xref: /dflybsd-src/sys/netinet/tcp_output.c (revision bc49aa1be5400e3bdd801519c6936e8947d5d432)
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 1.16 2004/07/17 20:31:31 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 /*
152  * Tcp output routine: figure out what should be sent and send it.
153  */
154 int
155 tcp_output(tp)
156 	struct tcpcb *tp;
157 {
158 	struct inpcb * const inp = tp->t_inpcb;
159 	struct socket *so = inp->inp_socket;
160 	long len, recvwin, sendwin;
161 	int off, flags, error;
162 	struct mbuf *m;
163 	struct ip *ip = NULL;
164 	struct ipovly *ipov = NULL;
165 	struct tcphdr *th;
166 	u_char opt[TCP_MAXOLEN];
167 	unsigned ipoptlen, optlen, hdrlen;
168 	int idle, sendalot;
169 	struct ip6_hdr *ip6 = NULL;
170 #ifdef INET6
171 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
172 #else
173 	const boolean_t isipv6 = FALSE;
174 #endif
175 	struct rmxp_tao *taop;
176 
177 	/*
178 	 * Determine length of data that should be transmitted,
179 	 * and flags that will be used.
180 	 * If there is some data or critical controls (SYN, RST)
181 	 * to send, then transmit; otherwise, investigate further.
182 	 */
183 	if ((tp->snd_max == tp->snd_una) &&
184 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
185 		/*
186 		 * We have been idle for "a while" and no acks are
187 		 * expected to clock out any data we send --
188 		 * slow start to get ack "clock" running again.
189 		 *
190 		 * Set the slow-start flight size depending on whether
191 		 * this is a local network or not.
192 		 */
193 		if ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
194 		    (!isipv6 && in_localaddr(inp->inp_faddr)))
195 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
196 		else
197 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
198 	}
199 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
200 	if (idle && (tp->t_flags & TF_MORETOCOME))
201 		tp->t_flags |= TF_LASTIDLE;
202 	else
203 		tp->t_flags &= ~TF_LASTIDLE;
204 
205 again:
206 	sendalot = 0;
207 	off = tp->snd_nxt - tp->snd_una;
208 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
209 	sendwin = min(sendwin, tp->snd_bwnd);
210 
211 	flags = tcp_outflags[tp->t_state];
212 	/*
213 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
214 	 * state flags.
215 	 */
216 	if (tp->t_flags & TF_NEEDFIN)
217 		flags |= TH_FIN;
218 	if (tp->t_flags & TF_NEEDSYN)
219 		flags |= TH_SYN;
220 
221 	/*
222 	 * If in persist timeout with window of 0, send 1 byte.
223 	 * Otherwise, if window is small but nonzero
224 	 * and timer expired, we will send what we can
225 	 * and go to transmit state.
226 	 */
227 	if (tp->t_flags & TF_FORCE) {
228 		if (sendwin == 0) {
229 			/*
230 			 * If we still have some data to send, then
231 			 * clear the FIN bit.  Usually this would
232 			 * happen below when it realizes that we
233 			 * aren't sending all the data.  However,
234 			 * if we have exactly 1 byte of unsent data,
235 			 * then it won't clear the FIN bit below,
236 			 * and if we are in persist state, we wind
237 			 * up sending the packet without recording
238 			 * that we sent the FIN bit.
239 			 *
240 			 * We can't just blindly clear the FIN bit,
241 			 * because if we don't have any more data
242 			 * to send then the probe will be the FIN
243 			 * itself.
244 			 */
245 			if (off < so->so_snd.sb_cc)
246 				flags &= ~TH_FIN;
247 			sendwin = 1;
248 		} else {
249 			callout_stop(tp->tt_persist);
250 			tp->t_rxtshift = 0;
251 		}
252 	}
253 
254 	/*
255 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
256 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
257 	 * a negative length.  This can also occur when TCP opens up
258 	 * its congestion window while receiving additional duplicate
259 	 * acks after fast-retransmit because TCP will reset snd_nxt
260 	 * to snd_max after the fast-retransmit.
261 	 *
262 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
263 	 * be set to snd_una, the offset will be 0, and the length may
264 	 * wind up 0.
