xref: /dflybsd-src/sys/netinet/tcp_output.c (revision a12ef770ba351628dde7e68bd7b722ae54dd554b)
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 1.24 2005/02/04 01:49:30 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 <netproto/ipsec/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 = 0;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
141 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
142 
143 static int avoid_pure_win_update = 1;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
145 	&avoid_pure_win_update, 1, "Avoid pure window updates when possible");
146 
147 /*
148  * Tcp output routine: figure out what should be sent and send it.
149  */
150 int
151 tcp_output(tp)
152 	struct tcpcb *tp;
153 {
154 	struct inpcb * const inp = tp->t_inpcb;
155 	struct socket *so = inp->inp_socket;
156 	long len, recvwin, sendwin;
157 	int off, flags, error;
158 	struct mbuf *m;
159 	struct ip *ip = NULL;
160 	struct ipovly *ipov = NULL;
161 	struct tcphdr *th;
162 	u_char opt[TCP_MAXOLEN];
163 	unsigned int ipoptlen, optlen, hdrlen;
164 	int idle;
165 	boolean_t sendalot;
166 	struct ip6_hdr *ip6 = NULL;
167 #ifdef INET6
168 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
169 #else
170 	const boolean_t isipv6 = FALSE;
171 #endif
172 	struct rmxp_tao *taop;
173 
174 	/*
175 	 * Determine length of data that should be transmitted,
176 	 * and flags that will be used.
177 	 * If there is some data or critical controls (SYN, RST)
178 	 * to send, then transmit; otherwise, investigate further.
179 	 */
180 	if ((tp->snd_max == tp->snd_una) &&
181 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
182 		/*
183 		 * We have been idle for "a while" and no acks are
184 		 * expected to clock out any data we send --
185 		 * slow start to get ack "clock" running again.
186 		 */
187 		tp->snd_cwnd = tp->t_maxseg;
188 	}
189 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
190 	if (idle && (tp->t_flags & TF_MORETOCOME))
191 		tp->t_flags |= TF_LASTIDLE;
192 	else
193 		tp->t_flags &= ~TF_LASTIDLE;
194 
195 again:
196 	/* Make use of SACK information when slow-starting after a RTO. */
197 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
198 	    !IN_FASTRECOVERY(tp))
199 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
200 
201 	sendalot = FALSE;
202 	off = tp->snd_nxt - tp->snd_una;
203 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
204 	sendwin = min(sendwin, tp->snd_bwnd);
205 
206 	flags = tcp_outflags[tp->t_state];
207 	/*
208 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
209 	 * state flags.
210 	 */
211 	if (tp->t_flags & TF_NEEDFIN)
212 		flags |= TH_FIN;
213 	if (tp->t_flags & TF_NEEDSYN)
214 		flags |= TH_SYN;
215 
216 	/*
217 	 * If in persist timeout with window of 0, send 1 byte.
218 	 * Otherwise, if window is small but nonzero
219 	 * and timer expired, we will send what we can
220 	 * and go to transmit state.
221 	 */
222 	if (tp->t_flags & TF_FORCE) {
223 		if (sendwin == 0) {
224 			/*
225 			 * If we still have some data to send, then
226 			 * clear the FIN bit.  Usually this would
227 			 * happen below when it realizes that we
228 			 * aren't sending all the data.  However,
229 			 * if we have exactly 1 byte of unsent data,
230 			 * then it won't clear the FIN bit below,
231 			 * and if we are in persist state, we wind
232 			 * up sending the packet without recording
233 			 * that we sent the FIN bit.
234 			 *
235 			 * We can't just blindly clear the FIN bit,
236 			 * because if we don't have any more data
237 			 * to send then the probe will be the FIN
238 			 * itself.
