xref: /freebsd/sys/netinet/tcp_output.c (revision 71fe318b852b8dfb3e799cb12ef184750f7f8eac)
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  * Tcp output routine: figure out what should be sent and send it.
110  */
111 int
112 tcp_output(struct tcpcb *tp)
113 {
114 	struct socket *so = tp->t_inpcb->inp_socket;
115 	long len, win;
116 	int off, flags, error;
117 	struct mbuf *m;
118 	struct ip *ip = NULL;
119 	struct ipovly *ipov = NULL;
120 	struct tcphdr *th;
121 	u_char opt[TCP_MAXOLEN];
122 	unsigned ipoptlen, optlen, hdrlen;
123 	int idle, sendalot;
124 #if 0
125 	int maxburst = TCP_MAXBURST;
126 #endif
127 	struct rmxp_tao *taop;
128 	struct rmxp_tao tao_noncached;
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 #ifndef INET6
137 	mtx_assert(&tp->t_inpcb->inp_mtx, MA_OWNED);
138 #endif
139 
140 	/*
141 	 * Determine length of data that should be transmitted,
142 	 * and flags that will be used.
143 	 * If there is some data or critical controls (SYN, RST)
144 	 * to send, then transmit; otherwise, investigate further.
145 	 */
146 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
147 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
148 		/*
149 		 * We have been idle for "a while" and no acks are
150 		 * expected to clock out any data we send --
151 		 * slow start to get ack "clock" running again.
152 		 *
153 		 * Set the slow-start flight size depending on whether
154 		 * this is a local network or not.
155 		 */
156 		int ss = ss_fltsz;
157 #ifdef INET6
158 		if (isipv6) {
159 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
160 				ss = ss_fltsz_local;
161 		} else
162 #endif /* INET6 */
163 		if (in_localaddr(tp->t_inpcb->inp_faddr))
164 			ss = ss_fltsz_local;
165 		tp->snd_cwnd = tp->t_maxseg * ss;
166 	}
167 	tp->t_flags &= ~TF_LASTIDLE;
168 	if (idle) {
169 		if (tp->t_flags & TF_MORETOCOME) {
170 			tp->t_flags |= TF_LASTIDLE;
171 			idle = 0;
172 		}
173 	}
174 again:
175 	sendalot = 0;
176 	off = tp->snd_nxt - tp->snd_una;
177 	win = min(tp->snd_wnd, tp->snd_cwnd);
178 	win = min(win, tp->snd_bwnd);
179 
180 	flags = tcp_outflags[tp->t_state];
181 	/*
182 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
183 	 * state flags.
184 	 */
185 	if (tp->t_flags & TF_NEEDFIN)
186 		flags |= TH_FIN;
187 	if (tp->t_flags & TF_NEEDSYN)
188 		flags |= TH_SYN;
189 
190 	/*
191 	 * If in persist timeout with window of 0, send 1 byte.
192 	 * Otherwise, if window is small but nonzero
193 	 * and timer expired, we will send what we can
194 	 * and go to transmit state.
195 	 */
196 	if (tp->t_force) {
197 		if (win == 0) {
198 			/*
199 			 * If we still have some data to send, then
200 			 * clear the FIN bit.  Usually this would
201 			 * happen below when it realizes that we
202 			 * aren't sending all the data.  However,
203 			 * if we have exactly 1 byte of unsent data,
204 			 * then it won't clear the FIN bit below,
205 			 * and if we are in persist state, we wind
206 			 * up sending the packet without recording
207 			 * that we sent the FIN bit.
208 			 *
209 			 * We can't just blindly clear the FIN bit,
210 			 * because if we don't have any more data
211 			 * to send then the probe will be the FIN
212 			 * itself.
