xref: /netbsd-src/sys/netinet/tcp_output.c (revision a536ee5124e62c9a0051a252f7833dc8f50f44c9)
1 /*	$NetBSD: tcp_output.c,v 1.174 2012/03/22 20:34:39 drochner 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 1.174 2012/03/22 20:34:39 drochner 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 
186 #include <netinet/tcp.h>
187 #define	TCPOUTFLAGS
188 #include <netinet/tcp_fsm.h>
189 #include <netinet/tcp_seq.h>
190 #include <netinet/tcp_timer.h>
191 #include <netinet/tcp_var.h>
192 #include <netinet/tcp_private.h>
193 #include <netinet/tcp_congctl.h>
194 #include <netinet/tcpip.h>
195 #include <netinet/tcp_debug.h>
196 #include <netinet/in_offload.h>
197 #include <netinet6/in6_offload.h>
198 
199 #ifdef notyet
200 extern struct mbuf *m_copypack();
201 #endif
202 
203 /*
204  * Knob to enable Congestion Window Monitoring, and control
205  * the burst size it allows.  Default burst is 4 packets, per
206  * the Internet draft.
207  */
208 int	tcp_cwm = 0;
209 int	tcp_cwm_burstsize = 4;
210 
211 int	tcp_do_autosndbuf = 1;
212 int	tcp_autosndbuf_inc = 8 * 1024;
213 int	tcp_autosndbuf_max = 256 * 1024;
214 
215 #ifdef TCP_OUTPUT_COUNTERS
216 #include <sys/device.h>
217 
218 extern struct evcnt tcp_output_bigheader;
219 extern struct evcnt tcp_output_predict_hit;
220 extern struct evcnt tcp_output_predict_miss;
221 extern struct evcnt tcp_output_copysmall;
222 extern struct evcnt tcp_output_copybig;
223 extern struct evcnt tcp_output_refbig;
224 
225 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
226 #else
227 
228 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
229 
230 #endif /* TCP_OUTPUT_COUNTERS */
231 
232 static
233 #ifndef GPROF
234 inline
235 #endif
236 int
237 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
238     bool *alwaysfragp)
239 {
240 #ifdef INET
241 	struct inpcb *inp = tp->t_inpcb;
242 #endif
243 #ifdef INET6
244 	struct in6pcb *in6p = tp->t_in6pcb;
245 #endif
246 	struct socket *so = NULL;
247 	struct rtentry *rt;
248 	struct ifnet *ifp;
249 	int size;
250 	int hdrlen;
251 	int optlen;
252 
253 	*alwaysfragp = false;
254 
255 #ifdef DIAGNOSTIC
256 	if (tp->t_inpcb && tp->t_in6pcb)
257 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
258 #endif
259 	switch (tp->t_family) {
260 #ifdef INET
261 	case AF_INET:
262 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
263 		break;
264 #endif
265 #ifdef INET6
266 	case AF_INET6:
267 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
268 		break;
269 #endif
270 	default:
271 		size = tcp_mssdflt;
272 		goto out;
273 	}
274 
275 	rt = NULL;
276 #ifdef INET
277 	if (inp) {
278 		rt = in_pcbrtentry(inp);
279 		so = inp->inp_socket;
280 	}
281 #endif
282 #ifdef INET6
283 	if (in6p) {
284 		rt = in6_pcbrtentry(in6p);
285 		so = in6p->in6p_socket;
286 	}
287 #endif
288 	if (rt == NULL) {
289 		size = tcp_mssdflt;
290 		goto out;
291 	}
292 
293 	ifp = rt->rt_ifp;
294 
295 	size = tcp_mssdflt;
296 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
297 #ifdef INET6
298 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
299 			/*
300 			 * RFC2460 section 5, last paragraph: if path MTU is
301 			 * smaller than 1280, use 1280 as packet size and
302 			 * attach fragment header.
303 			 */
304 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
305 			*alwaysfragp = true;
306 		} else
307 			size = rt->rt_rmx.rmx_mtu - hdrlen;
308 #else
309 		size = rt->rt_rmx.rmx_mtu - hdrlen;
310 #endif
311 	} else if (ifp->if_flags & IFF_LOOPBACK)
312 		size = ifp->if_mtu - hdrlen;
313 #ifdef INET
314 	else if (inp && tp->t_mtudisc)
315 		size = ifp->if_mtu - hdrlen;
316 	else if (inp && in_localaddr(inp->inp_faddr))
317 		size = ifp->if_mtu - hdrlen;
318 #endif
319 #ifdef INET6
320 	else if (in6p) {
321 #ifdef INET
322 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
323 			/* mapped addr case */
324 			struct in_addr d;
325 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
326 			if (tp->t_mtudisc || in_localaddr(d))
327 				size = ifp->if_mtu - hdrlen;
328 		} else
329 #endif
330 		{
331 			/*
332 			 * for IPv6, path MTU discovery is always turned on,
333 			 * or the node must use packet size <= 1280.
334 			 */
335 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
336 			size -= hdrlen;
337 		}
338 	}
339 #endif
340  out:
341 	/*
342 	 * Now we must make room for whatever extra TCP/IP options are in
343 	 * the packet.
344 	 */
345 	optlen = tcp_optlen(tp);
346 
347 	/*
348 	 * XXX tp->t_ourmss should have the right size, but without this code
349 	 * fragmentation will occur... need more investigation
350 	 */
351 #ifdef INET
352 	if (inp) {
353 #if defined(FAST_IPSEC)
354 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
355 			optlen += ipsec4_hdrsiz_tcp(tp);
356 #endif
357 		optlen += ip_optlen(inp);
358 	}
359 #endif
360 #ifdef INET6
361 #ifdef INET
362 	if (in6p && tp->t_family == AF_INET) {
363 #if defined(FAST_IPSEC)
364 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
365 			optlen += ipsec4_hdrsiz_tcp(tp);
366 #endif
367 		/* XXX size -= ip_optlen(in6p); */
368 	} else
369 #endif
370 	if (in6p && tp->t_family == AF_INET6) {
371 #if defined(FAST_IPSEC)
372 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
373 			optlen += ipsec6_hdrsiz_tcp(tp);
374 #endif
375 		optlen += ip6_optlen(in6p);
376 	}
377 #endif
378 	size -= optlen;
379 
380 	/* there may not be any room for data if mtu is too small */
381 	if (size < 0)
382 		return (EMSGSIZE);
383 
384 	/*
385 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
386 	 * assumes that path MTU is the same for both ways).  this is only
387 	 * for silly window avoidance, do not use the value for other purposes.
