xref: /freebsd/sys/netinet/tcp_output.c (revision eca8a663d442468f64e21ed869817b9048ab5a7b)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
34  * $FreeBSD$
35  */
36 
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_mac.h"
40 #include "opt_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/mac.h>
48 #include <sys/mbuf.h>
49 #include <sys/mutex.h>
50 #include <sys/protosw.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
54 
55 #include <net/route.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/ip_var.h>
62 #ifdef INET6
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/ip6.h>
65 #include <netinet6/ip6_var.h>
66 #endif
67 #include <netinet/tcp.h>
68 #define	TCPOUTFLAGS
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
74 #ifdef TCPDEBUG
75 #include <netinet/tcp_debug.h>
76 #endif
77 
78 #ifdef IPSEC
79 #include <netinet6/ipsec.h>
80 #endif /*IPSEC*/
81 
82 #ifdef FAST_IPSEC
83 #include <netipsec/ipsec.h>
84 #define	IPSEC
85 #endif /*FAST_IPSEC*/
86 
87 #include <machine/in_cksum.h>
88 
89 #ifdef notyet
90 extern struct mbuf *m_copypack();
91 #endif
92 
93 int path_mtu_discovery = 1;
94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
95 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
96 
97 int ss_fltsz = 1;
98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
99 	&ss_fltsz, 1, "Slow start flight size");
100 
101 int ss_fltsz_local = 4;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
103 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
104 
105 int     tcp_do_newreno = 1;
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
107         0, "Enable NewReno Algorithms");
108 
109 /*
110  * Tcp output routine: figure out what should be sent and send it.
111  */
112 int
113 tcp_output(struct tcpcb *tp)
114 {
115 	struct socket *so = tp->t_inpcb->inp_socket;
116 	long len, win;
117 	int off, flags, error;
118 	struct mbuf *m;
119 	struct ip *ip = NULL;
120 	struct ipovly *ipov = NULL;
121 	struct tcphdr *th;
122 	u_char opt[TCP_MAXOLEN];
123 	unsigned ipoptlen, optlen, hdrlen;
124 	int idle, sendalot;
125 #if 0
126 	int maxburst = TCP_MAXBURST;
127 #endif
128 	struct rmxp_tao *taop;
129 #ifdef INET6
130 	struct ip6_hdr *ip6 = NULL;
131 	int isipv6;
132 
133 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
134 #endif
135 
136 	INP_LOCK_ASSERT(tp->t_inpcb);
137 
138 	/*
139 	 * Determine length of data that should be transmitted,
140 	 * and flags that will be used.
141 	 * If there is some data or critical controls (SYN, RST)
142 	 * to send, then transmit; otherwise, investigate further.
143 	 */
144 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
145 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
146 		/*
147 		 * We have been idle for "a while" and no acks are
148 		 * expected to clock out any data we send --
149 		 * slow start to get ack "clock" running again.
150 		 *
151 		 * Set the slow-start flight size depending on whether
152 		 * this is a local network or not.
153 		 */
154 		int ss = ss_fltsz;
155 #ifdef INET6
156 		if (isipv6) {
157 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
158 				ss = ss_fltsz_local;
159 		} else
160 #endif /* INET6 */
161 		if (in_localaddr(tp->t_inpcb->inp_faddr))
162 			ss = ss_fltsz_local;
163 		tp->snd_cwnd = tp->t_maxseg * ss;
164 	}
165 	tp->t_flags &= ~TF_LASTIDLE;
166 	if (idle) {
167 		if (tp->t_flags & TF_MORETOCOME) {
168 			tp->t_flags |= TF_LASTIDLE;
169 			idle = 0;
170 		}
171 	}
172 again:
173 	sendalot = 0;
174 	off = tp->snd_nxt - tp->snd_una;
175 	win = min(tp->snd_wnd, tp->snd_cwnd);
176 	win = min(win, tp->snd_bwnd);
177 
178 	flags = tcp_outflags[tp->t_state];
179 	/*
180 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
181 	 * state flags.
