xref: /netbsd-src/sys/netinet/tcp_output.c (revision f89f6560d453f5e37386cc7938c072d2f528b9fa)
1 /*	$NetBSD: tcp_output.c,v 1.180 2015/02/14 12:57:53 he 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 1.180 2015/02/14 12:57:53 he 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 		tp->t_sndrexmitpack++;
443 		tcps[TCP_STAT_SNDREXMITPACK]++;
444 		tcps[TCP_STAT_SNDREXMITBYTE] += len;
445 	} else {
446 		tcps[TCP_STAT_SNDPACK]++;
447 		tcps[TCP_STAT_SNDBYTE] += len;
448 	}
449 	TCP_STAT_PUTREF();
450 #ifdef notyet
451 	if ((m = m_copypack(so->so_snd.sb_mb, off,
452 	    (int)len, max_linkhdr + hdrlen)) == 0)
453 		return (ENOBUFS);
454 	/*
455 	 * m_copypack left space for our hdr; use it.
456 	 */
457 	m->m_len += hdrlen;
458 	m->m_data -= hdrlen;
459 #else
460 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
461 	if (__predict_false(m == NULL))
462 		return (ENOBUFS);
463 	MCLAIM(m, &tcp_tx_mowner);
464 
465 	/*
466 	 * XXX Because other code assumes headers will fit in
467 	 * XXX one header mbuf.
468 	 *
469 	 * (This code should almost *never* be run.)
470 	 */
471 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
472 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
473 		MCLGET(m, M_DONTWAIT);
474 		if ((m->m_flags & M_EXT) == 0) {
475 			m_freem(m);
476 			return (ENOBUFS);
477 		}
478 	}
479 
480 	m->m_data += max_linkhdr;
481 	m->m_len = hdrlen;
482 
483 	/*
484 	 * To avoid traversing the whole sb_mb chain for correct
485 	 * data to send, remember last sent mbuf, its offset and
486 	 * the sent size.  When called the next time, see if the
487 	 * data to send is directly following the previous transfer.
488 	 * This is important for large TCP windows.
489 	 */
490 	if (off == 0 || tp->t_lastm == NULL ||
491 	    (tp->t_lastoff + tp->t_lastlen) != off) {
492 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
493 		/*
494 		 * Either a new packet or a retransmit.
495 		 * Start from the beginning.
496 		 */
497 		tp->t_lastm = so->so_snd.sb_mb;
498 		tp->t_inoff = off;
499 	} else {
500 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
501 		tp->t_inoff += tp->t_lastlen;
502 	}
503 
504 	/* Traverse forward to next packet */
505 	while (tp->t_inoff > 0) {
506 		if (tp->t_lastm == NULL)
507 			panic("tp->t_lastm == NULL");
508 		if (tp->t_inoff < tp->t_lastm->m_len)
509 			break;
510 		tp->t_inoff -= tp->t_lastm->m_len;
511 		tp->t_lastm = tp->t_lastm->m_next;
512 	}
513 
514 	tp->t_lastoff = off;
515 	tp->t_lastlen = len;
516 	m0 = tp->t_lastm;
517 	off = tp->t_inoff;
518 
519 	if (len <= M_TRAILINGSPACE(m)) {
520 		m_copydata(m0, off, (int) len, mtod(m, char *) + hdrlen);
521 		m->m_len += len;
522 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
523 	} else {
524 		m->m_next = m_copym(m0, off, (int) len, M_DONTWAIT);
525 		if (m->m_next == NULL) {
526 			m_freem(m);
527 			return (ENOBUFS);
528 		}
529 #ifdef TCP_OUTPUT_COUNTERS
530 		if (m->m_next->m_flags & M_EXT)
531 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
532 		else
533 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
534 #endif /* TCP_OUTPUT_COUNTERS */
535 	}
536 #endif
537 
538 	*mp = m;
539 	return (0);
540 }
541 
542 /*
543  * Tcp output routine: figure out what should be sent and send it.
544  */
545 int
546 tcp_output(struct tcpcb *tp)
547 {
548 	struct rtentry *rt;
549 	struct socket *so;
550 	struct route *ro;
551 	long len, win;
552 	int off, flags, error;
553 	struct mbuf *m;
554 	struct ip *ip;
555 #ifdef INET6
556 	struct ip6_hdr *ip6;
557 #endif
558 	struct tcphdr *th;
559 	u_char opt[MAX_TCPOPTLEN];
560 #define OPT_FITS(more)	((optlen + (more)) < sizeof(opt))
561 	unsigned optlen, hdrlen, packetlen;
562 	unsigned int sack_numblks;
563 	int idle, sendalot, txsegsize, rxsegsize;
564 	int txsegsize_nosack;
565 	int maxburst = TCP_MAXBURST;
566 	int af;		/* address family on the wire */
567 	int iphdrlen;
568 	int has_tso4, has_tso6;
569 	int has_tso, use_tso;
570 	bool alwaysfrag;
571 	int sack_rxmit;
572 	int sack_bytes_rxmt;
573 	int ecn_tos;
574 	struct sackhole *p;
575 #ifdef TCP_SIGNATURE
576 	int sigoff = 0;
577 #endif
578 	uint64_t *tcps;
579 
580 #ifdef DIAGNOSTIC
581 	if (tp->t_inpcb && tp->t_in6pcb)
582 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
583 #endif
584 	so = NULL;
585 	ro = NULL;
586 	if (tp->t_inpcb) {
587 		so = tp->t_inpcb->inp_socket;
588 		ro = &tp->t_inpcb->inp_route;
589 	}
590 #ifdef INET6
591 	else if (tp->t_in6pcb) {
592 		so = tp->t_in6pcb->in6p_socket;
593 		ro = &tp->t_in6pcb->in6p_route;
594 	}
595 #endif
596 
597 	switch (af = tp->t_family) {
598 #ifdef INET
599 	case AF_INET:
600 		if (tp->t_inpcb)
601 			break;
602 #ifdef INET6
603 		/* mapped addr case */
604 		if (tp->t_in6pcb)