388 	 *
389 	 * ipseclen is subtracted from both sides, this may not be right.
390 	 * I'm not quite sure about this (could someone comment).
391 	 */
392 	*txsegsizep = min(tp->t_peermss - optlen, size);
393 	/*
394 	 * Never send more than half a buffer full.  This insures that we can
395 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
396 	 * therefore acks will never be delayed unless we run out of data to
397 	 * transmit.
398 	 */
399 	if (so)
400 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
401 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
402 
403 	if (*txsegsizep != tp->t_segsz) {
404 		/*
405 		 * If the new segment size is larger, we don't want to
406 		 * mess up the congestion window, but if it is smaller
407 		 * we'll have to reduce the congestion window to ensure
408 		 * that we don't get into trouble with initial windows
409 		 * and the rest.  In any case, if the segment size
410 		 * has changed, chances are the path has, too, and
411 		 * our congestion window will be different.
412 		 */
413 		if (*txsegsizep < tp->t_segsz) {
414 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
415 					   * *txsegsizep, *txsegsizep);
416 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
417 						* *txsegsizep, *txsegsizep);
418 		}
419 		tp->t_segsz = *txsegsizep;
420 	}
421 
422 	return (0);
423 }
424 
425 static
426 #ifndef GPROF
427 inline
428 #endif
429 int
430 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
431     long len, int hdrlen, struct mbuf **mp)
432 {
433 	struct mbuf *m, *m0;
434 	uint64_t *tcps;
435 
436 	tcps = TCP_STAT_GETREF();
437 	if (tp->t_force && len == 1)
438 		tcps[TCP_STAT_SNDPROBE]++;
439 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
440 		tcps[TCP_STAT_SNDREXMITPACK]++;
441 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
442 	} else {
443 		tcps[TCP_STAT_SNDPACK]++;
444 		tcps[TCP_STAT_SNDBYTE] += len;
445 	}
446 	TCP_STAT_PUTREF();
447 #ifdef notyet
448 	if ((m = m_copypack(so->so_snd.sb_mb, off,
449 	    (int)len, max_linkhdr + hdrlen)) == 0)
450 		return (ENOBUFS);
451 	/*
452 	 * m_copypack left space for our hdr; use it.
453 	 */
454 	m->m_len += hdrlen;
455 	m->m_data -= hdrlen;
456 #else
457 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
458 	if (__predict_false(m == NULL))
459 		return (ENOBUFS);
460 	MCLAIM(m, &tcp_tx_mowner);
461 
462 	/*
463 	 * XXX Because other code assumes headers will fit in
464 	 * XXX one header mbuf.
465 	 *
466 	 * (This code should almost *never* be run.)
467 	 */
468 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
469 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
470 		MCLGET(m, M_DONTWAIT);
471 		if ((m->m_flags & M_EXT) == 0) {
472 			m_freem(m);
473 			return (ENOBUFS);
474 		}
475 	}
476 
477 	m->m_data += max_linkhdr;
478 	m->m_len = hdrlen;
479 
480 	/*
481 	 * To avoid traversing the whole sb_mb chain for correct
482 	 * data to send, remember last sent mbuf, its offset and
483 	 * the sent size.  When called the next time, see if the
484 	 * data to send is directly following the previous transfer.
485 	 * This is important for large TCP windows.
486 	 */
487 	if (off == 0 || tp->t_lastm == NULL ||
488 	    (tp->t_lastoff + tp->t_lastlen) != off) {
489 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
490 		/*
491 		 * Either a new packet or a retransmit.
492 		 * Start from the beginning.
493 		 */
494 		tp->t_lastm = so->so_snd.sb_mb;
495 		tp->t_inoff = off;
496 	} else {
497 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
498 		tp->t_inoff += tp->t_lastlen;
499 	}
500 
501 	/* Traverse forward to next packet */
502 	while (tp->t_inoff > 0) {
503 		if (tp->t_lastm == NULL)
504 			panic("tp->t_lastm == NULL");
505 		if (tp->t_inoff < tp->t_lastm->m_len)
506 			break;
507 		tp->t_inoff -= tp->t_lastm->m_len;
508 		tp->t_lastm = tp->t_lastm->m_next;
509 	}
510 
511 	tp->t_lastoff = off;
512 	tp->t_lastlen = len;
513 	m0 = tp->t_lastm;
514 	off = tp->t_inoff;
515 
516 	if (len <= M_TRAILINGSPACE(m)) {
517 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
518 		m->m_len += len;
519 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
520 	} else {
521 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
522 		if (m->m_next == NULL) {
523 			m_freem(m);
524 			return (ENOBUFS);
525 		}
526 #ifdef TCP_OUTPUT_COUNTERS
527 		if (m->m_next->m_flags & M_EXT)
528 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
529 		else
530 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
531 #endif /* TCP_OUTPUT_COUNTERS */
532 	}
533 #endif
534 
535 	*mp = m;
536 	return (0);
537 }
538 
539 /*
540  * Tcp output routine: figure out what should be sent and send it.