182 	 */
183 	if (tp->t_flags & TF_NEEDFIN)
184 		flags |= TH_FIN;
185 	if (tp->t_flags & TF_NEEDSYN)
186 		flags |= TH_SYN;
187 
188 	/*
189 	 * If in persist timeout with window of 0, send 1 byte.
190 	 * Otherwise, if window is small but nonzero
191 	 * and timer expired, we will send what we can
192 	 * and go to transmit state.
193 	 */
194 	if (tp->t_force) {
195 		if (win == 0) {
196 			/*
197 			 * If we still have some data to send, then
198 			 * clear the FIN bit.  Usually this would
199 			 * happen below when it realizes that we
200 			 * aren't sending all the data.  However,
201 			 * if we have exactly 1 byte of unsent data,
202 			 * then it won't clear the FIN bit below,
203 			 * and if we are in persist state, we wind
204 			 * up sending the packet without recording
205 			 * that we sent the FIN bit.
206 			 *
207 			 * We can't just blindly clear the FIN bit,
208 			 * because if we don't have any more data
209 			 * to send then the probe will be the FIN
210 			 * itself.
211 			 */
212 			if (off < so->so_snd.sb_cc)
213 				flags &= ~TH_FIN;
214 			win = 1;
215 		} else {
216 			callout_stop(tp->tt_persist);
217 			tp->t_rxtshift = 0;
218 		}
219 	}
220 
221 	/*
222 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
223 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
224 	 * a negative length.  This can also occur when tcp opens up
225 	 * its congestion window while receiving additional duplicate
226 	 * acks after fast-retransmit because TCP will reset snd_nxt
227 	 * to snd_max after the fast-retransmit.
228 	 *
229 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
230 	 * be set to snd_una, the offset will be 0, and the length may
231 	 * wind up 0.
232 	 */
233 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
234 
235 	taop = tcp_gettaocache(&tp->t_inpcb->inp_inc);
236 
237 	/*
238 	 * Lop off SYN bit if it has already been sent.  However, if this
239 	 * is SYN-SENT state and if segment contains data and if we don't
240 	 * know that foreign host supports TAO, suppress sending segment.
241 	 */
242 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
243 		flags &= ~TH_SYN;
244 		off--, len++;
245 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
246 		    (taop == NULL || taop->tao_ccsent == 0))
247 			return 0;
248 	}
249 
250 	/*
251 	 * Be careful not to send data and/or FIN on SYN segments
252 	 * in cases when no CC option will be sent.
253 	 * This measure is needed to prevent interoperability problems
254 	 * with not fully conformant TCP implementations.
255 	 */
256 	if ((flags & TH_SYN) &&
257 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
258 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
259 		len = 0;
260 		flags &= ~TH_FIN;
261 	}
262 
263 	if (len < 0) {
264 		/*
265 		 * If FIN has been sent but not acked,
266 		 * but we haven't been called to retransmit,
267 		 * len will be < 0.  Otherwise, window shrank
268 		 * after we sent into it.  If window shrank to 0,
269 		 * cancel pending retransmit, pull snd_nxt back
270 		 * to (closed) window, and set the persist timer
271 		 * if it isn't already going.  If the window didn't
272 		 * close completely, just wait for an ACK.
273 		 */
274 		len = 0;
275 		if (win == 0) {
276 			callout_stop(tp->tt_rexmt);
277 			tp->t_rxtshift = 0;
278 			tp->snd_nxt = tp->snd_una;
279 			if (!callout_active(tp->tt_persist))
280 				tcp_setpersist(tp);
281 		}
282 	}
283 
284 	/*
285 	 * len will be >= 0 after this point.  Truncate to the maximum
286 	 * segment length and ensure that FIN is removed if the length
287 	 * no longer contains the last data byte.