605 			break;
606 #endif
607 		return (EINVAL);
608 #endif
609 #ifdef INET6
610 	case AF_INET6:
611 		if (tp->t_in6pcb)
612 			break;
613 		return (EINVAL);
614 #endif
615 	default:
616 		return (EAFNOSUPPORT);
617 	}
618 
619 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
620 		return (EMSGSIZE);
621 
622 	idle = (tp->snd_max == tp->snd_una);
623 
624 	/*
625 	 * Determine if we can use TCP segmentation offload:
626 	 * - If we're using IPv4
627 	 * - If there is not an IPsec policy that prevents it
628 	 * - If the interface can do it
629 	 */
630 	has_tso4 = has_tso6 = false;
631 #if defined(INET)
632 	has_tso4 = tp->t_inpcb != NULL &&
633 #if defined(IPSEC)
634 	    (!ipsec_used || IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
635 	    IPSEC_DIR_OUTBOUND)) &&
636 #endif
637 	    (rt = rtcache_validate(&tp->t_inpcb->inp_route)) != NULL &&
638 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv4) != 0;
639 #endif /* defined(INET) */
640 #if defined(INET6)
641 	has_tso6 = tp->t_in6pcb != NULL &&
642 #if defined(IPSEC)
643 	    (!ipsec_used || IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
644 	    IPSEC_DIR_OUTBOUND)) &&
645 #endif
646 	    (rt = rtcache_validate(&tp->t_in6pcb->in6p_route)) != NULL &&
647 	    (rt->rt_ifp->if_capenable & IFCAP_TSOv6) != 0;
648 #endif /* defined(INET6) */
649 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
650 
651 	/*
652 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
653 	 *
654 	 *	Optionally, a TCP MAY set the restart window to the
655 	 *	minimum of the value used for the initial window and
656 	 *	the current value of cwnd (in other words, using a
657 	 *	larger value for the restart window should never increase
658 	 *	the size of cwnd).
659 	 */
660 	if (tcp_cwm) {
661 		/*
662 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
663 		 * Count the number of packets currently pending
664 		 * acknowledgement, and limit our congestion window
665 		 * to a pre-determined allowed burst size plus that count.
666 		 * This prevents bursting once all pending packets have
667 		 * been acknowledged (i.e. transmission is idle).
668 		 *
669 		 * XXX Link this to Initial Window?
670 		 */
671 		tp->snd_cwnd = min(tp->snd_cwnd,
672 		    (tcp_cwm_burstsize * txsegsize) +
673 		    (tp->snd_nxt - tp->snd_una));
674 	} else {
675 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
676 			/*
677 			 * We have been idle for "a while" and no acks are
678 			 * expected to clock out any data we send --
679 			 * slow start to get ack "clock" running again.
680 			 */
681 			int ss = tcp_init_win;
682 #ifdef INET
683 			if (tp->t_inpcb &&
684 			    in_localaddr(tp->t_inpcb->inp_faddr))
685 				ss = tcp_init_win_local;
686 #endif
687 #ifdef INET6
688 			if (tp->t_in6pcb &&
689 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
690 				ss = tcp_init_win_local;
691 #endif
692 			tp->snd_cwnd = min(tp->snd_cwnd,
693 			    TCP_INITIAL_WINDOW(ss, txsegsize));
694 		}
695 	}
696 
697 	txsegsize_nosack = txsegsize;
698 again:
699 	ecn_tos = 0;
700 	use_tso = has_tso;
701 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
702 		/* don't duplicate CWR/ECE. */
703 		use_tso = 0;
704 	}
705 	TCP_REASS_LOCK(tp);
706 	sack_numblks = tcp_sack_numblks(tp);
707 	if (sack_numblks) {
708 		int sackoptlen;
709 
710 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
711 		if (sackoptlen > txsegsize_nosack) {
712 			sack_numblks = 0; /* give up SACK */
713 			txsegsize = txsegsize_nosack;
714 		} else {
715 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
716 				/* don't duplicate D-SACK. */
717 				use_tso = 0;
718 			}
719 			txsegsize = txsegsize_nosack - sackoptlen;
720 		}
721 	} else {
722 		txsegsize = txsegsize_nosack;
723 	}
724 
725 	/*
726 	 * Determine length of data that should be transmitted, and
727 	 * flags that should be used.  If there is some data or critical
728 	 * controls (SYN, RST) to send, then transmit; otherwise,
729 	 * investigate further.
730 	 *
731 	 * Readjust SACK information to avoid resending duplicate data.
732 	 */
733 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
734 		tcp_sack_adjust(tp);
735 	sendalot = 0;
736 	off = tp->snd_nxt - tp->snd_una;
737 	win = min(tp->snd_wnd, tp->snd_cwnd);
738 
739 	flags = tcp_outflags[tp->t_state];
740 
741 	/*
742 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
743 	 * to send out new data (when sendalot is 1), bypass this function.
744 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
745 	 * we're replacing a (future) new transmission with a retransmission
746 	 * now, and we previously incremented snd_cwnd in tcp_input().
747 	 */
748 	/*
749 	 * Still in sack recovery , reset rxmit flag to zero.
