xref: /dflybsd-src/sys/netinet/tcp_output.c (revision 7d1ab061891a94bc40e4954c2851bb1691ea3d54)
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 1.25 2005/03/09 06:54:34 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 nsacked = 0;
158 	int off, flags, error;
159 	struct mbuf *m;
160 	struct ip *ip = NULL;
161 	struct ipovly *ipov = NULL;
162 	struct tcphdr *th;
163 	u_char opt[TCP_MAXOLEN];
164 	unsigned int ipoptlen, optlen, hdrlen;
165 	int idle;
166 	boolean_t sendalot;
167 	struct ip6_hdr *ip6 = NULL;
168 #ifdef INET6
169 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
170 #else
171 	const boolean_t isipv6 = FALSE;
172 #endif
173 	struct rmxp_tao *taop;
174 
175 	/*
176 	 * Determine length of data that should be transmitted,
177 	 * and flags that will be used.
178 	 * If there is some data or critical controls (SYN, RST)
179 	 * to send, then transmit; otherwise, investigate further.
180 	 */
181 	if ((tp->snd_max == tp->snd_una) &&
182 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
183 		/*
184 		 * We have been idle for "a while" and no acks are
185 		 * expected to clock out any data we send --
186 		 * slow start to get ack "clock" running again.
187 		 */
188 		tp->snd_cwnd = tp->t_maxseg;
189 	}
190 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
191 	if (idle && (tp->t_flags & TF_MORETOCOME))
192 		tp->t_flags |= TF_LASTIDLE;
193 	else
194 		tp->t_flags &= ~TF_LASTIDLE;
195 
196 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
197 	    !IN_FASTRECOVERY(tp))
198 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
199 
200 again:
201 	/* Make use of SACK information when slow-starting after a RTO. */
202 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
203 	    !IN_FASTRECOVERY(tp)) {
204 		tcp_seq old_snd_nxt = tp->snd_nxt;
205 
206 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
207 		nsacked += tp->snd_nxt - old_snd_nxt;
208 	}
209 
210 	sendalot = FALSE;
211 	off = tp->snd_nxt - tp->snd_una;
212 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
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 = TRUE;
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 		/*
387 		 * This ack case typically occurs when the user has drained
388 		 * the TCP socket buffer sufficiently to warrent an ack
389 		 * containing a 'pure window update'... that is, an ack that
390 		 * ONLY updates the tcp window.
391 		 *
392 		 * It is unclear why we would need to do a pure window update
393 		 * past 2 segments if we are going to do one at 1/2 the high
394 		 * water mark anyway, especially since under normal conditions
395 		 * the user program will drain the socket buffer quickly.
396 		 * The 2-segment pure window update will often add a large
397 		 * number of extra, unnecessary acks to the stream.
398 		 *
399 		 * avoid_pure_win_update now defaults to 1.
400 		 */
401 		if (avoid_pure_win_update == 0) {
402 			if (adv >= (long) (2 * tp->t_maxseg))
403 				goto send;
404 		}
405 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
406 			goto send;
407 	}
408 
409 	/*
410 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
411 	 * is also a catch-all for the retransmit timer timeout case.
412 	 */
413 	if (tp->t_flags & TF_ACKNOW)
414 		goto send;
415 	if ((flags & TH_RST) ||
416 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
417 		goto send;
418 	if (SEQ_GT(tp->snd_up, tp->snd_una))
419 		goto send;
420 	/*
421 	 * If our state indicates that FIN should be sent
422 	 * and we have not yet done so, then we need to send.
423 	 */
424 	if (flags & TH_FIN &&
425 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
426 		goto send;
427 
428 	/*
429 	 * TCP window updates are not reliable, rather a polling protocol
430 	 * using ``persist'' packets is used to insure receipt of window
431 	 * updates.  The three ``states'' for the output side are:
432 	 *	idle			not doing retransmits or persists
433 	 *	persisting		to move a small or zero window
434 	 *	(re)transmitting	and thereby not persisting
435 	 *
436 	 * callout_active(tp->tt_persist)
437 	 *	is true when we are in persist state.
438 	 * The TF_FORCE flag in tp->t_flags
439 	 *	is set when we are called to send a persist packet.
