xref: /dflybsd-src/sys/netinet/tcp_output.c (revision 42a4cbd9acaeee7f105704842b5ba26eed0ac151)
1 /*
2  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
36  *
37  * License terms: all terms for the DragonFly license above plus the following:
38  *
39  * 4. All advertising materials mentioning features or use of this software
40  *    must display the following acknowledgement:
41  *
42  *	This product includes software developed by Jeffrey M. Hsu
43  *	for the DragonFly Project.
44  *
45  *    This requirement may be waived with permission from Jeffrey Hsu.
46  *    This requirement will sunset and may be removed on July 8 2005,
47  *    after which the standard DragonFly license (as shown above) will
48  *    apply.
49  */
50 
51 /*
52  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
53  *	The Regents of the University of California.  All rights reserved.
54  *
55  * Redistribution and use in source and binary forms, with or without
56  * modification, are permitted provided that the following conditions
57  * are met:
58  * 1. Redistributions of source code must retain the above copyright
59  *    notice, this list of conditions and the following disclaimer.
60  * 2. Redistributions in binary form must reproduce the above copyright
61  *    notice, this list of conditions and the following disclaimer in the
62  *    documentation and/or other materials provided with the distribution.
63  * 3. All advertising materials mentioning features or use of this software
64  *    must display the following acknowledgement:
65  *	This product includes software developed by the University of
66  *	California, Berkeley and its contributors.
67  * 4. Neither the name of the University nor the names of its contributors
68  *    may be used to endorse or promote products derived from this software
69  *    without specific prior written permission.
70  *
71  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
72  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
73  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
74  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
75  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
76  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
77  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
78  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
79  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
80  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
81  * SUCH DAMAGE.
82  *
83  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
84  * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
85  * $DragonFly: src/sys/netinet/tcp_output.c,v 1.15 2004/07/08 22:07:35 hsu Exp $
86  */
87 
88 #include "opt_inet6.h"
89 #include "opt_ipsec.h"
90 #include "opt_tcpdebug.h"
91 
92 #include <sys/param.h>
93 #include <sys/systm.h>
94 #include <sys/kernel.h>
95 #include <sys/sysctl.h>
96 #include <sys/mbuf.h>
97 #include <sys/domain.h>
98 #include <sys/protosw.h>
99 #include <sys/socket.h>
100 #include <sys/socketvar.h>
101 #include <sys/in_cksum.h>
102 #include <sys/thread.h>
103 #include <sys/globaldata.h>
104 
105 #include <net/route.h>
106 
107 #include <netinet/in.h>
108 #include <netinet/in_systm.h>
109 #include <netinet/ip.h>
110 #include <netinet/in_pcb.h>
111 #include <netinet/ip_var.h>
112 #include <netinet6/in6_pcb.h>
113 #include <netinet/ip6.h>
114 #include <netinet6/ip6_var.h>
115 #include <netinet/tcp.h>
116 #define	TCPOUTFLAGS
117 #include <netinet/tcp_fsm.h>
118 #include <netinet/tcp_seq.h>
119 #include <netinet/tcp_timer.h>
120 #include <netinet/tcp_var.h>
121 #include <netinet/tcpip.h>
122 #ifdef TCPDEBUG
123 #include <netinet/tcp_debug.h>
124 #endif
125 
126 #ifdef IPSEC
127 #include <netinet6/ipsec.h>
128 #endif /*IPSEC*/
129 
130 #ifdef FAST_IPSEC
131 #include <netipsec/ipsec.h>
132 #define	IPSEC
133 #endif /*FAST_IPSEC*/
134 
135 #ifdef notyet
136 extern struct mbuf *m_copypack();
137 #endif
138 
139 int path_mtu_discovery = 1;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
141 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
142 
143 int ss_fltsz = 1;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
145 	&ss_fltsz, 1, "Slow start flight size");
146 
147 int ss_fltsz_local = 4;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
149 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
150 
151 int	tcp_do_newreno = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
153 	0, "Enable NewReno Algorithms");
154 
155 /*
156  * Tcp output routine: figure out what should be sent and send it.