750 	 */
751 	sack_rxmit = 0;
752 	sack_bytes_rxmt = 0;
753 	len = 0;
754 	p = NULL;
755 	do {
756 		long cwin;
757 		if (!TCP_SACK_ENABLED(tp))
758 			break;
759 		if (tp->t_partialacks < 0)
760 			break;
761 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
762 		if (p == NULL)
763 			break;
764 
765 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
766 		if (cwin < 0)
767 			cwin = 0;
768 		/* Do not retransmit SACK segments beyond snd_recover */
769 		if (SEQ_GT(p->end, tp->snd_recover)) {
770 			/*
771 			 * (At least) part of sack hole extends beyond
772 			 * snd_recover. Check to see if we can rexmit data
773 			 * for this hole.
774 			 */
775 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
776 				/*
777 				 * Can't rexmit any more data for this hole.
778 				 * That data will be rexmitted in the next
779 				 * sack recovery episode, when snd_recover
780 				 * moves past p->rxmit.
781 				 */
782 				p = NULL;
783 				break;
784 			}
785 			/* Can rexmit part of the current hole */
786 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
787 		} else
788 			len = ((long)ulmin(cwin, p->end - p->rxmit));
789 		off = p->rxmit - tp->snd_una;
790 		if (off + len > so->so_snd.sb_cc) {
791 			/* 1 for TH_FIN */
792 			KASSERT(off + len == so->so_snd.sb_cc + 1);
793 			KASSERT(p->rxmit + len == tp->snd_max);
794 			len = so->so_snd.sb_cc - off;
795 		}
796 		if (len > 0) {
797 			sack_rxmit = 1;
798 			sendalot = 1;
799 		}
800 	} while (/*CONSTCOND*/0);
801 
802 	/*
803 	 * If in persist timeout with window of 0, send 1 byte.
804 	 * Otherwise, if window is small but nonzero
805 	 * and timer expired, we will send what we can
806 	 * and go to transmit state.
807 	 */
808 	if (tp->t_force) {
809 		if (win == 0) {
810 			/*
811 			 * If we still have some data to send, then
812 			 * clear the FIN bit.  Usually this would
813 			 * happen below when it realizes that we
814 			 * aren't sending all the data.  However,
815 			 * if we have exactly 1 byte of unset data,
816 			 * then it won't clear the FIN bit below,
817 			 * and if we are in persist state, we wind
818 			 * up sending the packet without recording
819 			 * that we sent the FIN bit.
820 			 *
821 			 * We can't just blindly clear the FIN bit,
822 			 * because if we don't have any more data
823 			 * to send then the probe will be the FIN
824 			 * itself.
825 			 */
826 			if (off < so->so_snd.sb_cc)
827 				flags &= ~TH_FIN;
828 			win = 1;
829 		} else {
830 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
831 			tp->t_rxtshift = 0;
832 		}
833 	}
834 
835 	if (sack_rxmit == 0) {
836 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
837 			long cwin;
838 
839 			/*
840 			 * We are inside of a SACK recovery episode and are
841 			 * sending new data, having retransmitted all the
842 			 * data possible in the scoreboard.
843 			 */
844 			if (tp->snd_wnd < so->so_snd.sb_cc) {
845 				len = tp->snd_wnd - off;
846 				flags &= ~TH_FIN;
847 			} else {
848 				len = so->so_snd.sb_cc - off;
849 			}
850 
851 			/*
852 			 * From FreeBSD:
853 			 *  Don't remove this (len > 0) check !
854 			 *  We explicitly check for len > 0 here (although it
855 			 *  isn't really necessary), to work around a gcc
856 			 *  optimization issue - to force gcc to compute
857 			 *  len above. Without this check, the computation
858 			 *  of len is bungled by the optimizer.
859 			 */
860 			if (len > 0) {
861 				cwin = tp->snd_cwnd -
862 				    (tp->snd_nxt - tp->sack_newdata) -
863 				    sack_bytes_rxmt;
864 				if (cwin < 0)
865 					cwin = 0;
866 				if (cwin < len) {
867 					len = cwin;
868 					flags &= ~TH_FIN;
869 				}
870 			}
871 		} else if (win < so->so_snd.sb_cc) {
872 			len = win - off;
873 			flags &= ~TH_FIN;
874 		} else {
875 			len = so->so_snd.sb_cc - off;
876 		}
877 	}
878 
879 	if (len < 0) {
880 		/*
881 		 * If FIN has been sent but not acked,
882 		 * but we haven't been called to retransmit,
883 		 * len will be -1.  Otherwise, window shrank
884 		 * after we sent into it.  If window shrank to 0,
885 		 * cancel pending retransmit, pull snd_nxt back
886 		 * to (closed) window, and set the persist timer
887 		 * if it isn't already going.  If the window didn't
888 		 * close completely, just wait for an ACK.
889 		 *
890 		 * If we have a pending FIN, either it has already been
891 		 * transmitted or it is outside the window, so drop it.
892 		 * If the FIN has been transmitted, but this is not a
893 		 * retransmission, then len must be -1.  Therefore we also
894 		 * prevent here the sending of `gratuitous FINs'.  This
895 		 * eliminates the need to check for that case below (e.g.
896 		 * to back up snd_nxt before the FIN so that the sequence
897 		 * number is correct).
898 		 */
899 		len = 0;
900 		flags &= ~TH_FIN;
901 		if (win == 0) {
902 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
903 			tp->t_rxtshift = 0;
904 			tp->snd_nxt = tp->snd_una;
905 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
906 				tcp_setpersist(tp);
907 		}
908 	}
909 
910 	/*
911 	 * Automatic sizing enables the performance of large buffers
912 	 * and most of the efficiency of small ones by only allocating
913 	 * space when it is needed.