239 			 */
240 			if (off < so->so_snd.sb_cc)
241 				flags &= ~TH_FIN;
242 			sendwin = 1;
243 		} else {
244 			callout_stop(tp->tt_persist);
245 			tp->t_rxtshift = 0;
246 		}
247 	}
248 
249 	/*
250 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
251 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
252 	 * a negative length.  This can also occur when TCP opens up
253 	 * its congestion window while receiving additional duplicate
254 	 * acks after fast-retransmit because TCP will reset snd_nxt
255 	 * to snd_max after the fast-retransmit.
256 	 *
257 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
258 	 * be set to snd_una, the offset will be 0, and the length may
259 	 * wind up 0.
260 	 */
261 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
262 
263 	/*
264 	 * Lop off SYN bit if it has already been sent.  However, if this
265 	 * is SYN-SENT state and if segment contains data and if we don't
266 	 * know that foreign host supports TAO, suppress sending segment.
267 	 */
268 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
269 		flags &= ~TH_SYN;
270 		off--, len++;
271 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
272 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
273 		     taop->tao_ccsent == 0))
274 			return 0;
275 	}
276 
277 	/*
278 	 * Be careful not to send data and/or FIN on SYN segments
279 	 * in cases when no CC option will be sent.
280 	 * This measure is needed to prevent interoperability problems
281 	 * with not fully conformant TCP implementations.
282 	 */
283 	if ((flags & TH_SYN) &&
284 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
285 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
286 		len = 0;
287 		flags &= ~TH_FIN;
288 	}
289 
290 	if (len < 0) {
291 		/*
292 		 * If FIN has been sent but not acked,
293 		 * but we haven't been called to retransmit,
294 		 * len will be < 0.  Otherwise, window shrank
295 		 * after we sent into it.  If window shrank to 0,
296 		 * cancel pending retransmit, pull snd_nxt back
297 		 * to (closed) window, and set the persist timer
298 		 * if it isn't already going.  If the window didn't
299 		 * close completely, just wait for an ACK.
300 		 */
301 		len = 0;
302 		if (sendwin == 0) {
303 			callout_stop(tp->tt_rexmt);
304 			tp->t_rxtshift = 0;
305 			tp->snd_nxt = tp->snd_una;
306 			if (!callout_active(tp->tt_persist))
307 				tcp_setpersist(tp);
308 		}
309 	}
310 
311 	/*
312 	 * len will be >= 0 after this point.  Truncate to the maximum
313 	 * segment length and ensure that FIN is removed if the length
314 	 * no longer contains the last data byte.
315 	 */
316 	if (len > tp->t_maxseg) {
317 		len = tp->t_maxseg;
318 		sendalot = TRUE;
319 	}
320 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
321 		flags &= ~TH_FIN;
322 
323 	recvwin = sbspace(&so->so_rcv);
324 
325 	/*
326 	 * Sender silly window avoidance.   We transmit under the following
327 	 * conditions when len is non-zero:
328 	 *
329 	 *	- We have a full segment
330 	 *	- This is the last buffer in a write()/send() and we are
331 	 *	  either idle or running NODELAY
332 	 *	- we've timed out (e.g. persist timer)
333 	 *	- we have more then 1/2 the maximum send window's worth of
334 	 *	  data (receiver may be limited the window size)
335 	 *	- we need to retransmit
336 	 */
337 	if (len) {
338 		if (len == tp->t_maxseg)
339 			goto send;
340 		/*
341 		 * NOTE! on localhost connections an 'ack' from the remote
342 		 * end may occur synchronously with the output and cause
343 		 * us to flush a buffer queued with moretocome.  XXX
344 		 *
345 		 * note: the len + off check is almost certainly unnecessary.
346 		 */
347 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
348 		    (idle || (tp->t_flags & TF_NODELAY)) &&
349 		    len + off >= so->so_snd.sb_cc &&
350 		    !(tp->t_flags & TF_NOPUSH)) {
351 			goto send;
352 		}
353 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
354 			goto send;
355 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
356 			goto send;
357 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
358 			goto send;
359 	}
360 
361 	/*
362 	 * Compare available window to amount of window
363 	 * known to peer (as advertised window less
364 	 * next expected input).  If the difference is at least two
365 	 * max size segments, or at least 50% of the maximum possible
366 	 * window, then want to send a window update to peer.