213 			 */
214 			if (off < so->so_snd.sb_cc)
215 				flags &= ~TH_FIN;
216 			win = 1;
217 		} else {
218 			callout_stop(tp->tt_persist);
219 			tp->t_rxtshift = 0;
220 		}
221 	}
222 
223 	/*
224 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
225 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
226 	 * a negative length.  This can also occur when tcp opens up
227 	 * its congestion window while receiving additional duplicate
228 	 * acks after fast-retransmit because TCP will reset snd_nxt
229 	 * to snd_max after the fast-retransmit.
230 	 *
231 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
232 	 * be set to snd_una, the offset will be 0, and the length may
233 	 * wind up 0.
234 	 */
235 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
236 
237 	if ((taop = tcp_gettaocache(&tp->t_inpcb->inp_inc)) == NULL) {
238 		taop = &tao_noncached;
239 		bzero(taop, sizeof(*taop));
240 	}
241 
242 	/*
243 	 * Lop off SYN bit if it has already been sent.  However, if this
244 	 * is SYN-SENT state and if segment contains data and if we don't
245 	 * know that foreign host supports TAO, suppress sending segment.
246 	 */
247 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
248 		flags &= ~TH_SYN;
249 		off--, len++;
250 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
251 		    taop->tao_ccsent == 0)
252 			return 0;
253 	}
254 
255 	/*
256 	 * Be careful not to send data and/or FIN on SYN segments
257 	 * in cases when no CC option will be sent.
258 	 * This measure is needed to prevent interoperability problems
259 	 * with not fully conformant TCP implementations.
260 	 */
261 	if ((flags & TH_SYN) &&
262 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
263 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
264 		len = 0;
265 		flags &= ~TH_FIN;
266 	}
267 
268 	if (len < 0) {
269 		/*
270 		 * If FIN has been sent but not acked,
271 		 * but we haven't been called to retransmit,
272 		 * len will be < 0.  Otherwise, window shrank
273 		 * after we sent into it.  If window shrank to 0,
274 		 * cancel pending retransmit, pull snd_nxt back
275 		 * to (closed) window, and set the persist timer
276 		 * if it isn't already going.  If the window didn't
277 		 * close completely, just wait for an ACK.
278 		 */
279 		len = 0;
280 		if (win == 0) {
281 			callout_stop(tp->tt_rexmt);
282 			tp->t_rxtshift = 0;
283 			tp->snd_nxt = tp->snd_una;
284 			if (!callout_active(tp->tt_persist))
285 				tcp_setpersist(tp);
286 		}
287 	}
288 
289 	/*
290 	 * len will be >= 0 after this point.  Truncate to the maximum
291 	 * segment length and ensure that FIN is removed if the length
292 	 * no longer contains the last data byte.
293 	 */
294 	if (len > tp->t_maxseg) {
295 		len = tp->t_maxseg;
296 		sendalot = 1;
297 	}
298 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
299 		flags &= ~TH_FIN;
300 
301 	win = sbspace(&so->so_rcv);
302 
303 	/*
304 	 * Sender silly window avoidance.   We transmit under the following
305 	 * conditions when len is non-zero:
306 	 *
307 	 *	- We have a full segment
308 	 *	- This is the last buffer in a write()/send() and we are
309 	 *	  either idle or running NODELAY
310 	 *	- we've timed out (e.g. persist timer)
311 	 *	- we have more then 1/2 the maximum send window's worth of
312 	 *	  data (receiver may be limited the window size)
313 	 *	- we need to retransmit
314 	 */
315 	if (len) {
316 		if (len == tp->t_maxseg)
317 			goto send;
318 		/*
319 		 * NOTE! on localhost connections an 'ack' from the remote
320 		 * end may occur synchronously with the output and cause
321 		 * us to flush a buffer queued with moretocome.  XXX
322 		 *
323 		 * note: the len + off check is almost certainly unnecessary.
