xref: /dflybsd-src/sys/netinet/tcp_output.c (revision e9cb6d995373cab0661e6e4f3f2cd2c8b6459c11)
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 1.28 2005/05/10 15:48:10 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 
182 	/*
183 	 * If we have been idle for a while, the send congestion window
184 	 * could be no longer representative of the current state of the link.
185 	 * So unless we are expecting more acks to come in, slow-start from
186 	 * scratch to re-determine the send congestion window.
187 	 */
188 	if (tp->snd_max == tp->snd_una &&
189 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
190 		if (tcp_do_rfc3390) {
191 			int initial_cwnd =
192 			    min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380));
193 
194 			tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd);
195 		} else {
196 			tp->snd_cwnd = tp->t_maxseg;
197 		}
198 		tp->snd_wacked = 0;
199 	}
200 
201 	/*
202 	 * Calculate whether the transmit stream was previously idle
203 	 * and adjust TF_LASTIDLE for the next time.
204 	 */
205 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
206 	if (idle && (tp->t_flags & TF_MORETOCOME))
207 		tp->t_flags |= TF_LASTIDLE;
208 	else
209 		tp->t_flags &= ~TF_LASTIDLE;
210 
211 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
212 	    !IN_FASTRECOVERY(tp))
213 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
214 
215 again:
216 	/* Make use of SACK information when slow-starting after a RTO. */
217 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
218 	    !IN_FASTRECOVERY(tp)) {
219 		tcp_seq old_snd_nxt = tp->snd_nxt;
220 
221 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
222 		nsacked += tp->snd_nxt - old_snd_nxt;
223 	}
224 
225 	sendalot = FALSE;
226 	off = tp->snd_nxt - tp->snd_una;
227 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
228 	sendwin = min(sendwin, tp->snd_bwnd);
229 
230 	flags = tcp_outflags[tp->t_state];
231 	/*
232 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
233 	 * state flags.
234 	 */
235 	if (tp->t_flags & TF_NEEDFIN)
236 		flags |= TH_FIN;
237 	if (tp->t_flags & TF_NEEDSYN)
238 		flags |= TH_SYN;
239 
240 	/*
241 	 * If in persist timeout with window of 0, send 1 byte.
242 	 * Otherwise, if window is small but nonzero
243 	 * and timer expired, we will send what we can
244 	 * and go to transmit state.
245 	 */
246 	if (tp->t_flags & TF_FORCE) {
247 		if (sendwin == 0) {
248 			/*
249 			 * If we still have some data to send, then
250 			 * clear the FIN bit.  Usually this would
251 			 * happen below when it realizes that we
252 			 * aren't sending all the data.  However,
253 			 * if we have exactly 1 byte of unsent data,
254 			 * then it won't clear the FIN bit below,
255 			 * and if we are in persist state, we wind
256 			 * up sending the packet without recording
257 			 * that we sent the FIN bit.
258 			 *
259 			 * We can't just blindly clear the FIN bit,
260 			 * because if we don't have any more data
261 			 * to send then the probe will be the FIN
262 			 * itself.
263 			 */
264 			if (off < so->so_snd.sb_cc)
265 				flags &= ~TH_FIN;
266 			sendwin = 1;
267 		} else {
268 			callout_stop(tp->tt_persist);
269 			tp->t_rxtshift = 0;
270 		}
271 	}
272 
273 	/*
274 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
275 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
276 	 * a negative length.  This can also occur when TCP opens up
277 	 * its congestion window while receiving additional duplicate
278 	 * acks after fast-retransmit because TCP will reset snd_nxt
279 	 * to snd_max after the fast-retransmit.
280 	 *
281 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
282 	 * be set to snd_una, the offset will be 0, and the length may
283 	 * wind up 0.
284 	 */
285 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
286 
287 	/*
288 	 * Lop off SYN bit if it has already been sent.  However, if this
289 	 * is SYN-SENT state and if segment contains data and if we don't
290 	 * know that foreign host supports TAO, suppress sending segment.
291 	 */
292 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
293 		flags &= ~TH_SYN;
294 		off--, len++;
295 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
296 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
297 		     taop->tao_ccsent == 0))
298 			return 0;
299 	}
300 
301 	/*
302 	 * Be careful not to send data and/or FIN on SYN segments
303 	 * in cases when no CC option will be sent.
304 	 * This measure is needed to prevent interoperability problems
305 	 * with not fully conformant TCP implementations.