157  */
158 int
159 tcp_output(tp)
160 	struct tcpcb *tp;
161 {
162 	struct inpcb * const inp = tp->t_inpcb;
163 	struct socket *so = inp->inp_socket;
164 	long len, recvwin, sendwin;
165 	int off, flags, error;
166 	struct mbuf *m;
167 	struct ip *ip = NULL;
168 	struct ipovly *ipov = NULL;
169 	struct tcphdr *th;
170 	u_char opt[TCP_MAXOLEN];
171 	unsigned ipoptlen, optlen, hdrlen;
172 	int idle, sendalot;
173 	struct ip6_hdr *ip6 = NULL;
174 #ifdef INET6
175 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
176 #else
177 	const boolean_t isipv6 = FALSE;
178 #endif
179 	struct rmxp_tao *taop;
180 
181 	/*
182 	 * Determine length of data that should be transmitted,
183 	 * and flags that will be used.
184 	 * If there is some data or critical controls (SYN, RST)
185 	 * to send, then transmit; otherwise, investigate further.
186 	 */
187 	if ((tp->snd_max == tp->snd_una) &&
188 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
189 		/*
190 		 * We have been idle for "a while" and no acks are
191 		 * expected to clock out any data we send --
192 		 * slow start to get ack "clock" running again.
193 		 *
194 		 * Set the slow-start flight size depending on whether
195 		 * this is a local network or not.
196 		 */
197 		if ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
198 		    (!isipv6 && in_localaddr(inp->inp_faddr)))
199 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
200 		else
201 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
202 	}
203 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
204 	if (idle && (tp->t_flags & TF_MORETOCOME))
205 		tp->t_flags |= TF_LASTIDLE;
206 	else
207 		tp->t_flags &= ~TF_LASTIDLE;
208 
209 again:
210 	sendalot = 0;
211 	off = tp->snd_nxt - tp->snd_una;
212 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
213 	sendwin = min(sendwin, tp->snd_bwnd);
214 
215 	flags = tcp_outflags[tp->t_state];
216 	/*
217 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
218 	 * state flags.
219 	 */
220 	if (tp->t_flags & TF_NEEDFIN)
221 		flags |= TH_FIN;
222 	if (tp->t_flags & TF_NEEDSYN)
223 		flags |= TH_SYN;
224 
225 	/*
226 	 * If in persist timeout with window of 0, send 1 byte.
227 	 * Otherwise, if window is small but nonzero
228 	 * and timer expired, we will send what we can
229 	 * and go to transmit state.
230 	 */
231 	if (tp->t_flags & TF_FORCE) {
232 		if (sendwin == 0) {
233 			/*
234 			 * If we still have some data to send, then
235 			 * clear the FIN bit.  Usually this would
236 			 * happen below when it realizes that we
237 			 * aren't sending all the data.  However,
238 			 * if we have exactly 1 byte of unsent data,
239 			 * then it won't clear the FIN bit below,
240 			 * and if we are in persist state, we wind
241 			 * up sending the packet without recording
242 			 * that we sent the FIN bit.
243 			 *
244 			 * We can't just blindly clear the FIN bit,
245 			 * because if we don't have any more data
246 			 * to send then the probe will be the FIN
247 			 * itself.
248 			 */
249 			if (off < so->so_snd.sb_cc)
250 				flags &= ~TH_FIN;
251 			sendwin = 1;
252 		} else {
253 			callout_stop(tp->tt_persist);
254 			tp->t_rxtshift = 0;
255 		}
256 	}
257 
258 	/*
259 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
260 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
261 	 * a negative length.  This can also occur when TCP opens up
262 	 * its congestion window while receiving additional duplicate
263 	 * acks after fast-retransmit because TCP will reset snd_nxt
264 	 * to snd_max after the fast-retransmit.
265 	 *
266 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
267 	 * be set to snd_una, the offset will be 0, and the length may
268 	 * wind up 0.
269 	 */
270 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
271 
272 	/*
273 	 * Lop off SYN bit if it has already been sent.  However, if this
274 	 * is SYN-SENT state and if segment contains data and if we don't
275 	 * know that foreign host supports TAO, suppress sending segment.
276 	 */
277 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
278 		flags &= ~TH_SYN;
279 		off--, len++;
280 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
281 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
282 		     taop->tao_ccsent == 0))
283 			return 0;
284 	}
285 
286 	/*
287 	 * Be careful not to send data and/or FIN on SYN segments
288 	 * in cases when no CC option will be sent.
