xref: /netbsd-src/sys/netinet/tcp_output.c (revision 5aefcfdc06931dd97e76246d2fe0302f7b3fe094)
1 /*	$NetBSD: tcp_output.c,v 1.62 2000/11/06 00:50:12 itojun Exp $	*/
2 
3 /*
4 %%% portions-copyright-nrl-95
5 Portions of this software are Copyright 1995-1998 by Randall Atkinson,
6 Ronald Lee, Daniel McDonald, Bao Phan, and Chris Winters. All Rights
7 Reserved. All rights under this copyright have been assigned to the US
8 Naval Research Laboratory (NRL). The NRL Copyright Notice and License
9 Agreement Version 1.1 (January 17, 1995) applies to these portions of the
10 software.
11 You should have received a copy of the license with this software. If you
12 didn't get a copy, you may request one from <license@ipv6.nrl.navy.mil>.
13 
14 */
15 
16 /*
17  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
18  * All rights reserved.
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  * 1. Redistributions of source code must retain the above copyright
24  *    notice, this list of conditions and the following disclaimer.
25  * 2. Redistributions in binary form must reproduce the above copyright
26  *    notice, this list of conditions and the following disclaimer in the
27  *    documentation and/or other materials provided with the distribution.
28  * 3. Neither the name of the project nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  */
44 
45 /*-
46  * Copyright (c) 1997, 1998 The NetBSD Foundation, Inc.
47  * All rights reserved.
48  *
49  * This code is derived from software contributed to The NetBSD Foundation
50  * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation
51  * Facility, NASA Ames Research Center.
52  *
53  * Redistribution and use in source and binary forms, with or without
54  * modification, are permitted provided that the following conditions
55  * are met:
56  * 1. Redistributions of source code must retain the above copyright
57  *    notice, this list of conditions and the following disclaimer.
58  * 2. Redistributions in binary form must reproduce the above copyright
59  *    notice, this list of conditions and the following disclaimer in the
60  *    documentation and/or other materials provided with the distribution.
61  * 3. All advertising materials mentioning features or use of this software
62  *    must display the following acknowledgement:
63  *	This product includes software developed by the NetBSD
64  *	Foundation, Inc. and its contributors.
65  * 4. Neither the name of The NetBSD Foundation nor the names of its
66  *    contributors may be used to endorse or promote products derived
67  *    from this software without specific prior written permission.
68  *
69  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
70  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
71  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
72  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
73  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
74  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
75  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
76  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
77  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
78  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
79  * POSSIBILITY OF SUCH DAMAGE.
80  */
81 
82 /*
83  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
84  *	The Regents of the University of California.  All rights reserved.
85  *
86  * Redistribution and use in source and binary forms, with or without
87  * modification, are permitted provided that the following conditions
88  * are met:
89  * 1. Redistributions of source code must retain the above copyright
90  *    notice, this list of conditions and the following disclaimer.
91  * 2. Redistributions in binary form must reproduce the above copyright
92  *    notice, this list of conditions and the following disclaimer in the
93  *    documentation and/or other materials provided with the distribution.
94  * 3. All advertising materials mentioning features or use of this software
95  *    must display the following acknowledgement:
96  *	This product includes software developed by the University of
97  *	California, Berkeley and its contributors.
98  * 4. Neither the name of the University nor the names of its contributors
99  *    may be used to endorse or promote products derived from this software
100  *    without specific prior written permission.
101  *
102  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
103  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
104  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
105  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
106  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
107  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
108  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
109  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
110  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
111  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
112  * SUCH DAMAGE.
113  *
114  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
115  */
116 
117 #include "opt_inet.h"
118 #include "opt_ipsec.h"
119 
120 #include <sys/param.h>
121 #include <sys/systm.h>
122 #include <sys/malloc.h>
123 #include <sys/mbuf.h>
124 #include <sys/protosw.h>
125 #include <sys/socket.h>
126 #include <sys/socketvar.h>
127 #include <sys/errno.h>
128 #include <sys/domain.h>
129 
130 #include <net/if.h>
131 #include <net/route.h>
132 
133 #include <netinet/in.h>
134 #include <netinet/in_systm.h>
135 #include <netinet/ip.h>
136 #include <netinet/in_pcb.h>
137 #include <netinet/ip_var.h>
138 
139 #ifdef INET6
140 #ifndef INET
141 #include <netinet/in.h>
142 #endif
143 #include <netinet/ip6.h>
144 #include <netinet6/in6_pcb.h>
145 #include <netinet6/ip6_var.h>
146 #endif
147 
148 #include <netinet/tcp.h>
149 #define	TCPOUTFLAGS
150 #include <netinet/tcp_fsm.h>
151 #include <netinet/tcp_seq.h>
152 #include <netinet/tcp_timer.h>
153 #include <netinet/tcp_var.h>
154 #include <netinet/tcpip.h>
155 #include <netinet/tcp_debug.h>
156 
157 #ifdef notyet
158 extern struct mbuf *m_copypack();
159 #endif
160 
161 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
162 
163 /*
164  * Knob to enable Congestion Window Monitoring, and control the
165  * the burst size it allows.  Default burst is 4 packets, per
166  * the Internet draft.
