xref: /dflybsd-src/sys/netinet/tcp_output.c (revision ad63697b1227b985f99dd96f67aaafb238b3d148)
1 /*
2  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
68  * $DragonFly: src/sys/netinet/tcp_output.c,v 1.34 2007/04/22 01:13:14 dillon Exp $
69  */
70 
71 #include "opt_inet.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
80 #include <sys/mbuf.h>
81 #include <sys/domain.h>
82 #include <sys/protosw.h>
83 #include <sys/socket.h>
84 #include <sys/socketvar.h>
85 #include <sys/in_cksum.h>
86 #include <sys/thread.h>
87 #include <sys/globaldata.h>
88 
89 #include <net/route.h>
90 
91 #include <netinet/in.h>
92 #include <netinet/in_systm.h>
93 #include <netinet/ip.h>
94 #include <netinet/in_pcb.h>
95 #include <netinet/ip_var.h>
96 #include <netinet6/in6_pcb.h>
97 #include <netinet/ip6.h>
98 #include <netinet6/ip6_var.h>
99 #include <netinet/tcp.h>
100 #define	TCPOUTFLAGS
101 #include <netinet/tcp_fsm.h>
102 #include <netinet/tcp_seq.h>
103 #include <netinet/tcp_timer.h>
104 #include <netinet/tcp_timer2.h>
105 #include <netinet/tcp_var.h>
106 #include <netinet/tcpip.h>
107 #ifdef TCPDEBUG
108 #include <netinet/tcp_debug.h>
109 #endif
110 
111 #ifdef IPSEC
112 #include <netinet6/ipsec.h>
113 #endif /*IPSEC*/
114 
115 #ifdef FAST_IPSEC
116 #include <netproto/ipsec/ipsec.h>
117 #define	IPSEC
118 #endif /*FAST_IPSEC*/
119 
120 #ifdef notyet
121 extern struct mbuf *m_copypack();
122 #endif
123 
124 int path_mtu_discovery = 0;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
126 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
127 
128 static int avoid_pure_win_update = 1;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
130 	&avoid_pure_win_update, 1, "Avoid pure window updates when possible");
131 
132 int tcp_do_autosndbuf = 1;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
134     &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
135 
136 int tcp_autosndbuf_inc = 8*1024;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
138     &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
139 
140 int tcp_autosndbuf_max = 2*1024*1024;
141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
142     &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
143 
144 static int tcp_idle_cwv = 1;
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
146     &tcp_idle_cwv, 0,
147     "Congestion window validation after idle period (part of RFC2861)");
148 
149 static int tcp_idle_restart = 1;
150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
151     &tcp_idle_restart, 0, "Reset congestion window after idle period");
152 
153 static void	tcp_idle_cwnd_validate(struct tcpcb *);
154 
155 /*
156  * Tcp output routine: figure out what should be sent and send it.
157  */
158 int
159 tcp_output(struct tcpcb *tp)
160 {
161 	struct inpcb * const inp = tp->t_inpcb;
162 	struct socket *so = inp->inp_socket;
163 	long len, recvwin, sendwin;
164 	int nsacked = 0;
165 	int off, flags, error = 0;
166 #ifdef TCP_SIGNATURE
167 	int sigoff = 0;
168 #endif
169 	struct mbuf *m;
170 	struct ip *ip;
171 	struct ipovly *ipov;
172 	struct tcphdr *th;
173 	u_char opt[TCP_MAXOLEN];
174 	unsigned int ipoptlen, optlen, hdrlen;
175 	int idle, idle_cwv = 0;
176 	boolean_t sendalot;
177 	struct ip6_hdr *ip6;
178 #ifdef INET6
179 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
180 #else
181 	const boolean_t isipv6 = FALSE;
182 #endif
183 
184 	KKASSERT(so->so_port == &curthread->td_msgport);
185 
186 	/*
187 	 * Determine length of data that should be transmitted,
188 	 * and flags that will be used.
189 	 * If there is some data or critical controls (SYN, RST)
190 	 * to send, then transmit; otherwise, investigate further.
191 	 */
192 
193 	/*
194 	 * If we have been idle for a while, the send congestion window
195 	 * could be no longer representative of the current state of the
196 	 * link; need to validate congestion window.  However, we should
197 	 * not perform congestion window validation here, since we could
198 	 * be asked to send pure ACK.
199 	 */
200 	if (tp->snd_max == tp->snd_una &&
201 	    (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
202 		idle_cwv = 1;
203 
204 	/*
205 	 * Calculate whether the transmit stream was previously idle
206 	 * and adjust TF_LASTIDLE for the next time.