541  */
542 int
543 tcp_output(struct tcpcb *tp)
544 {
545 	struct rtentry *rt;
546 	struct socket *so;
547 	struct route *ro;
548 	long len, win;
549 	int off, flags, error;
550 	struct mbuf *m;
551 	struct ip *ip;
552 #ifdef INET6
553 	struct ip6_hdr *ip6;
554 #endif
555 	struct tcphdr *th;
556 	u_char opt[MAX_TCPOPTLEN];
557 	unsigned optlen, hdrlen, packetlen;
558 	unsigned int sack_numblks;
559 	int idle, sendalot, txsegsize, rxsegsize;
560 	int txsegsize_nosack;
561 	int maxburst = TCP_MAXBURST;
562 	int af;		/* address family on the wire */
563 	int iphdrlen;
564 	int has_tso4, has_tso6;
565 	int has_tso, use_tso;
566 	bool alwaysfrag;
567 	int sack_rxmit;
568 	int sack_bytes_rxmt;
569 	int ecn_tos;
570 	struct sackhole *p;
571 #ifdef TCP_SIGNATURE
572 	int sigoff = 0;
573 #endif
574 	uint64_t *tcps;
575 
576 #ifdef DIAGNOSTIC
577 	if (tp->t_inpcb && tp->t_in6pcb)
578 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
579 #endif
580 	so = NULL;
581 	ro = NULL;
582 	if (tp->t_inpcb) {
583 		so = tp->t_inpcb->inp_socket;
584 		ro = &tp->t_inpcb->inp_route;
585 	}
586 #ifdef INET6
587 	else if (tp->t_in6pcb) {
588 		so = tp->t_in6pcb->in6p_socket;
589 		ro = &tp->t_in6pcb->in6p_route;
590 	}
591 #endif
592 
593 	switch (af = tp->t_family) {
594 #ifdef INET
595 	case AF_INET:
596 		if (tp->t_inpcb)
597 			break;
598 #ifdef INET6
599 		/* mapped addr case */
600 		if (tp->t_in6pcb)
601 			break;
602 #endif
603 		return (EINVAL);
604 #endif
605 #ifdef INET6
606 	case AF_INET6:
607 		if (tp->t_in6pcb)
608 			break;
609 		return (EINVAL);
610 #endif
611 	default:
612 		return (EAFNOSUPPORT);
613 	}
614 
615 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
616 		return (EMSGSIZE);
617 
618 	idle = (tp->snd_max == tp->snd_una);
619 
620 	/*
621 	 * Determine if we can use TCP segmentation offload:
622 	 * - If we're using IPv4
623 	 * - If there is not an IPsec policy that prevents it
624 	 * - If the interface can do it
625 	 */
626 	has_tso4 = has_tso6 = false;
627 #if defined(INET)
628 	has_tso4 = tp->t_inpcb != NULL &&
629 #if defined(FAST_IPSEC)
630 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
631 		  		       IPSEC_DIR_OUTBOUND) &&
632 #endif
633 		  (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
634 		  (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
635 #endif /* defined(INET) */
636 #if defined(INET6)
637 	has_tso6 = tp->t_in6pcb != NULL &&
638 #if defined(FAST_IPSEC)
639 		  IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
640 		  		       IPSEC_DIR_OUTBOUND) &&
641 #endif
642 		  (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
643 		  (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
644 #endif /* defined(INET6) */
645 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
646 
647 	/*
648 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
649 	 *
650 	 *	Optionally, a TCP MAY set the restart window to the
651 	 *	minimum of the value used for the initial window and
652 	 *	the current value of cwnd (in other words, using a
653 	 *	larger value for the restart window should never increase
654 	 *	the size of cwnd).
655 	 */
656 	if (tcp_cwm) {
657 		/*
658 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
659 		 * Count the number of packets currently pending
660 		 * acknowledgement, and limit our congestion window
661 		 * to a pre-determined allowed burst size plus that count.
662 		 * This prevents bursting once all pending packets have
663 		 * been acknowledged (i.e. transmission is idle).
664 		 *
665 		 * XXX Link this to Initial Window?
666 		 */
667 		tp->snd_cwnd = min(tp->snd_cwnd,
668 		    (tcp_cwm_burstsize * txsegsize) +
669 		    (tp->snd_nxt - tp->snd_una));
670 	} else {
671 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
672 			/*
673 			 * We have been idle for "a while" and no acks are
674 			 * expected to clock out any data we send --
675 			 * slow start to get ack "clock" running again.
676 			 */
677 			int ss = tcp_init_win;
678 #ifdef INET
679 			if (tp->t_inpcb &&
680 			    in_localaddr(tp->t_inpcb->inp_faddr))
681 				ss = tcp_init_win_local;
682 #endif
683 #ifdef INET6
684 			if (tp->t_in6pcb &&
685 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
686 				ss = tcp_init_win_local;
687 #endif
688 			tp->snd_cwnd = min(tp->snd_cwnd,
689 			    TCP_INITIAL_WINDOW(ss, txsegsize));
690 		}
691 	}
692 
693 	txsegsize_nosack = txsegsize;
694 again:
695 	ecn_tos = 0;
696 	use_tso = has_tso;
697 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
698 		/* don't duplicate CWR/ECE. */
699 		use_tso = 0;
700 	}
701 	TCP_REASS_LOCK(tp);
702 	sack_numblks = tcp_sack_numblks(tp);
703 	if (sack_numblks) {
704 		int sackoptlen;
705 
706 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
707 		if (sackoptlen > txsegsize_nosack) {
708 			sack_numblks = 0; /* give up SACK */
709 			txsegsize = txsegsize_nosack;
710 		} else {
711 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
712 				/* don't duplicate D-SACK. */
713 				use_tso = 0;
714 			}
715 			txsegsize = txsegsize_nosack - sackoptlen;
716 		}
717 	} else {
718 		txsegsize = txsegsize_nosack;
719 	}
720 
721 	/*
722 	 * Determine length of data that should be transmitted, and
723 	 * flags that should be used.  If there is some data or critical
724 	 * controls (SYN, RST) to send, then transmit; otherwise,
725 	 * investigate further.
726 	 *
727 	 * Readjust SACK information to avoid resending duplicate data.
728 	 */
729 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
730 		tcp_sack_adjust(tp);
731 	sendalot = 0;
732 	off = tp->snd_nxt - tp->snd_una;
733 	win = min(tp->snd_wnd, tp->snd_cwnd);
734 
735 	flags = tcp_outflags[tp->t_state];
736 
737 	/*
738 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
739 	 * to send out new data (when sendalot is 1), bypass this function.
740 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
741 	 * we're replacing a (future) new transmission with a retransmission
742 	 * now, and we previously incremented snd_cwnd in tcp_input().
743 	 */
744 	/*
745 	 * Still in sack recovery , reset rxmit flag to zero.
746 	 */
747 	sack_rxmit = 0;
748 	sack_bytes_rxmt = 0;
749 	len = 0;
750 	p = NULL;
751 	do {
752 		long cwin;
753 		if (!TCP_SACK_ENABLED(tp))
754 			break;
755 		if (tp->t_partialacks < 0)
756 			break;
757 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
758 		if (p == NULL)
759 			break;
760 
761 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
762 		if (cwin < 0)
763 			cwin = 0;
764 		/* Do not retransmit SACK segments beyond snd_recover */
765 		if (SEQ_GT(p->end, tp->snd_recover)) {
766 			/*
767 			 * (At least) part of sack hole extends beyond
768 			 * snd_recover. Check to see if we can rexmit data
769 			 * for this hole.