288 	 */
289 	if (len > tp->t_maxseg) {
290 		len = tp->t_maxseg;
291 		sendalot = 1;
292 	}
293 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
294 		flags &= ~TH_FIN;
295 
296 	win = sbspace(&so->so_rcv);
297 
298 	/*
299 	 * Sender silly window avoidance.   We transmit under the following
300 	 * conditions when len is non-zero:
301 	 *
302 	 *	- We have a full segment
303 	 *	- This is the last buffer in a write()/send() and we are
304 	 *	  either idle or running NODELAY
305 	 *	- we've timed out (e.g. persist timer)
306 	 *	- we have more then 1/2 the maximum send window's worth of
307 	 *	  data (receiver may be limited the window size)
308 	 *	- we need to retransmit
309 	 */
310 	if (len) {
311 		if (len == tp->t_maxseg)
312 			goto send;
313 		/*
314 		 * NOTE! on localhost connections an 'ack' from the remote
315 		 * end may occur synchronously with the output and cause
316 		 * us to flush a buffer queued with moretocome.  XXX
317 		 *
318 		 * note: the len + off check is almost certainly unnecessary.
319 		 */
320 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
321 		    (idle || (tp->t_flags & TF_NODELAY)) &&
322 		    len + off >= so->so_snd.sb_cc &&
323 		    (tp->t_flags & TF_NOPUSH) == 0) {
324 			goto send;
325 		}
326 		if (tp->t_force)			/* typ. timeout case */
327 			goto send;
328 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
329 			goto send;
330 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
331 			goto send;
332 	}
333 
334 	/*
335 	 * Compare available window to amount of window
336 	 * known to peer (as advertised window less
337 	 * next expected input).  If the difference is at least two
338 	 * max size segments, or at least 50% of the maximum possible
339 	 * window, then want to send a window update to peer.
340 	 * Skip this if the connection is in T/TCP half-open state.
341 	 */
342 	if (win > 0 && !(tp->t_flags & TF_NEEDSYN)) {
343 		/*
344 		 * "adv" is the amount we can increase the window,
345 		 * taking into account that we are limited by
346 		 * TCP_MAXWIN << tp->rcv_scale.
347 		 */
348 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
349 			(tp->rcv_adv - tp->rcv_nxt);
350 
351 		if (adv >= (long) (2 * tp->t_maxseg))
352 			goto send;
353 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
354 			goto send;
355 	}
356 
357 	/*
358 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
359 	 * is also a catch-all for the retransmit timer timeout case.
360 	 */
361 	if (tp->t_flags & TF_ACKNOW)
362 		goto send;
363 	if ((flags & TH_RST) ||
364 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
365 		goto send;
366 	if (SEQ_GT(tp->snd_up, tp->snd_una))
367 		goto send;
368 	/*
369 	 * If our state indicates that FIN should be sent
370 	 * and we have not yet done so, then we need to send.
371 	 */
372 	if (flags & TH_FIN &&
373 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
374 		goto send;
375 
376 	/*
377 	 * TCP window updates are not reliable, rather a polling protocol
378 	 * using ``persist'' packets is used to insure receipt of window
379 	 * updates.  The three ``states'' for the output side are:
380 	 *	idle			not doing retransmits or persists
381 	 *	persisting		to move a small or zero window
382 	 *	(re)transmitting	and thereby not persisting
383 	 *
384 	 * callout_active(tp->tt_persist)
385 	 *	is true when we are in persist state.
386 	 * tp->t_force
387 	 *	is set when we are called to send a persist packet.
388 	 * callout_active(tp->tt_rexmt)
389 	 *	is set when we are retransmitting
390 	 * The output side is idle when both timers are zero.
391 	 *
392 	 * If send window is too small, there is data to transmit, and no
393 	 * retransmit or persist is pending, then go to persist state.
394 	 * If nothing happens soon, send when timer expires:
395 	 * if window is nonzero, transmit what we can,
396 	 * otherwise force out a byte.
397 	 */
398 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
399 	    !callout_active(tp->tt_persist)) {
400 		tp->t_rxtshift = 0;
401 		tcp_setpersist(tp);
402 	}
403 
404 	/*
405 	 * No reason to send a segment, just return.
406 	 */
407 	return (0);
408 
409 send:
410 	/*
411 	 * Before ESTABLISHED, force sending of initial options
412 	 * unless TCP set not to do any options.
413 	 * NOTE: we assume that the IP/TCP header plus TCP options
414 	 * always fit in a single mbuf, leaving room for a maximum
415 	 * link header, i.e.