306 	 */
307 	if ((flags & TH_SYN) &&
308 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
309 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
310 		len = 0;
311 		flags &= ~TH_FIN;
312 	}
313 
314 	if (len < 0) {
315 		/*
316 		 * If FIN has been sent but not acked,
317 		 * but we haven't been called to retransmit,
318 		 * len will be < 0.  Otherwise, window shrank
319 		 * after we sent into it.  If window shrank to 0,
320 		 * cancel pending retransmit, pull snd_nxt back
321 		 * to (closed) window, and set the persist timer
322 		 * if it isn't already going.  If the window didn't
323 		 * close completely, just wait for an ACK.
324 		 */
325 		len = 0;
326 		if (sendwin == 0) {
327 			callout_stop(tp->tt_rexmt);
328 			tp->t_rxtshift = 0;
329 			tp->snd_nxt = tp->snd_una;
330 			if (!callout_active(tp->tt_persist))
331 				tcp_setpersist(tp);
332 		}
333 	}
334 
335 	/*
336 	 * len will be >= 0 after this point.  Truncate to the maximum
337 	 * segment length and ensure that FIN is removed if the length
338 	 * no longer contains the last data byte.
339 	 */
340 	if (len > tp->t_maxseg) {
341 		len = tp->t_maxseg;
342 		sendalot = TRUE;
343 	}
344 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
345 		flags &= ~TH_FIN;
346 
347 	recvwin = sbspace(&so->so_rcv);
348 
349 	/*
350 	 * Sender silly window avoidance.   We transmit under the following
351 	 * conditions when len is non-zero:
352 	 *
353 	 *	- We have a full segment
354 	 *	- This is the last buffer in a write()/send() and we are
355 	 *	  either idle or running NODELAY
356 	 *	- we've timed out (e.g. persist timer)
357 	 *	- we have more then 1/2 the maximum send window's worth of
358 	 *	  data (receiver may be limiting the window size)
359 	 *	- we need to retransmit
360 	 */
361 	if (len) {
362 		if (len == tp->t_maxseg)
363 			goto send;
364 		/*
365 		 * NOTE! on localhost connections an 'ack' from the remote
366 		 * end may occur synchronously with the output and cause
367 		 * us to flush a buffer queued with moretocome.  XXX
368 		 *
369 		 * note: the len + off check is almost certainly unnecessary.
370 		 */
371 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
372 		    (idle || (tp->t_flags & TF_NODELAY)) &&
373 		    len + off >= so->so_snd.sb_cc &&
374 		    !(tp->t_flags & TF_NOPUSH)) {
375 			goto send;
376 		}
377 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
378 			goto send;
379 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
380 			goto send;
381 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
382 			goto send;
383 	}
384 
385 	/*
386 	 * Compare available window to amount of window
387 	 * known to peer (as advertised window less
388 	 * next expected input).  If the difference is at least two
389 	 * max size segments, or at least 50% of the maximum possible
390 	 * window, then want to send a window update to peer.
391 	 */
392 	if (recvwin > 0) {
393 		/*
394 		 * "adv" is the amount we can increase the window,
395 		 * taking into account that we are limited by
396 		 * TCP_MAXWIN << tp->rcv_scale.
397 		 */
398 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
399 			(tp->rcv_adv - tp->rcv_nxt);
400 
401 		/*
402 		 * This ack case typically occurs when the user has drained
403 		 * the TCP socket buffer sufficiently to warrent an ack
404 		 * containing a 'pure window update'... that is, an ack that
405 		 * ONLY updates the tcp window.
406 		 *
407 		 * It is unclear why we would need to do a pure window update
408 		 * past 2 segments if we are going to do one at 1/2 the high
409 		 * water mark anyway, especially since under normal conditions
410 		 * the user program will drain the socket buffer quickly.
411 		 * The 2-segment pure window update will often add a large
412 		 * number of extra, unnecessary acks to the stream.
413 		 *
414 		 * avoid_pure_win_update now defaults to 1.
415 		 */
416 		if (avoid_pure_win_update == 0) {
417 			if (adv >= (long) (2 * tp->t_maxseg))
418 				goto send;
419 		}
420 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
421 			goto send;
422 	}
423 
424 	/*
425 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
426 	 * is also a catch-all for the retransmit timer timeout case.