770 			 */
771 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
772 				/*
773 				 * Can't rexmit any more data for this hole.
774 				 * That data will be rexmitted in the next
775 				 * sack recovery episode, when snd_recover
776 				 * moves past p->rxmit.
777 				 */
778 				p = NULL;
779 				break;
780 			}
781 			/* Can rexmit part of the current hole */
782 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
783 		} else
784 			len = ((long)ulmin(cwin, p->end - p->rxmit));
785 		off = p->rxmit - tp->snd_una;
786 		if (off + len > so->so_snd.sb_cc) {
787 			/* 1 for TH_FIN */
788 			KASSERT(off + len == so->so_snd.sb_cc + 1);
789 			KASSERT(p->rxmit + len == tp->snd_max);
790 			len = so->so_snd.sb_cc - off;
791 		}
792 		if (len > 0) {
793 			sack_rxmit = 1;
794 			sendalot = 1;
795 		}
796 	} while (/*CONSTCOND*/0);
797 
798 	/*
799 	 * If in persist timeout with window of 0, send 1 byte.
800 	 * Otherwise, if window is small but nonzero
801 	 * and timer expired, we will send what we can
802 	 * and go to transmit state.
803 	 */
804 	if (tp->t_force) {
805 		if (win == 0) {
806 			/*
807 			 * If we still have some data to send, then
808 			 * clear the FIN bit.  Usually this would
809 			 * happen below when it realizes that we
810 			 * aren't sending all the data.  However,
811 			 * if we have exactly 1 byte of unset data,
812 			 * then it won't clear the FIN bit below,
813 			 * and if we are in persist state, we wind
814 			 * up sending the packet without recording
815 			 * that we sent the FIN bit.
816 			 *
817 			 * We can't just blindly clear the FIN bit,
818 			 * because if we don't have any more data
819 			 * to send then the probe will be the FIN
820 			 * itself.
821 			 */
822 			if (off < so->so_snd.sb_cc)
823 				flags &= ~TH_FIN;
824 			win = 1;
825 		} else {
826 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
827 			tp->t_rxtshift = 0;
828 		}
829 	}
830 
831 	if (sack_rxmit == 0) {
832 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
833 			long cwin;
834 
835 			/*
836 			 * We are inside of a SACK recovery episode and are
837 			 * sending new data, having retransmitted all the
838 			 * data possible in the scoreboard.
839 			 */
840 			if (tp->snd_wnd < so->so_snd.sb_cc) {
841 				len = tp->snd_wnd - off;
842 				flags &= ~TH_FIN;
843 			} else {
844 				len = so->so_snd.sb_cc - off;
845 			}
846 
847 			/*
848 			 * From FreeBSD:
849 			 *  Don't remove this (len > 0) check !
850 			 *  We explicitly check for len > 0 here (although it
851 			 *  isn't really necessary), to work around a gcc
852 			 *  optimization issue - to force gcc to compute
853 			 *  len above. Without this check, the computation
854 			 *  of len is bungled by the optimizer.
855 			 */
856 			if (len > 0) {
857 				cwin = tp->snd_cwnd -
858 				    (tp->snd_nxt - tp->sack_newdata) -
859 				    sack_bytes_rxmt;
860 				if (cwin < 0)
861 					cwin = 0;
862 				if (cwin < len) {
863 					len = cwin;
864 					flags &= ~TH_FIN;
865 				}
866 			}
867 		} else if (win < so->so_snd.sb_cc) {
868 			len = win - off;
869 			flags &= ~TH_FIN;
870 		} else {
871 			len = so->so_snd.sb_cc - off;
872 		}
873 	}
874 
875 	if (len < 0) {
876 		/*
877 		 * If FIN has been sent but not acked,
878 		 * but we haven't been called to retransmit,
879 		 * len will be -1.  Otherwise, window shrank
880 		 * after we sent into it.  If window shrank to 0,
881 		 * cancel pending retransmit, pull snd_nxt back
882 		 * to (closed) window, and set the persist timer
883 		 * if it isn't already going.  If the window didn't
884 		 * close completely, just wait for an ACK.
885 		 *
886 		 * If we have a pending FIN, either it has already been
887 		 * transmitted or it is outside the window, so drop it.
888 		 * If the FIN has been transmitted, but this is not a
889 		 * retransmission, then len must be -1.  Therefore we also
890 		 * prevent here the sending of `gratuitous FINs'.  This
891 		 * eliminates the need to check for that case below (e.g.
892 		 * to back up snd_nxt before the FIN so that the sequence
893 		 * number is correct).
894 		 */
895 		len = 0;
896 		flags &= ~TH_FIN;
897 		if (win == 0) {
898 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
899 			tp->t_rxtshift = 0;
900 			tp->snd_nxt = tp->snd_una;
901 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
902 				tcp_setpersist(tp);
903 		}
904 	}
905 
906 	/*
907 	 * Automatic sizing enables the performance of large buffers
908 	 * and most of the efficiency of small ones by only allocating
909 	 * space when it is needed.
910 	 *
911 	 * The criteria to step up the send buffer one notch are:
912 	 *  1. receive window of remote host is larger than send buffer
913 	 *     (with a fudge factor of 5/4th);
914 	 *  2. send buffer is filled to 7/8th with data (so we actually
915 	 *     have data to make use of it);
916 	 *  3. send buffer fill has not hit maximal automatic size;
917 	 *  4. our send window (slow start and cogestion controlled) is
918 	 *     larger than sent but unacknowledged data in send buffer.
919 	 *
920 	 * The remote host receive window scaling factor may limit the
921 	 * growing of the send buffer before it reaches its allowed
922 	 * maximum.
923 	 *
924 	 * It scales directly with slow start or congestion window
925 	 * and does at most one step per received ACK.  This fast
926 	 * scaling has the drawback of growing the send buffer beyond
927 	 * what is strictly necessary to make full use of a given
928 	 * delay*bandwith product.  However testing has shown this not
929 	 * to be much of an problem.  At worst we are trading wasting
930 	 * of available bandwith (the non-use of it) for wasting some
931 	 * socket buffer memory.
932 	 *
933 	 * TODO: Shrink send buffer during idle periods together
934 	 * with congestion window.  Requires another timer.