207 	 */
208 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
209 	if (idle && (tp->t_flags & TF_MORETOCOME))
210 		tp->t_flags |= TF_LASTIDLE;
211 	else
212 		tp->t_flags &= ~TF_LASTIDLE;
213 
214 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
215 	    !IN_FASTRECOVERY(tp))
216 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
217 
218 again:
219 	m = NULL;
220 	ip = NULL;
221 	ipov = NULL;
222 	th = NULL;
223 	ip6 = NULL;
224 
225 	/* Make use of SACK information when slow-starting after a RTO. */
226 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
227 	    !IN_FASTRECOVERY(tp)) {
228 		tcp_seq old_snd_nxt = tp->snd_nxt;
229 
230 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
231 		nsacked += tp->snd_nxt - old_snd_nxt;
232 	}
233 
234 	sendalot = FALSE;
235 	off = tp->snd_nxt - tp->snd_una;
236 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
237 	sendwin = min(sendwin, tp->snd_bwnd);
238 
239 	flags = tcp_outflags[tp->t_state];
240 	/*
241 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
242 	 * state flags.
243 	 */
244 	if (tp->t_flags & TF_NEEDFIN)
245 		flags |= TH_FIN;
246 	if (tp->t_flags & TF_NEEDSYN)
247 		flags |= TH_SYN;
248 
249 	/*
250 	 * If in persist timeout with window of 0, send 1 byte.
251 	 * Otherwise, if window is small but nonzero
252 	 * and timer expired, we will send what we can
253 	 * and go to transmit state.
254 	 */
255 	if (tp->t_flags & TF_FORCE) {
256 		if (sendwin == 0) {
257 			/*
258 			 * If we still have some data to send, then
259 			 * clear the FIN bit.  Usually this would
260 			 * happen below when it realizes that we
261 			 * aren't sending all the data.  However,
262 			 * if we have exactly 1 byte of unsent data,
263 			 * then it won't clear the FIN bit below,
264 			 * and if we are in persist state, we wind
265 			 * up sending the packet without recording
266 			 * that we sent the FIN bit.
267 			 *
268 			 * We can't just blindly clear the FIN bit,
269 			 * because if we don't have any more data
270 			 * to send then the probe will be the FIN
271 			 * itself.
272 			 */
273 			if (off < so->so_snd.ssb_cc)
274 				flags &= ~TH_FIN;
275 			sendwin = 1;
276 		} else {
277 			tcp_callout_stop(tp, tp->tt_persist);
278 			tp->t_rxtshift = 0;
279 		}
280 	}
281 
282 	/*
283 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
284 	 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
285 	 * a negative length.  This can also occur when TCP opens up
286 	 * its congestion window while receiving additional duplicate
287 	 * acks after fast-retransmit because TCP will reset snd_nxt
288 	 * to snd_max after the fast-retransmit.
289 	 *
290 	 * A negative length can also occur when we are in the
291 	 * TCPS_SYN_RECEIVED state due to a simultanious connect where
292 	 * our SYN has not been acked yet.
293 	 *
294 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
295 	 * be set to snd_una, the offset will be 0, and the length may
296 	 * wind up 0.
297 	 */
298 	len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
299 
300 	/*
301 	 * Lop off SYN bit if it has already been sent.  However, if this
302 	 * is SYN-SENT state and if segment contains data, suppress sending
303 	 * segment (sending the segment would be an option if we still
304 	 * did TAO and the remote host supported it).
305 	 */
306 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
307 		flags &= ~TH_SYN;
308 		off--, len++;
309 		if (len > 0 && tp->t_state == TCPS_SYN_SENT)
310 			return 0;
311 	}
312 
313 	/*
314 	 * Be careful not to send data and/or FIN on SYN segments.
315 	 * This measure is needed to prevent interoperability problems
316 	 * with not fully conformant TCP implementations.
317 	 */
318 	if (flags & TH_SYN) {
319 		len = 0;
320 		flags &= ~TH_FIN;
321 	}
322 
323 	if (len < 0) {
324 		/*
325 		 * A negative len can occur if our FIN has been sent but not
326 		 * acked, or if we are in a simultanious connect in the
327 		 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
328 		 * acked.
329 		 *
330 		 * If our window has contracted to 0 in the FIN case
331 		 * (which can only occur if we have NOT been called to
332 		 * retransmit as per code a few paragraphs up) then we
333 		 * want to shift the retransmit timer over to the
334 		 * persist timer.
335 		 *
336 		 * However, if we are in the TCPS_SYN_RECEIVED state
337 		 * (the SYN case) we will be in a simultanious connect and
338 		 * the window may be zero degeneratively.  In this case we
339 		 * do not want to shift to the persist timer after the SYN
340 		 * or the SYN+ACK transmission.