324 		 */
325 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
326 		    (idle || (tp->t_flags & TF_NODELAY)) &&
327 		    len + off >= so->so_snd.sb_cc &&
328 		    (tp->t_flags & TF_NOPUSH) == 0) {
329 			goto send;
330 		}
331 		if (tp->t_force)			/* typ. timeout case */
332 			goto send;
333 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
334 			goto send;
335 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
336 			goto send;
337 	}
338 
339 	/*
340 	 * Compare available window to amount of window
341 	 * known to peer (as advertised window less
342 	 * next expected input).  If the difference is at least two
343 	 * max size segments, or at least 50% of the maximum possible
344 	 * window, then want to send a window update to peer.
345 	 */
346 	if (win > 0) {
347 		/*
348 		 * "adv" is the amount we can increase the window,
349 		 * taking into account that we are limited by
350 		 * TCP_MAXWIN << tp->rcv_scale.
351 		 */
352 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
353 			(tp->rcv_adv - tp->rcv_nxt);
354 
355 		if (adv >= (long) (2 * tp->t_maxseg))
356 			goto send;
357 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
358 			goto send;
359 	}
360 
361 	/*
362 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
363 	 * is also a catch-all for the retransmit timer timeout case.
364 	 */
365 	if (tp->t_flags & TF_ACKNOW)
366 		goto send;
367 	if ((flags & TH_RST) ||
368 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
369 		goto send;
370 	if (SEQ_GT(tp->snd_up, tp->snd_una))
371 		goto send;
372 	/*
373 	 * If our state indicates that FIN should be sent
374 	 * and we have not yet done so, then we need to send.
375 	 */
376 	if (flags & TH_FIN &&
377 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
378 		goto send;
379 
380 	/*
381 	 * TCP window updates are not reliable, rather a polling protocol
382 	 * using ``persist'' packets is used to insure receipt of window
383 	 * updates.  The three ``states'' for the output side are:
384 	 *	idle			not doing retransmits or persists
385 	 *	persisting		to move a small or zero window
386 	 *	(re)transmitting	and thereby not persisting
387 	 *
388 	 * callout_active(tp->tt_persist)
389 	 *	is true when we are in persist state.
390 	 * tp->t_force
391 	 *	is set when we are called to send a persist packet.
392 	 * callout_active(tp->tt_rexmt)
393 	 *	is set when we are retransmitting
394 	 * The output side is idle when both timers are zero.
395 	 *
396 	 * If send window is too small, there is data to transmit, and no
397 	 * retransmit or persist is pending, then go to persist state.
398 	 * If nothing happens soon, send when timer expires:
399 	 * if window is nonzero, transmit what we can,
400 	 * otherwise force out a byte.
401 	 */
402 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
403 	    !callout_active(tp->tt_persist)) {
404 		tp->t_rxtshift = 0;
405 		tcp_setpersist(tp);
406 	}
407 
408 	/*
409 	 * No reason to send a segment, just return.
410 	 */
411 	return (0);
412 
413 send:
414 	/*
415 	 * Before ESTABLISHED, force sending of initial options
416 	 * unless TCP set not to do any options.
417 	 * NOTE: we assume that the IP/TCP header plus TCP options
418 	 * always fit in a single mbuf, leaving room for a maximum
419 	 * link header, i.e.
420 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
421 	 */
422 	optlen = 0;
423 #ifdef INET6
424 	if (isipv6)
425 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
426 	else
427 #endif
428 	hdrlen = sizeof (struct tcpiphdr);
429 	if (flags & TH_SYN) {
430 		tp->snd_nxt = tp->iss;
431 		if ((tp->t_flags & TF_NOOPT) == 0) {
432 			u_short mss;
433 
434 			opt[0] = TCPOPT_MAXSEG;
435 			opt[1] = TCPOLEN_MAXSEG;
436 			mss = htons((u_short) tcp_mssopt(tp));
437 			(void)memcpy(opt + 2, &mss, sizeof(mss));
438 			optlen = TCPOLEN_MAXSEG;
439 
440 			if ((tp->t_flags & TF_REQ_SCALE) &&
441 			    ((flags & TH_ACK) == 0 ||
442 			    (tp->t_flags & TF_RCVD_SCALE))) {
443 				*((u_int32_t *)(opt + optlen)) = htonl(
444 					TCPOPT_NOP << 24 |
445 					TCPOPT_WINDOW << 16 |
446 					TCPOLEN_WINDOW << 8 |
447 					tp->request_r_scale);
448 				optlen += 4;
449 			}
450 		}
451  	}
452 
453  	/*
454 	 * Send a timestamp and echo-reply if this is a SYN and our side
455 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
456 	 * and our peer have sent timestamps in our SYN's.