416 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
417 	 */
418 	optlen = 0;
419 #ifdef INET6
420 	if (isipv6)
421 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
422 	else
423 #endif
424 	hdrlen = sizeof (struct tcpiphdr);
425 	if (flags & TH_SYN) {
426 		tp->snd_nxt = tp->iss;
427 		if ((tp->t_flags & TF_NOOPT) == 0) {
428 			u_short mss;
429 
430 			opt[0] = TCPOPT_MAXSEG;
431 			opt[1] = TCPOLEN_MAXSEG;
432 			mss = htons((u_short) tcp_mssopt(tp));
433 			(void)memcpy(opt + 2, &mss, sizeof(mss));
434 			optlen = TCPOLEN_MAXSEG;
435 
436 			if ((tp->t_flags & TF_REQ_SCALE) &&
437 			    ((flags & TH_ACK) == 0 ||
438 			    (tp->t_flags & TF_RCVD_SCALE))) {
439 				*((u_int32_t *)(opt + optlen)) = htonl(
440 					TCPOPT_NOP << 24 |
441 					TCPOPT_WINDOW << 16 |
442 					TCPOLEN_WINDOW << 8 |
443 					tp->request_r_scale);
444 				optlen += 4;
445 			}
446 		}
447  	}
448 
449  	/*
450 	 * Send a timestamp and echo-reply if this is a SYN and our side
451 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
452 	 * and our peer have sent timestamps in our SYN's.
453  	 */
454  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
455  	    (flags & TH_RST) == 0 &&
456 	    ((flags & TH_ACK) == 0 ||
457 	     (tp->t_flags & TF_RCVD_TSTMP))) {
458 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
459 
460  		/* Form timestamp option as shown in appendix A of RFC 1323. */
461  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
462  		*lp++ = htonl(ticks);
463  		*lp   = htonl(tp->ts_recent);
464  		optlen += TCPOLEN_TSTAMP_APPA;
465  	}
466 
467  	/*
468 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
469 	 * options are allowed (!TF_NOOPT) and it's not a RST.
470  	 */
471  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
472  	     (flags & TH_RST) == 0) {
473 		switch (flags & (TH_SYN|TH_ACK)) {
474 		/*
475 		 * This is a normal ACK, send CC if we received CC before
476 		 * from our peer.
477 		 */
478 		case TH_ACK:
479 			if (!(tp->t_flags & TF_RCVD_CC))
480 				break;
481 			/*FALLTHROUGH*/
482 
483 		/*
484 		 * We can only get here in T/TCP's SYN_SENT* state, when
485 		 * we're a sending a non-SYN segment without waiting for
486 		 * the ACK of our SYN.  A check above assures that we only
487 		 * do this if our peer understands T/TCP.
488 		 */
489 		case 0:
490 			opt[optlen++] = TCPOPT_NOP;
491 			opt[optlen++] = TCPOPT_NOP;
492 			opt[optlen++] = TCPOPT_CC;
493 			opt[optlen++] = TCPOLEN_CC;
494 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
495 
496 			optlen += 4;
497 			break;
498 
499 		/*
500 		 * This is our initial SYN, check whether we have to use
501 		 * CC or CC.new.
502 		 */
503 		case TH_SYN:
504 			opt[optlen++] = TCPOPT_NOP;
505 			opt[optlen++] = TCPOPT_NOP;
506 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
507 						TCPOPT_CCNEW : TCPOPT_CC;
508 			opt[optlen++] = TCPOLEN_CC;
509 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
510  			optlen += 4;
511 			break;
512 
513 		/*
514 		 * This is a SYN,ACK; send CC and CC.echo if we received
515 		 * CC from our peer.