265 	 */
266 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
267 
268 	/*
269 	 * Lop off SYN bit if it has already been sent.  However, if this
270 	 * is SYN-SENT state and if segment contains data and if we don't
271 	 * know that foreign host supports TAO, suppress sending segment.
272 	 */
273 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
274 		flags &= ~TH_SYN;
275 		off--, len++;
276 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
277 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
278 		     taop->tao_ccsent == 0))
279 			return 0;
280 	}
281 
282 	/*
283 	 * Be careful not to send data and/or FIN on SYN segments
284 	 * in cases when no CC option will be sent.
285 	 * This measure is needed to prevent interoperability problems
286 	 * with not fully conformant TCP implementations.
287 	 */
288 	if ((flags & TH_SYN) &&
289 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
290 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
291 		len = 0;
292 		flags &= ~TH_FIN;
293 	}
294 
295 	if (len < 0) {
296 		/*
297 		 * If FIN has been sent but not acked,
298 		 * but we haven't been called to retransmit,
299 		 * len will be < 0.  Otherwise, window shrank
300 		 * after we sent into it.  If window shrank to 0,
301 		 * cancel pending retransmit, pull snd_nxt back
302 		 * to (closed) window, and set the persist timer
303 		 * if it isn't already going.  If the window didn't
304 		 * close completely, just wait for an ACK.
305 		 */
306 		len = 0;
307 		if (sendwin == 0) {
308 			callout_stop(tp->tt_rexmt);
309 			tp->t_rxtshift = 0;
310 			tp->snd_nxt = tp->snd_una;
311 			if (!callout_active(tp->tt_persist))
312 				tcp_setpersist(tp);
313 		}
314 	}
315 
316 	/*
317 	 * len will be >= 0 after this point.  Truncate to the maximum
318 	 * segment length and ensure that FIN is removed if the length
319 	 * no longer contains the last data byte.
320 	 */
321 	if (len > tp->t_maxseg) {
322 		len = tp->t_maxseg;
323 		sendalot = 1;
324 	}
325 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
326 		flags &= ~TH_FIN;
327 
328 	recvwin = sbspace(&so->so_rcv);
329 
330 	/*
331 	 * Sender silly window avoidance.   We transmit under the following
332 	 * conditions when len is non-zero:
333 	 *
334 	 *	- We have a full segment
335 	 *	- This is the last buffer in a write()/send() and we are
336 	 *	  either idle or running NODELAY
337 	 *	- we've timed out (e.g. persist timer)
338 	 *	- we have more then 1/2 the maximum send window's worth of
339 	 *	  data (receiver may be limited the window size)
340 	 *	- we need to retransmit
341 	 */
342 	if (len) {
343 		if (len == tp->t_maxseg)
344 			goto send;
345 		/*
346 		 * NOTE! on localhost connections an 'ack' from the remote
347 		 * end may occur synchronously with the output and cause
348 		 * us to flush a buffer queued with moretocome.  XXX
349 		 *
350 		 * note: the len + off check is almost certainly unnecessary.
351 		 */
352 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
353 		    (idle || (tp->t_flags & TF_NODELAY)) &&
354 		    len + off >= so->so_snd.sb_cc &&
355 		    !(tp->t_flags & TF_NOPUSH)) {
356 			goto send;
357 		}
358 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
359 			goto send;
360 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
361 			goto send;
362 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
363 			goto send;
364 	}
365 
366 	/*
367 	 * Compare available window to amount of window
368 	 * known to peer (as advertised window less
369 	 * next expected input).  If the difference is at least two
370 	 * max size segments, or at least 50% of the maximum possible
371 	 * window, then want to send a window update to peer.
372 	 */
373 	if (recvwin > 0) {
374 		/*
375 		 * "adv" is the amount we can increase the window,
376 		 * taking into account that we are limited by
377 		 * TCP_MAXWIN << tp->rcv_scale.
378 		 */
379 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
380 			(tp->rcv_adv - tp->rcv_nxt);
381 
382 		if (adv >= (long) (2 * tp->t_maxseg))
383 			goto send;
384 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
385 			goto send;
386 	}
387 
388 	/*
389 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
390 	 * is also a catch-all for the retransmit timer timeout case.