457  	 */
458  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
459  	    (flags & TH_RST) == 0 &&
460 	    ((flags & TH_ACK) == 0 ||
461 	     (tp->t_flags & TF_RCVD_TSTMP))) {
462 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
463 
464  		/* Form timestamp option as shown in appendix A of RFC 1323. */
465  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
466  		*lp++ = htonl(ticks);
467  		*lp   = htonl(tp->ts_recent);
468  		optlen += TCPOLEN_TSTAMP_APPA;
469  	}
470 
471  	/*
472 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
473 	 * options are allowed (!TF_NOOPT) and it's not a RST.
474  	 */
475  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
476  	     (flags & TH_RST) == 0) {
477 		switch (flags & (TH_SYN|TH_ACK)) {
478 		/*
479 		 * This is a normal ACK, send CC if we received CC before
480 		 * from our peer.
481 		 */
482 		case TH_ACK:
483 			if (!(tp->t_flags & TF_RCVD_CC))
484 				break;
485 			/*FALLTHROUGH*/
486 
487 		/*
488 		 * We can only get here in T/TCP's SYN_SENT* state, when
489 		 * we're a sending a non-SYN segment without waiting for
490 		 * the ACK of our SYN.  A check above assures that we only
491 		 * do this if our peer understands T/TCP.
492 		 */
493 		case 0:
494 			opt[optlen++] = TCPOPT_NOP;
495 			opt[optlen++] = TCPOPT_NOP;
496 			opt[optlen++] = TCPOPT_CC;
497 			opt[optlen++] = TCPOLEN_CC;
498 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
499 
500 			optlen += 4;
501 			break;
502 
503 		/*
504 		 * This is our initial SYN, check whether we have to use
505 		 * CC or CC.new.
506 		 */
507 		case TH_SYN:
508 			opt[optlen++] = TCPOPT_NOP;
509 			opt[optlen++] = TCPOPT_NOP;
510 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
511 						TCPOPT_CCNEW : TCPOPT_CC;
512 			opt[optlen++] = TCPOLEN_CC;
513 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
514  			optlen += 4;
515 			break;
516 
517 		/*
518 		 * This is a SYN,ACK; send CC and CC.echo if we received
519 		 * CC from our peer.
520 		 */
521 		case (TH_SYN|TH_ACK):
522 			if (tp->t_flags & TF_RCVD_CC) {
523 				opt[optlen++] = TCPOPT_NOP;
524 				opt[optlen++] = TCPOPT_NOP;
525 				opt[optlen++] = TCPOPT_CC;
526 				opt[optlen++] = TCPOLEN_CC;
527 				*(u_int32_t *)&opt[optlen] =
528 					htonl(tp->cc_send);
529 				optlen += 4;
530 				opt[optlen++] = TCPOPT_NOP;
531 				opt[optlen++] = TCPOPT_NOP;
532 				opt[optlen++] = TCPOPT_CCECHO;
533 				opt[optlen++] = TCPOLEN_CC;
534 				*(u_int32_t *)&opt[optlen] =
535 					htonl(tp->cc_recv);
536 				optlen += 4;
537 			}
538 			break;
539 		}
540  	}
541 
542  	hdrlen += optlen;
543 
544 #ifdef INET6
545 	if (isipv6)
546 		ipoptlen = ip6_optlen(tp->t_inpcb);
547 	else
548 #endif
549 	if (tp->t_inpcb->inp_options)
550 		ipoptlen = tp->t_inpcb->inp_options->m_len -
551 				offsetof(struct ipoption, ipopt_list);
552 	else
553 		ipoptlen = 0;
554 #ifdef IPSEC
555 	ipoptlen += ipsec_hdrsiz_tcp(tp);
556 #endif
557 
558 	/*
559 	 * Adjust data length if insertion of options will
560 	 * bump the packet length beyond the t_maxopd length.