427 	 */
428 	if (tp->t_flags & TF_ACKNOW)
429 		goto send;
430 	if ((flags & TH_RST) ||
431 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
432 		goto send;
433 	if (SEQ_GT(tp->snd_up, tp->snd_una))
434 		goto send;
435 	/*
436 	 * If our state indicates that FIN should be sent
437 	 * and we have not yet done so, then we need to send.
438 	 */
439 	if (flags & TH_FIN &&
440 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
441 		goto send;
442 
443 	/*
444 	 * TCP window updates are not reliable, rather a polling protocol
445 	 * using ``persist'' packets is used to insure receipt of window
446 	 * updates.  The three ``states'' for the output side are:
447 	 *	idle			not doing retransmits or persists
448 	 *	persisting		to move a small or zero window
449 	 *	(re)transmitting	and thereby not persisting
450 	 *
451 	 * callout_active(tp->tt_persist)
452 	 *	is true when we are in persist state.
453 	 * The TF_FORCE flag in tp->t_flags
454 	 *	is set when we are called to send a persist packet.
455 	 * callout_active(tp->tt_rexmt)
456 	 *	is set when we are retransmitting
457 	 * The output side is idle when both timers are zero.
458 	 *
459 	 * If send window is too small, there is data to transmit, and no
460 	 * retransmit or persist is pending, then go to persist state.
461 	 * If nothing happens soon, send when timer expires:
462 	 * if window is nonzero, transmit what we can,
463 	 * otherwise force out a byte.
464 	 */
465 	if (so->so_snd.sb_cc > 0 &&
466 	    !callout_active(tp->tt_rexmt) && !callout_active(tp->tt_persist)) {
467 		tp->t_rxtshift = 0;
468 		tcp_setpersist(tp);
469 	}
470 
471 	/*
472 	 * No reason to send a segment, just return.
473 	 */
474 	return (0);
475 
476 send:
477 	/*
478 	 * Before ESTABLISHED, force sending of initial options
479 	 * unless TCP set not to do any options.
480 	 * NOTE: we assume that the IP/TCP header plus TCP options
481 	 * always fit in a single mbuf, leaving room for a maximum
482 	 * link header, i.e.
483 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
484 	 */
485 	optlen = 0;
486 	if (isipv6)
487 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
488 	else
489 		hdrlen = sizeof(struct tcpiphdr);
490 	if (flags & TH_SYN) {
491 		tp->snd_nxt = tp->iss;
492 		if (!(tp->t_flags & TF_NOOPT)) {
493 			u_short mss;
494 
495 			opt[0] = TCPOPT_MAXSEG;
496 			opt[1] = TCPOLEN_MAXSEG;
497 			mss = htons((u_short) tcp_mssopt(tp));
498 			memcpy(opt + 2, &mss, sizeof mss);
499 			optlen = TCPOLEN_MAXSEG;
500 
501 			if ((tp->t_flags & TF_REQ_SCALE) &&
502 			    (!(flags & TH_ACK) ||
503 			     (tp->t_flags & TF_RCVD_SCALE))) {
504 				*((u_int32_t *)(opt + optlen)) = htonl(
505 					TCPOPT_NOP << 24 |
506 					TCPOPT_WINDOW << 16 |
507 					TCPOLEN_WINDOW << 8 |
508 					tp->request_r_scale);
509 				optlen += 4;
510 			}
511 
512 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
513 			    tp->t_flags & TF_SACK_PERMITTED) {
514 				uint32_t *lp = (uint32_t *)(opt + optlen);
515 
516 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
517 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
518 			}
519 		}
520 	}
521 
522 	/*
523 	 * Send a timestamp and echo-reply if this is a SYN and our side
524 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
525 	 * and our peer have sent timestamps in our SYN's.
526 	 */
527 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
528 	    !(flags & TH_RST) &&
529 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
530 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
531 
532 		/* Form timestamp option as shown in appendix A of RFC 1323. */
533 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
534 		*lp++ = htonl(ticks);
535 		*lp   = htonl(tp->ts_recent);
536 		optlen += TCPOLEN_TSTAMP_APPA;
537 	}
538 
539 	/*
540 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
541 	 * options are allowed (!TF_NOOPT) and it's not a RST.
542 	 */
543 	if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
544 	     !(flags & TH_RST)) {
545 		switch (flags & (TH_SYN | TH_ACK)) {
546 		/*
547 		 * This is a normal ACK, send CC if we received CC before
548 		 * from our peer.