289 	 * This measure is needed to prevent interoperability problems
290 	 * with not fully conformant TCP implementations.
291 	 */
292 	if ((flags & TH_SYN) &&
293 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
294 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
295 		len = 0;
296 		flags &= ~TH_FIN;
297 	}
298 
299 	if (len < 0) {
300 		/*
301 		 * If FIN has been sent but not acked,
302 		 * but we haven't been called to retransmit,
303 		 * len will be < 0.  Otherwise, window shrank
304 		 * after we sent into it.  If window shrank to 0,
305 		 * cancel pending retransmit, pull snd_nxt back
306 		 * to (closed) window, and set the persist timer
307 		 * if it isn't already going.  If the window didn't
308 		 * close completely, just wait for an ACK.
309 		 */
310 		len = 0;
311 		if (sendwin == 0) {
312 			callout_stop(tp->tt_rexmt);
313 			tp->t_rxtshift = 0;
314 			tp->snd_nxt = tp->snd_una;
315 			if (!callout_active(tp->tt_persist))
316 				tcp_setpersist(tp);
317 		}
318 	}
319 
320 	/*
321 	 * len will be >= 0 after this point.  Truncate to the maximum
322 	 * segment length and ensure that FIN is removed if the length
323 	 * no longer contains the last data byte.
324 	 */
325 	if (len > tp->t_maxseg) {
326 		len = tp->t_maxseg;
327 		sendalot = 1;
328 	}
329 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
330 		flags &= ~TH_FIN;
331 
332 	recvwin = sbspace(&so->so_rcv);
333 
334 	/*
335 	 * Sender silly window avoidance.   We transmit under the following
336 	 * conditions when len is non-zero:
337 	 *
338 	 *	- We have a full segment
339 	 *	- This is the last buffer in a write()/send() and we are
340 	 *	  either idle or running NODELAY
341 	 *	- we've timed out (e.g. persist timer)
342 	 *	- we have more then 1/2 the maximum send window's worth of
343 	 *	  data (receiver may be limited the window size)
344 	 *	- we need to retransmit
345 	 */
346 	if (len) {
347 		if (len == tp->t_maxseg)
348 			goto send;
349 		/*
350 		 * NOTE! on localhost connections an 'ack' from the remote
351 		 * end may occur synchronously with the output and cause
352 		 * us to flush a buffer queued with moretocome.  XXX
353 		 *
354 		 * note: the len + off check is almost certainly unnecessary.
355 		 */
356 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
357 		    (idle || (tp->t_flags & TF_NODELAY)) &&
358 		    len + off >= so->so_snd.sb_cc &&
359 		    !(tp->t_flags & TF_NOPUSH)) {
360 			goto send;
361 		}
362 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
363 			goto send;
364 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
365 			goto send;
366 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
367 			goto send;
368 	}
369 
370 	/*
371 	 * Compare available window to amount of window
372 	 * known to peer (as advertised window less
373 	 * next expected input).  If the difference is at least two
374 	 * max size segments, or at least 50% of the maximum possible
375 	 * window, then want to send a window update to peer.
376 	 */
377 	if (recvwin > 0) {
378 		/*
379 		 * "adv" is the amount we can increase the window,
380 		 * taking into account that we are limited by
381 		 * TCP_MAXWIN << tp->rcv_scale.
382 		 */
383 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
384 			(tp->rcv_adv - tp->rcv_nxt);
385 
386 		if (adv >= (long) (2 * tp->t_maxseg))
387 			goto send;
388 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
389 			goto send;
390 	}
391 
392 	/*
393 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
394 	 * is also a catch-all for the retransmit timer timeout case.
395 	 */
396 	if (tp->t_flags & TF_ACKNOW)
397 		goto send;
398 	if ((flags & TH_RST) ||
399 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
400 		goto send;
401 	if (SEQ_GT(tp->snd_up, tp->snd_una))
402 		goto send;
403 	/*
404 	 * If our state indicates that FIN should be sent
405 	 * and we have not yet done so, then we need to send.