516 		 */
517 		case (TH_SYN|TH_ACK):
518 			if (tp->t_flags & TF_RCVD_CC) {
519 				opt[optlen++] = TCPOPT_NOP;
520 				opt[optlen++] = TCPOPT_NOP;
521 				opt[optlen++] = TCPOPT_CC;
522 				opt[optlen++] = TCPOLEN_CC;
523 				*(u_int32_t *)&opt[optlen] =
524 					htonl(tp->cc_send);
525 				optlen += 4;
526 				opt[optlen++] = TCPOPT_NOP;
527 				opt[optlen++] = TCPOPT_NOP;
528 				opt[optlen++] = TCPOPT_CCECHO;
529 				opt[optlen++] = TCPOLEN_CC;
530 				*(u_int32_t *)&opt[optlen] =
531 					htonl(tp->cc_recv);
532 				optlen += 4;
533 			}
534 			break;
535 		}
536  	}
537 
538  	hdrlen += optlen;
539 
540 #ifdef INET6
541 	if (isipv6)
542 		ipoptlen = ip6_optlen(tp->t_inpcb);
543 	else
544 #endif
545 	if (tp->t_inpcb->inp_options)
546 		ipoptlen = tp->t_inpcb->inp_options->m_len -
547 				offsetof(struct ipoption, ipopt_list);
548 	else
549 		ipoptlen = 0;
550 #ifdef IPSEC
551 	ipoptlen += ipsec_hdrsiz_tcp(tp);
552 #endif
553 
554 	/*
555 	 * Adjust data length if insertion of options will
556 	 * bump the packet length beyond the t_maxopd length.
557 	 * Clear the FIN bit because we cut off the tail of
558 	 * the segment.
559 	 */
560 	if (len + optlen + ipoptlen > tp->t_maxopd) {
561 		/*
562 		 * If there is still more to send, don't close the connection.
563 		 */
564 		flags &= ~TH_FIN;
565 		len = tp->t_maxopd - optlen - ipoptlen;
566 		sendalot = 1;
567 	}
568 
569 /*#ifdef DIAGNOSTIC*/
570 #ifdef INET6
571  	if (max_linkhdr + hdrlen > MCLBYTES)
572 #else
573  	if (max_linkhdr + hdrlen > MHLEN)
574 #endif
575 		panic("tcphdr too big");
576 /*#endif*/
577 
578 	/*
579 	 * Grab a header mbuf, attaching a copy of data to
580 	 * be transmitted, and initialize the header from
581 	 * the template for sends on this connection.
582 	 */
583 	if (len) {
584 		if (tp->t_force && len == 1)
585 			tcpstat.tcps_sndprobe++;
586 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
587 			tcpstat.tcps_sndrexmitpack++;
588 			tcpstat.tcps_sndrexmitbyte += len;
589 		} else {
590 			tcpstat.tcps_sndpack++;
591 			tcpstat.tcps_sndbyte += len;
592 		}
593 #ifdef notyet
594 		if ((m = m_copypack(so->so_snd.sb_mb, off,
595 		    (int)len, max_linkhdr + hdrlen)) == 0) {
596 			error = ENOBUFS;
597 			goto out;
598 		}
599 		/*
600 		 * m_copypack left space for our hdr; use it.
601 		 */
602 		m->m_len += hdrlen;
603 		m->m_data -= hdrlen;
604 #else
605 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
606 		if (m == NULL) {
607 			error = ENOBUFS;
608 			goto out;
609 		}
610 #ifdef INET6
611 		if (MHLEN < hdrlen + max_linkhdr) {
612 			MCLGET(m, M_DONTWAIT);
613 			if ((m->m_flags & M_EXT) == 0) {
614 				m_freem(m);
615 				error = ENOBUFS;
616 				goto out;
617 			}
618 		}
619 #endif
620 		m->m_data += max_linkhdr;
621 		m->m_len = hdrlen;
622 		if (len <= MHLEN - hdrlen - max_linkhdr) {
623 			m_copydata(so->so_snd.sb_mb, off, (int) len,
624 			    mtod(m, caddr_t) + hdrlen);
625 			m->m_len += len;
626 		} else {
627 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
628 			if (m->m_next == 0) {
629 				(void) m_free(m);
630 				error = ENOBUFS;
631 				goto out;
632 			}
633 		}
634 #endif
635 		/*
636 		 * If we're sending everything we've got, set PUSH.
637 		 * (This will keep happy those implementations which only
638 		 * give data to the user when a buffer fills or
639 		 * a PUSH comes in.)