914 	 *
915 	 * The criteria to step up the send buffer one notch are:
916 	 *  1. receive window of remote host is larger than send buffer
917 	 *     (with a fudge factor of 5/4th);
918 	 *  2. send buffer is filled to 7/8th with data (so we actually
919 	 *     have data to make use of it);
920 	 *  3. send buffer fill has not hit maximal automatic size;
921 	 *  4. our send window (slow start and cogestion controlled) is
922 	 *     larger than sent but unacknowledged data in send buffer.
923 	 *
924 	 * The remote host receive window scaling factor may limit the
925 	 * growing of the send buffer before it reaches its allowed
926 	 * maximum.
927 	 *
928 	 * It scales directly with slow start or congestion window
929 	 * and does at most one step per received ACK.  This fast
930 	 * scaling has the drawback of growing the send buffer beyond
931 	 * what is strictly necessary to make full use of a given
932 	 * delay*bandwith product.  However testing has shown this not
933 	 * to be much of an problem.  At worst we are trading wasting
934 	 * of available bandwith (the non-use of it) for wasting some
935 	 * socket buffer memory.
936 	 *
937 	 * TODO: Shrink send buffer during idle periods together
938 	 * with congestion window.  Requires another timer.
939 	 */
940 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
941 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
942 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
943 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
944 		    win >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
945 			if (!sbreserve(&so->so_snd,
946 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
947 			     tcp_autosndbuf_max), so))
948 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
949 		}
950 	}
951 
952 	if (len > txsegsize) {
953 		if (use_tso) {
954 			/*
955 			 * Truncate TSO transfers to IP_MAXPACKET, and make
956 			 * sure that we send equal size transfers down the
957 			 * stack (rather than big-small-big-small-...).
958 			 */
959 #ifdef INET6
960 			CTASSERT(IPV6_MAXPACKET == IP_MAXPACKET);
961 #endif
962 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
963 			if (len <= txsegsize) {
964 				use_tso = 0;
965 			}
966 		} else
967 			len = txsegsize;
968 		flags &= ~TH_FIN;
969 		sendalot = 1;
970 	} else
971 		use_tso = 0;
972 	if (sack_rxmit) {
973 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
974 			flags &= ~TH_FIN;
975 	}
976 
977 	win = sbspace(&so->so_rcv);
978 
979 	/*
980 	 * Sender silly window avoidance.  If connection is idle
981 	 * and can send all data, a maximum segment,
982 	 * at least a maximum default-size segment do it,
983 	 * or are forced, do it; otherwise don't bother.
984 	 * If peer's buffer is tiny, then send
985 	 * when window is at least half open.
986 	 * If retransmitting (possibly after persist timer forced us
987 	 * to send into a small window), then must resend.
988 	 */
989 	if (len) {
990 		if (len >= txsegsize)
991 			goto send;
992 		if ((so->so_state & SS_MORETOCOME) == 0 &&
993 		    ((idle || tp->t_flags & TF_NODELAY) &&
994 		     len + off >= so->so_snd.sb_cc))
995 			goto send;
996 		if (tp->t_force)
997 			goto send;
998 		if (len >= tp->max_sndwnd / 2)
999 			goto send;
1000 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1001 			goto send;
1002 		if (sack_rxmit)
1003 			goto send;
1004 	}
1005 
1006 	/*
1007 	 * Compare available window to amount of window known to peer
1008 	 * (as advertised window less next expected input).  If the
1009 	 * difference is at least twice the size of the largest segment
1010 	 * we expect to receive (i.e. two segments) or at least 50% of
1011 	 * the maximum possible window, then want to send a window update
1012 	 * to peer.
1013 	 */
1014 	if (win > 0) {
1015 		/*
1016 		 * "adv" is the amount we can increase the window,
1017 		 * taking into account that we are limited by
1018 		 * TCP_MAXWIN << tp->rcv_scale.
1019 		 */
1020 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
1021 			(tp->rcv_adv - tp->rcv_nxt);
1022 
1023 		if (adv >= (long) (2 * rxsegsize))
1024 			goto send;
1025 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
1026 			goto send;
1027 	}
1028 
1029 	/*
1030 	 * Send if we owe peer an ACK.
1031 	 */
1032 	if (tp->t_flags & TF_ACKNOW)
1033 		goto send;
1034 	if (flags & (TH_SYN|TH_FIN|TH_RST))
1035 		goto send;
1036 	if (SEQ_GT(tp->snd_up, tp->snd_una))
1037 		goto send;
1038 	/*
1039 	 * In SACK, it is possible for tcp_output to fail to send a segment
1040 	 * after the retransmission timer has been turned off.  Make sure
1041 	 * that the retransmission timer is set.
1042 	 */
1043 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1044 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1045 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1046 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1047 		goto just_return;
1048 	}
1049 
1050 	/*
1051 	 * TCP window updates are not reliable, rather a polling protocol
1052 	 * using ``persist'' packets is used to insure receipt of window
1053 	 * updates.  The three ``states'' for the output side are:
1054 	 *	idle			not doing retransmits or persists
1055 	 *	persisting		to move a small or zero window
1056 	 *	(re)transmitting	and thereby not persisting
1057 	 *
1058 	 * tp->t_timer[TCPT_PERSIST]
1059 	 *	is set when we are in persist state.
1060 	 * tp->t_force
1061 	 *	is set when we are called to send a persist packet.
1062 	 * tp->t_timer[TCPT_REXMT]
1063 	 *	is set when we are retransmitting
1064 	 * The output side is idle when both timers are zero.
1065 	 *
1066 	 * If send window is too small, there is data to transmit, and no
1067 	 * retransmit or persist is pending, then go to persist state.