406 	 */
407 	if (flags & TH_FIN &&
408 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
409 		goto send;
410 
411 	/*
412 	 * TCP window updates are not reliable, rather a polling protocol
413 	 * using ``persist'' packets is used to insure receipt of window
414 	 * updates.  The three ``states'' for the output side are:
415 	 *	idle			not doing retransmits or persists
416 	 *	persisting		to move a small or zero window
417 	 *	(re)transmitting	and thereby not persisting
418 	 *
419 	 * callout_active(tp->tt_persist)
420 	 *	is true when we are in persist state.
421 	 * The TF_FORCE flag in tp->t_flags
422 	 *	is set when we are called to send a persist packet.
423 	 * callout_active(tp->tt_rexmt)
424 	 *	is set when we are retransmitting
425 	 * The output side is idle when both timers are zero.
426 	 *
427 	 * If send window is too small, there is data to transmit, and no
428 	 * retransmit or persist is pending, then go to persist state.
429 	 * If nothing happens soon, send when timer expires:
430 	 * if window is nonzero, transmit what we can,
431 	 * otherwise force out a byte.
432 	 */
433 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
434 	    !callout_active(tp->tt_persist)) {
435 		tp->t_rxtshift = 0;
436 		tcp_setpersist(tp);
437 	}
438 
439 	/*
440 	 * No reason to send a segment, just return.
441 	 */
442 	return (0);
443 
444 send:
445 	/*
446 	 * Before ESTABLISHED, force sending of initial options
447 	 * unless TCP set not to do any options.
448 	 * NOTE: we assume that the IP/TCP header plus TCP options
449 	 * always fit in a single mbuf, leaving room for a maximum
450 	 * link header, i.e.
451 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
452 	 */
453 	optlen = 0;
454 	if (isipv6)
455 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
456 	else
457 		hdrlen = sizeof(struct tcpiphdr);
458 	if (flags & TH_SYN) {
459 		tp->snd_nxt = tp->iss;
460 		if (!(tp->t_flags & TF_NOOPT)) {
461 			u_short mss;
462 
463 			opt[0] = TCPOPT_MAXSEG;
464 			opt[1] = TCPOLEN_MAXSEG;
465 			mss = htons((u_short) tcp_mssopt(tp));
466 			(void)memcpy(opt + 2, &mss, sizeof(mss));
467 			optlen = TCPOLEN_MAXSEG;
468 
469 			if ((tp->t_flags & TF_REQ_SCALE) &&
470 			    (!(flags & TH_ACK) ||
471 			     (tp->t_flags & TF_RCVD_SCALE))) {
472 				*((u_int32_t *)(opt + optlen)) = htonl(
473 					TCPOPT_NOP << 24 |
474 					TCPOPT_WINDOW << 16 |
475 					TCPOLEN_WINDOW << 8 |
476 					tp->request_r_scale);
477 				optlen += 4;
478 			}
479 		}
480 	}
481 
482 	/*
483 	 * Send a timestamp and echo-reply if this is a SYN and our side
484 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
485 	 * and our peer have sent timestamps in our SYN's.
486 	 */
487 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
488 	    !(flags & TH_RST) &&
489 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
490 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
491 
492 		/* Form timestamp option as shown in appendix A of RFC 1323. */
493 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
494 		*lp++ = htonl(ticks);
495 		*lp   = htonl(tp->ts_recent);
496 		optlen += TCPOLEN_TSTAMP_APPA;
497 	}
498 
499 	/*
500 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
501 	 * options are allowed (!TF_NOOPT) and it's not a RST.
502 	 */
503 	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
504 	     !(flags & TH_RST)) {
505 		switch (flags & (TH_SYN|TH_ACK)) {
506 		/*
507 		 * This is a normal ACK, send CC if we received CC before
508 		 * from our peer.
509 		 */
510 		case TH_ACK:
511 			if (!(tp->t_flags & TF_RCVD_CC))
512 				break;
513 			/*FALLTHROUGH*/
514 
515 		/*
516 		 * We can only get here in T/TCP's SYN_SENT* state, when
517 		 * we're a sending a non-SYN segment without waiting for
518 		 * the ACK of our SYN.  A check above assures that we only
519 		 * do this if our peer understands T/TCP.