367 	 */
368 	if (recvwin > 0) {
369 		/*
370 		 * "adv" is the amount we can increase the window,
371 		 * taking into account that we are limited by
372 		 * TCP_MAXWIN << tp->rcv_scale.
373 		 */
374 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
375 			(tp->rcv_adv - tp->rcv_nxt);
376 
377 		/*
378 		 * This ack case typically occurs when the user has drained
379 		 * the TCP socket buffer sufficiently to warrent an ack
380 		 * containing a 'pure window update'... that is, an ack that
381 		 * ONLY updates the tcp window.
382 		 *
383 		 * It is unclear why we would need to do a pure window update
384 		 * past 2 segments if we are going to do one at 1/2 the high
385 		 * water mark anyway, especially since under normal conditions
386 		 * the user program will drain the socket buffer quickly.
387 		 * The 2-segment pure window update will often add a large
388 		 * number of extra, unnecessary acks to the stream.
389 		 *
390 		 * avoid_pure_win_update now defaults to 1.
391 		 */
392 		if (avoid_pure_win_update == 0) {
393 			if (adv >= (long) (2 * tp->t_maxseg))
394 				goto send;
395 		}
396 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
397 			goto send;
398 	}
399 
400 	/*
401 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
402 	 * is also a catch-all for the retransmit timer timeout case.
403 	 */
404 	if (tp->t_flags & TF_ACKNOW)
405 		goto send;
406 	if ((flags & TH_RST) ||
407 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
408 		goto send;
409 	if (SEQ_GT(tp->snd_up, tp->snd_una))
410 		goto send;
411 	/*
412 	 * If our state indicates that FIN should be sent
413 	 * and we have not yet done so, then we need to send.
414 	 */
415 	if (flags & TH_FIN &&
416 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
417 		goto send;
418 
419 	/*
420 	 * TCP window updates are not reliable, rather a polling protocol
421 	 * using ``persist'' packets is used to insure receipt of window
422 	 * updates.  The three ``states'' for the output side are:
423 	 *	idle			not doing retransmits or persists
424 	 *	persisting		to move a small or zero window
425 	 *	(re)transmitting	and thereby not persisting
426 	 *
427 	 * callout_active(tp->tt_persist)
428 	 *	is true when we are in persist state.
429 	 * The TF_FORCE flag in tp->t_flags
430 	 *	is set when we are called to send a persist packet.
431 	 * callout_active(tp->tt_rexmt)
432 	 *	is set when we are retransmitting
433 	 * The output side is idle when both timers are zero.
434 	 *
435 	 * If send window is too small, there is data to transmit, and no
436 	 * retransmit or persist is pending, then go to persist state.
437 	 * If nothing happens soon, send when timer expires:
438 	 * if window is nonzero, transmit what we can,
439 	 * otherwise force out a byte.
440 	 */
441 	if (so->so_snd.sb_cc > 0 &&
442 	    !callout_active(tp->tt_rexmt) && !callout_active(tp->tt_persist)) {
443 		tp->t_rxtshift = 0;
444 		tcp_setpersist(tp);
445 	}
446 
447 	/*
448 	 * No reason to send a segment, just return.
449 	 */
450 	return (0);
451 
452 send:
453 	/*
454 	 * Before ESTABLISHED, force sending of initial options
455 	 * unless TCP set not to do any options.
456 	 * NOTE: we assume that the IP/TCP header plus TCP options
457 	 * always fit in a single mbuf, leaving room for a maximum
458 	 * link header, i.e.