167  */
168 int	tcp_cwm = 0;
169 int	tcp_cwm_burstsize = 4;
170 
171 static __inline void tcp_segsize __P((struct tcpcb *, int *, int *));
172 static __inline void
173 tcp_segsize(tp, txsegsizep, rxsegsizep)
174 	struct tcpcb *tp;
175 	int *txsegsizep, *rxsegsizep;
176 {
177 #ifdef INET
178 	struct inpcb *inp = tp->t_inpcb;
179 #endif
180 #ifdef INET6
181 	struct in6pcb *in6p = tp->t_in6pcb;
182 #endif
183 	struct rtentry *rt;
184 	struct ifnet *ifp;
185 	int size;
186 	int iphlen;
187 
188 #ifdef DIAGNOSTIC
189 	if (tp->t_inpcb && tp->t_in6pcb)
190 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
191 #endif
192 	switch (tp->t_family) {
193 #ifdef INET
194 	case AF_INET:
195 		iphlen = sizeof(struct ip);
196 		break;
197 #endif
198 #ifdef INET6
199 	case AF_INET6:
200 		iphlen = sizeof(struct ip6_hdr);
201 		break;
202 #endif
203 	default:
204 		size = tcp_mssdflt;
205 		goto out;
206 	}
207 
208 	rt = NULL;
209 #ifdef INET
210 	if (inp)
211 		rt = in_pcbrtentry(inp);
212 #endif
213 #ifdef INET6
214 	if (in6p)
215 		rt = in6_pcbrtentry(in6p);
216 #endif
217 	if (rt == NULL) {
218 		size = tcp_mssdflt;
219 		goto out;
220 	}
221 
222 	ifp = rt->rt_ifp;
223 
224 	size = tcp_mssdflt;
225 	if (rt->rt_rmx.rmx_mtu != 0)
226 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
227 	else if (ifp->if_flags & IFF_LOOPBACK)
228 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
229 #ifdef INET
230 	else if (ip_mtudisc)
231 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
232 	else if (inp && in_localaddr(inp->inp_faddr))
233 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
234 #endif
235 #ifdef INET6
236 	else if (in6p) {
237 #ifdef INET
238 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
239 			/* mapped addr case */
240 			struct in_addr d;
241 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
242 			if (in_localaddr(d))
243 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
244 		} else
245 #endif
246 		{
247 			if (in6_localaddr(&in6p->in6p_faddr))
248 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
249 		}
250 	}
251 #endif
252 	size -= tcp_optlen(tp);
253 	/*
254 	 * XXX tp->t_ourmss should have the right size, but without this code
255 	 * fragmentation will occur... need more investigation
256 	 */
257 #ifdef INET
258 	if (inp) {
259 #ifdef IPSEC
260 		size -= ipsec4_hdrsiz_tcp(tp);
261 #endif
262 		size -= ip_optlen(inp);
263 	}
264 #endif
265 #ifdef INET6
266 #ifdef INET
267 	if (in6p && tp->t_family == AF_INET) {
268 #ifdef IPSEC
269 		size -= ipsec4_hdrsiz_tcp(tp);
270 #endif
271 		/* XXX size -= ip_optlen(in6p); */
272 	} else
273 #endif
274 	if (in6p && tp->t_family == AF_INET6) {
275 #ifdef IPSEC
276 		size -= ipsec6_hdrsiz_tcp(tp);
277 #endif
278 		size -= ip6_optlen(in6p);
279 	}
280 #endif
281 
282  out:
283 	/*
284 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
285 	 * assumes that path MTU is the same for both ways).  this is only
286 	 * for silly window avoidance, do not use the value for other purposes.
287 	 *
288 	 * ipseclen is subtracted from both sides, this may not be right.
289 	 * I'm not quite sure about this (could someone comment).