640 		 */
641 		if (off + len == so->so_snd.sb_cc)
642 			flags |= TH_PUSH;
643 	} else {
644 		if (tp->t_flags & TF_ACKNOW)
645 			tcpstat.tcps_sndacks++;
646 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
647 			tcpstat.tcps_sndctrl++;
648 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
649 			tcpstat.tcps_sndurg++;
650 		else
651 			tcpstat.tcps_sndwinup++;
652 
653 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
654 		if (m == NULL) {
655 			error = ENOBUFS;
656 			goto out;
657 		}
658 #ifdef INET6
659 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
660 		    MHLEN >= hdrlen) {
661 			MH_ALIGN(m, hdrlen);
662 		} else
663 #endif
664 		m->m_data += max_linkhdr;
665 		m->m_len = hdrlen;
666 	}
667 	m->m_pkthdr.rcvif = (struct ifnet *)0;
668 #ifdef MAC
669 	mac_create_mbuf_from_socket(so, m);
670 #endif
671 #ifdef INET6
672 	if (isipv6) {
673 		ip6 = mtod(m, struct ip6_hdr *);
674 		th = (struct tcphdr *)(ip6 + 1);
675 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
676 	} else
677 #endif /* INET6 */
678       {
679 	ip = mtod(m, struct ip *);
680 	ipov = (struct ipovly *)ip;
681 	th = (struct tcphdr *)(ip + 1);
682 	tcpip_fillheaders(tp->t_inpcb, ip, th);
683       }
684 
685 	/*
686 	 * Fill in fields, remembering maximum advertised
687 	 * window for use in delaying messages about window sizes.
688 	 * If resending a FIN, be sure not to use a new sequence number.
689 	 */
690 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
691 	    tp->snd_nxt == tp->snd_max)
692 		tp->snd_nxt--;
693 	/*
694 	 * If we are doing retransmissions, then snd_nxt will
695 	 * not reflect the first unsent octet.  For ACK only
696 	 * packets, we do not want the sequence number of the
697 	 * retransmitted packet, we want the sequence number
698 	 * of the next unsent octet.  So, if there is no data
699 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
700 	 * when filling in ti_seq.  But if we are in persist
701 	 * state, snd_max might reflect one byte beyond the
702 	 * right edge of the window, so use snd_nxt in that
703 	 * case, since we know we aren't doing a retransmission.
704 	 * (retransmit and persist are mutually exclusive...)
705 	 */
706 	if (len || (flags & (TH_SYN|TH_FIN))
707 	    || callout_active(tp->tt_persist))
708 		th->th_seq = htonl(tp->snd_nxt);
709 	else
710 		th->th_seq = htonl(tp->snd_max);
711 	th->th_ack = htonl(tp->rcv_nxt);
712 	if (optlen) {
713 		bcopy(opt, th + 1, optlen);
714 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
715 	}
716 	th->th_flags = flags;
717 	/*
718 	 * Calculate receive window.  Don't shrink window,
719 	 * but avoid silly window syndrome.
720 	 */
721 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
722 		win = 0;
723 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
724 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
725 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
726 		win = (long)TCP_MAXWIN << tp->rcv_scale;
727 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
728 
729 
730 	/*
731 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
732 	 * a 0 window.  This may cause the remote transmitter to stall.  This
733 	 * flag tells soreceive() to disable delayed acknowledgements when
734 	 * draining the buffer.  This can occur if the receiver is attempting
735 	 * to read more data then can be buffered prior to transmitting on
736 	 * the connection.
737 	 */
738 	if (win == 0)
739 		tp->t_flags |= TF_RXWIN0SENT;
740 	else
741 		tp->t_flags &= ~TF_RXWIN0SENT;
742 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
743 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
744 		th->th_flags |= TH_URG;
745 	} else
746 		/*
747 		 * If no urgent pointer to send, then we pull
748 		 * the urgent pointer to the left edge of the send window
749 		 * so that it doesn't drift into the send window on sequence
750 		 * number wraparound.
751 		 */
752 		tp->snd_up = tp->snd_una;		/* drag it along */
753 
754 	/*
755 	 * Put TCP length in extended header, and then
756 	 * checksum extended header and data.
757 	 */
758 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
759 #ifdef INET6
760 	if (isipv6)
761 		/*
762 		 * ip6_plen is not need to be filled now, and will be filled
763 		 * in ip6_output.