935 	 */
936 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
937 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
938 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
939 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
940 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
941 			if (!sbreserve(&so->so_snd,
942 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
943 			     tcp_autosndbuf_max), so))
944 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
945 		}
946 	}
947 
948 	if (len > txsegsize) {
949 		if (use_tso) {
950 			/*
951 			 * Truncate TSO transfers to IP_MAXPACKET, and make
952 			 * sure that we send equal size transfers down the
953 			 * stack (rather than big-small-big-small-...).
954 			 */
955 #ifdef INET6
956 			CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET);
957 #endif
958 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
959 			if (len <= txsegsize) {
960 				use_tso = 0;
961 			}
962 		} else
963 			len = txsegsize;
964 		flags &= ~TH_FIN;
965 		sendalot = 1;
966 	} else
967 		use_tso = 0;
968 	if (sack_rxmit) {
969 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
970 			flags &= ~TH_FIN;
971 	}
972 
973 	win = sbspace(&so->so_rcv);
974 
975 	/*
976 	 * Sender silly window avoidance.  If connection is idle
977 	 * and can send all data, a maximum segment,
978 	 * at least a maximum default-size segment do it,
979 	 * or are forced, do it; otherwise don't bother.
980 	 * If peer's buffer is tiny, then send
981 	 * when window is at least half open.
982 	 * If retransmitting (possibly after persist timer forced us
983 	 * to send into a small window), then must resend.
984 	 */
985 	if (len) {
986 		if (len >= txsegsize)
987 			goto send;
988 		if ((so->so_state & SS_MORETOCOME) == 0 &&
989 		    ((idle || tp->t_flags & TF_NODELAY) &&
990 		     len + off >= so->so_snd.sb_cc))
991 			goto send;
992 		if (tp->t_force)
993 			goto send;
994 		if (len >= tp->max_sndwnd / 2)
995 			goto send;
996 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
997 			goto send;
998 		if (sack_rxmit)
999 			goto send;
1000 	}
1001 
1002 	/*
1003 	 * Compare available window to amount of window known to peer
1004 	 * (as advertised window less next expected input).  If the
1005 	 * difference is at least twice the size of the largest segment
1006 	 * we expect to receive (i.e. two segments) or at least 50% of
1007 	 * the maximum possible window, then want to send a window update
1008 	 * to peer.
1009 	 */
1010 	if (win > 0) {
1011 		/*
1012 		 * "adv" is the amount we can increase the window,
1013 		 * taking into account that we are limited by
1014 		 * TCP_MAXWIN << tp->rcv_scale.
1015 		 */
1016 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1017 			(tp->rcv_adv - tp->rcv_nxt);
1018 
1019 		if (adv >= (long) (2 * rxsegsize))
1020 			goto send;
1021 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1022 			goto send;
1023 	}
1024 
1025 	/*
1026 	 * Send if we owe peer an ACK.
1027 	 */
1028 	if (tp->t_flags & TF_ACKNOW)
1029 		goto send;
1030 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1031 		goto send;
1032 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1033 		goto send;
1034 	/*
1035 	 * In SACK, it is possible for tcp_output to fail to send a segment
1036 	 * after the retransmission timer has been turned off.  Make sure
1037 	 * that the retransmission timer is set.
1038 	 */
1039 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1040 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1041 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1042 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1043 		goto just_return;
1044 	}
1045 
1046 	/*
1047 	 * TCP window updates are not reliable, rather a polling protocol
1048 	 * using ``persist'' packets is used to insure receipt of window
1049 	 * updates.  The three ``states'' for the output side are:
1050 	 *	idle			not doing retransmits or persists
1051 	 *	persisting		to move a small or zero window
1052 	 *	(re)transmitting	and thereby not persisting
1053 	 *
1054 	 * tp->t_timer[TCPT_PERSIST]
1055 	 *	is set when we are in persist state.
1056 	 * tp->t_force
1057 	 *	is set when we are called to send a persist packet.
1058 	 * tp->t_timer[TCPT_REXMT]
1059 	 *	is set when we are retransmitting
1060 	 * The output side is idle when both timers are zero.
1061 	 *
1062 	 * If send window is too small, there is data to transmit, and no
1063 	 * retransmit or persist is pending, then go to persist state.
1064 	 * If nothing happens soon, send when timer expires:
1065 	 * if window is nonzero, transmit what we can,
1066 	 * otherwise force out a byte.
1067 	 */
1068 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1069 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1070 		tp->t_rxtshift = 0;
1071 		tcp_setpersist(tp);
1072 	}
1073 
1074 	/*
1075 	 * No reason to send a segment, just return.
1076 	 */
1077 just_return:
1078 	TCP_REASS_UNLOCK(tp);
1079 	return (0);
1080 
1081 send:
1082 	/*
1083 	 * Before ESTABLISHED, force sending of initial options
1084 	 * unless TCP set not to do any options.
1085 	 * NOTE: we assume that the IP/TCP header plus TCP options
1086 	 * always fit in a single mbuf, leaving room for a maximum
1087 	 * link header, i.e.
1088 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1089 	 */
1090 	optlen = 0;
1091 	switch (af) {
1092 #ifdef INET
1093 	case AF_INET:
1094 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1095 		break;
1096 #endif
1097 #ifdef INET6
1098 	case AF_INET6:
1099 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1100 		break;
1101 #endif
1102 	default:	/*pacify gcc*/
1103 		iphdrlen = 0;
1104 		break;
1105 	}
1106 	hdrlen = iphdrlen;
1107 	if (flags & TH_SYN) {
1108 		struct rtentry *synrt;
1109 
1110 		synrt = NULL;
1111 #ifdef INET
1112 		if (tp->t_inpcb)
1113 			synrt = in_pcbrtentry(tp->t_inpcb);
1114 #endif
1115 #ifdef INET6
1116 		if (tp->t_in6pcb)
1117 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1118 #endif
1119 
1120 		tp->snd_nxt = tp->iss;
1121 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1122 						    synrt->rt_ifp : NULL, af);
1123 		if ((tp->t_flags & TF_NOOPT) == 0) {
1124 			opt[0] = TCPOPT_MAXSEG;
1125 			opt[1] = 4;
1126 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1127 			opt[3] = tp->t_ourmss & 0xff;
1128 			optlen = 4;
1129 
1130 			if ((tp->t_flags & TF_REQ_SCALE) &&
1131 			    ((flags & TH_ACK) == 0 ||
1132 			    (tp->t_flags & TF_RCVD_SCALE))) {
1133 				*((u_int32_t *) (opt + optlen)) = htonl(
1134 					TCPOPT_NOP << 24 |
1135 					TCPOPT_WINDOW << 16 |
1136 					TCPOLEN_WINDOW << 8 |
1137 					tp->request_r_scale);
1138 				optlen += 4;
1139 			}
1140 			if (tcp_do_sack) {
1141 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1142 
1143 				cp[0] = TCPOPT_SACK_PERMITTED;
1144 				cp[1] = 2;
1145 				cp[2] = TCPOPT_NOP;
1146 				cp[3] = TCPOPT_NOP;
1147 				optlen += 4;
1148 			}
1149 		}
1150 	}
1151 
1152 	/*
1153 	 * Send a timestamp and echo-reply if this is a SYN and our side
1154 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1155 	 * and our peer have sent timestamps in our SYN's.