290 	 */
291 	*txsegsizep = min(tp->t_peermss, size);
292 	*rxsegsizep = min(tp->t_ourmss, size);
293 
294 	if (*txsegsizep != tp->t_segsz) {
295 		/*
296 		 * If the new segment size is larger, we don't want to
297 		 * mess up the congestion window, but if it is smaller
298 		 * we'll have to reduce the congestion window to ensure
299 		 * that we don't get into trouble with initial windows
300 		 * and the rest.  In any case, if the segment size
301 		 * has changed, chances are the path has, too, and
302 		 * our congestion window will be different.
303 		 */
304 		if (*txsegsizep < tp->t_segsz) {
305 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
306 					   * *txsegsizep, *txsegsizep);
307 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
308 						* *txsegsizep, *txsegsizep);
309 		}
310 		tp->t_segsz = *txsegsizep;
311 	}
312 }
313 
314 /*
315  * Tcp output routine: figure out what should be sent and send it.
316  */
317 int
318 tcp_output(tp)
319 	struct tcpcb *tp;
320 {
321 	struct socket *so;
322 	struct route *ro;
323 	long len, win;
324 	int off, flags, error;
325 	struct mbuf *m;
326 	struct ip *ip;
327 #ifdef INET6
328 	struct ip6_hdr *ip6;
329 #endif
330 	struct tcphdr *th;
331 	u_char opt[MAX_TCPOPTLEN];
332 	unsigned optlen, hdrlen;
333 	int idle, sendalot, txsegsize, rxsegsize;
334 	int maxburst = TCP_MAXBURST;
335 	int af;		/* address family on the wire */
336 	int iphdrlen;
337 
338 #ifdef DIAGNOSTIC
339 	if (tp->t_inpcb && tp->t_in6pcb)
340 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
341 #endif
342 	so = NULL;
343 	ro = NULL;
344 	if (tp->t_inpcb) {
345 		so = tp->t_inpcb->inp_socket;
346 		ro = &tp->t_inpcb->inp_route;
347 	}
348 #ifdef INET6
349 	else if (tp->t_in6pcb) {
350 		so = tp->t_in6pcb->in6p_socket;
351 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
352 	}
353 #endif
354 
355 	switch (af = tp->t_family) {
356 #ifdef INET
357 	case AF_INET:
358 		if (tp->t_inpcb)
359 			break;
360 #ifdef INET6
361 		/* mapped addr case */
362 		if (tp->t_in6pcb)
363 			break;
364 #endif
365 		return EINVAL;
366 #endif
367 #ifdef INET6
368 	case AF_INET6:
369 		if (tp->t_in6pcb)
370 			break;
371 		return EINVAL;
372 #endif
373 	default:
374 		return EAFNOSUPPORT;
375 	}
376 
377 	tcp_segsize(tp, &txsegsize, &rxsegsize);
378 
379 	idle = (tp->snd_max == tp->snd_una);
380 
381 	/*
382 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
383 	 *
384 	 *	Optionally, a TCP MAY set the restart window to the
385 	 *	minimum of the value used for the initial window and
386 	 *	the current value of cwnd (in other words, using a
387 	 *	larger value for the restart window should never increase
388 	 *	the size of cwnd).
389 	 */
390 	if (tcp_cwm) {
391 		/*
392 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
393 		 * Count the number of packets currently pending
394 		 * acknowledgement, and limit our congestion window
395 		 * to a pre-determined allowed burst size plus that count.
396 		 * This prevents bursting once all pending packets have
397 		 * been acknowledged (i.e. transmission is idle).
398 		 *
399 		 * XXX Link this to Initial Window?
400 		 */
401 		tp->snd_cwnd = min(tp->snd_cwnd,
402 		    (tcp_cwm_burstsize * txsegsize) +
403 		    (tp->snd_nxt - tp->snd_una));
404 	} else {
405 		if (idle && tp->t_idle >= tp->t_rxtcur) {
406 			/*
407 			 * We have been idle for "a while" and no acks are
408 			 * expected to clock out any data we send --
409 			 * slow start to get ack "clock" running again.
410 			 */
411 			tp->snd_cwnd = min(tp->snd_cwnd,
412 			    TCP_INITIAL_WINDOW(tcp_init_win, txsegsize));
413 		}
414 	}
415 
416 again:
417 	/*
418 	 * Determine length of data that should be transmitted, and
419 	 * flags that should be used.  If there is some data or critical
420 	 * controls (SYN, RST) to send, then transmit; otherwise,
421 	 * investigate further.
422 	 */
423 	sendalot = 0;
424 	off = tp->snd_nxt - tp->snd_una;
425 	win = min(tp->snd_wnd, tp->snd_cwnd);
426 
427 	flags = tcp_outflags[tp->t_state];
428 	/*
429 	 * If in persist timeout with window of 0, send 1 byte.