561 	 * Clear the FIN bit because we cut off the tail of
562 	 * the segment.
563 	 */
564 	if (len + optlen + ipoptlen > tp->t_maxopd) {
565 		/*
566 		 * If there is still more to send, don't close the connection.
567 		 */
568 		flags &= ~TH_FIN;
569 		len = tp->t_maxopd - optlen - ipoptlen;
570 		sendalot = 1;
571 	}
572 
573 /*#ifdef DIAGNOSTIC*/
574 #ifdef INET6
575  	if (max_linkhdr + hdrlen > MCLBYTES)
576 #else
577  	if (max_linkhdr + hdrlen > MHLEN)
578 #endif
579 		panic("tcphdr too big");
580 /*#endif*/
581 
582 	/*
583 	 * Grab a header mbuf, attaching a copy of data to
584 	 * be transmitted, and initialize the header from
585 	 * the template for sends on this connection.
586 	 */
587 	if (len) {
588 		if (tp->t_force && len == 1)
589 			tcpstat.tcps_sndprobe++;
590 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
591 			tcpstat.tcps_sndrexmitpack++;
592 			tcpstat.tcps_sndrexmitbyte += len;
593 		} else {
594 			tcpstat.tcps_sndpack++;
595 			tcpstat.tcps_sndbyte += len;
596 		}
597 #ifdef notyet
598 		if ((m = m_copypack(so->so_snd.sb_mb, off,
599 		    (int)len, max_linkhdr + hdrlen)) == 0) {
600 			error = ENOBUFS;
601 			goto out;
602 		}
603 		/*
604 		 * m_copypack left space for our hdr; use it.
605 		 */
606 		m->m_len += hdrlen;
607 		m->m_data -= hdrlen;
608 #else
609 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
610 		if (m == NULL) {
611 			error = ENOBUFS;
612 			goto out;
613 		}
614 #ifdef INET6
615 		if (MHLEN < hdrlen + max_linkhdr) {
616 			MCLGET(m, M_DONTWAIT);
617 			if ((m->m_flags & M_EXT) == 0) {
618 				m_freem(m);
619 				error = ENOBUFS;
620 				goto out;
621 			}
622 		}
623 #endif
624 		m->m_data += max_linkhdr;
625 		m->m_len = hdrlen;
626 		if (len <= MHLEN - hdrlen - max_linkhdr) {
627 			m_copydata(so->so_snd.sb_mb, off, (int) len,
628 			    mtod(m, caddr_t) + hdrlen);
629 			m->m_len += len;
630 		} else {
631 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
632 			if (m->m_next == 0) {
633 				(void) m_free(m);
634 				error = ENOBUFS;
635 				goto out;
636 			}
637 		}
638 #endif
639 		/*
640 		 * If we're sending everything we've got, set PUSH.
641 		 * (This will keep happy those implementations which only
642 		 * give data to the user when a buffer fills or
643 		 * a PUSH comes in.)