440 	 * callout_active(tp->tt_rexmt)
441 	 *	is set when we are retransmitting
442 	 * The output side is idle when both timers are zero.
443 	 *
444 	 * If send window is too small, there is data to transmit, and no
445 	 * retransmit or persist is pending, then go to persist state.
446 	 * If nothing happens soon, send when timer expires:
447 	 * if window is nonzero, transmit what we can,
448 	 * otherwise force out a byte.
449 	 */
450 	if (so->so_snd.sb_cc > 0 &&
451 	    !callout_active(tp->tt_rexmt) && !callout_active(tp->tt_persist)) {
452 		tp->t_rxtshift = 0;
453 		tcp_setpersist(tp);
454 	}
455 
456 	/*
457 	 * No reason to send a segment, just return.
458 	 */
459 	return (0);
460 
461 send:
462 	/*
463 	 * Before ESTABLISHED, force sending of initial options
464 	 * unless TCP set not to do any options.
465 	 * NOTE: we assume that the IP/TCP header plus TCP options
466 	 * always fit in a single mbuf, leaving room for a maximum
467 	 * link header, i.e.
468 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
469 	 */
470 	optlen = 0;
471 	if (isipv6)
472 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
473 	else
474 		hdrlen = sizeof(struct tcpiphdr);
475 	if (flags & TH_SYN) {
476 		tp->snd_nxt = tp->iss;
477 		if (!(tp->t_flags & TF_NOOPT)) {
478 			u_short mss;
479 
480 			opt[0] = TCPOPT_MAXSEG;
481 			opt[1] = TCPOLEN_MAXSEG;
482 			mss = htons((u_short) tcp_mssopt(tp));
483 			memcpy(opt + 2, &mss, sizeof mss);
484 			optlen = TCPOLEN_MAXSEG;
485 
486 			if ((tp->t_flags & TF_REQ_SCALE) &&
487 			    (!(flags & TH_ACK) ||
488 			     (tp->t_flags & TF_RCVD_SCALE))) {
489 				*((u_int32_t *)(opt + optlen)) = htonl(
490 					TCPOPT_NOP << 24 |
491 					TCPOPT_WINDOW << 16 |
492 					TCPOLEN_WINDOW << 8 |
493 					tp->request_r_scale);
494 				optlen += 4;
495 			}
496 
497 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
498 			    tp->t_flags & TF_SACK_PERMITTED) {
499 				uint32_t *lp = (uint32_t *)(opt + optlen);
500 
501 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
502 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
503 			}
504 		}
505 	}
506 
507 	/*
508 	 * Send a timestamp and echo-reply if this is a SYN and our side
509 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
510 	 * and our peer have sent timestamps in our SYN's.
511 	 */
512 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
513 	    !(flags & TH_RST) &&
514 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
515 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
516 
517 		/* Form timestamp option as shown in appendix A of RFC 1323. */
518 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
519 		*lp++ = htonl(ticks);
520 		*lp   = htonl(tp->ts_recent);
521 		optlen += TCPOLEN_TSTAMP_APPA;
522 	}
523 
524 	/*
525 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
526 	 * options are allowed (!TF_NOOPT) and it's not a RST.
527 	 */
528 	if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
529 	     !(flags & TH_RST)) {
530 		switch (flags & (TH_SYN | TH_ACK)) {
531 		/*
532 		 * This is a normal ACK, send CC if we received CC before
533 		 * from our peer.
534 		 */
535 		case TH_ACK:
536 			if (!(tp->t_flags & TF_RCVD_CC))
537 				break;
538 			/*FALLTHROUGH*/
539 
540 		/*
541 		 * We can only get here in T/TCP's SYN_SENT* state, when
542 		 * we're a sending a non-SYN segment without waiting for
543 		 * the ACK of our SYN.  A check above assures that we only
544 		 * do this if our peer understands T/TCP.
545 		 */
546 		case 0:
547 			opt[optlen++] = TCPOPT_NOP;
548 			opt[optlen++] = TCPOPT_NOP;
549 			opt[optlen++] = TCPOPT_CC;
550 			opt[optlen++] = TCPOLEN_CC;
551 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
552 			optlen += 4;
553 			break;
554 
555 		/*
556 		 * This is our initial SYN, check whether we have to use
557 		 * CC or CC.new.