430 	 * Otherwise, if window is small but nonzero
431 	 * and timer expired, we will send what we can
432 	 * and go to transmit state.
433 	 */
434 	if (tp->t_force) {
435 		if (win == 0) {
436 			/*
437 			 * If we still have some data to send, then
438 			 * clear the FIN bit.  Usually this would
439 			 * happen below when it realizes that we
440 			 * aren't sending all the data.  However,
441 			 * if we have exactly 1 byte of unset data,
442 			 * then it won't clear the FIN bit below,
443 			 * and if we are in persist state, we wind
444 			 * up sending the packet without recording
445 			 * that we sent the FIN bit.
446 			 *
447 			 * We can't just blindly clear the FIN bit,
448 			 * because if we don't have any more data
449 			 * to send then the probe will be the FIN
450 			 * itself.
451 			 */
452 			if (off < so->so_snd.sb_cc)
453 				flags &= ~TH_FIN;
454 			win = 1;
455 		} else {
456 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
457 			tp->t_rxtshift = 0;
458 		}
459 	}
460 
461 	if (win < so->so_snd.sb_cc) {
462 		len = win - off;
463 		flags &= ~TH_FIN;
464 	} else
465 		len = so->so_snd.sb_cc - off;
466 
467 	if (len < 0) {
468 		/*
469 		 * If FIN has been sent but not acked,
470 		 * but we haven't been called to retransmit,
471 		 * len will be -1.  Otherwise, window shrank
472 		 * after we sent into it.  If window shrank to 0,
473 		 * cancel pending retransmit, pull snd_nxt back
474 		 * to (closed) window, and set the persist timer
475 		 * if it isn't already going.  If the window didn't
476 		 * close completely, just wait for an ACK.
477 		 *
478 		 * If we have a pending FIN, either it has already been
479 		 * transmitted or it is outside the window, so drop it.
480 		 * If the FIN has been transmitted, but this is not a
481 		 * retransmission, then len must be -1.  Therefore we also
482 		 * prevent here the sending of `gratuitous FINs'.  This
483 		 * eliminates the need to check for that case below (e.g.
484 		 * to back up snd_nxt before the FIN so that the sequence
485 		 * number is correct).
486 		 */
487 		len = 0;
488 		flags &= ~TH_FIN;
489 		if (win == 0) {
490 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
491 			tp->t_rxtshift = 0;
492 			tp->snd_nxt = tp->snd_una;
493 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
494 				tcp_setpersist(tp);
495 		}
496 	}
497 	if (len > txsegsize) {
498 		len = txsegsize;
499 		flags &= ~TH_FIN;
500 		sendalot = 1;
501 	}
502 
503 	win = sbspace(&so->so_rcv);
504 
505 	/*
506 	 * Sender silly window avoidance.  If connection is idle
507 	 * and can send all data, a maximum segment,
508 	 * at least a maximum default-size segment do it,
509 	 * or are forced, do it; otherwise don't bother.
510 	 * If peer's buffer is tiny, then send
511 	 * when window is at least half open.
512 	 * If retransmitting (possibly after persist timer forced us
513 	 * to send into a small window), then must resend.
514 	 */
515 	if (len) {
516 		if (len == txsegsize)
517 			goto send;
518 		if ((so->so_state & SS_MORETOCOME) == 0 &&
519 		    ((idle || tp->t_flags & TF_NODELAY) &&
520 		     len + off >= so->so_snd.sb_cc))
521 			goto send;
522 		if (tp->t_force)
523 			goto send;
524 		if (len >= tp->max_sndwnd / 2)
525 			goto send;
526 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
527 			goto send;
528 	}
529 
530 	/*
531 	 * Compare available window to amount of window known to peer
532 	 * (as advertised window less next expected input).  If the
533 	 * difference is at least twice the size of the largest segment
534 	 * we expect to receive (i.e. two segments) or at least 50% of
535 	 * the maximum possible window, then want to send a window update
536 	 * to peer.
537 	 */
538 	if (win > 0) {
539 		/*
540 		 * "adv" is the amount we can increase the window,
541 		 * taking into account that we are limited by
542 		 * TCP_MAXWIN << tp->rcv_scale.
543 		 */
544 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
545 			(tp->rcv_adv - tp->rcv_nxt);
546 
547 		if (adv >= (long) (2 * rxsegsize))
548 			goto send;
549 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
550 			goto send;
551 	}
552 
553 	/*
554 	 * Send if we owe peer an ACK.