341 		 */
342 		len = 0;
343 		if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
344 			tcp_callout_stop(tp, tp->tt_rexmt);
345 			tp->t_rxtshift = 0;
346 			tp->snd_nxt = tp->snd_una;
347 			if (!tcp_callout_active(tp, tp->tt_persist))
348 				tcp_setpersist(tp);
349 		}
350 	}
351 
352 	KASSERT(len >= 0, ("%s: len < 0", __func__));
353 	/*
354 	 * Automatic sizing of send socket buffer.  Often the send buffer
355 	 * size is not optimally adjusted to the actual network conditions
356 	 * at hand (delay bandwidth product).  Setting the buffer size too
357 	 * small limits throughput on links with high bandwidth and high
358 	 * delay (eg. trans-continental/oceanic links).  Setting the
359 	 * buffer size too big consumes too much real kernel memory,
360 	 * especially with many connections on busy servers.
361 	 *
362 	 * The criteria to step up the send buffer one notch are:
363 	 *  1. receive window of remote host is larger than send buffer
364 	 *     (with a fudge factor of 5/4th);
365 	 *  2. send buffer is filled to 7/8th with data (so we actually
366 	 *     have data to make use of it);
367 	 *  3. send buffer fill has not hit maximal automatic size;
368 	 *  4. our send window (slow start and cogestion controlled) is
369 	 *     larger than sent but unacknowledged data in send buffer.
370 	 *
371 	 * The remote host receive window scaling factor may limit the
372 	 * growing of the send buffer before it reaches its allowed
373 	 * maximum.
374 	 *
375 	 * It scales directly with slow start or congestion window
376 	 * and does at most one step per received ACK.  This fast
377 	 * scaling has the drawback of growing the send buffer beyond
378 	 * what is strictly necessary to make full use of a given
379 	 * delay*bandwith product.  However testing has shown this not
380 	 * to be much of an problem.  At worst we are trading wasting
381 	 * of available bandwith (the non-use of it) for wasting some
382 	 * socket buffer memory.
383 	 *
384 	 * TODO: Shrink send buffer during idle periods together
385 	 * with congestion window.  Requires another timer.  Has to
386 	 * wait for upcoming tcp timer rewrite.
387 	 */
388 	if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
389 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
390 		    so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
391 		    so->so_snd.ssb_cc < tcp_autosndbuf_max &&
392 		    sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
393 			u_long newsize;
394 
395 			newsize = ulmin(so->so_snd.ssb_hiwat +
396 					 tcp_autosndbuf_inc,
397 					tcp_autosndbuf_max);
398 			if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
399 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
400 			if (newsize >= (TCP_MAXWIN << tp->snd_scale))
401 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
402 		}
403 	}
404 
405 	/*
406 	 * Truncate to the maximum segment length and ensure that FIN is
407 	 * removed if the length no longer contains the last data byte.
408 	 */
409 	if (len > tp->t_maxseg) {
410 		len = tp->t_maxseg;
411 		sendalot = TRUE;
412 	}
413 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
414 		flags &= ~TH_FIN;
415 
416 	recvwin = ssb_space(&so->so_rcv);
417 
418 	/*
419 	 * Sender silly window avoidance.   We transmit under the following
420 	 * conditions when len is non-zero:
421 	 *
422 	 *	- We have a full segment
423 	 *	- This is the last buffer in a write()/send() and we are
424 	 *	  either idle or running NODELAY
425 	 *	- we've timed out (e.g. persist timer)
426 	 *	- we have more then 1/2 the maximum send window's worth of
427 	 *	  data (receiver may be limiting the window size)
428 	 *	- we need to retransmit
429 	 */
430 	if (len) {
431 		if (len == tp->t_maxseg)
432 			goto send;
433 		/*
434 		 * NOTE! on localhost connections an 'ack' from the remote
435 		 * end may occur synchronously with the output and cause
436 		 * us to flush a buffer queued with moretocome.  XXX
437 		 *
438 		 * note: the len + off check is almost certainly unnecessary.
439 		 */
440 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
441 		    (idle || (tp->t_flags & TF_NODELAY)) &&
442 		    len + off >= so->so_snd.ssb_cc &&
443 		    !(tp->t_flags & TF_NOPUSH)) {
444 			goto send;
445 		}
446 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
447 			goto send;
448 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
449 			goto send;
450 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
451 			goto send;
452 	}
453 
454 	/*
455 	 * Compare available window to amount of window
456 	 * known to peer (as advertised window less
457 	 * next expected input).  If the difference is at least two
458 	 * max size segments, or at least 50% of the maximum possible
459 	 * window, then want to send a window update to peer.
460 	 */
461 	if (recvwin > 0) {
462 		/*
463 		 * "adv" is the amount we can increase the window,
464 		 * taking into account that we are limited by
465 		 * TCP_MAXWIN << tp->rcv_scale.