459 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
460 	 */
461 	optlen = 0;
462 	if (isipv6)
463 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
464 	else
465 		hdrlen = sizeof(struct tcpiphdr);
466 	if (flags & TH_SYN) {
467 		tp->snd_nxt = tp->iss;
468 		if (!(tp->t_flags & TF_NOOPT)) {
469 			u_short mss;
470 
471 			opt[0] = TCPOPT_MAXSEG;
472 			opt[1] = TCPOLEN_MAXSEG;
473 			mss = htons((u_short) tcp_mssopt(tp));
474 			memcpy(opt + 2, &mss, sizeof mss);
475 			optlen = TCPOLEN_MAXSEG;
476 
477 			if ((tp->t_flags & TF_REQ_SCALE) &&
478 			    (!(flags & TH_ACK) ||
479 			     (tp->t_flags & TF_RCVD_SCALE))) {
480 				*((u_int32_t *)(opt + optlen)) = htonl(
481 					TCPOPT_NOP << 24 |
482 					TCPOPT_WINDOW << 16 |
483 					TCPOLEN_WINDOW << 8 |
484 					tp->request_r_scale);
485 				optlen += 4;
486 			}
487 
488 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
489 			    tp->t_flags & TF_SACK_PERMITTED) {
490 				uint32_t *lp = (uint32_t *)(opt + optlen);
491 
492 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
493 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
494 			}
495 		}
496 	}
497 
498 	/*
499 	 * Send a timestamp and echo-reply if this is a SYN and our side
500 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
501 	 * and our peer have sent timestamps in our SYN's.
502 	 */
503 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
504 	    !(flags & TH_RST) &&
505 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
506 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
507 
508 		/* Form timestamp option as shown in appendix A of RFC 1323. */
509 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
510 		*lp++ = htonl(ticks);
511 		*lp   = htonl(tp->ts_recent);
512 		optlen += TCPOLEN_TSTAMP_APPA;
513 	}
514 
515 	/*
516 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
517 	 * options are allowed (!TF_NOOPT) and it's not a RST.
518 	 */
519 	if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
520 	     !(flags & TH_RST)) {
521 		switch (flags & (TH_SYN | TH_ACK)) {
522 		/*
523 		 * This is a normal ACK, send CC if we received CC before
524 		 * from our peer.
525 		 */
526 		case TH_ACK:
527 			if (!(tp->t_flags & TF_RCVD_CC))
528 				break;
529 			/*FALLTHROUGH*/
530 
531 		/*
532 		 * We can only get here in T/TCP's SYN_SENT* state, when
533 		 * we're a sending a non-SYN segment without waiting for
534 		 * the ACK of our SYN.  A check above assures that we only
535 		 * do this if our peer understands T/TCP.
536 		 */
537 		case 0:
538 			opt[optlen++] = TCPOPT_NOP;
539 			opt[optlen++] = TCPOPT_NOP;
540 			opt[optlen++] = TCPOPT_CC;
541 			opt[optlen++] = TCPOLEN_CC;
542 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
543 			optlen += 4;
544 			break;
545 
546 		/*
547 		 * This is our initial SYN, check whether we have to use
548 		 * CC or CC.new.
549 		 */
550 		case TH_SYN:
551 			opt[optlen++] = TCPOPT_NOP;
552 			opt[optlen++] = TCPOPT_NOP;
553 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
554 						TCPOPT_CCNEW : TCPOPT_CC;
555 			opt[optlen++] = TCPOLEN_CC;
556 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
557 			optlen += 4;
558 			break;
559 
560 		/*
561 		 * This is a SYN,ACK; send CC and CC.echo if we received
562 		 * CC from our peer.
563 		 */
564 		case (TH_SYN | TH_ACK):
565 			if (tp->t_flags & TF_RCVD_CC) {
566 				opt[optlen++] = TCPOPT_NOP;
567 				opt[optlen++] = TCPOPT_NOP;
568 				opt[optlen++] = TCPOPT_CC;
569 				opt[optlen++] = TCPOLEN_CC;
570 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
571 				optlen += 4;
572 				opt[optlen++] = TCPOPT_NOP;
573 				opt[optlen++] = TCPOPT_NOP;
574 				opt[optlen++] = TCPOPT_CCECHO;
575 				opt[optlen++] = TCPOLEN_CC;
576 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv);
577 				optlen += 4;
578 			}
579 			break;
580 		}
581 	}
582 
583 	/*
584 	 * If this is a SACK connection and we have a block to report,
585 	 * fill in the SACK blocks in the TCP options.