644 		 */
645 		if (off + len == so->so_snd.sb_cc)
646 			flags |= TH_PUSH;
647 	} else {
648 		if (tp->t_flags & TF_ACKNOW)
649 			tcpstat.tcps_sndacks++;
650 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
651 			tcpstat.tcps_sndctrl++;
652 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
653 			tcpstat.tcps_sndurg++;
654 		else
655 			tcpstat.tcps_sndwinup++;
656 
657 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
658 		if (m == NULL) {
659 			error = ENOBUFS;
660 			goto out;
661 		}
662 #ifdef INET6
663 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
664 		    MHLEN >= hdrlen) {
665 			MH_ALIGN(m, hdrlen);
666 		} else
667 #endif
668 		m->m_data += max_linkhdr;
669 		m->m_len = hdrlen;
670 	}
671 	m->m_pkthdr.rcvif = (struct ifnet *)0;
672 #ifdef MAC
673 	mac_create_mbuf_from_socket(so, m);
674 #endif
675 #ifdef INET6
676 	if (isipv6) {
677 		ip6 = mtod(m, struct ip6_hdr *);
678 		th = (struct tcphdr *)(ip6 + 1);
679 		tcp_fillheaders(tp, ip6, th);
680 	} else
681 #endif /* INET6 */
682       {
683 	ip = mtod(m, struct ip *);
684 	ipov = (struct ipovly *)ip;
685 	th = (struct tcphdr *)(ip + 1);
686 	/* this picks up the pseudo header (w/o the length) */
687 	tcp_fillheaders(tp, ip, th);
688       }
689 
690 	/*
691 	 * Fill in fields, remembering maximum advertised
692 	 * window for use in delaying messages about window sizes.
693 	 * If resending a FIN, be sure not to use a new sequence number.
694 	 */
695 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
696 	    tp->snd_nxt == tp->snd_max)
697 		tp->snd_nxt--;
698 	/*
699 	 * If we are doing retransmissions, then snd_nxt will
700 	 * not reflect the first unsent octet.  For ACK only
701 	 * packets, we do not want the sequence number of the
702 	 * retransmitted packet, we want the sequence number
703 	 * of the next unsent octet.  So, if there is no data
704 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
705 	 * when filling in ti_seq.  But if we are in persist
706 	 * state, snd_max might reflect one byte beyond the
707 	 * right edge of the window, so use snd_nxt in that
708 	 * case, since we know we aren't doing a retransmission.
709 	 * (retransmit and persist are mutually exclusive...)
710 	 */
711 	if (len || (flags & (TH_SYN|TH_FIN))
712 	    || callout_active(tp->tt_persist))
713 		th->th_seq = htonl(tp->snd_nxt);
714 	else
715 		th->th_seq = htonl(tp->snd_max);
716 	th->th_ack = htonl(tp->rcv_nxt);
717 	if (optlen) {
718 		bcopy(opt, th + 1, optlen);
719 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
720 	}
721 	th->th_flags = flags;
722 	/*
723 	 * Calculate receive window.  Don't shrink window,
724 	 * but avoid silly window syndrome.
725 	 */
726 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
727 		win = 0;
728 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
729 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
730 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
731 		win = (long)TCP_MAXWIN << tp->rcv_scale;
732 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
733 
734 
735 	/*
736 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
737 	 * a 0 window.  This may cause the remote transmitter to stall.  This
738 	 * flag tells soreceive() to disable delayed acknowledgements when
739 	 * draining the buffer.  This can occur if the receiver is attempting
740 	 * to read more data then can be buffered prior to transmitting on
741 	 * the connection.
742 	 */
743 	if (win == 0)
744 		tp->t_flags |= TF_RXWIN0SENT;
745 	else
746 		tp->t_flags &= ~TF_RXWIN0SENT;
747 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
748 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
749 		th->th_flags |= TH_URG;
750 	} else
751 		/*
752 		 * If no urgent pointer to send, then we pull
753 		 * the urgent pointer to the left edge of the send window
754 		 * so that it doesn't drift into the send window on sequence
755 		 * number wraparound.
756 		 */
757 		tp->snd_up = tp->snd_una;		/* drag it along */
758 
759 	/*
760 	 * Put TCP length in extended header, and then
761 	 * checksum extended header and data.
762 	 */
763 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
764 #ifdef INET6
765 	if (isipv6)
766 		/*
767 		 * ip6_plen is not need to be filled now, and will be filled
768 		 * in ip6_output.
