xref: /dflybsd-src/sys/netinet/tcp_output.c (revision 48d201a5a8c1dab4aa7166b0812594c101fc43c3)
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 1.18 2004/08/08 06:33:24 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 static int avoid_pure_win_update = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
153 	&avoid_pure_win_update, 1, "Avoid pure window updates when possible");
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;
173 	boolean_t sendalot;
174 	struct ip6_hdr *ip6 = NULL;
175 #ifdef INET6
176 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
177 #else
178 	const boolean_t isipv6 = FALSE;
179 #endif
180 	struct rmxp_tao *taop;
181 
182 	/*
183 	 * Determine length of data that should be transmitted,
184 	 * and flags that will be used.
185 	 * If there is some data or critical controls (SYN, RST)
186 	 * to send, then transmit; otherwise, investigate further.
187 	 */
188 	if ((tp->snd_max == tp->snd_una) &&
189 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
190 		/*
191 		 * We have been idle for "a while" and no acks are
192 		 * expected to clock out any data we send --
193 		 * slow start to get ack "clock" running again.
194 		 *
195 		 * Set the slow-start flight size depending on whether
196 		 * this is a local network or not.
197 		 */
198 		if ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
199 		    (!isipv6 && in_localaddr(inp->inp_faddr)))
200 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
201 		else
202 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
203 	}
204 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
205 	if (idle && (tp->t_flags & TF_MORETOCOME))
206 		tp->t_flags |= TF_LASTIDLE;
207 	else
208 		tp->t_flags &= ~TF_LASTIDLE;
209 
210 again:
211 	sendalot = FALSE;
212 	off = tp->snd_nxt - tp->snd_una;
213 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
214 	sendwin = min(sendwin, tp->snd_bwnd);
215 
216 	flags = tcp_outflags[tp->t_state];
217 	/*
218 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
219 	 * state flags.
220 	 */
221 	if (tp->t_flags & TF_NEEDFIN)
222 		flags |= TH_FIN;
223 	if (tp->t_flags & TF_NEEDSYN)
224 		flags |= TH_SYN;
225 
226 	/*
227 	 * If in persist timeout with window of 0, send 1 byte.
228 	 * Otherwise, if window is small but nonzero
229 	 * and timer expired, we will send what we can
230 	 * and go to transmit state.
231 	 */
232 	if (tp->t_flags & TF_FORCE) {
233 		if (sendwin == 0) {
234 			/*
235 			 * If we still have some data to send, then
236 			 * clear the FIN bit.  Usually this would
237 			 * happen below when it realizes that we
238 			 * aren't sending all the data.  However,
239 			 * if we have exactly 1 byte of unsent data,
240 			 * then it won't clear the FIN bit below,
241 			 * and if we are in persist state, we wind
242 			 * up sending the packet without recording
243 			 * that we sent the FIN bit.
244 			 *
245 			 * We can't just blindly clear the FIN bit,
246 			 * because if we don't have any more data
247 			 * to send then the probe will be the FIN
248 			 * itself.
249 			 */
250 			if (off < so->so_snd.sb_cc)
251 				flags &= ~TH_FIN;
252 			sendwin = 1;
253 		} else {
254 			callout_stop(tp->tt_persist);
255 			tp->t_rxtshift = 0;
256 		}
257 	}
258 
259 	/*
260 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
261 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
262 	 * a negative length.  This can also occur when TCP opens up
263 	 * its congestion window while receiving additional duplicate
264 	 * acks after fast-retransmit because TCP will reset snd_nxt
265 	 * to snd_max after the fast-retransmit.
266 	 *
267 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
268 	 * be set to snd_una, the offset will be 0, and the length may
269 	 * wind up 0.
270 	 */
271 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
272 
273 	/*
274 	 * Lop off SYN bit if it has already been sent.  However, if this
275 	 * is SYN-SENT state and if segment contains data and if we don't
276 	 * know that foreign host supports TAO, suppress sending segment.
277 	 */
278 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
279 		flags &= ~TH_SYN;
280 		off--, len++;
281 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
282 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
283 		     taop->tao_ccsent == 0))
284 			return 0;
285 	}
286 
287 	/*
288 	 * Be careful not to send data and/or FIN on SYN segments
289 	 * in cases when no CC option will be sent.