1068 	 * If nothing happens soon, send when timer expires:
1069 	 * if window is nonzero, transmit what we can,
1070 	 * otherwise force out a byte.
1071 	 */
1072 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1073 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1074 		tp->t_rxtshift = 0;
1075 		tcp_setpersist(tp);
1076 	}
1077 
1078 	/*
1079 	 * No reason to send a segment, just return.
1080 	 */
1081 just_return:
1082 	TCP_REASS_UNLOCK(tp);
1083 	return (0);
1084 
1085 send:
1086 	/*
1087 	 * Before ESTABLISHED, force sending of initial options
1088 	 * unless TCP set not to do any options.
1089 	 * NOTE: we assume that the IP/TCP header plus TCP options
1090 	 * always fit in a single mbuf, leaving room for a maximum
1091 	 * link header, i.e.
1092 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1093 	 */
1094 	optlen = 0;
1095 	switch (af) {
1096 #ifdef INET
1097 	case AF_INET:
1098 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1099 		break;
1100 #endif
1101 #ifdef INET6
1102 	case AF_INET6:
1103 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1104 		break;
1105 #endif
1106 	default:	/*pacify gcc*/
1107 		iphdrlen = 0;
1108 		break;
1109 	}
1110 	hdrlen = iphdrlen;
1111 	if (flags & TH_SYN) {
1112 		struct rtentry *synrt;
1113 
1114 		synrt = NULL;
1115 #ifdef INET
1116 		if (tp->t_inpcb)
1117 			synrt = in_pcbrtentry(tp->t_inpcb);
1118 #endif
1119 #ifdef INET6
1120 		if (tp->t_in6pcb)
1121 			synrt = in6_pcbrtentry(tp->t_in6pcb);
1122 #endif
1123 
1124 		tp->snd_nxt = tp->iss;
1125 		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
1126 						    synrt->rt_ifp : NULL, af);
1127 		if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(4)) {
1128 			opt[0] = TCPOPT_MAXSEG;
1129 			opt[1] = 4;
1130 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1131 			opt[3] = tp->t_ourmss & 0xff;
1132 			optlen = 4;
1133 
1134 			if ((tp->t_flags & TF_REQ_SCALE) &&
1135 			    ((flags & TH_ACK) == 0 ||
1136 			    (tp->t_flags & TF_RCVD_SCALE)) &&
1137 			    OPT_FITS(4)) {
1138 				*((u_int32_t *) (opt + optlen)) = htonl(
1139 					TCPOPT_NOP << 24 |
1140 					TCPOPT_WINDOW << 16 |
1141 					TCPOLEN_WINDOW << 8 |
1142 					tp->request_r_scale);
1143 				optlen += 4;
1144 			}
1145 			if (tcp_do_sack && OPT_FITS(4)) {
1146 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1147 
1148 				cp[0] = TCPOPT_SACK_PERMITTED;
1149 				cp[1] = 2;
1150 				cp[2] = TCPOPT_NOP;
1151 				cp[3] = TCPOPT_NOP;
1152 				optlen += 4;
1153 			}
1154 		}
1155 	}
1156 
1157 	/*
1158 	 * Send a timestamp and echo-reply if this is a SYN and our side
1159 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1160 	 * and our peer have sent timestamps in our SYN's.
1161 	 */
1162 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1163 	     (flags & TH_RST) == 0 &&
1164 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1165 	     (tp->t_flags & TF_RCVD_TSTMP)) && OPT_FITS(TCPOLEN_TSTAMP_APPA)) {
1166 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1167 
1168 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1169 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1170 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1171 		*lp   = htonl(tp->ts_recent);
1172 		optlen += TCPOLEN_TSTAMP_APPA;
1173 
1174 		/* Set receive buffer autosizing timestamp. */
1175 		if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1176 			tp->rfbuf_ts = TCP_TIMESTAMP(tp);
1177 	}
1178 
1179 	/*
1180 	 * Tack on the SACK block if it is necessary.
1181 	 */
1182 	if (sack_numblks) {
1183 		int sack_len;
1184 		u_char *bp = (u_char *)(opt + optlen);
1185 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1186 		struct ipqent *tiqe;
1187 
1188 		sack_len = sack_numblks * 8 + 2;
1189 		if (OPT_FITS(sack_len + 2)) {
1190 			bp[0] = TCPOPT_NOP;
1191 			bp[1] = TCPOPT_NOP;
1192 			bp[2] = TCPOPT_SACK;
1193 			bp[3] = sack_len;
1194 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1195 				sack_numblks--;
1196 				*lp++ = htonl(tp->rcv_dsack_block.left);
1197 				*lp++ = htonl(tp->rcv_dsack_block.right);
1198 				tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1199 			}
1200 			for (tiqe = TAILQ_FIRST(&tp->timeq);
1201 			    sack_numblks > 0;
1202 			    tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1203 				KASSERT(tiqe != NULL);
1204 				sack_numblks--;
1205 				*lp++ = htonl(tiqe->ipqe_seq);
1206 				*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1207 				    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1208 			}
1209 			optlen += sack_len + 2;
1210 		}
1211 	}
1212 	TCP_REASS_UNLOCK(tp);
1213 
1214 #ifdef TCP_SIGNATURE
1215 	if ((tp->t_flags & TF_SIGNATURE) && OPT_FITS(TCPOLEN_SIGNATURE + 2)) {
1216 		u_char *bp;
1217 		/*
1218 		 * Initialize TCP-MD5 option (RFC2385)
1219 		 */
1220 		bp = (u_char *)opt + optlen;
1221 		*bp++ = TCPOPT_SIGNATURE;
1222 		*bp++ = TCPOLEN_SIGNATURE;
1223 		sigoff = optlen + 2;
1224 		memset(bp, 0, TCP_SIGLEN);
1225 		bp += TCP_SIGLEN;
1226 		optlen += TCPOLEN_SIGNATURE;
1227 		/*
1228 		 * Terminate options list and maintain 32-bit alignment.