391 	 */
392 	if (tp->t_flags & TF_ACKNOW)
393 		goto send;
394 	if ((flags & TH_RST) ||
395 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
396 		goto send;
397 	if (SEQ_GT(tp->snd_up, tp->snd_una))
398 		goto send;
399 	/*
400 	 * If our state indicates that FIN should be sent
401 	 * and we have not yet done so, then we need to send.
402 	 */
403 	if (flags & TH_FIN &&
404 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
405 		goto send;
406 
407 	/*
408 	 * TCP window updates are not reliable, rather a polling protocol
409 	 * using ``persist'' packets is used to insure receipt of window
410 	 * updates.  The three ``states'' for the output side are:
411 	 *	idle			not doing retransmits or persists
412 	 *	persisting		to move a small or zero window
413 	 *	(re)transmitting	and thereby not persisting
414 	 *
415 	 * callout_active(tp->tt_persist)
416 	 *	is true when we are in persist state.
417 	 * The TF_FORCE flag in tp->t_flags
418 	 *	is set when we are called to send a persist packet.
419 	 * callout_active(tp->tt_rexmt)
420 	 *	is set when we are retransmitting
421 	 * The output side is idle when both timers are zero.
422 	 *
423 	 * If send window is too small, there is data to transmit, and no
424 	 * retransmit or persist is pending, then go to persist state.
425 	 * If nothing happens soon, send when timer expires:
426 	 * if window is nonzero, transmit what we can,
427 	 * otherwise force out a byte.
428 	 */
429 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
430 	    !callout_active(tp->tt_persist)) {
431 		tp->t_rxtshift = 0;
432 		tcp_setpersist(tp);
433 	}
434 
435 	/*
436 	 * No reason to send a segment, just return.
437 	 */
438 	return (0);
439 
440 send:
441 	/*
442 	 * Before ESTABLISHED, force sending of initial options
443 	 * unless TCP set not to do any options.
444 	 * NOTE: we assume that the IP/TCP header plus TCP options
445 	 * always fit in a single mbuf, leaving room for a maximum
446 	 * link header, i.e.
447 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
448 	 */
449 	optlen = 0;
450 	if (isipv6)
451 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
452 	else
453 		hdrlen = sizeof(struct tcpiphdr);
454 	if (flags & TH_SYN) {
455 		tp->snd_nxt = tp->iss;
456 		if (!(tp->t_flags & TF_NOOPT)) {
457 			u_short mss;
458 
459 			opt[0] = TCPOPT_MAXSEG;
460 			opt[1] = TCPOLEN_MAXSEG;
461 			mss = htons((u_short) tcp_mssopt(tp));
462 			(void)memcpy(opt + 2, &mss, sizeof(mss));
463 			optlen = TCPOLEN_MAXSEG;
464 
465 			if ((tp->t_flags & TF_REQ_SCALE) &&
466 			    (!(flags & TH_ACK) ||
467 			     (tp->t_flags & TF_RCVD_SCALE))) {
468 				*((u_int32_t *)(opt + optlen)) = htonl(
469 					TCPOPT_NOP << 24 |
470 					TCPOPT_WINDOW << 16 |
471 					TCPOLEN_WINDOW << 8 |
472 					tp->request_r_scale);
473 				optlen += 4;
474 			}
475 		}
476 	}
477 
478 	/*
479 	 * Send a timestamp and echo-reply if this is a SYN and our side
480 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
481 	 * and our peer have sent timestamps in our SYN's.
482 	 */
483 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
484 	    !(flags & TH_RST) &&
485 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
486 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
487 
488 		/* Form timestamp option as shown in appendix A of RFC 1323. */
489 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
490 		*lp++ = htonl(ticks);
491 		*lp   = htonl(tp->ts_recent);
492 		optlen += TCPOLEN_TSTAMP_APPA;
493 	}
494 
495 	/*
496 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
497 	 * options are allowed (!TF_NOOPT) and it's not a RST.
498 	 */
499 	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
500 	     !(flags & TH_RST)) {
501 		switch (flags & (TH_SYN|TH_ACK)) {
502 		/*
503 		 * This is a normal ACK, send CC if we received CC before
504 		 * from our peer.