549 		 */
550 		case TH_ACK:
551 			if (!(tp->t_flags & TF_RCVD_CC))
552 				break;
553 			/*FALLTHROUGH*/
554 
555 		/*
556 		 * We can only get here in T/TCP's SYN_SENT* state, when
557 		 * we're a sending a non-SYN segment without waiting for
558 		 * the ACK of our SYN.  A check above assures that we only
559 		 * do this if our peer understands T/TCP.
560 		 */
561 		case 0:
562 			opt[optlen++] = TCPOPT_NOP;
563 			opt[optlen++] = TCPOPT_NOP;
564 			opt[optlen++] = TCPOPT_CC;
565 			opt[optlen++] = TCPOLEN_CC;
566 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
567 			optlen += 4;
568 			break;
569 
570 		/*
571 		 * This is our initial SYN, check whether we have to use
572 		 * CC or CC.new.
573 		 */
574 		case TH_SYN:
575 			opt[optlen++] = TCPOPT_NOP;
576 			opt[optlen++] = TCPOPT_NOP;
577 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
578 						TCPOPT_CCNEW : TCPOPT_CC;
579 			opt[optlen++] = TCPOLEN_CC;
580 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
581 			optlen += 4;
582 			break;
583 
584 		/*
585 		 * This is a SYN,ACK; send CC and CC.echo if we received
586 		 * CC from our peer.
587 		 */
588 		case (TH_SYN | TH_ACK):
589 			if (tp->t_flags & TF_RCVD_CC) {
590 				opt[optlen++] = TCPOPT_NOP;
591 				opt[optlen++] = TCPOPT_NOP;
592 				opt[optlen++] = TCPOPT_CC;
593 				opt[optlen++] = TCPOLEN_CC;
594 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
595 				optlen += 4;
596 				opt[optlen++] = TCPOPT_NOP;
597 				opt[optlen++] = TCPOPT_NOP;
598 				opt[optlen++] = TCPOPT_CCECHO;
599 				opt[optlen++] = TCPOLEN_CC;
600 				*(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv);
601 				optlen += 4;
602 			}
603 			break;
604 		}
605 	}
606 
607 	/*
608 	 * If this is a SACK connection and we have a block to report,
609 	 * fill in the SACK blocks in the TCP options.
610 	 */
611 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
612 		TF_SACK_PERMITTED &&
613 	    (!LIST_EMPTY(&tp->t_segq) ||
614 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
615 		tcp_sack_fill_report(tp, opt, &optlen);
616 
617 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
618 	hdrlen += optlen;
619 
620 	if (isipv6) {
621 		ipoptlen = ip6_optlen(inp);
622 	} else {
623 		if (inp->inp_options) {
624 			ipoptlen = inp->inp_options->m_len -
625 			    offsetof(struct ipoption, ipopt_list);
626 		} else {
627 			ipoptlen = 0;
628 		}
629 	}
630 #ifdef IPSEC
631 	ipoptlen += ipsec_hdrsiz_tcp(tp);
632 #endif
633 
634 	/*
635 	 * Adjust data length if insertion of options will
636 	 * bump the packet length beyond the t_maxopd length.
637 	 * Clear the FIN bit because we cut off the tail of
638 	 * the segment.
639 	 */
640 	if (len + optlen + ipoptlen > tp->t_maxopd) {
641 		/*
642 		 * If there is still more to send, don't close the connection.
643 		 */
644 		flags &= ~TH_FIN;
645 		len = tp->t_maxopd - optlen - ipoptlen;
646 		sendalot = TRUE;
647 	}
648 
649 #ifdef INET6
650 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
651 #else
652 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
653 #endif
654 
655 	/*
656 	 * Grab a header mbuf, attaching a copy of data to
657 	 * be transmitted, and initialize the header from
658 	 * the template for sends on this connection.
659 	 */
660 	if (len) {
661 		if ((tp->t_flags & TF_FORCE) && len == 1)
662 			tcpstat.tcps_sndprobe++;
663 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
664 			if (tp->snd_nxt == tp->snd_una)
665 				tp->snd_max_rexmt = tp->snd_max;
666 			tcpstat.tcps_sndrexmitpack++;
667 			tcpstat.tcps_sndrexmitbyte += len;
668 		} else {
669 			tcpstat.tcps_sndpack++;
670 			tcpstat.tcps_sndbyte += len;
671 		}
672 #ifdef notyet
673 		if ((m = m_copypack(so->so_snd.sb_mb, off, (int)len,
674 		    max_linkhdr + hdrlen)) == NULL) {
675 			error = ENOBUFS;
676 			goto out;
677 		}
678 		/*
679 		 * m_copypack left space for our hdr; use it.