586 	 */
587 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
588 		TF_SACK_PERMITTED &&
589 	    (!LIST_EMPTY(&tp->t_segq) ||
590 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
591 		tcp_sack_fill_report(tp, opt, &optlen);
592 
593 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
594 	hdrlen += optlen;
595 
596 	if (isipv6) {
597 		ipoptlen = ip6_optlen(inp);
598 	} else {
599 		if (inp->inp_options) {
600 			ipoptlen = inp->inp_options->m_len -
601 			    offsetof(struct ipoption, ipopt_list);
602 		} else {
603 			ipoptlen = 0;
604 		}
605 	}
606 #ifdef IPSEC
607 	ipoptlen += ipsec_hdrsiz_tcp(tp);
608 #endif
609 
610 	/*
611 	 * Adjust data length if insertion of options will
612 	 * bump the packet length beyond the t_maxopd length.
613 	 * Clear the FIN bit because we cut off the tail of
614 	 * the segment.
615 	 */
616 	if (len + optlen + ipoptlen > tp->t_maxopd) {
617 		/*
618 		 * If there is still more to send, don't close the connection.
619 		 */
620 		flags &= ~TH_FIN;
621 		len = tp->t_maxopd - optlen - ipoptlen;
622 		sendalot = TRUE;
623 	}
624 
625 #ifdef INET6
626 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
627 #else
628 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
629 #endif
630 
631 	/*
632 	 * Grab a header mbuf, attaching a copy of data to
633 	 * be transmitted, and initialize the header from
634 	 * the template for sends on this connection.
635 	 */
636 	if (len) {
637 		if ((tp->t_flags & TF_FORCE) && len == 1)
638 			tcpstat.tcps_sndprobe++;
639 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
640 			if (tp->snd_nxt == tp->snd_una)
641 				tp->snd_max_rexmt = tp->snd_max;
642 			tcpstat.tcps_sndrexmitpack++;
643 			tcpstat.tcps_sndrexmitbyte += len;
644 		} else {
645 			tcpstat.tcps_sndpack++;
646 			tcpstat.tcps_sndbyte += len;
647 		}
648 #ifdef notyet
649 		if ((m = m_copypack(so->so_snd.sb_mb, off, (int)len,
650 		    max_linkhdr + hdrlen)) == NULL) {
651 			error = ENOBUFS;
652 			goto out;
653 		}
654 		/*
655 		 * m_copypack left space for our hdr; use it.
656 		 */
657 		m->m_len += hdrlen;
658 		m->m_data -= hdrlen;
659 #else
660 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
661 		if (m == NULL) {
662 			error = ENOBUFS;
663 			goto out;
664 		}
665 #ifdef INET6
666 		if (MHLEN < hdrlen + max_linkhdr) {
667 			MCLGET(m, MB_DONTWAIT);
668 			if (!(m->m_flags & M_EXT)) {
669 				m_freem(m);
670 				error = ENOBUFS;
671 				goto out;
672 			}
673 		}
674 #endif
675 		m->m_data += max_linkhdr;
676 		m->m_len = hdrlen;
677 		if (len <= MHLEN - hdrlen - max_linkhdr) {
678 			m_copydata(so->so_snd.sb_mb, off, (int) len,
679 			    mtod(m, caddr_t) + hdrlen);
680 			m->m_len += len;
681 		} else {
682 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
683 			if (m->m_next == NULL) {
684 				m_free(m);
685 				error = ENOBUFS;
686 				goto out;
687 			}
688 		}
689 #endif
690 		/*
691 		 * If we're sending everything we've got, set PUSH.