1229  		 */
1230 		*bp++ = TCPOPT_NOP;
1231 		*bp++ = TCPOPT_EOL;
1232  		optlen += 2;
1233  	}
1234 #endif /* TCP_SIGNATURE */
1235 
1236 	hdrlen += optlen;
1237 
1238 #ifdef DIAGNOSTIC
1239 	if (!use_tso && len > txsegsize)
1240 		panic("tcp data to be sent is larger than segment");
1241 	else if (use_tso && len > IP_MAXPACKET)
1242 		panic("tcp data to be sent is larger than max TSO size");
1243 	if (max_linkhdr + hdrlen > MCLBYTES)
1244 		panic("tcphdr too big");
1245 #endif
1246 
1247 	/*
1248 	 * Grab a header mbuf, attaching a copy of data to
1249 	 * be transmitted, and initialize the header from
1250 	 * the template for sends on this connection.
1251 	 */
1252 	if (len) {
1253 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1254 		if (error)
1255 			goto out;
1256 		/*
1257 		 * If we're sending everything we've got, set PUSH.
1258 		 * (This will keep happy those implementations which only
1259 		 * give data to the user when a buffer fills or
1260 		 * a PUSH comes in.)
1261 		 */
1262 		if (off + len == so->so_snd.sb_cc)
1263 			flags |= TH_PUSH;
1264 	} else {
1265 		tcps = TCP_STAT_GETREF();
1266 		if (tp->t_flags & TF_ACKNOW)
1267 			tcps[TCP_STAT_SNDACKS]++;
1268 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1269 			tcps[TCP_STAT_SNDCTRL]++;
1270 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1271 			tcps[TCP_STAT_SNDURG]++;
1272 		else
1273 			tcps[TCP_STAT_SNDWINUP]++;
1274 		TCP_STAT_PUTREF();
1275 
1276 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1277 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1278 			MCLGET(m, M_DONTWAIT);
1279 			if ((m->m_flags & M_EXT) == 0) {
1280 				m_freem(m);
1281 				m = NULL;
1282 			}
1283 		}
1284 		if (m == NULL) {
1285 			error = ENOBUFS;
1286 			goto out;
1287 		}
1288 		MCLAIM(m, &tcp_tx_mowner);
1289 		m->m_data += max_linkhdr;
1290 		m->m_len = hdrlen;
1291 	}
1292 	m->m_pkthdr.rcvif = NULL;
1293 	switch (af) {
1294 #ifdef INET
1295 	case AF_INET:
1296 		ip = mtod(m, struct ip *);
1297 #ifdef INET6
1298 		ip6 = NULL;
1299 #endif
1300 		th = (struct tcphdr *)(ip + 1);
1301 		break;
1302 #endif
1303 #ifdef INET6
1304 	case AF_INET6:
1305 		ip = NULL;
1306 		ip6 = mtod(m, struct ip6_hdr *);
1307 		th = (struct tcphdr *)(ip6 + 1);
1308 		break;
1309 #endif
1310 	default:	/*pacify gcc*/
1311 		ip = NULL;
1312 #ifdef INET6
1313 		ip6 = NULL;
1314 #endif
1315 		th = NULL;
1316 		break;
1317 	}
1318 	if (tp->t_template == 0)
1319 		panic("tcp_output");
1320 	if (tp->t_template->m_len < iphdrlen)
1321 		panic("tcp_output");
1322 	bcopy(mtod(tp->t_template, void *), mtod(m, void *), iphdrlen);
1323 
1324 	/*
1325 	 * If we are starting a connection, send ECN setup
1326 	 * SYN packet. If we are on a retransmit, we may
1327 	 * resend those bits a number of times as per
1328 	 * RFC 3168.
1329 	 */
1330 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1331 		if (tp->t_flags & TF_SYN_REXMT) {
1332 			if (tp->t_ecn_retries--)
1333 				flags |= TH_ECE|TH_CWR;
1334 		} else {
1335 			flags |= TH_ECE|TH_CWR;
1336 			tp->t_ecn_retries = tcp_ecn_maxretries;
1337 		}
1338 	}
1339 
1340 	if (TCP_ECN_ALLOWED(tp)) {
1341 		/*
1342 		 * If the peer has ECN, mark data packets
1343 		 * ECN capable. Ignore pure ack packets, retransmissions
1344 		 * and window probes.
1345 		 */
1346 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1347 		    !(tp->t_force && len == 1)) {
1348 			ecn_tos = IPTOS_ECN_ECT0;
1349 			TCP_STATINC(TCP_STAT_ECN_ECT);
1350 		}
1351 
1352 		/*
1353 		 * Reply with proper ECN notifications.
1354 		 */
1355 		if (tp->t_flags & TF_ECN_SND_CWR) {
1356 			flags |= TH_CWR;
1357 			tp->t_flags &= ~TF_ECN_SND_CWR;
1358 		}
1359 		if (tp->t_flags & TF_ECN_SND_ECE) {
1360 			flags |= TH_ECE;
1361 		}
1362 	}
1363 
1364 
1365 	/*
1366 	 * If we are doing retransmissions, then snd_nxt will
1367 	 * not reflect the first unsent octet.  For ACK only
1368 	 * packets, we do not want the sequence number of the
1369 	 * retransmitted packet, we want the sequence number
1370 	 * of the next unsent octet.  So, if there is no data
1371 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1372 	 * when filling in ti_seq.  But if we are in persist
1373 	 * state, snd_max might reflect one byte beyond the
1374 	 * right edge of the window, so use snd_nxt in that
1375 	 * case, since we know we aren't doing a retransmission.