680 		 */
681 		m->m_len += hdrlen;
682 		m->m_data -= hdrlen;
683 #else
684 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
685 		if (m == NULL) {
686 			error = ENOBUFS;
687 			goto out;
688 		}
689 #ifdef INET6
690 		if (MHLEN < hdrlen + max_linkhdr) {
691 			MCLGET(m, MB_DONTWAIT);
692 			if (!(m->m_flags & M_EXT)) {
693 				m_freem(m);
694 				error = ENOBUFS;
695 				goto out;
696 			}
697 		}
698 #endif
699 		m->m_data += max_linkhdr;
700 		m->m_len = hdrlen;
701 		if (len <= MHLEN - hdrlen - max_linkhdr) {
702 			m_copydata(so->so_snd.sb_mb, off, (int) len,
703 			    mtod(m, caddr_t) + hdrlen);
704 			m->m_len += len;
705 		} else {
706 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
707 			if (m->m_next == NULL) {
708 				m_free(m);
709 				error = ENOBUFS;
710 				goto out;
711 			}
712 		}
713 #endif
714 		/*
715 		 * If we're sending everything we've got, set PUSH.
716 		 * (This will keep happy those implementations which only
717 		 * give data to the user when a buffer fills or
718 		 * a PUSH comes in.)
719 		 */
720 		if (off + len == so->so_snd.sb_cc)
721 			flags |= TH_PUSH;
722 	} else {
723 		if (tp->t_flags & TF_ACKNOW)
724 			tcpstat.tcps_sndacks++;
725 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
726 			tcpstat.tcps_sndctrl++;
727 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
728 			tcpstat.tcps_sndurg++;
729 		else
730 			tcpstat.tcps_sndwinup++;
731 
732 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
733 		if (m == NULL) {
734 			error = ENOBUFS;
735 			goto out;
736 		}
737 		if (isipv6 &&
738 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
739 			MH_ALIGN(m, hdrlen);
740 		else
741 			m->m_data += max_linkhdr;
742 		m->m_len = hdrlen;
743 	}
744 	m->m_pkthdr.rcvif = (struct ifnet *)0;
745 	if (isipv6) {
746 		ip6 = mtod(m, struct ip6_hdr *);
747 		th = (struct tcphdr *)(ip6 + 1);
748 		tcp_fillheaders(tp, ip6, th);
749 	} else {
750 		ip = mtod(m, struct ip *);
751 		ipov = (struct ipovly *)ip;
752 		th = (struct tcphdr *)(ip + 1);
753 		/* this picks up the pseudo header (w/o the length) */
754 		tcp_fillheaders(tp, ip, th);
755 	}
756 
757 	/*
758 	 * Fill in fields, remembering maximum advertised
759 	 * window for use in delaying messages about window sizes.
760 	 * If resending a FIN, be sure not to use a new sequence number.
761 	 */
762 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
763 	    tp->snd_nxt == tp->snd_max)
764 		tp->snd_nxt--;
765 	/*
766 	 * If we are doing retransmissions, then snd_nxt will
767 	 * not reflect the first unsent octet.  For ACK only
768 	 * packets, we do not want the sequence number of the
769 	 * retransmitted packet, we want the sequence number
770 	 * of the next unsent octet.  So, if there is no data
771 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
772 	 * when filling in ti_seq.  But if we are in persist
773 	 * state, snd_max might reflect one byte beyond the
774 	 * right edge of the window, so use snd_nxt in that
775 	 * case, since we know we aren't doing a retransmission.
776 	 * (retransmit and persist are mutually exclusive...)
777 	 */
778 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
779 		th->th_seq = htonl(tp->snd_nxt);
780 	else
781 		th->th_seq = htonl(tp->snd_max);
782 	th->th_ack = htonl(tp->rcv_nxt);
783 	if (optlen) {
784 		bcopy(opt, th + 1, optlen);
785 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
786 	}
787 	th->th_flags = flags;
788 	/*
789 	 * Calculate receive window.  Don't shrink window,
790 	 * but avoid silly window syndrome.