520 		 */
521 		case 0:
522 			opt[optlen++] = TCPOPT_NOP;
523 			opt[optlen++] = TCPOPT_NOP;
524 			opt[optlen++] = TCPOPT_CC;
525 			opt[optlen++] = TCPOLEN_CC;
526 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
527 
528 			optlen += 4;
529 			break;
530 
531 		/*
532 		 * This is our initial SYN, check whether we have to use
533 		 * CC or CC.new.
534 		 */
535 		case TH_SYN:
536 			opt[optlen++] = TCPOPT_NOP;
537 			opt[optlen++] = TCPOPT_NOP;
538 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
539 						TCPOPT_CCNEW : TCPOPT_CC;
540 			opt[optlen++] = TCPOLEN_CC;
541 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
542 			optlen += 4;
543 			break;
544 
545 		/*
546 		 * This is a SYN,ACK; send CC and CC.echo if we received
547 		 * CC from our peer.
548 		 */
549 		case (TH_SYN|TH_ACK):
550 			if (tp->t_flags & TF_RCVD_CC) {
551 				opt[optlen++] = TCPOPT_NOP;
552 				opt[optlen++] = TCPOPT_NOP;
553 				opt[optlen++] = TCPOPT_CC;
554 				opt[optlen++] = TCPOLEN_CC;
555 				*(u_int32_t *)&opt[optlen] =
556 					htonl(tp->cc_send);
557 				optlen += 4;
558 				opt[optlen++] = TCPOPT_NOP;
559 				opt[optlen++] = TCPOPT_NOP;
560 				opt[optlen++] = TCPOPT_CCECHO;
561 				opt[optlen++] = TCPOLEN_CC;
562 				*(u_int32_t *)&opt[optlen] =
563 					htonl(tp->cc_recv);
564 				optlen += 4;
565 			}
566 			break;
567 		}
568 	}
569 
570 	hdrlen += optlen;
571 
572 	if (isipv6)
573 		ipoptlen = ip6_optlen(inp);
574 	else {
575 		if (inp->inp_options) {
576 			ipoptlen = inp->inp_options->m_len -
577 			    offsetof(struct ipoption, ipopt_list);
578 		} else {
579 			ipoptlen = 0;
580 		}
581 	}
582 #ifdef IPSEC
583 	ipoptlen += ipsec_hdrsiz_tcp(tp);
584 #endif
585 
586 	/*
587 	 * Adjust data length if insertion of options will
588 	 * bump the packet length beyond the t_maxopd length.
589 	 * Clear the FIN bit because we cut off the tail of
590 	 * the segment.
591 	 */
592 	if (len + optlen + ipoptlen > tp->t_maxopd) {
593 		/*
594 		 * If there is still more to send, don't close the connection.
595 		 */
596 		flags &= ~TH_FIN;
597 		len = tp->t_maxopd - optlen - ipoptlen;
598 		sendalot = 1;
599 	}
600 
601 #ifdef INET6
602 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
603 #else
604 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
605 #endif
606 
607 	/*
608 	 * Grab a header mbuf, attaching a copy of data to
609 	 * be transmitted, and initialize the header from
610 	 * the template for sends on this connection.
611 	 */
612 	if (len) {
613 		if ((tp->t_flags & TF_FORCE) && len == 1)
614 			tcpstat.tcps_sndprobe++;
615 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
616 			tcpstat.tcps_sndrexmitpack++;
617 			tcpstat.tcps_sndrexmitbyte += len;
618 		} else {
619 			tcpstat.tcps_sndpack++;
620 			tcpstat.tcps_sndbyte += len;
621 		}
622 #ifdef notyet
623 		if ((m = m_copypack(so->so_snd.sb_mb, off,
624 		    (int)len, max_linkhdr + hdrlen)) == 0) {
625 			error = ENOBUFS;
626 			goto out;
627 		}
628 		/*
629 		 * m_copypack left space for our hdr; use it.