1376 	 * (retransmit and persist are mutually exclusive...)
1377 	 */
1378 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1379 		th->th_seq = htonl(p->rxmit);
1380 		p->rxmit += len;
1381 	} else {
1382 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1383 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1384 			th->th_seq = htonl(tp->snd_nxt);
1385 		else
1386 			th->th_seq = htonl(tp->snd_max);
1387 	}
1388 	th->th_ack = htonl(tp->rcv_nxt);
1389 	if (optlen) {
1390 		bcopy((void *)opt, (void *)(th + 1), optlen);
1391 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1392 	}
1393 	th->th_flags = flags;
1394 	/*
1395 	 * Calculate receive window.  Don't shrink window,
1396 	 * but avoid silly window syndrome.
1397 	 */
1398 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1399 		win = 0;
1400 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1401 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1402 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1403 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1404 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1405 	if (th->th_win == 0) {
1406 		tp->t_sndzerowin++;
1407 	}
1408 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1409 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1410 		if (urp > IP_MAXPACKET)
1411 			urp = IP_MAXPACKET;
1412 		th->th_urp = htons((u_int16_t)urp);
1413 		th->th_flags |= TH_URG;
1414 	} else
1415 		/*
1416 		 * If no urgent pointer to send, then we pull
1417 		 * the urgent pointer to the left edge of the send window
1418 		 * so that it doesn't drift into the send window on sequence
1419 		 * number wraparound.
1420 		 */
1421 		tp->snd_up = tp->snd_una;		/* drag it along */
1422 
1423 #ifdef TCP_SIGNATURE
1424 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1425 		struct secasvar *sav;
1426 		u_int8_t *sigp;
1427 
1428 		sav = tcp_signature_getsav(m, th);
1429 
1430 		if (sav == NULL) {
1431 			if (m)
1432 				m_freem(m);
1433 			return (EPERM);
1434 		}
1435 
1436 		m->m_pkthdr.len = hdrlen + len;
1437 		sigp = (char *)th + sizeof(*th) + sigoff;
1438 		tcp_signature(m, th, (char *)th - mtod(m, char *), sav, sigp);
1439 
1440 		key_sa_recordxfer(sav, m);
1441 		KEY_FREESAV(&sav);
1442 	}
1443 #endif
1444 
1445 	/*
1446 	 * Set ourselves up to be checksummed just before the packet
1447 	 * hits the wire.
1448 	 */
1449 	switch (af) {
1450 #ifdef INET
1451 	case AF_INET:
1452 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1453 		if (use_tso) {
1454 			m->m_pkthdr.segsz = txsegsize;
1455 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1456 		} else {
1457 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1458 			if (len + optlen) {
1459 				/* Fixup the pseudo-header checksum. */
1460 				/* XXXJRT Not IP Jumbogram safe. */
1461 				th->th_sum = in_cksum_addword(th->th_sum,
1462 				    htons((u_int16_t) (len + optlen)));
1463 			}
1464 		}
1465 		break;
1466 #endif
1467 #ifdef INET6
1468 	case AF_INET6:
1469 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1470 		if (use_tso) {
1471 			m->m_pkthdr.segsz = txsegsize;
1472 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1473 		} else {
1474 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1475 			if (len + optlen) {
1476 				/* Fixup the pseudo-header checksum. */
1477 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1478 				th->th_sum = in_cksum_addword(th->th_sum,
1479 				    htons((u_int16_t) (len + optlen)));
1480 			}
1481 		}
1482 		break;
1483 #endif
1484 	}
1485 
1486 	/*
1487 	 * In transmit state, time the transmission and arrange for
1488 	 * the retransmit.  In persist state, just set snd_max.
1489 	 */
1490 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1491 		tcp_seq startseq = tp->snd_nxt;
1492 
1493 		/*
1494 		 * Advance snd_nxt over sequence space of this segment.
1495 		 * There are no states in which we send both a SYN and a FIN,
1496 		 * so we collapse the tests for these flags.
1497 		 */
1498 		if (flags & (TH_SYN|TH_FIN))
1499 			tp->snd_nxt++;
1500 		if (sack_rxmit)
1501 			goto timer;
1502 		tp->snd_nxt += len;
1503 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1504 			tp->snd_max = tp->snd_nxt;
1505 			/*
1506 			 * Time this transmission if not a retransmission and
1507 			 * not currently timing anything.
1508 			 */
1509 			if (tp->t_rtttime == 0) {
1510 				tp->t_rtttime = tcp_now;
1511 				tp->t_rtseq = startseq;
1512 				TCP_STATINC(TCP_STAT_SEGSTIMED);
1513 			}
1514 		}
1515 
1516 		/*
1517 		 * Set retransmit timer if not currently set,
1518 		 * and not doing an ack or a keep-alive probe.
1519 		 * Initial value for retransmit timer is smoothed
1520 		 * round-trip time + 2 * round-trip time variance.
1521 		 * Initialize shift counter which is used for backoff
1522 		 * of retransmit time.
1523 		 */
1524 timer:
1525 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1526 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1527 		    tp->snd_nxt != tp->snd_una)) {
1528 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1529 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1530 				tp->t_rxtshift = 0;
1531 			}
1532 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1533 		}
1534 	} else
1535 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1536 			tp->snd_max = tp->snd_nxt + len;
1537 
1538 #ifdef TCP_DEBUG
1539 	/*
1540 	 * Trace.