558 		 */
559 		case TH_SYN:
560 			opt[optlen++] = TCPOPT_NOP;
561 			opt[optlen++] = TCPOPT_NOP;
562 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
563 						TCPOPT_CCNEW : TCPOPT_CC;
564 			opt[optlen++] = TCPOLEN_CC;
565 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
566 			optlen += 4;
567 			break;
568 
569 		/*
570 		 * This is a SYN,ACK; send CC and CC.echo if we received
571 		 * CC from our peer.
572 		 */
573 		case (TH_SYN | TH_ACK):
574 			if (tp->t_flags & TF_RCVD_CC) {
575 				opt[optlen++] = TCPOPT_NOP;
576 				opt[optlen++] = TCPOPT_NOP;
577 				opt[optlen++] = TCPOPT_CC;
578 				opt[optlen++] = TCPOLEN_CC;
579 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
580 				optlen += 4;
581 				opt[optlen++] = TCPOPT_NOP;
582 				opt[optlen++] = TCPOPT_NOP;
583 				opt[optlen++] = TCPOPT_CCECHO;
584 				opt[optlen++] = TCPOLEN_CC;
585 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv);
586 				optlen += 4;
587 			}
588 			break;
589 		}
590 	}
591 
592 	/*
593 	 * If this is a SACK connection and we have a block to report,
594 	 * fill in the SACK blocks in the TCP options.
595 	 */
596 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
597 		TF_SACK_PERMITTED &&
598 	    (!LIST_EMPTY(&tp->t_segq) ||
599 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
600 		tcp_sack_fill_report(tp, opt, &optlen);
601 
602 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
603 	hdrlen += optlen;
604 
605 	if (isipv6) {
606 		ipoptlen = ip6_optlen(inp);
607 	} else {
608 		if (inp->inp_options) {
609 			ipoptlen = inp->inp_options->m_len -
610 			    offsetof(struct ipoption, ipopt_list);
611 		} else {
612 			ipoptlen = 0;
613 		}
614 	}
615 #ifdef IPSEC
616 	ipoptlen += ipsec_hdrsiz_tcp(tp);
617 #endif
618 
619 	/*
620 	 * Adjust data length if insertion of options will
621 	 * bump the packet length beyond the t_maxopd length.
622 	 * Clear the FIN bit because we cut off the tail of
623 	 * the segment.
624 	 */
625 	if (len + optlen + ipoptlen > tp->t_maxopd) {
626 		/*
627 		 * If there is still more to send, don't close the connection.
628 		 */
629 		flags &= ~TH_FIN;
630 		len = tp->t_maxopd - optlen - ipoptlen;
631 		sendalot = TRUE;
632 	}
633 
634 #ifdef INET6
635 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
636 #else
637 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
638 #endif
639 
640 	/*
641 	 * Grab a header mbuf, attaching a copy of data to
642 	 * be transmitted, and initialize the header from
643 	 * the template for sends on this connection.
644 	 */
645 	if (len) {
646 		if ((tp->t_flags & TF_FORCE) && len == 1)
647 			tcpstat.tcps_sndprobe++;
648 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
649 			if (tp->snd_nxt == tp->snd_una)
650 				tp->snd_max_rexmt = tp->snd_max;
651 			tcpstat.tcps_sndrexmitpack++;
652 			tcpstat.tcps_sndrexmitbyte += len;
653 		} else {
654 			tcpstat.tcps_sndpack++;
655 			tcpstat.tcps_sndbyte += len;
656 		}
657 #ifdef notyet
658 		if ((m = m_copypack(so->so_snd.sb_mb, off, (int)len,
659 		    max_linkhdr + hdrlen)) == NULL) {
660 			error = ENOBUFS;
661 			goto out;
662 		}
663 		/*
664 		 * m_copypack left space for our hdr; use it.