555 	 */
556 	if (tp->t_flags & TF_ACKNOW)
557 		goto send;
558 	if (flags & (TH_SYN|TH_FIN|TH_RST))
559 		goto send;
560 	if (SEQ_GT(tp->snd_up, tp->snd_una))
561 		goto send;
562 
563 	/*
564 	 * TCP window updates are not reliable, rather a polling protocol
565 	 * using ``persist'' packets is used to insure receipt of window
566 	 * updates.  The three ``states'' for the output side are:
567 	 *	idle			not doing retransmits or persists
568 	 *	persisting		to move a small or zero window
569 	 *	(re)transmitting	and thereby not persisting
570 	 *
571 	 * tp->t_timer[TCPT_PERSIST]
572 	 *	is set when we are in persist state.
573 	 * tp->t_force
574 	 *	is set when we are called to send a persist packet.
575 	 * tp->t_timer[TCPT_REXMT]
576 	 *	is set when we are retransmitting
577 	 * The output side is idle when both timers are zero.
578 	 *
579 	 * If send window is too small, there is data to transmit, and no
580 	 * retransmit or persist is pending, then go to persist state.
581 	 * If nothing happens soon, send when timer expires:
582 	 * if window is nonzero, transmit what we can,
583 	 * otherwise force out a byte.
584 	 */
585 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
586 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
587 		tp->t_rxtshift = 0;
588 		tcp_setpersist(tp);
589 	}
590 
591 	/*
592 	 * No reason to send a segment, just return.
593 	 */
594 	return (0);
595 
596 send:
597 	/*
598 	 * Before ESTABLISHED, force sending of initial options
599 	 * unless TCP set not to do any options.
600 	 * NOTE: we assume that the IP/TCP header plus TCP options
601 	 * always fit in a single mbuf, leaving room for a maximum
602 	 * link header, i.e.
603 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
604 	 */
605 	optlen = 0;
606 	switch (af) {
607 #ifdef INET
608 	case AF_INET:
609 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
610 		break;
611 #endif
612 #ifdef INET6
613 	case AF_INET6:
614 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
615 		break;
616 #endif
617 	default:	/*pacify gcc*/
618 		iphdrlen = 0;
619 		break;
620 	}
621 	hdrlen = iphdrlen;
622 	if (flags & TH_SYN) {
623 		struct rtentry *rt;
624 
625 		rt = NULL;
626 #ifdef INET
627 		if (tp->t_inpcb)
628 			rt = in_pcbrtentry(tp->t_inpcb);
629 #endif
630 #ifdef INET6
631 		if (tp->t_in6pcb)
632 			rt = in6_pcbrtentry(tp->t_in6pcb);
633 #endif
634 
635 		tp->snd_nxt = tp->iss;
636 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
637 						    rt->rt_ifp : NULL, af);
638 		if ((tp->t_flags & TF_NOOPT) == 0) {
639 			opt[0] = TCPOPT_MAXSEG;
640 			opt[1] = 4;
641 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
642 			opt[3] = tp->t_ourmss & 0xff;
643 			optlen = 4;
644 
645 			if ((tp->t_flags & TF_REQ_SCALE) &&
646 			    ((flags & TH_ACK) == 0 ||
647 			    (tp->t_flags & TF_RCVD_SCALE))) {
648 				*((u_int32_t *) (opt + optlen)) = htonl(
649 					TCPOPT_NOP << 24 |
650 					TCPOPT_WINDOW << 16 |
651 					TCPOLEN_WINDOW << 8 |
652 					tp->request_r_scale);
653 				optlen += 4;
654 			}
655 		}
656  	}
657 
658  	/*
659 	 * Send a timestamp and echo-reply if this is a SYN and our side
660 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
661 	 * and our peer have sent timestamps in our SYN's.
662  	 */
663  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
664  	     (flags & TH_RST) == 0 &&
665  	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
666 	     (tp->t_flags & TF_RCVD_TSTMP))) {
667 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
668 
669  		/* Form timestamp option as shown in appendix A of RFC 1323. */
670  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
671  		*lp++ = htonl(tcp_now);
672  		*lp   = htonl(tp->ts_recent);
673  		optlen += TCPOLEN_TSTAMP_APPA;
674  	}
675 
676  	hdrlen += optlen;
677 
678 #ifdef DIAGNOSTIC
679 	if (len > txsegsize)
680 		panic("tcp data to be sent is larger than segment");
681  	if (max_linkhdr + hdrlen > MCLBYTES)
682 		panic("tcphdr too big");
683 #endif
684 
685 	/*
686 	 * Grab a header mbuf, attaching a copy of data to
687 	 * be transmitted, and initialize the header from
688 	 * the template for sends on this connection.