1156 	 */
1157 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1158 	     (flags & TH_RST) == 0 &&
1159 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1160 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1161 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1162 
1163 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1164 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1165 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1166 		*lp   = htonl(tp->ts_recent);
1167 		optlen += TCPOLEN_TSTAMP_APPA;
1168 
1169 		/* Set receive buffer autosizing timestamp. */
1170 		if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1171 			tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1172 	}
1173 
1174 	/*
1175 	 * Tack on the SACK block if it is necessary.
1176 	 */
1177 	if (sack_numblks) {
1178 		int sack_len;
1179 		u_char *bp = (u_char *)(opt + optlen);
1180 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1181 		struct ipqent *tiqe;
1182 
1183 		sack_len = sack_numblks * 8 + 2;
1184 		bp[0] = TCPOPT_NOP;
1185 		bp[1] = TCPOPT_NOP;
1186 		bp[2] = TCPOPT_SACK;
1187 		bp[3] = sack_len;
1188 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1189 			sack_numblks--;
1190 			*lp++ = htonl(tp->rcv_dsack_block.left);
1191 			*lp++ = htonl(tp->rcv_dsack_block.right);
1192 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1193 		}
1194 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1195 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1196 			KASSERT(tiqe != NULL);
1197 			sack_numblks--;
1198 			*lp++ = htonl(tiqe->ipqe_seq);
1199 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1200 			    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1201 		}
1202 		optlen += sack_len + 2;
1203 	}
1204 	TCP_REASS_UNLOCK(tp);
1205 
1206 #ifdef TCP_SIGNATURE
1207 	if (tp->t_flags & TF_SIGNATURE) {
1208 		u_char *bp;
1209 		/*
1210 		 * Initialize TCP-MD5 option (RFC2385)
1211 		 */
1212 		bp = (u_char *)opt + optlen;
1213 		*bp++ = TCPOPT_SIGNATURE;
1214 		*bp++ = TCPOLEN_SIGNATURE;
1215 		sigoff = optlen + 2;
1216 		memset(bp, 0, TCP_SIGLEN);
1217 		bp += TCP_SIGLEN;
1218 		optlen += TCPOLEN_SIGNATURE;
1219 		/*
1220 		 * Terminate options list and maintain 32-bit alignment.
1221  		 */
1222 		*bp++ = TCPOPT_NOP;
1223 		*bp++ = TCPOPT_EOL;
1224  		optlen += 2;
1225  	}
1226 #endif /* TCP_SIGNATURE */
1227 
1228 	hdrlen += optlen;
1229 
1230 #ifdef DIAGNOSTIC
1231 	if (!use_tso && len > txsegsize)
1232 		panic("tcp data to be sent is larger than segment");
1233 	else if (use_tso && len > IP_MAXPACKET)
1234 		panic("tcp data to be sent is larger than max TSO size");
1235 	if (max_linkhdr + hdrlen > MCLBYTES)
1236 		panic("tcphdr too big");
1237 #endif
1238 
1239 	/*
1240 	 * Grab a header mbuf, attaching a copy of data to
1241 	 * be transmitted, and initialize the header from
1242 	 * the template for sends on this connection.
1243 	 */
1244 	if (len) {
1245 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1246 		if (error)
1247 			goto out;
1248 		/*
1249 		 * If we're sending everything we've got, set PUSH.
1250 		 * (This will keep happy those implementations which only
1251 		 * give data to the user when a buffer fills or
1252 		 * a PUSH comes in.)
1253 		 */
1254 		if (off + len == so->so_snd.sb_cc)
1255 			flags |= TH_PUSH;
1256 	} else {
1257 		tcps = TCP_STAT_GETREF();
1258 		if (tp->t_flags & TF_ACKNOW)
1259 			tcps[TCP_STAT_SNDACKS]++;
1260 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1261 			tcps[TCP_STAT_SNDCTRL]++;
1262 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1263 			tcps[TCP_STAT_SNDURG]++;
1264 		else
1265 			tcps[TCP_STAT_SNDWINUP]++;
1266 		TCP_STAT_PUTREF();
1267 
1268 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1269 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1270 			MCLGET(m, M_DONTWAIT);
1271 			if ((m->m_flags & M_EXT) == 0) {
1272 				m_freem(m);
1273 				m = NULL;
1274 			}
1275 		}
1276 		if (m == NULL) {
1277 			error = ENOBUFS;
1278 			goto out;
1279 		}
1280 		MCLAIM(m, &tcp_tx_mowner);
1281 		m->m_data += max_linkhdr;
1282 		m->m_len = hdrlen;
1283 	}
1284 	m->m_pkthdr.rcvif = NULL;
1285 	switch (af) {
1286 #ifdef INET
1287 	case AF_INET:
1288 		ip = mtod(m, struct ip *);
1289 #ifdef INET6
1290 		ip6 = NULL;
1291 #endif
1292 		th = (struct tcphdr *)(ip + 1);
1293 		break;
1294 #endif
1295 #ifdef INET6
1296 	case AF_INET6:
1297 		ip = NULL;
1298 		ip6 = mtod(m, struct ip6_hdr *);
1299 		th = (struct tcphdr *)(ip6 + 1);
1300 		break;
1301 #endif
1302 	default:	/*pacify gcc*/
1303 		ip = NULL;
1304 #ifdef INET6
1305 		ip6 = NULL;
1306 #endif
1307 		th = NULL;
1308 		break;
1309 	}
1310 	if (tp->t_template == 0)
1311 		panic("tcp_output");
1312 	if (tp->t_template->m_len < iphdrlen)
1313 		panic("tcp_output");
1314 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1315 
1316 	/*
1317 	 * If we are starting a connection, send ECN setup
1318 	 * SYN packet. If we are on a retransmit, we may
1319 	 * resend those bits a number of times as per
1320 	 * RFC 3168.