665 		 */
666 		m->m_len += hdrlen;
667 		m->m_data -= hdrlen;
668 #else
669 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
670 		if (m == NULL) {
671 			error = ENOBUFS;
672 			goto out;
673 		}
674 #ifdef INET6
675 		if (MHLEN < hdrlen + max_linkhdr) {
676 			MCLGET(m, MB_DONTWAIT);
677 			if (!(m->m_flags & M_EXT)) {
678 				m_freem(m);
679 				error = ENOBUFS;
680 				goto out;
681 			}
682 		}
683 #endif
684 		m->m_data += max_linkhdr;
685 		m->m_len = hdrlen;
686 		if (len <= MHLEN - hdrlen - max_linkhdr) {
687 			m_copydata(so->so_snd.sb_mb, off, (int) len,
688 			    mtod(m, caddr_t) + hdrlen);
689 			m->m_len += len;
690 		} else {
691 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
692 			if (m->m_next == NULL) {
693 				m_free(m);
694 				error = ENOBUFS;
695 				goto out;
696 			}
697 		}
698 #endif
699 		/*
700 		 * If we're sending everything we've got, set PUSH.
701 		 * (This will keep happy those implementations which only
702 		 * give data to the user when a buffer fills or
703 		 * a PUSH comes in.)
704 		 */
705 		if (off + len == so->so_snd.sb_cc)
706 			flags |= TH_PUSH;
707 	} else {
708 		if (tp->t_flags & TF_ACKNOW)
709 			tcpstat.tcps_sndacks++;
710 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
711 			tcpstat.tcps_sndctrl++;
712 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
713 			tcpstat.tcps_sndurg++;
714 		else
715 			tcpstat.tcps_sndwinup++;
716 
717 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
718 		if (m == NULL) {
719 			error = ENOBUFS;
720 			goto out;
721 		}
722 		if (isipv6 &&
723 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
724 			MH_ALIGN(m, hdrlen);
725 		else
726 			m->m_data += max_linkhdr;
727 		m->m_len = hdrlen;
728 	}
729 	m->m_pkthdr.rcvif = (struct ifnet *)0;
730 	if (isipv6) {
731 		ip6 = mtod(m, struct ip6_hdr *);
732 		th = (struct tcphdr *)(ip6 + 1);
733 		tcp_fillheaders(tp, ip6, th);
734 	} else {
735 		ip = mtod(m, struct ip *);
736 		ipov = (struct ipovly *)ip;
737 		th = (struct tcphdr *)(ip + 1);
738 		/* this picks up the pseudo header (w/o the length) */
739 		tcp_fillheaders(tp, ip, th);
740 	}
741 
742 	/*
743 	 * Fill in fields, remembering maximum advertised
744 	 * window for use in delaying messages about window sizes.
745 	 * If resending a FIN, be sure not to use a new sequence number.
746 	 */
747 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
748 	    tp->snd_nxt == tp->snd_max)
749 		tp->snd_nxt--;
750 	/*
751 	 * If we are doing retransmissions, then snd_nxt will
752 	 * not reflect the first unsent octet.  For ACK only
753 	 * packets, we do not want the sequence number of the
754 	 * retransmitted packet, we want the sequence number
755 	 * of the next unsent octet.  So, if there is no data
756 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
757 	 * when filling in ti_seq.  But if we are in persist
758 	 * state, snd_max might reflect one byte beyond the
759 	 * right edge of the window, so use snd_nxt in that
760 	 * case, since we know we aren't doing a retransmission.
761 	 * (retransmit and persist are mutually exclusive...)
762 	 */
763 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
764 		th->th_seq = htonl(tp->snd_nxt);
765 	else
766 		th->th_seq = htonl(tp->snd_max);
767 	th->th_ack = htonl(tp->rcv_nxt);
768 	if (optlen) {
769 		bcopy(opt, th + 1, optlen);
770 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
771 	}
772 	th->th_flags = flags;
773 	/*
774 	 * Calculate receive window.  Don't shrink window,
775 	 * but avoid silly window syndrome.
776 	 */
777 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
778 	    recvwin < (long)tp->t_maxseg)
779 		recvwin = 0;
780 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
781 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
782 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
783 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
784 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
785 
786 	/*
787 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
788 	 * a 0 window.  This may cause the remote transmitter to stall.  This
789 	 * flag tells soreceive() to disable delayed acknowledgements when
790 	 * draining the buffer.  This can occur if the receiver is attempting
791 	 * to read more data then can be buffered prior to transmitting on
792 	 * the connection.