689 	 */
690 	if (len) {
691 		if (tp->t_force && len == 1)
692 			tcpstat.tcps_sndprobe++;
693 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
694 			tcpstat.tcps_sndrexmitpack++;
695 			tcpstat.tcps_sndrexmitbyte += len;
696 		} else {
697 			tcpstat.tcps_sndpack++;
698 			tcpstat.tcps_sndbyte += len;
699 		}
700 #ifdef notyet
701 		if ((m = m_copypack(so->so_snd.sb_mb, off,
702 		    (int)len, max_linkhdr + hdrlen)) == 0) {
703 			error = ENOBUFS;
704 			goto out;
705 		}
706 		/*
707 		 * m_copypack left space for our hdr; use it.
708 		 */
709 		m->m_len += hdrlen;
710 		m->m_data -= hdrlen;
711 #else
712 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
713 		if (m != NULL &&
714 		    (max_linkhdr + hdrlen > MHLEN ||
715 		     max_linkhdr + hdrlen + len <= MCLBYTES)) {
716 			MCLGET(m, M_DONTWAIT);
717 			if ((m->m_flags & M_EXT) == 0) {
718 				m_freem(m);
719 				m = NULL;
720 			}
721 		}
722 		if (m == NULL) {
723 			error = ENOBUFS;
724 			goto out;
725 		}
726 		m->m_data += max_linkhdr;
727 		m->m_len = hdrlen;
728 		if (len <= M_TRAILINGSPACE(m)) {
729 			m_copydata(so->so_snd.sb_mb, off, (int) len,
730 			    mtod(m, caddr_t) + hdrlen);
731 			m->m_len += len;
732 		} else {
733 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
734 			if (m->m_next == NULL) {
735 				m_freem(m);
736 				error = ENOBUFS;
737 				goto out;
738 			}
739 		}
740 #endif
741 		/*
742 		 * If we're sending everything we've got, set PUSH.
743 		 * (This will keep happy those implementations which only
744 		 * give data to the user when a buffer fills or
745 		 * a PUSH comes in.)
746 		 */
747 		if (off + len == so->so_snd.sb_cc)
748 			flags |= TH_PUSH;
749 	} else {
750 		if (tp->t_flags & TF_ACKNOW)
751 			tcpstat.tcps_sndacks++;
752 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
753 			tcpstat.tcps_sndctrl++;
754 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
755 			tcpstat.tcps_sndurg++;
756 		else
757 			tcpstat.tcps_sndwinup++;
758 
759 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
760 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
761 			MCLGET(m, M_DONTWAIT);
762 			if ((m->m_flags & M_EXT) == 0) {
763 				m_freem(m);
764 				m = NULL;
765 			}
766 		}
767 		if (m == NULL) {
768 			error = ENOBUFS;
769 			goto out;
770 		}
771 		m->m_data += max_linkhdr;
772 		m->m_len = hdrlen;
773 	}
774 	m->m_pkthdr.rcvif = (struct ifnet *)0;
775 	switch (af) {
776 #ifdef INET
777 	case AF_INET:
778 		ip = mtod(m, struct ip *);
779 #ifdef INET6
780 		ip6 = NULL;
781 #endif
782 		th = (struct tcphdr *)(ip + 1);
783 		break;
784 #endif
785 #ifdef INET6
786 	case AF_INET6:
787 		ip = NULL;
788 		ip6 = mtod(m, struct ip6_hdr *);
789 		th = (struct tcphdr *)(ip6 + 1);
790 		break;
791 #endif
792 	default:	/*pacify gcc*/
793 		ip = NULL;
794 #ifdef INET6
795 		ip6 = NULL;
796 #endif
797 		th = NULL;
798 		break;
799 	}
800 	if (tp->t_template == 0)
801 		panic("tcp_output");
802 	if (tp->t_template->m_len < iphdrlen)
803 		panic("tcp_output");
804 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
805 
806 	/*
807 	 * If we are doing retransmissions, then snd_nxt will
808 	 * not reflect the first unsent octet.  For ACK only
809 	 * packets, we do not want the sequence number of the
810 	 * retransmitted packet, we want the sequence number
811 	 * of the next unsent octet.  So, if there is no data
812 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
813 	 * when filling in ti_seq.  But if we are in persist
814 	 * state, snd_max might reflect one byte beyond the
815 	 * right edge of the window, so use snd_nxt in that
816 	 * case, since we know we aren't doing a retransmission.