466 		 */
467 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
468 			(tp->rcv_adv - tp->rcv_nxt);
469 		long hiwat;
470 
471 		/*
472 		 * This ack case typically occurs when the user has drained
473 		 * the TCP socket buffer sufficiently to warrent an ack
474 		 * containing a 'pure window update'... that is, an ack that
475 		 * ONLY updates the tcp window.
476 		 *
477 		 * It is unclear why we would need to do a pure window update
478 		 * past 2 segments if we are going to do one at 1/2 the high
479 		 * water mark anyway, especially since under normal conditions
480 		 * the user program will drain the socket buffer quickly.
481 		 * The 2-segment pure window update will often add a large
482 		 * number of extra, unnecessary acks to the stream.
483 		 *
484 		 * avoid_pure_win_update now defaults to 1.
485 		 */
486 		if (avoid_pure_win_update == 0 ||
487 		    (tp->t_flags & TF_RXRESIZED)) {
488 			if (adv >= (long) (2 * tp->t_maxseg)) {
489 				goto send;
490 			}
491 		}
492 		hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
493 		if (hiwat > (long)so->so_rcv.ssb_hiwat)
494 			hiwat = (long)so->so_rcv.ssb_hiwat;
495 		if (adv >= hiwat / 2)
496 			goto send;
497 	}
498 
499 	/*
500 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
501 	 * is also a catch-all for the retransmit timer timeout case.
502 	 */
503 	if (tp->t_flags & TF_ACKNOW)
504 		goto send;
505 	if ((flags & TH_RST) ||
506 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
507 		goto send;
508 	if (SEQ_GT(tp->snd_up, tp->snd_una))
509 		goto send;
510 	/*
511 	 * If our state indicates that FIN should be sent
512 	 * and we have not yet done so, then we need to send.
513 	 */
514 	if ((flags & TH_FIN) &&
515 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
516 		goto send;
517 
518 	/*
519 	 * TCP window updates are not reliable, rather a polling protocol
520 	 * using ``persist'' packets is used to insure receipt of window
521 	 * updates.  The three ``states'' for the output side are:
522 	 *	idle			not doing retransmits or persists
523 	 *	persisting		to move a small or zero window
524 	 *	(re)transmitting	and thereby not persisting
525 	 *
526 	 * tcp_callout_active(tp, tp->tt_persist)
527 	 *	is true when we are in persist state.
528 	 * The TF_FORCE flag in tp->t_flags
529 	 *	is set when we are called to send a persist packet.
530 	 * tcp_callout_active(tp, tp->tt_rexmt)
531 	 *	is set when we are retransmitting
532 	 * The output side is idle when both timers are zero.
533 	 *
534 	 * If send window is too small, there is data to transmit, and no
535 	 * retransmit or persist is pending, then go to persist state.
536 	 *
537 	 * If nothing happens soon, send when timer expires:
538 	 * if window is nonzero, transmit what we can, otherwise force out
539 	 * a byte.
540 	 *
541 	 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
542 	 * with data pending.  This situation can occur during a
543 	 * simultanious connect.
544 	 */
545 	if (so->so_snd.ssb_cc > 0 &&
546 	    tp->t_state != TCPS_SYN_RECEIVED &&
547 	    !tcp_callout_active(tp, tp->tt_rexmt) &&
548 	    !tcp_callout_active(tp, tp->tt_persist)) {
549 		tp->t_rxtshift = 0;
550 		tcp_setpersist(tp);
551 	}
552 
553 	/*
554 	 * No reason to send a segment, just return.
555 	 */
556 	return (0);
557 
558 send:
559 	/*
560 	 * Before ESTABLISHED, force sending of initial options
561 	 * unless TCP set not to do any options.
562 	 * NOTE: we assume that the IP/TCP header plus TCP options
563 	 * always fit in a single mbuf, leaving room for a maximum
564 	 * link header, i.e.
565 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
566 	 */
567 	optlen = 0;
568 	if (isipv6)
569 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
570 	else
571 		hdrlen = sizeof(struct tcpiphdr);
572 	if (flags & TH_SYN) {
573 		tp->snd_nxt = tp->iss;
574 		if (!(tp->t_flags & TF_NOOPT)) {
575 			u_short mss;
576 
577 			opt[0] = TCPOPT_MAXSEG;
578 			opt[1] = TCPOLEN_MAXSEG;
579 			mss = htons((u_short) tcp_mssopt(tp));
580 			memcpy(opt + 2, &mss, sizeof mss);
581 			optlen = TCPOLEN_MAXSEG;
582 
583 			if ((tp->t_flags & TF_REQ_SCALE) &&
584 			    (!(flags & TH_ACK) ||
585 			     (tp->t_flags & TF_RCVD_SCALE))) {
586 				*((u_int32_t *)(opt + optlen)) = htonl(
587 					TCPOPT_NOP << 24 |
588 					TCPOPT_WINDOW << 16 |
589 					TCPOLEN_WINDOW << 8 |
590 					tp->request_r_scale);
591 				optlen += 4;
592 			}
593 
594 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
595 			    tp->t_flags & TF_SACK_PERMITTED) {
596 				uint32_t *lp = (uint32_t *)(opt + optlen);
597 
598 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
599 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
600 			}
601 		}
602 	}
603 
604 	/*
605 	 * Send a timestamp and echo-reply if this is a SYN and our side
606 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
607 	 * and our peer have sent timestamps in our SYN's.