764 		 */
765 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
766 				       sizeof(struct tcphdr) + optlen + len);
767 	else
768 #endif /* INET6 */
769       {
770 	m->m_pkthdr.csum_flags = CSUM_TCP;
771 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
772 	th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
773 	    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
774 
775 	/* IP version must be set here for ipv4/ipv6 checking later */
776 	KASSERT(ip->ip_v == IPVERSION,
777 	    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
778       }
779 
780 	/*
781 	 * In transmit state, time the transmission and arrange for
782 	 * the retransmit.  In persist state, just set snd_max.
783 	 */
784 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
785 		tcp_seq startseq = tp->snd_nxt;
786 
787 		/*
788 		 * Advance snd_nxt over sequence space of this segment.
789 		 */
790 		if (flags & (TH_SYN|TH_FIN)) {
791 			if (flags & TH_SYN)
792 				tp->snd_nxt++;
793 			if (flags & TH_FIN) {
794 				tp->snd_nxt++;
795 				tp->t_flags |= TF_SENTFIN;
796 			}
797 		}
798 		tp->snd_nxt += len;
799 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
800 			tp->snd_max = tp->snd_nxt;
801 			/*
802 			 * Time this transmission if not a retransmission and
803 			 * not currently timing anything.
804 			 */
805 			if (tp->t_rtttime == 0) {
806 				tp->t_rtttime = ticks;
807 				tp->t_rtseq = startseq;
808 				tcpstat.tcps_segstimed++;
809 			}
810 		}
811 
812 		/*
813 		 * Set retransmit timer if not currently set,
814 		 * and not doing a pure ack or a keep-alive probe.
815 		 * Initial value for retransmit timer is smoothed
816 		 * round-trip time + 2 * round-trip time variance.
817 		 * Initialize shift counter which is used for backoff
818 		 * of retransmit time.
819 		 */
820 		if (!callout_active(tp->tt_rexmt) &&
821 		    tp->snd_nxt != tp->snd_una) {
822 			if (callout_active(tp->tt_persist)) {
823 				callout_stop(tp->tt_persist);
824 				tp->t_rxtshift = 0;
825 			}
826 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
827 				      tcp_timer_rexmt, tp);
828 		}
829 	} else {
830 		/*
831 		 * Persist case, update snd_max but since we are in
832 		 * persist mode (no window) we do not update snd_nxt.
833 		 */
834 		int xlen = len;
835 		if (flags & TH_SYN)
836 			++xlen;
837 		if (flags & TH_FIN) {
838 			++xlen;
839 			tp->t_flags |= TF_SENTFIN;
840 		}
841 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
842 			tp->snd_max = tp->snd_nxt + len;
843 	}
844 
845 #ifdef TCPDEBUG
846 	/*
847 	 * Trace.
848 	 */
849 	if (so->so_options & SO_DEBUG) {
850 		u_short save = ipov->ih_len;
851 		ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
852 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
853 		ipov->ih_len = save;
854 	}
855 #endif
856 
857 	/*
858 	 * Fill in IP length and desired time to live and
859 	 * send to IP level.  There should be a better way
860 	 * to handle ttl and tos; we could keep them in
861 	 * the template, but need a way to checksum without them.
862 	 */
863 	/*
864 	 * m->m_pkthdr.len should have been set before cksum calcuration,
865 	 * because in6_cksum() need it.
866 	 */
867 #ifdef INET6
868 	if (isipv6) {
869 		/*
870 		 * we separately set hoplimit for every segment, since the
871 		 * user might want to change the value via setsockopt.
872 		 * Also, desired default hop limit might be changed via
873 		 * Neighbor Discovery.
874 		 */
875 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
876 					       tp->t_inpcb->in6p_route.ro_rt ?