290 	 * This measure is needed to prevent interoperability problems
291 	 * with not fully conformant TCP implementations.
292 	 */
293 	if ((flags & TH_SYN) &&
294 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
295 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
296 		len = 0;
297 		flags &= ~TH_FIN;
298 	}
299 
300 	if (len < 0) {
301 		/*
302 		 * If FIN has been sent but not acked,
303 		 * but we haven't been called to retransmit,
304 		 * len will be < 0.  Otherwise, window shrank
305 		 * after we sent into it.  If window shrank to 0,
306 		 * cancel pending retransmit, pull snd_nxt back
307 		 * to (closed) window, and set the persist timer
308 		 * if it isn't already going.  If the window didn't
309 		 * close completely, just wait for an ACK.
310 		 */
311 		len = 0;
312 		if (sendwin == 0) {
313 			callout_stop(tp->tt_rexmt);
314 			tp->t_rxtshift = 0;
315 			tp->snd_nxt = tp->snd_una;
316 			if (!callout_active(tp->tt_persist))
317 				tcp_setpersist(tp);
318 		}
319 	}
320 
321 	/*
322 	 * len will be >= 0 after this point.  Truncate to the maximum
323 	 * segment length and ensure that FIN is removed if the length
324 	 * no longer contains the last data byte.
325 	 */
326 	if (len > tp->t_maxseg) {
327 		len = tp->t_maxseg;
328 		sendalot = TRUE;
329 	}
330 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
331 		flags &= ~TH_FIN;
332 
333 	recvwin = sbspace(&so->so_rcv);
334 
335 	/*
336 	 * Sender silly window avoidance.   We transmit under the following
337 	 * conditions when len is non-zero:
338 	 *
339 	 *	- We have a full segment
340 	 *	- This is the last buffer in a write()/send() and we are
341 	 *	  either idle or running NODELAY
342 	 *	- we've timed out (e.g. persist timer)
343 	 *	- we have more then 1/2 the maximum send window's worth of
344 	 *	  data (receiver may be limited the window size)
345 	 *	- we need to retransmit
346 	 */
347 	if (len) {
348 		if (len == tp->t_maxseg)
349 			goto send;
350 		/*
351 		 * NOTE! on localhost connections an 'ack' from the remote
352 		 * end may occur synchronously with the output and cause
353 		 * us to flush a buffer queued with moretocome.  XXX
354 		 *
355 		 * note: the len + off check is almost certainly unnecessary.
356 		 */
357 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
358 		    (idle || (tp->t_flags & TF_NODELAY)) &&
359 		    len + off >= so->so_snd.sb_cc &&
360 		    !(tp->t_flags & TF_NOPUSH)) {
361 			goto send;
362 		}
363 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
364 			goto send;
365 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
366 			goto send;
367 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
368 			goto send;
369 	}
370 
371 	/*
372 	 * Compare available window to amount of window
373 	 * known to peer (as advertised window less
374 	 * next expected input).  If the difference is at least two
375 	 * max size segments, or at least 50% of the maximum possible
376 	 * window, then want to send a window update to peer.
377 	 */
378 	if (recvwin > 0) {
379 		/*
380 		 * "adv" is the amount we can increase the window,
381 		 * taking into account that we are limited by
382 		 * TCP_MAXWIN << tp->rcv_scale.
383 		 */
384 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
385 			(tp->rcv_adv - tp->rcv_nxt);
386 
387 		/*
388 		 * This ack case typically occurs when the user has drained
389 		 * the TCP socket buffer sufficiently to warrent an ack
390 		 * containing a 'pure window update'... that is, an ack that
391 		 * ONLY updates the tcp window.
392 		 *
393 		 * It is unclear why we would need to do a pure window update
394 		 * past 2 segments if we are going to do one at 1/2 the high
395 		 * water mark anyway, especially since under normal conditions
396 		 * the user program will drain the socket buffer quickly.
397 		 * The 2-segment pure window update will often add a large
398 		 * number of extra, unnecessary acks to the stream.
399 		 *
400 		 * avoid_pure_win_update now defaults to 1.
401 		 */
402 		if (avoid_pure_win_update == 0) {
403 			if (adv >= (long) (2 * tp->t_maxseg))
404 				goto send;
405 		}
406 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
407 			goto send;
408 	}
409 
410 	/*
411 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
412 	 * is also a catch-all for the retransmit timer timeout case.