505 		 */
506 		case TH_ACK:
507 			if (!(tp->t_flags & TF_RCVD_CC))
508 				break;
509 			/*FALLTHROUGH*/
510 
511 		/*
512 		 * We can only get here in T/TCP's SYN_SENT* state, when
513 		 * we're a sending a non-SYN segment without waiting for
514 		 * the ACK of our SYN.  A check above assures that we only
515 		 * do this if our peer understands T/TCP.
516 		 */
517 		case 0:
518 			opt[optlen++] = TCPOPT_NOP;
519 			opt[optlen++] = TCPOPT_NOP;
520 			opt[optlen++] = TCPOPT_CC;
521 			opt[optlen++] = TCPOLEN_CC;
522 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
523 
524 			optlen += 4;
525 			break;
526 
527 		/*
528 		 * This is our initial SYN, check whether we have to use
529 		 * CC or CC.new.
530 		 */
531 		case TH_SYN:
532 			opt[optlen++] = TCPOPT_NOP;
533 			opt[optlen++] = TCPOPT_NOP;
534 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
535 						TCPOPT_CCNEW : TCPOPT_CC;
536 			opt[optlen++] = TCPOLEN_CC;
537 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
538 			optlen += 4;
539 			break;
540 
541 		/*
542 		 * This is a SYN,ACK; send CC and CC.echo if we received
543 		 * CC from our peer.
544 		 */
545 		case (TH_SYN|TH_ACK):
546 			if (tp->t_flags & TF_RCVD_CC) {
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] =
552 					htonl(tp->cc_send);
553 				optlen += 4;
554 				opt[optlen++] = TCPOPT_NOP;
555 				opt[optlen++] = TCPOPT_NOP;
556 				opt[optlen++] = TCPOPT_CCECHO;
557 				opt[optlen++] = TCPOLEN_CC;
558 				*(u_int32_t *)&opt[optlen] =
559 					htonl(tp->cc_recv);
560 				optlen += 4;
561 			}
562 			break;
563 		}
564 	}
565 
566 	hdrlen += optlen;
567 
568 	if (isipv6)
569 		ipoptlen = ip6_optlen(inp);
570 	else {
571 		if (inp->inp_options) {
572 			ipoptlen = inp->inp_options->m_len -
573 			    offsetof(struct ipoption, ipopt_list);
574 		} else {
575 			ipoptlen = 0;
576 		}
577 	}
578 #ifdef IPSEC
579 	ipoptlen += ipsec_hdrsiz_tcp(tp);
580 #endif
581 
582 	/*
583 	 * Adjust data length if insertion of options will
584 	 * bump the packet length beyond the t_maxopd length.
585 	 * Clear the FIN bit because we cut off the tail of
586 	 * the segment.
587 	 */
588 	if (len + optlen + ipoptlen > tp->t_maxopd) {
589 		/*
590 		 * If there is still more to send, don't close the connection.
591 		 */
592 		flags &= ~TH_FIN;
593 		len = tp->t_maxopd - optlen - ipoptlen;
594 		sendalot = 1;
595 	}
596 
597 #ifdef INET6
598 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
599 #else
600 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
601 #endif
602 
603 	/*
604 	 * Grab a header mbuf, attaching a copy of data to
605 	 * be transmitted, and initialize the header from
606 	 * the template for sends on this connection.
607 	 */
608 	if (len) {
609 		if ((tp->t_flags & TF_FORCE) && len == 1)
610 			tcpstat.tcps_sndprobe++;
611 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
612 			tcpstat.tcps_sndrexmitpack++;
613 			tcpstat.tcps_sndrexmitbyte += len;
614 		} else {
615 			tcpstat.tcps_sndpack++;
616 			tcpstat.tcps_sndbyte += len;
617 		}
618 #ifdef notyet
619 		if ((m = m_copypack(so->so_snd.sb_mb, off,
620 		    (int)len, max_linkhdr + hdrlen)) == 0) {
621 			error = ENOBUFS;
622 			goto out;
623 		}
624 		/*
625 		 * m_copypack left space for our hdr; use it.