692 		 * (This will keep happy those implementations which only
693 		 * give data to the user when a buffer fills or
694 		 * a PUSH comes in.)
695 		 */
696 		if (off + len == so->so_snd.sb_cc)
697 			flags |= TH_PUSH;
698 	} else {
699 		if (tp->t_flags & TF_ACKNOW)
700 			tcpstat.tcps_sndacks++;
701 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
702 			tcpstat.tcps_sndctrl++;
703 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
704 			tcpstat.tcps_sndurg++;
705 		else
706 			tcpstat.tcps_sndwinup++;
707 
708 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
709 		if (m == NULL) {
710 			error = ENOBUFS;
711 			goto out;
712 		}
713 		if (isipv6 &&
714 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
715 			MH_ALIGN(m, hdrlen);
716 		else
717 			m->m_data += max_linkhdr;
718 		m->m_len = hdrlen;
719 	}
720 	m->m_pkthdr.rcvif = (struct ifnet *)0;
721 	if (isipv6) {
722 		ip6 = mtod(m, struct ip6_hdr *);
723 		th = (struct tcphdr *)(ip6 + 1);
724 		tcp_fillheaders(tp, ip6, th);
725 	} else {
726 		ip = mtod(m, struct ip *);
727 		ipov = (struct ipovly *)ip;
728 		th = (struct tcphdr *)(ip + 1);
729 		/* this picks up the pseudo header (w/o the length) */
730 		tcp_fillheaders(tp, ip, th);
731 	}
732 
733 	/*
734 	 * Fill in fields, remembering maximum advertised
735 	 * window for use in delaying messages about window sizes.
736 	 * If resending a FIN, be sure not to use a new sequence number.
737 	 */
738 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
739 	    tp->snd_nxt == tp->snd_max)
740 		tp->snd_nxt--;
741 	/*
742 	 * If we are doing retransmissions, then snd_nxt will
743 	 * not reflect the first unsent octet.  For ACK only
744 	 * packets, we do not want the sequence number of the
745 	 * retransmitted packet, we want the sequence number
746 	 * of the next unsent octet.  So, if there is no data
747 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
748 	 * when filling in ti_seq.  But if we are in persist
749 	 * state, snd_max might reflect one byte beyond the
750 	 * right edge of the window, so use snd_nxt in that
751 	 * case, since we know we aren't doing a retransmission.
752 	 * (retransmit and persist are mutually exclusive...)
753 	 */
754 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
755 		th->th_seq = htonl(tp->snd_nxt);
756 	else
757 		th->th_seq = htonl(tp->snd_max);
758 	th->th_ack = htonl(tp->rcv_nxt);
759 	if (optlen) {
760 		bcopy(opt, th + 1, optlen);
761 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
762 	}
763 	th->th_flags = flags;
764 	/*
765 	 * Calculate receive window.  Don't shrink window,
766 	 * but avoid silly window syndrome.
767 	 */
768 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
769 	    recvwin < (long)tp->t_maxseg)
770 		recvwin = 0;
771 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
772 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
773 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
774 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
775 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
776 
777 	/*
778 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
779 	 * a 0 window.  This may cause the remote transmitter to stall.  This
780 	 * flag tells soreceive() to disable delayed acknowledgements when
781 	 * draining the buffer.  This can occur if the receiver is attempting
782 	 * to read more data then can be buffered prior to transmitting on
783 	 * the connection.
784 	 */
785 	if (recvwin == 0)
786 		tp->t_flags |= TF_RXWIN0SENT;
787 	else
788 		tp->t_flags &= ~TF_RXWIN0SENT;
789 
790 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
791 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
792 		th->th_flags |= TH_URG;
793 	} else {
794 		/*
795 		 * If no urgent pointer to send, then we pull
796 		 * the urgent pointer to the left edge of the send window
797 		 * so that it doesn't drift into the send window on sequence
798 		 * number wraparound.