793 	 */
794 	if (recvwin == 0)
795 		tp->t_flags |= TF_RXWIN0SENT;
796 	else
797 		tp->t_flags &= ~TF_RXWIN0SENT;
798 
799 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
800 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
801 		th->th_flags |= TH_URG;
802 	} else {
803 		/*
804 		 * If no urgent pointer to send, then we pull
805 		 * the urgent pointer to the left edge of the send window
806 		 * so that it doesn't drift into the send window on sequence
807 		 * number wraparound.
808 		 */
809 		tp->snd_up = tp->snd_una;		/* drag it along */
810 	}
811 
812 	/*
813 	 * Put TCP length in extended header, and then
814 	 * checksum extended header and data.
815 	 */
816 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
817 	if (isipv6) {
818 		/*
819 		 * ip6_plen is not need to be filled now, and will be filled
820 		 * in ip6_output().
821 		 */
822 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
823 				       sizeof(struct tcphdr) + optlen + len);
824 	} else {
825 		m->m_pkthdr.csum_flags = CSUM_TCP;
826 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
827 		if (len + optlen)
828 			th->th_sum = in_addword(th->th_sum,
829 						htons((u_short)(optlen + len)));
830 
831 		/* IP version must be set here for ipv4/ipv6 checking later */
832 		KASSERT(ip->ip_v == IPVERSION,
833 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
834 	}
835 
836 	/*
837 	 * In transmit state, time the transmission and arrange for
838 	 * the retransmit.  In persist state, just set snd_max.
839 	 */
840 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
841 		tcp_seq startseq = tp->snd_nxt;
842 
843 		/*
844 		 * Advance snd_nxt over sequence space of this segment.
845 		 */
846 		if (flags & (TH_SYN | TH_FIN)) {
847 			if (flags & TH_SYN)
848 				tp->snd_nxt++;
849 			if (flags & TH_FIN) {
850 				tp->snd_nxt++;
851 				tp->t_flags |= TF_SENTFIN;
852 			}
853 		}
854 		tp->snd_nxt += len;
855 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
856 			tp->snd_max = tp->snd_nxt;
857 			/*
858 			 * Time this transmission if not a retransmission and
859 			 * not currently timing anything.
860 			 */
861 			if (tp->t_rtttime == 0) {
862 				tp->t_rtttime = ticks;
863 				tp->t_rtseq = startseq;
864 				tcpstat.tcps_segstimed++;
865 			}
866 		}
867 
868 		/*
869 		 * Set retransmit timer if not currently set,
870 		 * and not doing a pure ack or a keep-alive probe.
871 		 * Initial value for retransmit timer is smoothed
872 		 * round-trip time + 2 * round-trip time variance.
873 		 * Initialize shift counter which is used for backoff
874 		 * of retransmit time.
875 		 */
876 		if (!callout_active(tp->tt_rexmt) &&
877 		    tp->snd_nxt != tp->snd_una) {
878 			if (callout_active(tp->tt_persist)) {
879 				callout_stop(tp->tt_persist);
880 				tp->t_rxtshift = 0;
881 			}
882 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
883 				      tcp_timer_rexmt, tp);
884 		}
885 	} else {
886 		/*
887 		 * Persist case, update snd_max but since we are in
888 		 * persist mode (no window) we do not update snd_nxt.
889 		 */
890 		int xlen = len;
891 		if (flags & TH_SYN)
892 			++xlen;
893 		if (flags & TH_FIN) {
894 			++xlen;
895 			tp->t_flags |= TF_SENTFIN;
896 		}
897 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
898 			tp->snd_max = tp->snd_nxt + xlen;
899 	}
900 
901 #ifdef TCPDEBUG
902 	/*
903 	 * Trace.
904 	 */
905 	if (so->so_options & SO_DEBUG)
906 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
907 #endif
908 
909 	/*
910 	 * Fill in IP length and desired time to live and
911 	 * send to IP level.  There should be a better way
912 	 * to handle ttl and tos; we could keep them in
913 	 * the template, but need a way to checksum without them.
914 	 */
915 	/*
916 	 * m->m_pkthdr.len should have been set before cksum calcuration,
917 	 * because in6_cksum() need it.