1321 	 */
1322 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1323 		if (tp->t_flags & TF_SYN_REXMT) {
1324 			if (tp->t_ecn_retries--)
1325 				flags |= TH_ECE|TH_CWR;
1326 		} else {
1327 			flags |= TH_ECE|TH_CWR;
1328 			tp->t_ecn_retries = tcp_ecn_maxretries;
1329 		}
1330 	}
1331 
1332 	if (TCP_ECN_ALLOWED(tp)) {
1333 		/*
1334 		 * If the peer has ECN, mark data packets
1335 		 * ECN capable. Ignore pure ack packets, retransmissions
1336 		 * and window probes.
1337 		 */
1338 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1339 		    !(tp->t_force && len == 1)) {
1340 			ecn_tos = IPTOS_ECN_ECT0;
1341 			TCP_STATINC(TCP_STAT_ECN_ECT);
1342 		}
1343 
1344 		/*
1345 		 * Reply with proper ECN notifications.
1346 		 */
1347 		if (tp->t_flags & TF_ECN_SND_CWR) {
1348 			flags |= TH_CWR;
1349 			tp->t_flags &= ~TF_ECN_SND_CWR;
1350 		}
1351 		if (tp->t_flags & TF_ECN_SND_ECE) {
1352 			flags |= TH_ECE;
1353 		}
1354 	}
1355 
1356 
1357 	/*
1358 	 * If we are doing retransmissions, then snd_nxt will
1359 	 * not reflect the first unsent octet.  For ACK only
1360 	 * packets, we do not want the sequence number of the
1361 	 * retransmitted packet, we want the sequence number
1362 	 * of the next unsent octet.  So, if there is no data
1363 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1364 	 * when filling in ti_seq.  But if we are in persist
1365 	 * state, snd_max might reflect one byte beyond the
1366 	 * right edge of the window, so use snd_nxt in that
1367 	 * case, since we know we aren't doing a retransmission.
1368 	 * (retransmit and persist are mutually exclusive...)
1369 	 */
1370 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1371 		th->th_seq = htonl(p->rxmit);
1372 		p->rxmit += len;
1373 	} else {
1374 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1375 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1376 			th->th_seq = htonl(tp->snd_nxt);
1377 		else
1378 			th->th_seq = htonl(tp->snd_max);
1379 	}
1380 	th->th_ack = htonl(tp->rcv_nxt);
1381 	if (optlen) {
1382 		bcopy((void *)opt, (void *)(th + 1), optlen);
1383 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1384 	}
1385 	th->th_flags = flags;
1386 	/*
1387 	 * Calculate receive window.  Don't shrink window,
1388 	 * but avoid silly window syndrome.
1389 	 */
1390 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1391 		win = 0;
1392 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1393 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1394 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1395 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1396 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1397 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1398 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1399 		if (urp > IP_MAXPACKET)
1400 			urp = IP_MAXPACKET;
1401 		th->th_urp = htons((u_int16_t)urp);
1402 		th->th_flags |= TH_URG;
1403 	} else
1404 		/*
1405 		 * If no urgent pointer to send, then we pull
1406 		 * the urgent pointer to the left edge of the send window
1407 		 * so that it doesn't drift into the send window on sequence
1408 		 * number wraparound.
1409 		 */
1410 		tp->snd_up = tp->snd_una;		/* drag it along */
1411 
1412 #ifdef TCP_SIGNATURE
1413 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1414 		struct secasvar *sav;
1415 		u_int8_t *sigp;
1416 
1417 		sav = tcp_signature_getsav(m, th);
1418 
1419 		if (sav == NULL) {
1420 			if (m)
1421 				m_freem(m);
1422 			return (EPERM);
1423 		}
1424 
1425 		m->m_pkthdr.len = hdrlen + len;
1426 		sigp = (char *)th + sizeof(*th) + sigoff;
1427 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1428 
1429 		key_sa_recordxfer(sav, m);
1430 #ifdef FAST_IPSEC
1431 		KEY_FREESAV(&sav);
1432 #else
1433 		key_freesav(sav);
1434 #endif
1435 	}
1436 #endif
1437 
1438 	/*
1439 	 * Set ourselves up to be checksummed just before the packet
1440 	 * hits the wire.
1441 	 */
1442 	switch (af) {
1443 #ifdef INET
1444 	case AF_INET:
1445 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1446 		if (use_tso) {
1447 			m->m_pkthdr.segsz = txsegsize;
1448 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1449 		} else {
1450 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1451 			if (len + optlen) {
1452 				/* Fixup the pseudo-header checksum. */
1453 				/* XXXJRT Not IP Jumbogram safe. */
1454 				th->th_sum = in_cksum_addword(th->th_sum,
1455 				    htons((u_int16_t) (len + optlen)));
1456 			}
1457 		}
1458 		break;
1459 #endif
1460 #ifdef INET6
1461 	case AF_INET6:
1462 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1463 		if (use_tso) {
1464 			m->m_pkthdr.segsz = txsegsize;
1465 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1466 		} else {
1467 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1468 			if (len + optlen) {
1469 				/* Fixup the pseudo-header checksum. */
1470 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1471 				th->th_sum = in_cksum_addword(th->th_sum,
1472 				    htons((u_int16_t) (len + optlen)));
1473 			}
1474 		}
1475 		break;
1476 #endif
1477 	}
1478 
1479 	/*
1480 	 * In transmit state, time the transmission and arrange for
1481 	 * the retransmit.  In persist state, just set snd_max.
1482 	 */
1483 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1484 		tcp_seq startseq = tp->snd_nxt;
1485 
1486 		/*
1487 		 * Advance snd_nxt over sequence space of this segment.
1488 		 * There are no states in which we send both a SYN and a FIN,
1489 		 * so we collapse the tests for these flags.
1490 		 */
1491 		if (flags & (TH_SYN|TH_FIN))
1492 			tp->snd_nxt++;
1493 		if (sack_rxmit)
1494 			goto timer;
1495 		tp->snd_nxt += len;
1496 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1497 			tp->snd_max = tp->snd_nxt;
1498 			/*
1499 			 * Time this transmission if not a retransmission and
1500 			 * not currently timing anything.