791 	 */
792 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
793 	    recvwin < (long)tp->t_maxseg)
794 		recvwin = 0;
795 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
796 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
797 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
798 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
799 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
800 
801 	/*
802 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
803 	 * a 0 window.  This may cause the remote transmitter to stall.  This
804 	 * flag tells soreceive() to disable delayed acknowledgements when
805 	 * draining the buffer.  This can occur if the receiver is attempting
806 	 * to read more data then can be buffered prior to transmitting on
807 	 * the connection.
808 	 */
809 	if (recvwin == 0)
810 		tp->t_flags |= TF_RXWIN0SENT;
811 	else
812 		tp->t_flags &= ~TF_RXWIN0SENT;
813 
814 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
815 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
816 		th->th_flags |= TH_URG;
817 	} else {
818 		/*
819 		 * If no urgent pointer to send, then we pull
820 		 * the urgent pointer to the left edge of the send window
821 		 * so that it doesn't drift into the send window on sequence
822 		 * number wraparound.
823 		 */
824 		tp->snd_up = tp->snd_una;		/* drag it along */
825 	}
826 
827 	/*
828 	 * Put TCP length in extended header, and then
829 	 * checksum extended header and data.
830 	 */
831 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
832 	if (isipv6) {
833 		/*
834 		 * ip6_plen is not need to be filled now, and will be filled
835 		 * in ip6_output().
836 		 */
837 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
838 				       sizeof(struct tcphdr) + optlen + len);
839 	} else {
840 		m->m_pkthdr.csum_flags = CSUM_TCP;
841 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
842 		if (len + optlen)
843 			th->th_sum = in_addword(th->th_sum,
844 						htons((u_short)(optlen + len)));
845 
846 		/* IP version must be set here for ipv4/ipv6 checking later */
847 		KASSERT(ip->ip_v == IPVERSION,
848 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
849 	}
850 
851 	/*
852 	 * In transmit state, time the transmission and arrange for
853 	 * the retransmit.  In persist state, just set snd_max.
854 	 */
855 	if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
856 		tcp_seq startseq = tp->snd_nxt;
857 
858 		/*
859 		 * Advance snd_nxt over sequence space of this segment.
860 		 */
861 		if (flags & (TH_SYN | TH_FIN)) {
862 			if (flags & TH_SYN)
863 				tp->snd_nxt++;
864 			if (flags & TH_FIN) {
865 				tp->snd_nxt++;
866 				tp->t_flags |= TF_SENTFIN;
867 			}
868 		}
869 		tp->snd_nxt += len;
870 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
871 			tp->snd_max = tp->snd_nxt;
872 			/*
873 			 * Time this transmission if not a retransmission and
874 			 * not currently timing anything.
875 			 */
876 			if (tp->t_rtttime == 0) {
877 				tp->t_rtttime = ticks;
878 				tp->t_rtseq = startseq;
879 				tcpstat.tcps_segstimed++;
880 			}
881 		}
882 
883 		/*
884 		 * Set retransmit timer if not currently set,
885 		 * and not doing a pure ack or a keep-alive probe.
886 		 * Initial value for retransmit timer is smoothed
887 		 * round-trip time + 2 * round-trip time variance.
888 		 * Initialize shift counter which is used for backoff
889 		 * of retransmit time.
890 		 */
891 		if (!callout_active(tp->tt_rexmt) &&
892 		    tp->snd_nxt != tp->snd_una) {
893 			if (callout_active(tp->tt_persist)) {
894 				callout_stop(tp->tt_persist);
895 				tp->t_rxtshift = 0;
896 			}
897 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
898 				      tcp_timer_rexmt, tp);
899 		}
900 	} else {
901 		/*
902 		 * Persist case, update snd_max but since we are in
903 		 * persist mode (no window) we do not update snd_nxt.
904 		 */
905 		int xlen = len;
906 		if (flags & TH_SYN)
907 			++xlen;
908 		if (flags & TH_FIN) {
909 			++xlen;
910 			tp->t_flags |= TF_SENTFIN;
911 		}
912 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
913 			tp->snd_max = tp->snd_nxt + xlen;
914 	}
915 
916 #ifdef TCPDEBUG
917 	/*
918 	 * Trace.
919 	 */
920 	if (so->so_options & SO_DEBUG)
921 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
922 #endif
923 
924 	/*
925 	 * Fill in IP length and desired time to live and
926 	 * send to IP level.  There should be a better way
927 	 * to handle ttl and tos; we could keep them in
928 	 * the template, but need a way to checksum without them.