608 	 */
609 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
610 	    !(flags & TH_RST) &&
611 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
612 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
613 
614 		/* Form timestamp option as shown in appendix A of RFC 1323. */
615 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
616 		*lp++ = htonl(ticks);
617 		*lp   = htonl(tp->ts_recent);
618 		optlen += TCPOLEN_TSTAMP_APPA;
619 	}
620 
621 	/* Set receive buffer autosizing timestamp. */
622 	if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
623 		tp->rfbuf_ts = ticks;
624 
625 	/*
626 	 * If this is a SACK connection and we have a block to report,
627 	 * fill in the SACK blocks in the TCP options.
628 	 */
629 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
630 		TF_SACK_PERMITTED &&
631 	    (!LIST_EMPTY(&tp->t_segq) ||
632 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
633 		tcp_sack_fill_report(tp, opt, &optlen);
634 
635 #ifdef TCP_SIGNATURE
636 	if (tp->t_flags & TF_SIGNATURE) {
637 		int i;
638 		u_char *bp;
639 		/*
640 		 * Initialize TCP-MD5 option (RFC2385)
641 		 */
642 		bp = (u_char *)opt + optlen;
643 		*bp++ = TCPOPT_SIGNATURE;
644 		*bp++ = TCPOLEN_SIGNATURE;
645 		sigoff = optlen + 2;
646 		for (i = 0; i < TCP_SIGLEN; i++)
647 			*bp++ = 0;
648 		optlen += TCPOLEN_SIGNATURE;
649 		/*
650 		 * Terminate options list and maintain 32-bit alignment.
651 		 */
652 		*bp++ = TCPOPT_NOP;
653 		*bp++ = TCPOPT_EOL;
654 		optlen += 2;
655 	}
656 #endif /* TCP_SIGNATURE */
657 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
658 	hdrlen += optlen;
659 
660 	if (isipv6) {
661 		ipoptlen = ip6_optlen(inp);
662 	} else {
663 		if (inp->inp_options) {
664 			ipoptlen = inp->inp_options->m_len -
665 			    offsetof(struct ipoption, ipopt_list);
666 		} else {
667 			ipoptlen = 0;
668 		}
669 	}
670 #ifdef IPSEC
671 	ipoptlen += ipsec_hdrsiz_tcp(tp);
672 #endif
673 
674 	/*
675 	 * Adjust data length if insertion of options will bump the packet
676 	 * length beyond the t_maxopd length.  Clear FIN to prevent premature
677 	 * closure since there is still more data to send after this (now
678 	 * truncated) packet.
679 	 *
680 	 * If just the options do not fit we are in a no-win situation and
681 	 * we treat it as an unreachable host.
682 	 */
683 	if (len + optlen + ipoptlen > tp->t_maxopd) {
684 		if (tp->t_maxopd <= optlen + ipoptlen) {
685 			static time_t last_optlen_report;
686 
687 			if (last_optlen_report != time_second) {
688 				last_optlen_report = time_second;
689 				kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd);
690 			}
691 			error = EHOSTUNREACH;
692 			goto out;
693 		} else {
694 			flags &= ~TH_FIN;
695 			len = tp->t_maxopd - optlen - ipoptlen;
696 			sendalot = TRUE;
697 		}
698 	}
699 
700 #ifdef INET6
701 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
702 #else
703 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
704 #endif
705 
706 	/*
707 	 * Grab a header mbuf, attaching a copy of data to
708 	 * be transmitted, and initialize the header from
709 	 * the template for sends on this connection.
710 	 */
711 	if (len) {
712 		if ((tp->t_flags & TF_FORCE) && len == 1)
713 			tcpstat.tcps_sndprobe++;
714 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
715 			if (tp->snd_nxt == tp->snd_una)
716 				tp->snd_max_rexmt = tp->snd_max;
717 			if (nsacked) {
718 				tcpstat.tcps_sndsackrtopack++;
719 				tcpstat.tcps_sndsackrtobyte += len;
720 			}
721 			tcpstat.tcps_sndrexmitpack++;
722 			tcpstat.tcps_sndrexmitbyte += len;
723 		} else {
724 			tcpstat.tcps_sndpack++;
725 			tcpstat.tcps_sndbyte += len;
726 		}
727 		if (idle_cwv) {
728 			idle_cwv = 0;
729 			tcp_idle_cwnd_validate(tp);
730 		}
731 		/* Update last send time after CWV */
732 		tp->snd_last = ticks;
733 #ifdef notyet
734 		if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
735 		    max_linkhdr + hdrlen)) == NULL) {
736 			error = ENOBUFS;
737 			goto after_th;
738 		}
739 		/*
740 		 * m_copypack left space for our hdr; use it.