413 	 */
414 	if (tp->t_flags & TF_ACKNOW)
415 		goto send;
416 	if ((flags & TH_RST) ||
417 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
418 		goto send;
419 	if (SEQ_GT(tp->snd_up, tp->snd_una))
420 		goto send;
421 	/*
422 	 * If our state indicates that FIN should be sent
423 	 * and we have not yet done so, then we need to send.
424 	 */
425 	if (flags & TH_FIN &&
426 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
427 		goto send;
428 
429 	/*
430 	 * TCP window updates are not reliable, rather a polling protocol
431 	 * using ``persist'' packets is used to insure receipt of window
432 	 * updates.  The three ``states'' for the output side are:
433 	 *	idle			not doing retransmits or persists
434 	 *	persisting		to move a small or zero window
435 	 *	(re)transmitting	and thereby not persisting
436 	 *
437 	 * callout_active(tp->tt_persist)
438 	 *	is true when we are in persist state.
439 	 * The TF_FORCE flag in tp->t_flags
440 	 *	is set when we are called to send a persist packet.
441 	 * callout_active(tp->tt_rexmt)
442 	 *	is set when we are retransmitting
443 	 * The output side is idle when both timers are zero.
444 	 *
445 	 * If send window is too small, there is data to transmit, and no
446 	 * retransmit or persist is pending, then go to persist state.
447 	 * If nothing happens soon, send when timer expires:
448 	 * if window is nonzero, transmit what we can,
449 	 * otherwise force out a byte.
450 	 */
451 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
452 	    !callout_active(tp->tt_persist)) {
453 		tp->t_rxtshift = 0;
454 		tcp_setpersist(tp);
455 	}
456 
457 	/*
458 	 * No reason to send a segment, just return.
459 	 */
460 	return (0);
461 
462 send:
463 	/*
464 	 * Before ESTABLISHED, force sending of initial options
465 	 * unless TCP set not to do any options.
466 	 * NOTE: we assume that the IP/TCP header plus TCP options
467 	 * always fit in a single mbuf, leaving room for a maximum
468 	 * link header, i.e.
469 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
470 	 */
471 	optlen = 0;
472 	if (isipv6)
473 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
474 	else
475 		hdrlen = sizeof(struct tcpiphdr);
476 	if (flags & TH_SYN) {
477 		tp->snd_nxt = tp->iss;
478 		if (!(tp->t_flags & TF_NOOPT)) {
479 			u_short mss;
480 
481 			opt[0] = TCPOPT_MAXSEG;
482 			opt[1] = TCPOLEN_MAXSEG;
483 			mss = htons((u_short) tcp_mssopt(tp));
484 			(void)memcpy(opt + 2, &mss, sizeof(mss));
485 			optlen = TCPOLEN_MAXSEG;
486 
487 			if ((tp->t_flags & TF_REQ_SCALE) &&
488 			    (!(flags & TH_ACK) ||
489 			     (tp->t_flags & TF_RCVD_SCALE))) {
490 				*((u_int32_t *)(opt + optlen)) = htonl(
491 					TCPOPT_NOP << 24 |
492 					TCPOPT_WINDOW << 16 |
493 					TCPOLEN_WINDOW << 8 |
494 					tp->request_r_scale);
495 				optlen += 4;
496 			}
497 		}
498 	}
499 
500 	/*
501 	 * Send a timestamp and echo-reply if this is a SYN and our side
502 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
503 	 * and our peer have sent timestamps in our SYN's.
504 	 */
505 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
506 	    !(flags & TH_RST) &&
507 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
508 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
509 
510 		/* Form timestamp option as shown in appendix A of RFC 1323. */
511 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
512 		*lp++ = htonl(ticks);
513 		*lp   = htonl(tp->ts_recent);
514 		optlen += TCPOLEN_TSTAMP_APPA;
515 	}
516 
517 	/*
518 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
519 	 * options are allowed (!TF_NOOPT) and it's not a RST.
520 	 */
521 	if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
522 	     !(flags & TH_RST)) {
523 		switch (flags & (TH_SYN | TH_ACK)) {
524 		/*
525 		 * This is a normal ACK, send CC if we received CC before
526 		 * from our peer.