626 		 */
627 		m->m_len += hdrlen;
628 		m->m_data -= hdrlen;
629 #else
630 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
631 		if (m == NULL) {
632 			error = ENOBUFS;
633 			goto out;
634 		}
635 #ifdef INET6
636 		if (MHLEN < hdrlen + max_linkhdr) {
637 			MCLGET(m, MB_DONTWAIT);
638 			if (!(m->m_flags & M_EXT)) {
639 				m_freem(m);
640 				error = ENOBUFS;
641 				goto out;
642 			}
643 		}
644 #endif
645 		m->m_data += max_linkhdr;
646 		m->m_len = hdrlen;
647 		if (len <= MHLEN - hdrlen - max_linkhdr) {
648 			m_copydata(so->so_snd.sb_mb, off, (int) len,
649 			    mtod(m, caddr_t) + hdrlen);
650 			m->m_len += len;
651 		} else {
652 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
653 			if (m->m_next == 0) {
654 				(void) m_free(m);
655 				error = ENOBUFS;
656 				goto out;
657 			}
658 		}
659 #endif
660 		/*
661 		 * If we're sending everything we've got, set PUSH.
662 		 * (This will keep happy those implementations which only
663 		 * give data to the user when a buffer fills or
664 		 * a PUSH comes in.)
665 		 */
666 		if (off + len == so->so_snd.sb_cc)
667 			flags |= TH_PUSH;
668 	} else {
669 		if (tp->t_flags & TF_ACKNOW)
670 			tcpstat.tcps_sndacks++;
671 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
672 			tcpstat.tcps_sndctrl++;
673 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
674 			tcpstat.tcps_sndurg++;
675 		else
676 			tcpstat.tcps_sndwinup++;
677 
678 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
679 		if (m == NULL) {
680 			error = ENOBUFS;
681 			goto out;
682 		}
683 		if (isipv6 &&
684 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
685 			MH_ALIGN(m, hdrlen);
686 		else
687 			m->m_data += max_linkhdr;
688 		m->m_len = hdrlen;
689 	}
690 	m->m_pkthdr.rcvif = (struct ifnet *)0;
691 	if (isipv6) {
692 		ip6 = mtod(m, struct ip6_hdr *);
693 		th = (struct tcphdr *)(ip6 + 1);
694 		tcp_fillheaders(tp, ip6, th);
695 	} else {
696 		ip = mtod(m, struct ip *);
697 		ipov = (struct ipovly *)ip;
698 		th = (struct tcphdr *)(ip + 1);
699 		/* this picks up the pseudo header (w/o the length) */
700 		tcp_fillheaders(tp, ip, th);
701 	}
702 
703 	/*
704 	 * Fill in fields, remembering maximum advertised
705 	 * window for use in delaying messages about window sizes.
706 	 * If resending a FIN, be sure not to use a new sequence number.
707 	 */
708 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
709 	    tp->snd_nxt == tp->snd_max)
710 		tp->snd_nxt--;
711 	/*
712 	 * If we are doing retransmissions, then snd_nxt will
713 	 * not reflect the first unsent octet.  For ACK only
714 	 * packets, we do not want the sequence number of the
715 	 * retransmitted packet, we want the sequence number
716 	 * of the next unsent octet.  So, if there is no data
717 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
718 	 * when filling in ti_seq.  But if we are in persist
719 	 * state, snd_max might reflect one byte beyond the
720 	 * right edge of the window, so use snd_nxt in that
721 	 * case, since we know we aren't doing a retransmission.
722 	 * (retransmit and persist are mutually exclusive...)
723 	 */
724 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
725 		th->th_seq = htonl(tp->snd_nxt);
726 	else
727 		th->th_seq = htonl(tp->snd_max);
728 	th->th_ack = htonl(tp->rcv_nxt);
729 	if (optlen) {
730 		bcopy(opt, th + 1, optlen);
731 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
732 	}
733 	th->th_flags = flags;
734 	/*
735 	 * Calculate receive window.  Don't shrink window,
736 	 * but avoid silly window syndrome.
737 	 */
738 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
739 	    recvwin < (long)tp->t_maxseg)
740 		recvwin = 0;
741 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
742 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
743 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
744 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
745 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
746 
747 	/*
748 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
749 	 * a 0 window.  This may cause the remote transmitter to stall.  This
750 	 * flag tells soreceive() to disable delayed acknowledgements when
751 	 * draining the buffer.  This can occur if the receiver is attempting
752 	 * to read more data then can be buffered prior to transmitting on
753 	 * the connection.