741 		 */
742 		m->m_len += hdrlen;
743 		m->m_data -= hdrlen;
744 #else
745 #ifndef INET6
746 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
747 #else
748 		m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
749 			   M_PKTHDR, NULL);
750 #endif
751 		if (m == NULL) {
752 			error = ENOBUFS;
753 			goto after_th;
754 		}
755 		m->m_data += max_linkhdr;
756 		m->m_len = hdrlen;
757 		if (len <= MHLEN - hdrlen - max_linkhdr) {
758 			m_copydata(so->so_snd.ssb_mb, off, (int) len,
759 			    mtod(m, caddr_t) + hdrlen);
760 			m->m_len += len;
761 		} else {
762 			m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
763 			if (m->m_next == NULL) {
764 				m_free(m);
765 				m = NULL;
766 				error = ENOBUFS;
767 				goto after_th;
768 			}
769 		}
770 #endif
771 		/*
772 		 * If we're sending everything we've got, set PUSH.
773 		 * (This will keep happy those implementations which only
774 		 * give data to the user when a buffer fills or
775 		 * a PUSH comes in.)
776 		 */
777 		if (off + len == so->so_snd.ssb_cc)
778 			flags |= TH_PUSH;
779 	} else {
780 		if (tp->t_flags & TF_ACKNOW)
781 			tcpstat.tcps_sndacks++;
782 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
783 			tcpstat.tcps_sndctrl++;
784 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
785 			tcpstat.tcps_sndurg++;
786 		else
787 			tcpstat.tcps_sndwinup++;
788 
789 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
790 		if (m == NULL) {
791 			error = ENOBUFS;
792 			goto after_th;
793 		}
794 		if (isipv6 &&
795 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
796 			MH_ALIGN(m, hdrlen);
797 		else
798 			m->m_data += max_linkhdr;
799 		m->m_len = hdrlen;
800 	}
801 	m->m_pkthdr.rcvif = NULL;
802 	if (isipv6) {
803 		ip6 = mtod(m, struct ip6_hdr *);
804 		th = (struct tcphdr *)(ip6 + 1);
805 		tcp_fillheaders(tp, ip6, th);
806 	} else {
807 		ip = mtod(m, struct ip *);
808 		ipov = (struct ipovly *)ip;
809 		th = (struct tcphdr *)(ip + 1);
810 		/* this picks up the pseudo header (w/o the length) */
811 		tcp_fillheaders(tp, ip, th);
812 	}
813 after_th:
814 	/*
815 	 * Fill in fields, remembering maximum advertised
816 	 * window for use in delaying messages about window sizes.
817 	 * If resending a FIN, be sure not to use a new sequence number.
818 	 */
819 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
820 	    tp->snd_nxt == tp->snd_max)
821 		tp->snd_nxt--;
822 
823 	if (th != NULL) {
824 		/*
825 		 * If we are doing retransmissions, then snd_nxt will
826 		 * not reflect the first unsent octet.  For ACK only
827 		 * packets, we do not want the sequence number of the
828 		 * retransmitted packet, we want the sequence number
829 		 * of the next unsent octet.  So, if there is no data
830 		 * (and no SYN or FIN), use snd_max instead of snd_nxt
831 		 * when filling in ti_seq.  But if we are in persist
832 		 * state, snd_max might reflect one byte beyond the
833 		 * right edge of the window, so use snd_nxt in that
834 		 * case, since we know we aren't doing a retransmission.
835 		 * (retransmit and persist are mutually exclusive...)
836 		 */
837 		if (len || (flags & (TH_SYN|TH_FIN)) ||
838 		    tcp_callout_active(tp, tp->tt_persist))
839 			th->th_seq = htonl(tp->snd_nxt);
840 		else
841 			th->th_seq = htonl(tp->snd_max);
842 		th->th_ack = htonl(tp->rcv_nxt);
843 		if (optlen) {
844 			bcopy(opt, th + 1, optlen);
845 			th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
846 		}
847 		th->th_flags = flags;
848 	}
849 
850 	/*
851 	 * Calculate receive window.  Don't shrink window, but avoid
852 	 * silly window syndrome by sending a 0 window if the actual
853 	 * window is less then one segment.