630 		 */
631 		m->m_len += hdrlen;
632 		m->m_data -= hdrlen;
633 #else
634 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
635 		if (m == NULL) {
636 			error = ENOBUFS;
637 			goto out;
638 		}
639 #ifdef INET6
640 		if (MHLEN < hdrlen + max_linkhdr) {
641 			MCLGET(m, MB_DONTWAIT);
642 			if (!(m->m_flags & M_EXT)) {
643 				m_freem(m);
644 				error = ENOBUFS;
645 				goto out;
646 			}
647 		}
648 #endif
649 		m->m_data += max_linkhdr;
650 		m->m_len = hdrlen;
651 		if (len <= MHLEN - hdrlen - max_linkhdr) {
652 			m_copydata(so->so_snd.sb_mb, off, (int) len,
653 			    mtod(m, caddr_t) + hdrlen);
654 			m->m_len += len;
655 		} else {
656 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
657 			if (m->m_next == 0) {
658 				(void) m_free(m);
659 				error = ENOBUFS;
660 				goto out;
661 			}
662 		}
663 #endif
664 		/*
665 		 * If we're sending everything we've got, set PUSH.
666 		 * (This will keep happy those implementations which only
667 		 * give data to the user when a buffer fills or
668 		 * a PUSH comes in.)
669 		 */
670 		if (off + len == so->so_snd.sb_cc)
671 			flags |= TH_PUSH;
672 	} else {
673 		if (tp->t_flags & TF_ACKNOW)
674 			tcpstat.tcps_sndacks++;
675 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
676 			tcpstat.tcps_sndctrl++;
677 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
678 			tcpstat.tcps_sndurg++;
679 		else
680 			tcpstat.tcps_sndwinup++;
681 
682 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
683 		if (m == NULL) {
684 			error = ENOBUFS;
685 			goto out;
686 		}
687 		if (isipv6 &&
688 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
689 			MH_ALIGN(m, hdrlen);
690 		else
691 			m->m_data += max_linkhdr;
692 		m->m_len = hdrlen;
693 	}
694 	m->m_pkthdr.rcvif = (struct ifnet *)0;
695 	if (isipv6) {
696 		ip6 = mtod(m, struct ip6_hdr *);
697 		th = (struct tcphdr *)(ip6 + 1);
698 		tcp_fillheaders(tp, ip6, th);
699 	} else {
700 		ip = mtod(m, struct ip *);
701 		ipov = (struct ipovly *)ip;
702 		th = (struct tcphdr *)(ip + 1);
703 		/* this picks up the pseudo header (w/o the length) */
704 		tcp_fillheaders(tp, ip, th);
705 	}
706 
707 	/*
708 	 * Fill in fields, remembering maximum advertised
709 	 * window for use in delaying messages about window sizes.
710 	 * If resending a FIN, be sure not to use a new sequence number.
711 	 */
712 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
713 	    tp->snd_nxt == tp->snd_max)
714 		tp->snd_nxt--;
715 	/*
716 	 * If we are doing retransmissions, then snd_nxt will
717 	 * not reflect the first unsent octet.  For ACK only
718 	 * packets, we do not want the sequence number of the
719 	 * retransmitted packet, we want the sequence number
720 	 * of the next unsent octet.  So, if there is no data
721 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
722 	 * when filling in ti_seq.  But if we are in persist
723 	 * state, snd_max might reflect one byte beyond the
724 	 * right edge of the window, so use snd_nxt in that
725 	 * case, since we know we aren't doing a retransmission.
726 	 * (retransmit and persist are mutually exclusive...)
727 	 */
728 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
729 		th->th_seq = htonl(tp->snd_nxt);
730 	else
731 		th->th_seq = htonl(tp->snd_max);
732 	th->th_ack = htonl(tp->rcv_nxt);
733 	if (optlen) {
734 		bcopy(opt, th + 1, optlen);
735 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
736 	}
737 	th->th_flags = flags;
738 	/*
739 	 * Calculate receive window.  Don't shrink window,
740 	 * but avoid silly window syndrome.
741 	 */
742 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
743 	    recvwin < (long)tp->t_maxseg)
744 		recvwin = 0;
745 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
746 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
747 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
748 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
749 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
750 
751 	/*
752 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
753 	 * a 0 window.  This may cause the remote transmitter to stall.  This
754 	 * flag tells soreceive() to disable delayed acknowledgements when
755 	 * draining the buffer.  This can occur if the receiver is attempting
756 	 * to read more data then can be buffered prior to transmitting on
757 	 * the connection.