1541 	 */
1542 	if (so->so_options & SO_DEBUG)
1543 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1544 #endif
1545 
1546 	/*
1547 	 * Fill in IP length and desired time to live and
1548 	 * send to IP level.  There should be a better way
1549 	 * to handle ttl and tos; we could keep them in
1550 	 * the template, but need a way to checksum without them.
1551 	 */
1552 	m->m_pkthdr.len = hdrlen + len;
1553 
1554 	switch (af) {
1555 #ifdef INET
1556 	case AF_INET:
1557 		ip->ip_len = htons(m->m_pkthdr.len);
1558 		packetlen = m->m_pkthdr.len;
1559 		if (tp->t_inpcb) {
1560 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1561 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos | ecn_tos;
1562 		}
1563 #ifdef INET6
1564 		else if (tp->t_in6pcb) {
1565 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1566 			ip->ip_tos = ecn_tos;	/*XXX*/
1567 		}
1568 #endif
1569 		break;
1570 #endif
1571 #ifdef INET6
1572 	case AF_INET6:
1573 		packetlen = m->m_pkthdr.len;
1574 		ip6->ip6_nxt = IPPROTO_TCP;
1575 		if (tp->t_in6pcb) {
1576 			/*
1577 			 * we separately set hoplimit for every segment, since
1578 			 * the user might want to change the value via
1579 			 * setsockopt. Also, desired default hop limit might
1580 			 * be changed via Neighbor Discovery.
1581 			 */
1582 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1583 				(rt = rtcache_validate(ro)) != NULL ? rt->rt_ifp
1584 				                                    : NULL);
1585 		}
1586 		ip6->ip6_flow |= htonl(ecn_tos << 20);
1587 		/* ip6->ip6_flow = ??? (from template) */
1588 		/* ip6_plen will be filled in ip6_output(). */
1589 		break;
1590 #endif
1591 	default:	/*pacify gcc*/
1592 		packetlen = 0;
1593 		break;
1594 	}
1595 
1596 	switch (af) {
1597 #ifdef INET
1598 	case AF_INET:
1599 	    {
1600 		struct mbuf *opts;
1601 
1602 		if (tp->t_inpcb)
1603 			opts = tp->t_inpcb->inp_options;
1604 		else
1605 			opts = NULL;
1606 		error = ip_output(m, opts, ro,
1607 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1608 			(so->so_options & SO_DONTROUTE), NULL, so);
1609 		break;
1610 	    }
1611 #endif
1612 #ifdef INET6
1613 	case AF_INET6:
1614 	    {
1615 		struct ip6_pktopts *opts;
1616 
1617 		if (tp->t_in6pcb)
1618 			opts = tp->t_in6pcb->in6p_outputopts;
1619 		else
1620 			opts = NULL;
1621 		error = ip6_output(m, opts, ro, so->so_options & SO_DONTROUTE,
1622 			NULL, so, NULL);
1623 		break;
1624 	    }
1625 #endif
1626 	default:
1627 		error = EAFNOSUPPORT;
1628 		break;
1629 	}
1630 	if (error) {
1631 out:
1632 		if (error == ENOBUFS) {
1633 			TCP_STATINC(TCP_STAT_SELFQUENCH);
1634 #ifdef INET
1635 			if (tp->t_inpcb)
1636 				tcp_quench(tp->t_inpcb, 0);
1637 #endif
1638 #ifdef INET6
1639 			if (tp->t_in6pcb)
1640 				tcp6_quench(tp->t_in6pcb, 0);
1641 #endif
1642 			error = 0;
1643 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1644 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1645 			tp->t_softerror = error;
1646 			error = 0;
1647 		}
1648 
1649 		/* Back out the seqence number advance. */
1650 		if (sack_rxmit)
1651 			p->rxmit -= len;
1652 
1653 		/* Restart the delayed ACK timer, if necessary. */
1654 		if (tp->t_flags & TF_DELACK)
1655 			TCP_RESTART_DELACK(tp);
1656 
1657 		return (error);
1658 	}
1659 
1660 	if (packetlen > tp->t_pmtud_mtu_sent)
1661 		tp->t_pmtud_mtu_sent = packetlen;
1662 
1663 	tcps = TCP_STAT_GETREF();
1664 	tcps[TCP_STAT_SNDTOTAL]++;
1665 	if (tp->t_flags & TF_DELACK)
1666 		tcps[TCP_STAT_DELACK]++;
1667 	TCP_STAT_PUTREF();
1668 
1669 	/*
1670 	 * Data sent (as far as we can tell).
1671 	 * If this advertises a larger window than any other segment,
1672 	 * then remember the size of the advertised window.
1673 	 * Any pending ACK has now been sent.
1674 	 */
1675 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1676 		tp->rcv_adv = tp->rcv_nxt + win;
1677 	tp->last_ack_sent = tp->rcv_nxt;
1678 	tp->t_flags &= ~TF_ACKNOW;
1679 	TCP_CLEAR_DELACK(tp);
1680 #ifdef DIAGNOSTIC
1681 	if (maxburst < 0)
1682 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1683 #endif
1684 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1685 		goto again;
1686 	return (0);
1687 }
1688 
1689 void
1690 tcp_setpersist(struct tcpcb *tp)
1691 {
1692 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1693 	int nticks;
1694 
1695 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1696 		panic("tcp_output REXMT");
1697 	/*
1698 	 * Start/restart persistance timer.
1699 	 */
1700 	if (t < tp->t_rttmin)
1701 		t = tp->t_rttmin;
1702 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1703 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1704 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1705 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1706 		tp->t_rxtshift++;
1707 }
1708