854 	 */
855 	if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
856 	    recvwin < (long)tp->t_maxseg)
857 		recvwin = 0;
858 	if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
859 		recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
860 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
861 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
862 
863 	/*
864 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
865 	 * a 0 window.  This may cause the remote transmitter to stall.  This
866 	 * flag tells soreceive() to disable delayed acknowledgements when
867 	 * draining the buffer.  This can occur if the receiver is attempting
868 	 * to read more data then can be buffered prior to transmitting on
869 	 * the connection.
870 	 */
871 	if (recvwin == 0)
872 		tp->t_flags |= TF_RXWIN0SENT;
873 	else
874 		tp->t_flags &= ~TF_RXWIN0SENT;
875 
876 	if (th != NULL)
877 		th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
878 
879 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
880 		if (th != NULL) {
881 			th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
882 			th->th_flags |= TH_URG;
883 		}
884 	} else {
885 		/*
886 		 * If no urgent pointer to send, then we pull
887 		 * the urgent pointer to the left edge of the send window
888 		 * so that it doesn't drift into the send window on sequence
889 		 * number wraparound.
890 		 */
891 		tp->snd_up = tp->snd_una;		/* drag it along */
892 	}
893 
894 	if (th != NULL) {
895 #ifdef TCP_SIGNATURE
896 		if (tp->t_flags & TF_SIGNATURE) {
897 			tcpsignature_compute(m, len, optlen,
898 			    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
899 		}
900 #endif /* TCP_SIGNATURE */
901 
902 		/*
903 		 * Put TCP length in extended header, and then
904 		 * checksum extended header and data.
905 		 */
906 		m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
907 		if (isipv6) {
908 			/*
909 			 * ip6_plen is not need to be filled now, and will be
910 			 * filled in ip6_output().
911 			 */
912 			th->th_sum = in6_cksum(m, IPPROTO_TCP,
913 			    sizeof(struct ip6_hdr),
914 			    sizeof(struct tcphdr) + optlen + len);
915 		} else {
916 			m->m_pkthdr.csum_flags = CSUM_TCP;
917 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
918 			if (len + optlen) {
919 				th->th_sum = in_addword(th->th_sum,
920 				    htons((u_short)(optlen + len)));
921 			}
922 
923 			/*
924 			 * IP version must be set here for ipv4/ipv6 checking
925 			 * later
926 			 */
927 			KASSERT(ip->ip_v == IPVERSION,
928 			    ("%s: IP version incorrect: %d",
929 			     __func__, ip->ip_v));
930 		}
931 	}
932 
933 	/*
934 	 * In transmit state, time the transmission and arrange for
935 	 * the retransmit.  In persist state, just set snd_max.
936 	 */
937 	if (!(tp->t_flags & TF_FORCE) ||
938 	    !tcp_callout_active(tp, tp->tt_persist)) {
939 		tcp_seq startseq = tp->snd_nxt;
940 
941 		/*
942 		 * Advance snd_nxt over sequence space of this segment.
943 		 */
944 		if (flags & (TH_SYN | TH_FIN)) {
945 			if (flags & TH_SYN)
946 				tp->snd_nxt++;
947 			if (flags & TH_FIN) {
948 				tp->snd_nxt++;
949 				tp->t_flags |= TF_SENTFIN;
950 			}
951 		}
952 		tp->snd_nxt += len;
953 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
954 			tp->snd_max = tp->snd_nxt;
955 			/*
956 			 * Time this transmission if not a retransmission and
957 			 * not currently timing anything.
958 			 */
959 			if (tp->t_rtttime == 0) {
960 				tp->t_rtttime = ticks;
961 				tp->t_rtseq = startseq;
962 				tcpstat.tcps_segstimed++;
963 			}
964 		}
965 
966 		/*
967 		 * Set retransmit timer if not currently set,
968 		 * and not doing a pure ack or a keep-alive probe.
969 		 * Initial value for retransmit timer is smoothed
970 		 * round-trip time + 2 * round-trip time variance.
971 		 * Initialize shift counter which is used for backoff
972 		 * of retransmit time.
973 		 */
974 		if (!tcp_callout_active(tp, tp->tt_rexmt) &&
975 		    tp->snd_nxt != tp->snd_una) {
976 			if (tcp_callout_active(tp, tp->tt_persist)) {
977 				tcp_callout_stop(tp, tp->tt_persist);
978 				tp->t_rxtshift = 0;
979 			}
980 			tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
981 			    tcp_timer_rexmt);
982 		}
983 	} else {
984 		/*
985 		 * Persist case, update snd_max but since we are in
986 		 * persist mode (no window) we do not update snd_nxt.