527 		 */
528 		case TH_ACK:
529 			if (!(tp->t_flags & TF_RCVD_CC))
530 				break;
531 			/*FALLTHROUGH*/
532 
533 		/*
534 		 * We can only get here in T/TCP's SYN_SENT* state, when
535 		 * we're a sending a non-SYN segment without waiting for
536 		 * the ACK of our SYN.  A check above assures that we only
537 		 * do this if our peer understands T/TCP.
538 		 */
539 		case 0:
540 			opt[optlen++] = TCPOPT_NOP;
541 			opt[optlen++] = TCPOPT_NOP;
542 			opt[optlen++] = TCPOPT_CC;
543 			opt[optlen++] = TCPOLEN_CC;
544 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
545 			optlen += 4;
546 			break;
547 
548 		/*
549 		 * This is our initial SYN, check whether we have to use
550 		 * CC or CC.new.
551 		 */
552 		case TH_SYN:
553 			opt[optlen++] = TCPOPT_NOP;
554 			opt[optlen++] = TCPOPT_NOP;
555 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
556 						TCPOPT_CCNEW : TCPOPT_CC;
557 			opt[optlen++] = TCPOLEN_CC;
558 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
559 			optlen += 4;
560 			break;
561 
562 		/*
563 		 * This is a SYN,ACK; send CC and CC.echo if we received
564 		 * CC from our peer.
565 		 */
566 		case (TH_SYN | TH_ACK):
567 			if (tp->t_flags & TF_RCVD_CC) {
568 				opt[optlen++] = TCPOPT_NOP;
569 				opt[optlen++] = TCPOPT_NOP;
570 				opt[optlen++] = TCPOPT_CC;
571 				opt[optlen++] = TCPOLEN_CC;
572 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
573 				optlen += 4;
574 				opt[optlen++] = TCPOPT_NOP;
575 				opt[optlen++] = TCPOPT_NOP;
576 				opt[optlen++] = TCPOPT_CCECHO;
577 				opt[optlen++] = TCPOLEN_CC;
578 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv);
579 				optlen += 4;
580 			}
581 			break;
582 		}
583 	}
584 
585 	hdrlen += optlen;
586 
587 	if (isipv6) {
588 		ipoptlen = ip6_optlen(inp);
589 	} else {
590 		if (inp->inp_options) {
591 			ipoptlen = inp->inp_options->m_len -
592 			    offsetof(struct ipoption, ipopt_list);
593 		} else {
594 			ipoptlen = 0;
595 		}
596 	}
597 #ifdef IPSEC
598 	ipoptlen += ipsec_hdrsiz_tcp(tp);
599 #endif
600 
601 	/*
602 	 * Adjust data length if insertion of options will
603 	 * bump the packet length beyond the t_maxopd length.
604 	 * Clear the FIN bit because we cut off the tail of
605 	 * the segment.
606 	 */
607 	if (len + optlen + ipoptlen > tp->t_maxopd) {
608 		/*
609 		 * If there is still more to send, don't close the connection.
610 		 */
611 		flags &= ~TH_FIN;
612 		len = tp->t_maxopd - optlen - ipoptlen;
613 		sendalot = TRUE;
614 	}
615 
616 #ifdef INET6
617 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
618 #else
619 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
620 #endif
621 
622 	/*
623 	 * Grab a header mbuf, attaching a copy of data to
624 	 * be transmitted, and initialize the header from
625 	 * the template for sends on this connection.
626 	 */
627 	if (len) {
628 		if ((tp->t_flags & TF_FORCE) && len == 1)
629 			tcpstat.tcps_sndprobe++;
630 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
631 			tcpstat.tcps_sndrexmitpack++;
632 			tcpstat.tcps_sndrexmitbyte += len;
633 		} else {
634 			tcpstat.tcps_sndpack++;
635 			tcpstat.tcps_sndbyte += len;
636 		}
637 #ifdef notyet
638 		if ((m = m_copypack(so->so_snd.sb_mb, off, (int)len,
639 		    max_linkhdr + hdrlen)) == NULL) {
640 			error = ENOBUFS;
641 			goto out;
642 		}
643 		/*
644 		 * m_copypack left space for our hdr; use it.