769 		 */
770 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
771 				       sizeof(struct tcphdr) + optlen + len);
772 	else
773 #endif /* INET6 */
774       {
775 	m->m_pkthdr.csum_flags = CSUM_TCP;
776 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
777 	if (len + optlen)
778 		th->th_sum = in_addword(th->th_sum,
779 		    htons((u_short)(optlen + len)));
780 
781 	/* IP version must be set here for ipv4/ipv6 checking later */
782 	KASSERT(ip->ip_v == IPVERSION,
783 	    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
784       }
785 
786 	/*
787 	 * In transmit state, time the transmission and arrange for
788 	 * the retransmit.  In persist state, just set snd_max.
789 	 */
790 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
791 		tcp_seq startseq = tp->snd_nxt;
792 
793 		/*
794 		 * Advance snd_nxt over sequence space of this segment.
795 		 */
796 		if (flags & (TH_SYN|TH_FIN)) {
797 			if (flags & TH_SYN)
798 				tp->snd_nxt++;
799 			if (flags & TH_FIN) {
800 				tp->snd_nxt++;
801 				tp->t_flags |= TF_SENTFIN;
802 			}
803 		}
804 		tp->snd_nxt += len;
805 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
806 			tp->snd_max = tp->snd_nxt;
807 			/*
808 			 * Time this transmission if not a retransmission and
809 			 * not currently timing anything.
810 			 */
811 			if (tp->t_rtttime == 0) {
812 				tp->t_rtttime = ticks;
813 				tp->t_rtseq = startseq;
814 				tcpstat.tcps_segstimed++;
815 			}
816 		}
817 
818 		/*
819 		 * Set retransmit timer if not currently set,
820 		 * and not doing a pure ack or a keep-alive probe.
821 		 * Initial value for retransmit timer is smoothed
822 		 * round-trip time + 2 * round-trip time variance.
823 		 * Initialize shift counter which is used for backoff
824 		 * of retransmit time.
825 		 */
826 		if (!callout_active(tp->tt_rexmt) &&
827 		    tp->snd_nxt != tp->snd_una) {
828 			if (callout_active(tp->tt_persist)) {
829 				callout_stop(tp->tt_persist);
830 				tp->t_rxtshift = 0;
831 			}
832 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
833 				      tcp_timer_rexmt, tp);
834 		}
835 	} else {
836 		/*
837 		 * Persist case, update snd_max but since we are in
838 		 * persist mode (no window) we do not update snd_nxt.
839 		 */
840 		int xlen = len;
841 		if (flags & TH_SYN)
842 			++xlen;
843 		if (flags & TH_FIN) {
844 			++xlen;
845 			tp->t_flags |= TF_SENTFIN;
846 		}
847 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
848 			tp->snd_max = tp->snd_nxt + len;
849 	}
850 
851 #ifdef TCPDEBUG
852 	/*
853 	 * Trace.
854 	 */
855 	if (so->so_options & SO_DEBUG)
856 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
857 #endif
858 
859 	/*
860 	 * Fill in IP length and desired time to live and
861 	 * send to IP level.  There should be a better way
862 	 * to handle ttl and tos; we could keep them in
863 	 * the template, but need a way to checksum without them.
864 	 */
865 	/*
866 	 * m->m_pkthdr.len should have been set before cksum calcuration,
867 	 * because in6_cksum() need it.
868 	 */
869 #ifdef INET6
870 	if (isipv6) {
871 		/*
872 		 * we separately set hoplimit for every segment, since the
873 		 * user might want to change the value via setsockopt.
874 		 * Also, desired default hop limit might be changed via
875 		 * Neighbor Discovery.
876 		 */
877 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
878 					       tp->t_inpcb->in6p_route.ro_rt ?