758 	 */
759 	if (recvwin == 0)
760 		tp->t_flags |= TF_RXWIN0SENT;
761 	else
762 		tp->t_flags &= ~TF_RXWIN0SENT;
763 
764 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
765 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
766 		th->th_flags |= TH_URG;
767 	} else {
768 		/*
769 		 * If no urgent pointer to send, then we pull
770 		 * the urgent pointer to the left edge of the send window
771 		 * so that it doesn't drift into the send window on sequence
772 		 * number wraparound.
773 		 */
774 		tp->snd_up = tp->snd_una;		/* drag it along */
775 	}
776 
777 	/*
778 	 * Put TCP length in extended header, and then
779 	 * checksum extended header and data.
780 	 */
781 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
782 	if (isipv6) {
783 		/*
784 		 * ip6_plen is not need to be filled now, and will be filled
785 		 * in ip6_output.
786 		 */
787 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
788 				       sizeof(struct tcphdr) + optlen + len);
789 	} else {
790 		m->m_pkthdr.csum_flags = CSUM_TCP;
791 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
792 		if (len + optlen)
793 			th->th_sum = in_addword(th->th_sum,
794 						htons((u_short)(optlen + len)));
795 
796 		/* IP version must be set here for ipv4/ipv6 checking later */
797 		KASSERT(ip->ip_v == IPVERSION,
798 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
799 	}
800 
801 	/*
802 	 * In transmit state, time the transmission and arrange for
803 	 * the retransmit.  In persist state, just set snd_max.
804 	 */
805 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
806 		tcp_seq startseq = tp->snd_nxt;
807 
808 		/*
809 		 * Advance snd_nxt over sequence space of this segment.
810 		 */
811 		if (flags & (TH_SYN|TH_FIN)) {
812 			if (flags & TH_SYN)
813 				tp->snd_nxt++;
814 			if (flags & TH_FIN) {
815 				tp->snd_nxt++;
816 				tp->t_flags |= TF_SENTFIN;
817 			}
818 		}
819 		tp->snd_nxt += len;
820 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
821 			tp->snd_max = tp->snd_nxt;
822 			/*
823 			 * Time this transmission if not a retransmission and
824 			 * not currently timing anything.
825 			 */
826 			if (tp->t_rtttime == 0) {
827 				tp->t_rtttime = ticks;
828 				tp->t_rtseq = startseq;
829 				tcpstat.tcps_segstimed++;
830 			}
831 		}
832 
833 		/*
834 		 * Set retransmit timer if not currently set,
835 		 * and not doing a pure ack or a keep-alive probe.
836 		 * Initial value for retransmit timer is smoothed
837 		 * round-trip time + 2 * round-trip time variance.
838 		 * Initialize shift counter which is used for backoff
839 		 * of retransmit time.
840 		 */
841 		if (!callout_active(tp->tt_rexmt) &&
842 		    tp->snd_nxt != tp->snd_una) {
843 			if (callout_active(tp->tt_persist)) {
844 				callout_stop(tp->tt_persist);
845 				tp->t_rxtshift = 0;
846 			}
847 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
848 				      tcp_timer_rexmt, tp);
849 		}
850 	} else {
851 		/*
852 		 * Persist case, update snd_max but since we are in
853 		 * persist mode (no window) we do not update snd_nxt.
854 		 */
855 		int xlen = len;
856 		if (flags & TH_SYN)
857 			++xlen;
858 		if (flags & TH_FIN) {
859 			++xlen;
860 			tp->t_flags |= TF_SENTFIN;
861 		}
862 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
863 			tp->snd_max = tp->snd_nxt + xlen;
864 	}
865 
866 #ifdef TCPDEBUG
867 	/*
868 	 * Trace.
869 	 */
870 	if (so->so_options & SO_DEBUG)
871 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
872 #endif
873 
874 	/*
875 	 * Fill in IP length and desired time to live and
876 	 * send to IP level.  There should be a better way
877 	 * to handle ttl and tos; we could keep them in
878 	 * the template, but need a way to checksum without them.
879 	 */
880 	/*
881 	 * m->m_pkthdr.len should have been set before cksum calcuration,
882 	 * because in6_cksum() need it.