645 		 */
646 		m->m_len += hdrlen;
647 		m->m_data -= hdrlen;
648 #else
649 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
650 		if (m == NULL) {
651 			error = ENOBUFS;
652 			goto out;
653 		}
654 #ifdef INET6
655 		if (MHLEN < hdrlen + max_linkhdr) {
656 			MCLGET(m, MB_DONTWAIT);
657 			if (!(m->m_flags & M_EXT)) {
658 				m_freem(m);
659 				error = ENOBUFS;
660 				goto out;
661 			}
662 		}
663 #endif
664 		m->m_data += max_linkhdr;
665 		m->m_len = hdrlen;
666 		if (len <= MHLEN - hdrlen - max_linkhdr) {
667 			m_copydata(so->so_snd.sb_mb, off, (int) len,
668 			    mtod(m, caddr_t) + hdrlen);
669 			m->m_len += len;
670 		} else {
671 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
672 			if (m->m_next == NULL) {
673 				(void) m_free(m);
674 				error = ENOBUFS;
675 				goto out;
676 			}
677 		}
678 #endif
679 		/*
680 		 * If we're sending everything we've got, set PUSH.
681 		 * (This will keep happy those implementations which only
682 		 * give data to the user when a buffer fills or
683 		 * a PUSH comes in.)
684 		 */
685 		if (off + len == so->so_snd.sb_cc)
686 			flags |= TH_PUSH;
687 	} else {
688 		if (tp->t_flags & TF_ACKNOW)
689 			tcpstat.tcps_sndacks++;
690 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
691 			tcpstat.tcps_sndctrl++;
692 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
693 			tcpstat.tcps_sndurg++;
694 		else
695 			tcpstat.tcps_sndwinup++;
696 
697 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
698 		if (m == NULL) {
699 			error = ENOBUFS;
700 			goto out;
701 		}
702 		if (isipv6 &&
703 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
704 			MH_ALIGN(m, hdrlen);
705 		else
706 			m->m_data += max_linkhdr;
707 		m->m_len = hdrlen;
708 	}
709 	m->m_pkthdr.rcvif = (struct ifnet *)0;
710 	if (isipv6) {
711 		ip6 = mtod(m, struct ip6_hdr *);
712 		th = (struct tcphdr *)(ip6 + 1);
713 		tcp_fillheaders(tp, ip6, th);
714 	} else {
715 		ip = mtod(m, struct ip *);
716 		ipov = (struct ipovly *)ip;
717 		th = (struct tcphdr *)(ip + 1);
718 		/* this picks up the pseudo header (w/o the length) */
719 		tcp_fillheaders(tp, ip, th);
720 	}
721 
722 	/*
723 	 * Fill in fields, remembering maximum advertised
724 	 * window for use in delaying messages about window sizes.
725 	 * If resending a FIN, be sure not to use a new sequence number.
726 	 */
727 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
728 	    tp->snd_nxt == tp->snd_max)
729 		tp->snd_nxt--;
730 	/*
731 	 * If we are doing retransmissions, then snd_nxt will
732 	 * not reflect the first unsent octet.  For ACK only
733 	 * packets, we do not want the sequence number of the
734 	 * retransmitted packet, we want the sequence number
735 	 * of the next unsent octet.  So, if there is no data
736 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
737 	 * when filling in ti_seq.  But if we are in persist
738 	 * state, snd_max might reflect one byte beyond the
739 	 * right edge of the window, so use snd_nxt in that
740 	 * case, since we know we aren't doing a retransmission.
741 	 * (retransmit and persist are mutually exclusive...)
742 	 */
743 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
744 		th->th_seq = htonl(tp->snd_nxt);
745 	else
746 		th->th_seq = htonl(tp->snd_max);
747 	th->th_ack = htonl(tp->rcv_nxt);
748 	if (optlen) {
749 		bcopy(opt, th + 1, optlen);
750 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
751 	}
752 	th->th_flags = flags;
753 	/*
754 	 * Calculate receive window.  Don't shrink window,
755 	 * but avoid silly window syndrome.
756 	 */
757 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
758 	    recvwin < (long)tp->t_maxseg)
759 		recvwin = 0;
760 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
761 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
762 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
763 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
764 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
765 
766 	/*
767 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
768 	 * a 0 window.  This may cause the remote transmitter to stall.  This
769 	 * flag tells soreceive() to disable delayed acknowledgements when
770 	 * draining the buffer.  This can occur if the receiver is attempting
771 	 * to read more data then can be buffered prior to transmitting on
772 	 * the connection.