879 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
880 					       : NULL);
881 
882 		/* TODO: IPv6 IP6TOS_ECT bit on */
883 		error = ip6_output(m,
884 			    tp->t_inpcb->in6p_outputopts,
885 			    &tp->t_inpcb->in6p_route,
886 			    (so->so_options & SO_DONTROUTE), NULL, NULL,
887 			    tp->t_inpcb);
888 	} else
889 #endif /* INET6 */
890     {
891 	struct rtentry *rt;
892 	ip->ip_len = m->m_pkthdr.len;
893 #ifdef INET6
894  	if (INP_CHECK_SOCKAF(so, AF_INET6))
895  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb,
896  					    tp->t_inpcb->in6p_route.ro_rt ?
897  					    tp->t_inpcb->in6p_route.ro_rt->rt_ifp
898  					    : NULL);
899  	else
900 #endif /* INET6 */
901 	ip->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
902 	ip->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
903 	/*
904 	 * See if we should do MTU discovery.  We do it only if the following
905 	 * are true:
906 	 *	1) we have a valid route to the destination
907 	 *	2) the MTU is not locked (if it is, then discovery has been
908 	 *	   disabled)
909 	 */
910 	if (path_mtu_discovery
911 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
912 	    && rt->rt_flags & RTF_UP
913 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
914 		ip->ip_off |= IP_DF;
915 	}
916 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
917 	    (so->so_options & SO_DONTROUTE), 0, tp->t_inpcb);
918     }
919 	if (error) {
920 
921 		/*
922 		 * We know that the packet was lost, so back out the
923 		 * sequence number advance, if any.
924 		 */
925 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
926 			/*
927 			 * No need to check for TH_FIN here because
928 			 * the TF_SENTFIN flag handles that case.
929 			 */
930 			if ((flags & TH_SYN) == 0)
931 				tp->snd_nxt -= len;
932 		}
933 
934 out:
935 		if (error == ENOBUFS) {
936 	                if (!callout_active(tp->tt_rexmt) &&
937                             !callout_active(tp->tt_persist))
938 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
939                                       tcp_timer_rexmt, tp);
940 			tcp_quench(tp->t_inpcb, 0);
941 			return (0);
942 		}
943 		if (error == EMSGSIZE) {
944 			/*
945 			 * ip_output() will have already fixed the route
946 			 * for us.  tcp_mtudisc() will, as its last action,
947 			 * initiate retransmission, so it is important to
948 			 * not do so here.
949 			 */
950 			tcp_mtudisc(tp->t_inpcb, 0);
951 			return 0;
952 		}
953 		if ((error == EHOSTUNREACH || error == ENETDOWN)
954 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
955 			tp->t_softerror = error;
956 			return (0);
957 		}
958 		return (error);
959 	}
960 	tcpstat.tcps_sndtotal++;
961 
962 	/*
963 	 * Data sent (as far as we can tell).
964 	 * If this advertises a larger window than any other segment,
965 	 * then remember the size of the advertised window.
966 	 * Any pending ACK has now been sent.
967 	 */
968 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
969 		tp->rcv_adv = tp->rcv_nxt + win;
970 	tp->last_ack_sent = tp->rcv_nxt;
971 	tp->t_flags &= ~TF_ACKNOW;
972 	if (tcp_delack_enabled)
973 		callout_stop(tp->tt_delack);
974 #if 0
975 	/*
976 	 * This completely breaks TCP if newreno is turned on.  What happens
977 	 * is that if delayed-acks are turned on on the receiver, this code
978 	 * on the transmitter effectively destroys the TCP window, forcing
979 	 * it to four packets (1.5Kx4 = 6K window).
980 	 */
981 	if (sendalot && (!tcp_do_newreno || --maxburst))
982 		goto again;
983 #endif
984 	if (sendalot)
985 		goto again;
986 	return (0);
987 }
988 
989 void
990 tcp_setpersist(tp)
991 	register struct tcpcb *tp;
992 {
993 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
994 	int tt;
995 
996 	if (callout_active(tp->tt_rexmt))
997 		panic("tcp_setpersist: retransmit pending");
998 	/*
999 	 * Start/restart persistance timer.
1000 	 */
1001 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1002 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1003 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1004 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1005 		tp->t_rxtshift++;
1006 }
1007