1501 			 */
1502 			if (tp->t_rtttime == 0) {
1503 				tp->t_rtttime = tcp_now;
1504 				tp->t_rtseq = startseq;
1505 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1506 			}
1507 		}
1508 
1509 		/*
1510 		 * Set retransmit timer if not currently set,
1511 		 * and not doing an ack or a keep-alive probe.
1512 		 * Initial value for retransmit timer is smoothed
1513 		 * round-trip time + 2 * round-trip time variance.
1514 		 * Initialize shift counter which is used for backoff
1515 		 * of retransmit time.
1516 		 */
1517 timer:
1518 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1519 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1520 		    tp->snd_nxt != tp->snd_una)) {
1521 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1522 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1523 				tp->t_rxtshift = 0;
1524 			}
1525 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1526 		}
1527 	} else
1528 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1529 			tp->snd_max = tp->snd_nxt + len;
1530 
1531 #ifdef TCP_DEBUG
1532 	/*
1533 	 * Trace.
1534 	 */
1535 	if (so->so_options & SO_DEBUG)
1536 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1537 #endif
1538 
1539 	/*
1540 	 * Fill in IP length and desired time to live and
1541 	 * send to IP level.  There should be a better way
1542 	 * to handle ttl and tos; we could keep them in
1543 	 * the template, but need a way to checksum without them.
1544 	 */
1545 	m->m_pkthdr.len = hdrlen + len;
1546 
1547 	switch (af) {
1548 #ifdef INET
1549 	case AF_INET:
1550 		ip->ip_len = htons(m->m_pkthdr.len);
1551 		packetlen = m->m_pkthdr.len;
1552 		if (tp->t_inpcb) {
1553 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1554 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos;
1555 		}
1556 #ifdef INET6
1557 		else if (tp->t_in6pcb) {
1558 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1559 			ip->ip_tos = ecn_tos;	/*XXX*/
1560 		}
1561 #endif
1562 		break;
1563 #endif
1564 #ifdef INET6
1565 	case AF_INET6:
1566 		packetlen = m->m_pkthdr.len;
1567 		ip6->ip6_nxt = IPPROTO_TCP;
1568 		if (tp->t_in6pcb) {
1569 			/*
1570 			 * we separately set hoplimit for every segment, since
1571 			 * the user might want to change the value via
1572 			 * setsockopt. Also, desired default hop limit might
1573 			 * be changed via Neighbor Discovery.
1574 			 */
1575 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1576 				(rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp
1577 				                                    : NULL);
1578 		}
1579 		ip6->ip6_flow |= htonl(ecn_tos << 20);
1580 		/* ip6->ip6_flow = ??? (from template) */
1581 		/* ip6_plen will be filled in ip6_output(). */
1582 		break;
1583 #endif
1584 	default:	/*pacify gcc*/
1585 		packetlen = 0;
1586 		break;
1587 	}
1588 
1589 	switch (af) {
1590 #ifdef INET
1591 	case AF_INET:
1592 	    {
1593 		struct mbuf *opts;
1594 
1595 		if (tp->t_inpcb)
1596 			opts = tp->t_inpcb->inp_options;
1597 		else
1598 			opts = NULL;
1599 		error = ip_output(m, opts, ro,
1600 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1601 			(so->so_options & SO_DONTROUTE), NULL, so);
1602 		break;
1603 	    }
1604 #endif
1605 #ifdef INET6
1606 	case AF_INET6:
1607 	    {
1608 		struct ip6_pktopts *opts;
1609 
1610 		if (tp->t_in6pcb)
1611 			opts = tp->t_in6pcb->in6p_outputopts;
1612 		else
1613 			opts = NULL;
1614 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1615 			NULL, so, NULL);
1616 		break;
1617 	    }
1618 #endif
1619 	default:
1620 		error = EAFNOSUPPORT;
1621 		break;
1622 	}
1623 	if (error) {
1624 out:
1625 		if (error == ENOBUFS) {
1626 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1627 #ifdef INET
1628 			if (tp->t_inpcb)
1629 				tcp_quench(tp->t_inpcb, 0);
1630 #endif
1631 #ifdef INET6
1632 			if (tp->t_in6pcb)
1633 				tcp6_quench(tp->t_in6pcb, 0);
1634 #endif
1635 			error = 0;
1636 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1637 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1638 			tp->t_softerror = error;
1639 			error = 0;
1640 		}
1641 
1642 		/* Back out the seqence number advance. */
1643 		if (sack_rxmit)
1644 			p->rxmit -= len;
1645 
1646 		/* Restart the delayed ACK timer, if necessary. */
1647 		if (tp->t_flags & TF_DELACK)
1648 			TCP_RESTART_DELACK(tp);
1649 
1650 		return (error);
1651 	}
1652 
1653 	if (packetlen > tp->t_pmtud_mtu_sent)
1654 		tp->t_pmtud_mtu_sent = packetlen;
1655 
1656 	tcps = TCP_STAT_GETREF();
1657 	tcps[TCP_STAT_SNDTOTAL]++;
1658 	if (tp->t_flags & TF_DELACK)
1659 		tcps[TCP_STAT_DELACK]++;
1660 	TCP_STAT_PUTREF();
1661 
1662 	/*
1663 	 * Data sent (as far as we can tell).
1664 	 * If this advertises a larger window than any other segment,
1665 	 * then remember the size of the advertised window.
1666 	 * Any pending ACK has now been sent.
1667 	 */
1668 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1669 		tp->rcv_adv = tp->rcv_nxt + win;
1670 	tp->last_ack_sent = tp->rcv_nxt;
1671 	tp->t_flags &= ~TF_ACKNOW;
1672 	TCP_CLEAR_DELACK(tp);
1673 #ifdef DIAGNOSTIC
1674 	if (maxburst < 0)
1675 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1676 #endif
1677 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1678 		goto again;
1679 	return (0);
1680 }
1681 
1682 void
1683 tcp_setpersist(struct tcpcb *tp)
1684 {
1685 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1686 	int nticks;
1687 
1688 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1689 		panic("tcp_output REXMT");
1690 	/*
1691 	 * Start/restart persistance timer.
1692 	 */
1693 	if (t < tp->t_rttmin)
1694 		t = tp->t_rttmin;
1695 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1696 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1697 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1698 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1699 		tp->t_rxtshift++;
1700 }
1701