773 	 */
774 	if (recvwin == 0)
775 		tp->t_flags |= TF_RXWIN0SENT;
776 	else
777 		tp->t_flags &= ~TF_RXWIN0SENT;
778 
779 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
780 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
781 		th->th_flags |= TH_URG;
782 	} else {
783 		/*
784 		 * If no urgent pointer to send, then we pull
785 		 * the urgent pointer to the left edge of the send window
786 		 * so that it doesn't drift into the send window on sequence
787 		 * number wraparound.
788 		 */
789 		tp->snd_up = tp->snd_una;		/* drag it along */
790 	}
791 
792 	/*
793 	 * Put TCP length in extended header, and then
794 	 * checksum extended header and data.
795 	 */
796 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
797 	if (isipv6) {
798 		/*
799 		 * ip6_plen is not need to be filled now, and will be filled
800 		 * in ip6_output.
801 		 */
802 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
803 				       sizeof(struct tcphdr) + optlen + len);
804 	} else {
805 		m->m_pkthdr.csum_flags = CSUM_TCP;
806 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
807 		if (len + optlen)
808 			th->th_sum = in_addword(th->th_sum,
809 						htons((u_short)(optlen + len)));
810 
811 		/* IP version must be set here for ipv4/ipv6 checking later */
812 		KASSERT(ip->ip_v == IPVERSION,
813 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
814 	}
815 
816 	/*
817 	 * In transmit state, time the transmission and arrange for
818 	 * the retransmit.  In persist state, just set snd_max.
819 	 */
820 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
821 		tcp_seq startseq = tp->snd_nxt;
822 
823 		/*
824 		 * Advance snd_nxt over sequence space of this segment.
825 		 */
826 		if (flags & (TH_SYN | TH_FIN)) {
827 			if (flags & TH_SYN)
828 				tp->snd_nxt++;
829 			if (flags & TH_FIN) {
830 				tp->snd_nxt++;
831 				tp->t_flags |= TF_SENTFIN;
832 			}
833 		}
834 		tp->snd_nxt += len;
835 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
836 			tp->snd_max = tp->snd_nxt;
837 			/*
838 			 * Time this transmission if not a retransmission and
839 			 * not currently timing anything.
840 			 */
841 			if (tp->t_rtttime == 0) {
842 				tp->t_rtttime = ticks;
843 				tp->t_rtseq = startseq;
844 				tcpstat.tcps_segstimed++;
845 			}
846 		}
847 
848 		/*
849 		 * Set retransmit timer if not currently set,
850 		 * and not doing a pure ack or a keep-alive probe.
851 		 * Initial value for retransmit timer is smoothed
852 		 * round-trip time + 2 * round-trip time variance.
853 		 * Initialize shift counter which is used for backoff
854 		 * of retransmit time.
855 		 */
856 		if (!callout_active(tp->tt_rexmt) &&
857 		    tp->snd_nxt != tp->snd_una) {
858 			if (callout_active(tp->tt_persist)) {
859 				callout_stop(tp->tt_persist);
860 				tp->t_rxtshift = 0;
861 			}
862 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
863 				      tcp_timer_rexmt, tp);
864 		}
865 	} else {
866 		/*
867 		 * Persist case, update snd_max but since we are in
868 		 * persist mode (no window) we do not update snd_nxt.
869 		 */
870 		int xlen = len;
871 		if (flags & TH_SYN)
872 			++xlen;
873 		if (flags & TH_FIN) {
874 			++xlen;
875 			tp->t_flags |= TF_SENTFIN;
876 		}
877 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
878 			tp->snd_max = tp->snd_nxt + xlen;
879 	}
880 
881 #ifdef TCPDEBUG
882 	/*
883 	 * Trace.
884 	 */
885 	if (so->so_options & SO_DEBUG)
886 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
887 #endif
888 
889 	/*
890 	 * Fill in IP length and desired time to live and
891 	 * send to IP level.  There should be a better way
892 	 * to handle ttl and tos; we could keep them in
893 	 * the template, but need a way to checksum without them.
894 	 */
895 	/*
896 	 * m->m_pkthdr.len should have been set before cksum calcuration,
897 	 * because in6_cksum() need it.