754 	 */
755 	if (recvwin == 0)
756 		tp->t_flags |= TF_RXWIN0SENT;
757 	else
758 		tp->t_flags &= ~TF_RXWIN0SENT;
759 
760 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
761 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
762 		th->th_flags |= TH_URG;
763 	} else {
764 		/*
765 		 * If no urgent pointer to send, then we pull
766 		 * the urgent pointer to the left edge of the send window
767 		 * so that it doesn't drift into the send window on sequence
768 		 * number wraparound.
769 		 */
770 		tp->snd_up = tp->snd_una;		/* drag it along */
771 	}
772 
773 	/*
774 	 * Put TCP length in extended header, and then
775 	 * checksum extended header and data.
776 	 */
777 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
778 	if (isipv6) {
779 		/*
780 		 * ip6_plen is not need to be filled now, and will be filled
781 		 * in ip6_output.
782 		 */
783 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
784 				       sizeof(struct tcphdr) + optlen + len);
785 	} else {
786 		m->m_pkthdr.csum_flags = CSUM_TCP;
787 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
788 		if (len + optlen)
789 			th->th_sum = in_addword(th->th_sum,
790 						htons((u_short)(optlen + len)));
791 
792 		/* IP version must be set here for ipv4/ipv6 checking later */
793 		KASSERT(ip->ip_v == IPVERSION,
794 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
795 	}
796 
797 	/*
798 	 * In transmit state, time the transmission and arrange for
799 	 * the retransmit.  In persist state, just set snd_max.
800 	 */
801 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
802 		tcp_seq startseq = tp->snd_nxt;
803 
804 		/*
805 		 * Advance snd_nxt over sequence space of this segment.
806 		 */
807 		if (flags & (TH_SYN|TH_FIN)) {
808 			if (flags & TH_SYN)
809 				tp->snd_nxt++;
810 			if (flags & TH_FIN) {
811 				tp->snd_nxt++;
812 				tp->t_flags |= TF_SENTFIN;
813 			}
814 		}
815 		tp->snd_nxt += len;
816 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
817 			tp->snd_max = tp->snd_nxt;
818 			/*
819 			 * Time this transmission if not a retransmission and
820 			 * not currently timing anything.
821 			 */
822 			if (tp->t_rtttime == 0) {
823 				tp->t_rtttime = ticks;
824 				tp->t_rtseq = startseq;
825 				tcpstat.tcps_segstimed++;
826 			}
827 		}
828 
829 		/*
830 		 * Set retransmit timer if not currently set,
831 		 * and not doing a pure ack or a keep-alive probe.
832 		 * Initial value for retransmit timer is smoothed
833 		 * round-trip time + 2 * round-trip time variance.
834 		 * Initialize shift counter which is used for backoff
835 		 * of retransmit time.
836 		 */
837 		if (!callout_active(tp->tt_rexmt) &&
838 		    tp->snd_nxt != tp->snd_una) {
839 			if (callout_active(tp->tt_persist)) {
840 				callout_stop(tp->tt_persist);
841 				tp->t_rxtshift = 0;
842 			}
843 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
844 				      tcp_timer_rexmt, tp);
845 		}
846 	} else {
847 		/*
848 		 * Persist case, update snd_max but since we are in
849 		 * persist mode (no window) we do not update snd_nxt.
850 		 */
851 		int xlen = len;
852 		if (flags & TH_SYN)
853 			++xlen;
854 		if (flags & TH_FIN) {
855 			++xlen;
856 			tp->t_flags |= TF_SENTFIN;
857 		}
858 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
859 			tp->snd_max = tp->snd_nxt + xlen;
860 	}
861 
862 #ifdef TCPDEBUG
863 	/*
864 	 * Trace.
865 	 */
866 	if (so->so_options & SO_DEBUG)
867 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
868 #endif
869 
870 	/*
871 	 * Fill in IP length and desired time to live and
872 	 * send to IP level.  There should be a better way
873 	 * to handle ttl and tos; we could keep them in
874 	 * the template, but need a way to checksum without them.
875 	 */
876 	/*
877 	 * m->m_pkthdr.len should have been set before cksum calcuration,
878 	 * because in6_cksum() need it.
