin;
1698 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1699 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1700 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1701 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1702 		tp->t_rxtshift++;
1703 }
1704