883 	 */
884 	if (isipv6) {
885 		/*
886 		 * we separately set hoplimit for every segment, since the
887 		 * user might want to change the value via setsockopt.
888 		 * Also, desired default hop limit might be changed via
889 		 * Neighbor Discovery.
890 		 */
891 		ip6->ip6_hlim = in6_selecthlim(inp,
892 		    (inp->in6p_route.ro_rt ?
893 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
894 
895 		/* TODO: IPv6 IP6TOS_ECT bit on */
896 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
897 		    (so->so_options & SO_DONTROUTE), NULL, NULL, inp);
898 	} else {
899 		struct rtentry *rt;
900 		ip->ip_len = m->m_pkthdr.len;
901 #ifdef INET6
902 		if (INP_CHECK_SOCKAF(so, AF_INET6))
903 			ip->ip_ttl = in6_selecthlim(inp,
904 			    (inp->in6p_route.ro_rt ?
905 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
906 		else
907 #endif
908 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
909 
910 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
911 		/*
912 		 * See if we should do MTU discovery.
913 		 * We do it only if the following are true:
914 		 *	1) we have a valid route to the destination
915 		 *	2) the MTU is not locked (if it is,
916 		 *	   then discovery has been disabled)
917 		 */
918 		if (path_mtu_discovery &&
919 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
920 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
921 			ip->ip_off |= IP_DF;
922 
923 		error = ip_output(m, inp->inp_options, &inp->inp_route,
924 		    (so->so_options & SO_DONTROUTE), NULL, inp);
925 	}
926 	if (error) {
927 
928 		/*
929 		 * We know that the packet was lost, so back out the
930 		 * sequence number advance, if any.
931 		 */
932 		if (!(tp->t_flags & TF_FORCE) ||
933 		    !callout_active(tp->tt_persist)) {
934 			/*
935 			 * No need to check for TH_FIN here because
936 			 * the TF_SENTFIN flag handles that case.
937 			 */
938 			if (!(flags & TH_SYN))
939 				tp->snd_nxt -= len;
940 		}
941 
942 out:
943 		if (error == ENOBUFS) {
944 			/*
945 			 * If we can't send, make sure there is something
946 			 * to get us going again later.  Persist state
947 			 * is not necessarily right, but it is close enough.
948 			 */
949 			if (!callout_active(tp->tt_rexmt) &&
950 			    !callout_active(tp->tt_persist)) {
951 				tp->t_rxtshift = 0;
952 				tcp_setpersist(tp);
953 			}
954 			tcp_quench(inp, 0);
955 			return (0);
956 		}
957 		if (error == EMSGSIZE) {
958 			/*
959 			 * ip_output() will have already fixed the route
960 			 * for us.  tcp_mtudisc() will, as its last action,
961 			 * initiate retransmission, so it is important to
962 			 * not do so here.
963 			 */
964 			tcp_mtudisc(inp, 0);
965 			return 0;
966 		}
967 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
968 		    TCPS_HAVERCVDSYN(tp->t_state)) {
969 			tp->t_softerror = error;
970 			return (0);
971 		}
972 		return (error);
973 	}
974 	tcpstat.tcps_sndtotal++;
975 
976 	/*
977 	 * Data sent (as far as we can tell).
978 	 * If this advertises a larger window than any other segment,
979 	 * then remember the size of the advertised window.
980 	 * Any pending ACK has now been sent.
981 	 */
982 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
983 		tp->rcv_adv = tp->rcv_nxt + recvwin;
984 	tp->last_ack_sent = tp->rcv_nxt;
985 	tp->t_flags &= ~TF_ACKNOW;
986 	if (tcp_delack_enabled)
987 		callout_stop(tp->tt_delack);
988 	if (sendalot)
989 		goto again;
990 	return (0);
991 }
992 
993 void
994 tcp_setpersist(tp)
995 	struct tcpcb *tp;
996 {
997 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
998 	int tt;
999 
1000 	if (callout_active(tp->tt_rexmt))
1001 		panic("tcp_setpersist: retransmit pending");
1002 	/*
1003 	 * Start/restart persistance timer.
1004 	 */
1005 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1006 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1007 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1008 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1009 		tp->t_rxtshift++;
1010 }
1011