817 	 * (retransmit and persist are mutually exclusive...)
818 	 */
819 	if (len || (flags & (TH_SYN|TH_FIN)) ||
820 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
821 		th->th_seq = htonl(tp->snd_nxt);
822 	else
823 		th->th_seq = htonl(tp->snd_max);
824 	th->th_ack = htonl(tp->rcv_nxt);
825 	if (optlen) {
826 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
827 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
828 	}
829 	th->th_flags = flags;
830 	/*
831 	 * Calculate receive window.  Don't shrink window,
832 	 * but avoid silly window syndrome.
833 	 */
834 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
835 		win = 0;
836 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
837 		win = (long)TCP_MAXWIN << tp->rcv_scale;
838 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
839 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
840 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
841 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
842 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
843 		if (urp > IP_MAXPACKET)
844 			urp = IP_MAXPACKET;
845 		th->th_urp = htons((u_int16_t)urp);
846 		th->th_flags |= TH_URG;
847 	} else
848 		/*
849 		 * If no urgent pointer to send, then we pull
850 		 * the urgent pointer to the left edge of the send window
851 		 * so that it doesn't drift into the send window on sequence
852 		 * number wraparound.
853 		 */
854 		tp->snd_up = tp->snd_una;		/* drag it along */
855 
856 	/*
857 	 * Put TCP length in extended header, and then
858 	 * checksum extended header and data.
859 	 */
860 	switch (af) {
861 #ifdef INET
862 	case AF_INET:
863 	    {
864 		struct ipovly *ipov = (struct ipovly *)ip;
865 		if (len + optlen)
866 			ipov->ih_len = htons((u_int16_t)(sizeof(struct tcphdr) +
867 			    optlen + len));
868 		bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
869 		th->th_sum = 0;
870 		th->th_sum = in_cksum(m, (int)(hdrlen + len));
871 		break;
872 	    }
873 #endif
874 #ifdef INET6
875 	case AF_INET6:
876 		/* equals to hdrlen + len */
877 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
878 			+ sizeof(struct tcphdr) + optlen + len;
879 		th->th_sum = 0;
880 		th->th_sum = in6_cksum(m, IPPROTO_TCP,
881 				sizeof(struct ip6_hdr),
882 				sizeof(struct tcphdr) + optlen + len);
883 		break;
884 #endif
885 	}
886 
887 	/*
888 	 * In transmit state, time the transmission and arrange for
889 	 * the retransmit.  In persist state, just set snd_max.
890 	 */
891 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
892 		tcp_seq startseq = tp->snd_nxt;
893 
894 		/*
895 		 * Advance snd_nxt over sequence space of this segment.
896 		 * There are no states in which we send both a SYN and a FIN,
897 		 * so we collapse the tests for these flags.
898 		 */
899 		if (flags & (TH_SYN|TH_FIN))
900 			tp->snd_nxt++;
901 		tp->snd_nxt += len;
902 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
903 			tp->snd_max = tp->snd_nxt;
904 			/*
905 			 * Time this transmission if not a retransmission and
906 			 * not currently timing anything.
907 			 */
908 			if (tp->t_rtt == 0) {
909 				tp->t_rtt = 1;
910 				tp->t_rtseq = startseq;
911 				tcpstat.tcps_segstimed++;
912 			}
913 		}
914 
915 		/*
916 		 * Set retransmit timer if not currently set,
917 		 * and not doing an ack or a keep-alive probe.
918 		 * Initial value for retransmit timer is smoothed
919 		 * round-trip time + 2 * round-trip time variance.
920 		 * Initialize shift counter which is used for backoff
921 		 * of retransmit time.
922 		 */
923 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
924 		    tp->snd_nxt != tp->snd_una) {
925 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
926 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
927 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
928 				tp->t_rxtshift = 0;
929 			}
930 		}
931 	} else
932 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
933 			tp->snd_max = tp->snd_nxt + len;
934 
935 	/*
936 	 * Trace.
937 	 */
938 	if (so->so_options & SO_DEBUG) {
939 		/*
940 		 * need to recover version # field, which was overwritten
941 		 * on ip_cksum computation.
942 		 */
943 		struct ip *sip;
944 		sip = mtod(m, struct ip *);
945 		switch (af) {
946 #ifdef INET
947 		case AF_INET:
948 			sip->ip_v = 4;
949 			break;
950 #endif
951 #ifdef INET6
952 		case AF_INET6:
953 			sip->ip_v = 6;
954 			break;
955 #endif
956 		}
957 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
958 	}
959 
960 	/*
961 	 * Fill in IP length and desired time to live and
962 	 * send to IP level.  There should be a better way
963 	 * to handle ttl and tos; we could keep them in
964 	 * the template, but need a way to checksum without them.