877 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
878 					       : NULL);
879 
880 		/* TODO: IPv6 IP6TOS_ECT bit on */
881 #if defined(IPSEC) && !defined(FAST_IPSEC)
882 		if (ipsec_setsocket(m, so) != 0) {
883 			m_freem(m);
884 			error = ENOBUFS;
885 			goto out;
886 		}
887 #endif /*IPSEC*/
888 		error = ip6_output(m,
889 			    tp->t_inpcb->in6p_outputopts,
890 			    &tp->t_inpcb->in6p_route,
891 			    (so->so_options & SO_DONTROUTE), NULL, NULL,
892 			    tp->t_inpcb);
893 	} else
894 #endif /* INET6 */
895     {
896 	struct rtentry *rt;
897 	ip->ip_len = m->m_pkthdr.len;
898 #ifdef INET6
899  	if (INP_CHECK_SOCKAF(so, AF_INET6))
900  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb,
901  					    tp->t_inpcb->in6p_route.ro_rt ?
902  					    tp->t_inpcb->in6p_route.ro_rt->rt_ifp
903  					    : NULL);
904 #endif /* INET6 */
905 	/*
906 	 * See if we should do MTU discovery.  We do it only if the following
907 	 * are true:
908 	 *	1) we have a valid route to the destination
909 	 *	2) the MTU is not locked (if it is, then discovery has been
910 	 *	   disabled)
911 	 */
912 	if (path_mtu_discovery
913 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
914 	    && rt->rt_flags & RTF_UP
915 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
916 		ip->ip_off |= IP_DF;
917 	}
918 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
919 	    (so->so_options & SO_DONTROUTE), 0, tp->t_inpcb);
920     }
921 	if (error) {
922 
923 		/*
924 		 * We know that the packet was lost, so back out the
925 		 * sequence number advance, if any.
926 		 */
927 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
928 			/*
929 			 * No need to check for TH_FIN here because
930 			 * the TF_SENTFIN flag handles that case.
931 			 */
932 			if ((flags & TH_SYN) == 0)
933 				tp->snd_nxt -= len;
934 		}
935 
936 out:
937 		if (error == ENOBUFS) {
938 	                if (!callout_active(tp->tt_rexmt) &&
939                             !callout_active(tp->tt_persist))
940 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
941                                       tcp_timer_rexmt, tp);
942 			tcp_quench(tp->t_inpcb, 0);
943 			return (0);
944 		}
945 		if (error == EMSGSIZE) {
946 			/*
947 			 * ip_output() will have already fixed the route
948 			 * for us.  tcp_mtudisc() will, as its last action,
949 			 * initiate retransmission, so it is important to
950 			 * not do so here.
951 			 */
952 			tcp_mtudisc(tp->t_inpcb, 0);
953 			return 0;
954 		}
955 		if ((error == EHOSTUNREACH || error == ENETDOWN)
956 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
957 			tp->t_softerror = error;
958 			return (0);
959 		}
960 		return (error);
961 	}
962 	tcpstat.tcps_sndtotal++;
963 
964 	/*
965 	 * Data sent (as far as we can tell).
966 	 * If this advertises a larger window than any other segment,
967 	 * then remember the size of the advertised window.
968 	 * Any pending ACK has now been sent.
969 	 */
970 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
971 		tp->rcv_adv = tp->rcv_nxt + win;
972 	tp->last_ack_sent = tp->rcv_nxt;
973 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
974 	if (callout_active(tp->tt_delack))
975 		callout_stop(tp->tt_delack);
976 #if 0
977 	/*
978 	 * This completely breaks TCP if newreno is turned on.  What happens
979 	 * is that if delayed-acks are turned on on the receiver, this code
980 	 * on the transmitter effectively destroys the TCP window, forcing
981 	 * it to four packets (1.5Kx4 = 6K window).
982 	 */
983 	if (sendalot && (!tcp_do_newreno || --maxburst))
984 		goto again;
985 #endif
986 	if (sendalot)
987 		goto again;
988 	return (0);
989 }
990 
991 void
992 tcp_setpersist(tp)
993 	register struct tcpcb *tp;
994 {
995 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
996 	int tt;
997 
998 	if (callout_active(tp->tt_rexmt))
999 		panic("tcp_setpersist: retransmit pending");
1000 	/*
1001 	 * Start/restart persistance timer.
1002 	 */
1003 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1004 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1005 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1006 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1007 		tp->t_rxtshift++;
1008 }
1009