918 	 */
919 	if (isipv6) {
920 		/*
921 		 * we separately set hoplimit for every segment, since the
922 		 * user might want to change the value via setsockopt.
923 		 * Also, desired default hop limit might be changed via
924 		 * Neighbor Discovery.
925 		 */
926 		ip6->ip6_hlim = in6_selecthlim(inp,
927 		    (inp->in6p_route.ro_rt ?
928 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
929 
930 		/* TODO: IPv6 IP6TOS_ECT bit on */
931 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
932 				   (so->so_options & SO_DONTROUTE), NULL, NULL,
933 				   inp);
934 	} else {
935 		struct rtentry *rt;
936 		ip->ip_len = m->m_pkthdr.len;
937 #ifdef INET6
938 		if (INP_CHECK_SOCKAF(so, AF_INET6))
939 			ip->ip_ttl = in6_selecthlim(inp,
940 			    (inp->in6p_route.ro_rt ?
941 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
942 		else
943 #endif
944 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
945 
946 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
947 		/*
948 		 * See if we should do MTU discovery.
949 		 * We do it only if the following are true:
950 		 *	1) we have a valid route to the destination
951 		 *	2) the MTU is not locked (if it is,
952 		 *	   then discovery has been disabled)
953 		 */
954 		if (path_mtu_discovery &&
955 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
956 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
957 			ip->ip_off |= IP_DF;
958 
959 		error = ip_output(m, inp->inp_options, &inp->inp_route,
960 				  (so->so_options & SO_DONTROUTE), NULL, inp);
961 	}
962 	if (error) {
963 
964 		/*
965 		 * We know that the packet was lost, so back out the
966 		 * sequence number advance, if any.
967 		 */
968 		if (!(tp->t_flags & TF_FORCE) ||
969 		    !callout_active(tp->tt_persist)) {
970 			/*
971 			 * No need to check for TH_FIN here because
972 			 * the TF_SENTFIN flag handles that case.
973 			 */
974 			if (!(flags & TH_SYN))
975 				tp->snd_nxt -= len;
976 		}
977 
978 out:
979 		if (error == ENOBUFS) {
980 			/*
981 			 * If we can't send, make sure there is something
982 			 * to get us going again later.  Persist state
983 			 * is not necessarily right, but it is close enough.
984 			 */
985 			if (!callout_active(tp->tt_rexmt) &&
986 			    !callout_active(tp->tt_persist)) {
987 				tp->t_rxtshift = 0;
988 				tcp_setpersist(tp);
989 			}
990 			tcp_quench(inp, 0);
991 			return (0);
992 		}
993 		if (error == EMSGSIZE) {
994 			/*
995 			 * ip_output() will have already fixed the route
996 			 * for us.  tcp_mtudisc() will, as its last action,
997 			 * initiate retransmission, so it is important to
998 			 * not do so here.
999 			 */
1000 			tcp_mtudisc(inp, 0);
1001 			return 0;
1002 		}
1003 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1004 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1005 			tp->t_softerror = error;
1006 			return (0);
1007 		}
1008 		return (error);
1009 	}
1010 	tcpstat.tcps_sndtotal++;
1011 
1012 	/*
1013 	 * Data sent (as far as we can tell).
1014 	 * If this advertises a larger window than any other segment,
1015 	 * then remember the size of the advertised window.
1016 	 * Any pending ACK has now been sent.
1017 	 */
1018 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
1019 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1020 	tp->last_ack_sent = tp->rcv_nxt;
1021 	tp->t_flags &= ~TF_ACKNOW;
1022 	if (tcp_delack_enabled)
1023 		callout_stop(tp->tt_delack);
1024 	if (sendalot)
1025 		goto again;
1026 	return (0);
1027 }
1028 
1029 void
1030 tcp_setpersist(tp)
1031 	struct tcpcb *tp;
1032 {
1033 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1034 	int tt;
1035 
1036 	if (callout_active(tp->tt_rexmt))
1037 		panic("tcp_setpersist: retransmit pending");
1038 	/*
1039 	 * Start/restart persistance timer.
1040 	 */
1041 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1042 		      TCPTV_PERSMAX);
1043 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1044 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1045 		tp->t_rxtshift++;
1046 }
1047