965 	 */
966 	m->m_pkthdr.len = hdrlen + len;
967 
968 	switch (af) {
969 #ifdef INET
970 	case AF_INET:
971 		ip->ip_len = m->m_pkthdr.len;
972 		if (tp->t_inpcb) {
973 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
974 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
975 		}
976 #ifdef INET6
977 		else if (tp->t_in6pcb) {
978 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
979 			ip->ip_tos = 0;	/*XXX*/
980 		}
981 #endif
982 		break;
983 #endif
984 #ifdef INET6
985 	case AF_INET6:
986 		ip6->ip6_nxt = IPPROTO_TCP;
987 		if (tp->t_in6pcb) {
988 			/*
989 			 * we separately set hoplimit for every segment, since
990 			 * the user might want to change the value via
991 			 * setsockopt. Also, desired default hop limit might
992 			 * be changed via Neighbor Discovery.
993 			 */
994 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
995 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
996 		}
997 		/* ip6->ip6_flow = ??? */
998 		/* ip6_plen will be filled in ip6_output(). */
999 		break;
1000 #endif
1001 	}
1002 
1003 #ifdef IPSEC
1004 	ipsec_setsocket(m, so);
1005 #endif /*IPSEC*/
1006 
1007 	switch (af) {
1008 #ifdef INET
1009 	case AF_INET:
1010 	    {
1011 		struct mbuf *opts;
1012 
1013 		if (tp->t_inpcb)
1014 			opts = tp->t_inpcb->inp_options;
1015 		else
1016 			opts = NULL;
1017 		error = ip_output(m, opts, ro,
1018 			(ip_mtudisc ? IP_MTUDISC : 0) |
1019 			(so->so_options & SO_DONTROUTE),
1020 			0);
1021 		break;
1022 	    }
1023 #endif
1024 #ifdef INET6
1025 	case AF_INET6:
1026 	    {
1027 		struct ip6_pktopts *opts;
1028 
1029 		if (tp->t_in6pcb)
1030 			opts = tp->t_in6pcb->in6p_outputopts;
1031 		else
1032 			opts = NULL;
1033 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1034 			so->so_options & SO_DONTROUTE, 0, NULL);
1035 		break;
1036 	    }
1037 #endif
1038 	default:
1039 		error = EAFNOSUPPORT;
1040 		break;
1041 	}
1042 	if (error) {
1043 out:
1044 		if (error == ENOBUFS) {
1045 #ifdef INET
1046 			if (tp->t_inpcb)
1047 				tcp_quench(tp->t_inpcb, 0);
1048 #endif
1049 #ifdef INET6
1050 			if (tp->t_in6pcb)
1051 				tcp6_quench(tp->t_in6pcb, 0);
1052 #endif
1053 			return (0);
1054 		}
1055 		if ((error == EHOSTUNREACH || error == ENETDOWN)
1056 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
1057 			tp->t_softerror = error;
1058 			return (0);
1059 		}
1060 		return (error);
1061 	}
1062 	tcpstat.tcps_sndtotal++;
1063 	if (tp->t_flags & TF_DELACK)
1064 		tcpstat.tcps_delack++;
1065 
1066 	/*
1067 	 * Data sent (as far as we can tell).
1068 	 * If this advertises a larger window than any other segment,
1069 	 * then remember the size of the advertised window.
1070 	 * Any pending ACK has now been sent.
1071 	 */
1072 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1073 		tp->rcv_adv = tp->rcv_nxt + win;
1074 	tp->last_ack_sent = tp->rcv_nxt;
1075 	tp->t_flags &= ~TF_ACKNOW;
1076 	TCP_CLEAR_DELACK(tp);
1077 #ifdef DIAGNOSTIC
1078 	if (maxburst < 0)
1079 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1080 #endif
1081 	if (sendalot && (!tcp_do_newreno || --maxburst))
1082 		goto again;
1083 	return (0);
1084 }
1085 
1086 void
1087 tcp_setpersist(tp)
1088 	struct tcpcb *tp;
1089 {
1090 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1091 	int nticks;
1092 
1093 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1094 		panic("tcp_output REXMT");
1095 	/*
1096 	 * Start/restart persistance timer.
1097 	 */
1098 	if (t < tp->t_rttmin)
1099 		t = tp->t_rttmin;
1100 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1101 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1102 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1103 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1104 		tp->t_rxtshift++;
1105 }
1106