799 		 */
800 		tp->snd_up = tp->snd_una;		/* drag it along */
801 	}
802 
803 	/*
804 	 * Put TCP length in extended header, and then
805 	 * checksum extended header and data.
806 	 */
807 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
808 	if (isipv6) {
809 		/*
810 		 * ip6_plen is not need to be filled now, and will be filled
811 		 * in ip6_output().
812 		 */
813 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
814 				       sizeof(struct tcphdr) + optlen + len);
815 	} else {
816 		m->m_pkthdr.csum_flags = CSUM_TCP;
817 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
818 		if (len + optlen)
819 			th->th_sum = in_addword(th->th_sum,
820 						htons((u_short)(optlen + len)));
821 
822 		/* IP version must be set here for ipv4/ipv6 checking later */
823 		KASSERT(ip->ip_v == IPVERSION,
824 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
825 	}
826 
827 	/*
828 	 * In transmit state, time the transmission and arrange for
829 	 * the retransmit.  In persist state, just set snd_max.
830 	 */
831 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
832 		tcp_seq startseq = tp->snd_nxt;
833 
834 		/*
835 		 * Advance snd_nxt over sequence space of this segment.
836 		 */
837 		if (flags & (TH_SYN | TH_FIN)) {
838 			if (flags & TH_SYN)
839 				tp->snd_nxt++;
840 			if (flags & TH_FIN) {
841 				tp->snd_nxt++;
842 				tp->t_flags |= TF_SENTFIN;
843 			}
844 		}
845 		tp->snd_nxt += len;
846 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
847 			tp->snd_max = tp->snd_nxt;
848 			/*
849 			 * Time this transmission if not a retransmission and
850 			 * not currently timing anything.
851 			 */
852 			if (tp->t_rtttime == 0) {
853 				tp->t_rtttime = ticks;
854 				tp->t_rtseq = startseq;
855 				tcpstat.tcps_segstimed++;
856 			}
857 		}
858 
859 		/*
860 		 * Set retransmit timer if not currently set,
861 		 * and not doing a pure ack or a keep-alive probe.
862 		 * Initial value for retransmit timer is smoothed
863 		 * round-trip time + 2 * round-trip time variance.
864 		 * Initialize shift counter which is used for backoff
865 		 * of retransmit time.
866 		 */
867 		if (!callout_active(tp->tt_rexmt) &&
868 		    tp->snd_nxt != tp->snd_una) {
869 			if (callout_active(tp->tt_persist)) {
870 				callout_stop(tp->tt_persist);
871 				tp->t_rxtshift = 0;
872 			}
873 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
874 				      tcp_timer_rexmt, tp);
875 		}
876 	} else {
877 		/*
878 		 * Persist case, update snd_max but since we are in
879 		 * persist mode (no window) we do not update snd_nxt.
880 		 */
881 		int xlen = len;
882 		if (flags & TH_SYN)
883 			++xlen;
884 		if (flags & TH_FIN) {
885 			++xlen;
886 			tp->t_flags |= TF_SENTFIN;
887 		}
888 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
889 			tp->snd_max = tp->snd_nxt + xlen;
890 	}
891 
892 #ifdef TCPDEBUG
893 	/*
894 	 * Trace.
895 	 */
896 	if (so->so_options & SO_DEBUG)
897 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
898 #endif
899 
900 	/*
901 	 * Fill in IP length and desired time to live and
902 	 * send to IP level.  There should be a better way
903 	 * to handle ttl and tos; we could keep them in
904 	 * the template, but need a way to checksum without them.
905 	 */
906 	/*
907 	 * m->m_pkthdr.len should have been set before cksum calcuration,
908 	 * because in6_cksum() need it.
909 	 */
910 	if (isipv6) {
911 		/*
912 		 * we separately set hoplimit for every segment, since the
913 		 * user might want to change the value via setsockopt.
914 		 * Also, desired default hop limit might be changed via
915 		 * Neighbor Discovery.
916 		 */
917 		ip6->ip6_hlim = in6_selecthlim(inp,
918 		    (inp->in6p_route.ro_rt ?