879 	 */
880 	if (isipv6) {
881 		/*
882 		 * we separately set hoplimit for every segment, since the
883 		 * user might want to change the value via setsockopt.
884 		 * Also, desired default hop limit might be changed via
885 		 * Neighbor Discovery.
886 		 */
887 		ip6->ip6_hlim = in6_selecthlim(inp,
888 		    (inp->in6p_route.ro_rt ?
889 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
890 
891 		/* TODO: IPv6 IP6TOS_ECT bit on */
892 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
893 		    (so->so_options & SO_DONTROUTE), NULL, NULL, inp);
894 	} else {
895 		struct rtentry *rt;
896 		ip->ip_len = m->m_pkthdr.len;
897 #ifdef INET6
898 		if (INP_CHECK_SOCKAF(so, AF_INET6))
899 			ip->ip_ttl = in6_selecthlim(inp,
900 			    (inp->in6p_route.ro_rt ?
901 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
902 		else
903 #endif
904 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
905 
906 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
907 		/*
908 		 * See if we should do MTU discovery.
909 		 * We do it only if the following are true:
910 		 *	1) we have a valid route to the destination
911 		 *	2) the MTU is not locked (if it is,
912 		 *	   then discovery has been disabled)
913 		 */
914 		if (path_mtu_discovery &&
915 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
916 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
917 			ip->ip_off |= IP_DF;
918 
919 		error = ip_output(m, inp->inp_options, &inp->inp_route,
920 		    (so->so_options & SO_DONTROUTE), NULL, inp);
921 	}
922 	if (error) {
923 
924 		/*
925 		 * We know that the packet was lost, so back out the
926 		 * sequence number advance, if any.
927 		 */
928 		if (!(tp->t_flags & TF_FORCE) ||
929 		    !callout_active(tp->tt_persist)) {
930 			/*
931 			 * No need to check for TH_FIN here because
932 			 * the TF_SENTFIN flag handles that case.
933 			 */
934 			if (!(flags & TH_SYN))
935 				tp->snd_nxt -= len;
936 		}
937 
938 out:
939 		if (error == ENOBUFS) {
940 			/*
941 			 * If we can't send, make sure there is something
942 			 * to get us going again later.  Persist state
943 			 * is not necessarily right, but it is close enough.
944 			 */
945 			if (!callout_active(tp->tt_rexmt) &&
946 			    !callout_active(tp->tt_persist)) {
947 				tp->t_rxtshift = 0;
948 				tcp_setpersist(tp);
949 			}
950 			tcp_quench(inp, 0);
951 			return (0);
952 		}
953 		if (error == EMSGSIZE) {
954 			/*
955 			 * ip_output() will have already fixed the route
956 			 * for us.  tcp_mtudisc() will, as its last action,
957 			 * initiate retransmission, so it is important to
958 			 * not do so here.
959 			 */
960 			tcp_mtudisc(inp, 0);
961 			return 0;
962 		}
963 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
964 		    TCPS_HAVERCVDSYN(tp->t_state)) {
965 			tp->t_softerror = error;
966 			return (0);
967 		}
968 		return (error);
969 	}
970 	tcpstat.tcps_sndtotal++;
971 
972 	/*
973 	 * Data sent (as far as we can tell).
974 	 * If this advertises a larger window than any other segment,
975 	 * then remember the size of the advertised window.
976 	 * Any pending ACK has now been sent.
977 	 */
978 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
979 		tp->rcv_adv = tp->rcv_nxt + recvwin;
980 	tp->last_ack_sent = tp->rcv_nxt;
981 	tp->t_flags &= ~TF_ACKNOW;
982 	if (tcp_delack_enabled)
983 		callout_stop(tp->tt_delack);
984 	if (sendalot)
985 		goto again;
986 	return (0);
987 }
988 
989 void
990 tcp_setpersist(tp)
991 	struct tcpcb *tp;
992 {
993 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
994 	int tt;
995 
996 	if (callout_active(tp->tt_rexmt))
997 		panic("tcp_setpersist: retransmit pending");
998 	/*
999 	 * Start/restart persistance timer.
1000 	 */
1001 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1002 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1003 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1004 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1005 		tp->t_rxtshift++;
1006 }
1007