929 	 */
930 	/*
931 	 * m->m_pkthdr.len should have been set before cksum calcuration,
932 	 * because in6_cksum() need it.
933 	 */
934 	if (isipv6) {
935 		/*
936 		 * we separately set hoplimit for every segment, since the
937 		 * user might want to change the value via setsockopt.
938 		 * Also, desired default hop limit might be changed via
939 		 * Neighbor Discovery.
940 		 */
941 		ip6->ip6_hlim = in6_selecthlim(inp,
942 		    (inp->in6p_route.ro_rt ?
943 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
944 
945 		/* TODO: IPv6 IP6TOS_ECT bit on */
946 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
947 				   (so->so_options & SO_DONTROUTE), NULL, NULL,
948 				   inp);
949 	} else {
950 		struct rtentry *rt;
951 		ip->ip_len = m->m_pkthdr.len;
952 #ifdef INET6
953 		if (INP_CHECK_SOCKAF(so, AF_INET6))
954 			ip->ip_ttl = in6_selecthlim(inp,
955 			    (inp->in6p_route.ro_rt ?
956 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
957 		else
958 #endif
959 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
960 
961 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
962 		/*
963 		 * See if we should do MTU discovery.
964 		 * We do it only if the following are true:
965 		 *	1) we have a valid route to the destination
966 		 *	2) the MTU is not locked (if it is,
967 		 *	   then discovery has been disabled)
968 		 */
969 		if (path_mtu_discovery &&
970 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
971 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
972 			ip->ip_off |= IP_DF;
973 
974 		error = ip_output(m, inp->inp_options, &inp->inp_route,
975 				  (so->so_options & SO_DONTROUTE), NULL, inp);
976 	}
977 	if (error) {
978 
979 		/*
980 		 * We know that the packet was lost, so back out the
981 		 * sequence number advance, if any.
982 		 */
983 		if (!(tp->t_flags & TF_FORCE) ||
984 		    !callout_active(tp->tt_persist)) {
985 			/*
986 			 * No need to check for TH_FIN here because
987 			 * the TF_SENTFIN flag handles that case.
988 			 */
989 			if (!(flags & TH_SYN))
990 				tp->snd_nxt -= len;
991 		}
992 
993 out:
994 		if (error == ENOBUFS) {
995 			/*
996 			 * If we can't send, make sure there is something
997 			 * to get us going again later.  Persist state
998 			 * is not necessarily right, but it is close enough.
999 			 */
1000 			if (!callout_active(tp->tt_rexmt) &&
1001 			    !callout_active(tp->tt_persist)) {
1002 				tp->t_rxtshift = 0;
1003 				tcp_setpersist(tp);
1004 			}
1005 			tcp_quench(inp, 0);
1006 			return (0);
1007 		}
1008 		if (error == EMSGSIZE) {
1009 			/*
1010 			 * ip_output() will have already fixed the route
1011 			 * for us.  tcp_mtudisc() will, as its last action,
1012 			 * initiate retransmission, so it is important to
1013 			 * not do so here.
1014 			 */
1015 			tcp_mtudisc(inp, 0);
1016 			return 0;
1017 		}
1018 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1019 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1020 			tp->t_softerror = error;
1021 			return (0);
1022 		}
1023 		return (error);
1024 	}
1025 	tcpstat.tcps_sndtotal++;
1026 
1027 	/*
1028 	 * Data sent (as far as we can tell).
1029 	 * If this advertises a larger window than any other segment,
1030 	 * then remember the size of the advertised window.
1031 	 * Any pending ACK has now been sent.
1032 	 */
1033 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
1034 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1035 	tp->last_ack_sent = tp->rcv_nxt;
1036 	tp->t_flags &= ~TF_ACKNOW;
1037 	if (tcp_delack_enabled)
1038 		callout_stop(tp->tt_delack);
1039 	if (sendalot)
1040 		goto again;
1041 	return (0);
1042 }
1043 
1044 void
1045 tcp_setpersist(tp)
1046 	struct tcpcb *tp;
1047 {
1048 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1049 	int tt;
1050 
1051 	if (callout_active(tp->tt_rexmt))
1052 		panic("tcp_setpersist: retransmit pending");
1053 	/*
1054 	 * Start/restart persistance timer.
1055 	 */
1056 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1057 		      TCPTV_PERSMAX);
1058 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1059 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1060 		tp->t_rxtshift++;
1061 }
1062