919 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
920 
921 		/* TODO: IPv6 IP6TOS_ECT bit on */
922 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
923 				   (so->so_options & SO_DONTROUTE), NULL, NULL,
924 				   inp);
925 	} else {
926 		struct rtentry *rt;
927 		ip->ip_len = m->m_pkthdr.len;
928 #ifdef INET6
929 		if (INP_CHECK_SOCKAF(so, AF_INET6))
930 			ip->ip_ttl = in6_selecthlim(inp,
931 			    (inp->in6p_route.ro_rt ?
932 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
933 		else
934 #endif
935 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
936 
937 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
938 		/*
939 		 * See if we should do MTU discovery.
940 		 * We do it only if the following are true:
941 		 *	1) we have a valid route to the destination
942 		 *	2) the MTU is not locked (if it is,
943 		 *	   then discovery has been disabled)
944 		 */
945 		if (path_mtu_discovery &&
946 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
947 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
948 			ip->ip_off |= IP_DF;
949 
950 		error = ip_output(m, inp->inp_options, &inp->inp_route,
951 				  (so->so_options & SO_DONTROUTE), NULL, inp);
952 	}
953 	if (error) {
954 
955 		/*
956 		 * We know that the packet was lost, so back out the
957 		 * sequence number advance, if any.
958 		 */
959 		if (!(tp->t_flags & TF_FORCE) ||
960 		    !callout_active(tp->tt_persist)) {
961 			/*
962 			 * No need to check for TH_FIN here because
963 			 * the TF_SENTFIN flag handles that case.
964 			 */
965 			if (!(flags & TH_SYN))
966 				tp->snd_nxt -= len;
967 		}
968 
969 out:
970 		if (error == ENOBUFS) {
971 			/*
972 			 * If we can't send, make sure there is something
973 			 * to get us going again later.  Persist state
974 			 * is not necessarily right, but it is close enough.
975 			 */
976 			if (!callout_active(tp->tt_rexmt) &&
977 			    !callout_active(tp->tt_persist)) {
978 				tp->t_rxtshift = 0;
979 				tcp_setpersist(tp);
980 			}
981 			tcp_quench(inp, 0);
982 			return (0);
983 		}
984 		if (error == EMSGSIZE) {
985 			/*
986 			 * ip_output() will have already fixed the route
987 			 * for us.  tcp_mtudisc() will, as its last action,
988 			 * initiate retransmission, so it is important to
989 			 * not do so here.
990 			 */
991 			tcp_mtudisc(inp, 0);
992 			return 0;
993 		}
994 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
995 		    TCPS_HAVERCVDSYN(tp->t_state)) {
996 			tp->t_softerror = error;
997 			return (0);
998 		}
999 		return (error);
1000 	}
1001 	tcpstat.tcps_sndtotal++;
1002 
1003 	/*
1004 	 * Data sent (as far as we can tell).
1005 	 * If this advertises a larger window than any other segment,
1006 	 * then remember the size of the advertised window.
1007 	 * Any pending ACK has now been sent.
1008 	 */
1009 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
1010 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1011 	tp->last_ack_sent = tp->rcv_nxt;
1012 	tp->t_flags &= ~TF_ACKNOW;
1013 	if (tcp_delack_enabled)
1014 		callout_stop(tp->tt_delack);
1015 	if (sendalot)
1016 		goto again;
1017 	return (0);
1018 }
1019 
1020 void
1021 tcp_setpersist(tp)
1022 	struct tcpcb *tp;
1023 {
1024 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1025 	int tt;
1026 
1027 	if (callout_active(tp->tt_rexmt))
1028 		panic("tcp_setpersist: retransmit pending");
1029 	/*
1030 	 * Start/restart persistance timer.
1031 	 */
1032 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1033 		      TCPTV_PERSMAX);
1034 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1035 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1036 		tp->t_rxtshift++;
1037 }
1038