898 	 */
899 	if (isipv6) {
900 		/*
901 		 * we separately set hoplimit for every segment, since the
902 		 * user might want to change the value via setsockopt.
903 		 * Also, desired default hop limit might be changed via
904 		 * Neighbor Discovery.
905 		 */
906 		ip6->ip6_hlim = in6_selecthlim(inp,
907 		    (inp->in6p_route.ro_rt ?
908 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
909 
910 		/* TODO: IPv6 IP6TOS_ECT bit on */
911 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
912 		    (so->so_options & SO_DONTROUTE), NULL, NULL, inp);
913 	} else {
914 		struct rtentry *rt;
915 		ip->ip_len = m->m_pkthdr.len;
916 #ifdef INET6
917 		if (INP_CHECK_SOCKAF(so, AF_INET6))
918 			ip->ip_ttl = in6_selecthlim(inp,
919 			    (inp->in6p_route.ro_rt ?
920 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
921 		else
922 #endif
923 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
924 
925 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
926 		/*
927 		 * See if we should do MTU discovery.
928 		 * We do it only if the following are true:
929 		 *	1) we have a valid route to the destination
930 		 *	2) the MTU is not locked (if it is,
931 		 *	   then discovery has been disabled)
932 		 */
933 		if (path_mtu_discovery &&
934 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
935 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
936 			ip->ip_off |= IP_DF;
937 
938 		error = ip_output(m, inp->inp_options, &inp->inp_route,
939 		    (so->so_options & SO_DONTROUTE), NULL, inp);
940 	}
941 	if (error) {
942 
943 		/*
944 		 * We know that the packet was lost, so back out the
945 		 * sequence number advance, if any.
946 		 */
947 		if (!(tp->t_flags & TF_FORCE) ||
948 		    !callout_active(tp->tt_persist)) {
949 			/*
950 			 * No need to check for TH_FIN here because
951 			 * the TF_SENTFIN flag handles that case.
952 			 */
953 			if (!(flags & TH_SYN))
954 				tp->snd_nxt -= len;
955 		}
956 
957 out:
958 		if (error == ENOBUFS) {
959 			/*
960 			 * If we can't send, make sure there is something
961 			 * to get us going again later.  Persist state
962 			 * is not necessarily right, but it is close enough.
963 			 */
964 			if (!callout_active(tp->tt_rexmt) &&
965 			    !callout_active(tp->tt_persist)) {
966 				tp->t_rxtshift = 0;
967 				tcp_setpersist(tp);
968 			}
969 			tcp_quench(inp, 0);
970 			return (0);
971 		}
972 		if (error == EMSGSIZE) {
973 			/*
974 			 * ip_output() will have already fixed the route
975 			 * for us.  tcp_mtudisc() will, as its last action,
976 			 * initiate retransmission, so it is important to
977 			 * not do so here.
978 			 */
979 			tcp_mtudisc(inp, 0);
980 			return 0;
981 		}
982 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
983 		    TCPS_HAVERCVDSYN(tp->t_state)) {
984 			tp->t_softerror = error;
985 			return (0);
986 		}
987 		return (error);
988 	}
989 	tcpstat.tcps_sndtotal++;
990 
991 	/*
992 	 * Data sent (as far as we can tell).
993 	 * If this advertises a larger window than any other segment,
994 	 * then remember the size of the advertised window.
995 	 * Any pending ACK has now been sent.
996 	 */
997 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
998 		tp->rcv_adv = tp->rcv_nxt + recvwin;
999 	tp->last_ack_sent = tp->rcv_nxt;
1000 	tp->t_flags &= ~TF_ACKNOW;
1001 	if (tcp_delack_enabled)
1002 		callout_stop(tp->tt_delack);
1003 	if (sendalot)
1004 		goto again;
1005 	return (0);
1006 }
1007 
1008 void
1009 tcp_setpersist(tp)
1010 	struct tcpcb *tp;
1011 {
1012 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1013 	int tt;
1014 
1015 	if (callout_active(tp->tt_rexmt))
1016 		panic("tcp_setpersist: retransmit pending");
1017 	/*
1018 	 * Start/restart persistance timer.
1019 	 */
1020 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1021 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1022 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1023 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1024 		tp->t_rxtshift++;
1025 }
1026