987 		 */
988 		int xlen = len;
989 		if (flags & TH_SYN)
990 			panic("tcp_output: persist timer to send SYN\n");
991 		if (flags & TH_FIN) {
992 			++xlen;
993 			tp->t_flags |= TF_SENTFIN;
994 		}
995 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
996 			tp->snd_max = tp->snd_nxt + xlen;
997 	}
998 
999 	if (th != NULL) {
1000 #ifdef TCPDEBUG
1001 		/* Trace. */
1002 		if (so->so_options & SO_DEBUG) {
1003 			tcp_trace(TA_OUTPUT, tp->t_state, tp,
1004 			    mtod(m, void *), th, 0);
1005 		}
1006 #endif
1007 
1008 		/*
1009 		 * Fill in IP length and desired time to live and
1010 		 * send to IP level.  There should be a better way
1011 		 * to handle ttl and tos; we could keep them in
1012 		 * the template, but need a way to checksum without them.
1013 		 */
1014 		/*
1015 		 * m->m_pkthdr.len should have been set before cksum
1016 		 * calcuration, because in6_cksum() need it.
1017 		 */
1018 		if (isipv6) {
1019 			/*
1020 			 * we separately set hoplimit for every segment,
1021 			 * since the user might want to change the value
1022 			 * via setsockopt.  Also, desired default hop
1023 			 * limit might be changed via Neighbor Discovery.
1024 			 */
1025 			ip6->ip6_hlim = in6_selecthlim(inp,
1026 			    (inp->in6p_route.ro_rt ?
1027 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
1028 
1029 			/* TODO: IPv6 IP6TOS_ECT bit on */
1030 			error = ip6_output(m, inp->in6p_outputopts,
1031 			    &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1032 			    NULL, NULL, inp);
1033 		} else {
1034 			struct rtentry *rt;
1035 			ip->ip_len = m->m_pkthdr.len;
1036 #ifdef INET6
1037 			if (INP_CHECK_SOCKAF(so, AF_INET6))
1038 				ip->ip_ttl = in6_selecthlim(inp,
1039 				    (inp->in6p_route.ro_rt ?
1040 				     inp->in6p_route.ro_rt->rt_ifp : NULL));
1041 			else
1042 #endif
1043 				ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
1044 
1045 			ip->ip_tos = inp->inp_ip_tos;	/* XXX */
1046 			/*
1047 			 * See if we should do MTU discovery.
1048 			 * We do it only if the following are true:
1049 			 *	1) we have a valid route to the destination
1050 			 *	2) the MTU is not locked (if it is,
1051 			 *	   then discovery has been disabled)
1052 			 */
1053 			if (path_mtu_discovery &&
1054 			    (rt = inp->inp_route.ro_rt) &&
1055 			    (rt->rt_flags & RTF_UP) &&
1056 			    !(rt->rt_rmx.rmx_locks & RTV_MTU))
1057 				ip->ip_off |= IP_DF;
1058 
1059 			error = ip_output(m, inp->inp_options, &inp->inp_route,
1060 					  (so->so_options & SO_DONTROUTE) |
1061 					  IP_DEBUGROUTE, NULL, inp);
1062 		}
1063 	} else {
1064 		KASSERT(error != 0, ("no error, but th not set\n"));
1065 	}
1066 	if (error) {
1067 
1068 		/*
1069 		 * We know that the packet was lost, so back out the
1070 		 * sequence number advance, if any.
1071 		 */
1072 		if (!(tp->t_flags & TF_FORCE) ||
1073 		    !tcp_callout_active(tp, tp->tt_persist)) {
1074 			/*
1075 			 * No need to check for TH_FIN here because
1076 			 * the TF_SENTFIN flag handles that case.
1077 			 */
1078 			if (!(flags & TH_SYN))
1079 				tp->snd_nxt -= len;
1080 		}
1081 
1082 out:
1083 		if (error == ENOBUFS) {
1084 			/*
1085 			 * If we can't send, make sure there is something
1086 			 * to get us going again later.
1087 			 *
1088 			 * The persist timer isn't necessarily allowed in all
1089 			 * states, use the rexmt timer.
1090 			 */
1091 			if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1092 			    !tcp_callout_active(tp, tp->tt_persist)) {
1093 				tcp_callout_reset(tp, tp->tt_rexmt,
1094 						  tp->t_rxtcur,
1095 						  tcp_timer_rexmt);
1096 #if 0
1097 				tp->t_rxtshift = 0;
1098 				tcp_setpersist(tp);
1099 #endif
1100 			}
1101 			tcp_quench(inp, 0);
1102 			return (0);
1103 		}
1104 		if (error == EMSGSIZE) {
1105 			/*
1106 			 * ip_output() will have already fixed the route
1107 			 * for us.  tcp_mtudisc() will, as its last action,
1108 			 * initiate retransmission, so it is important to
1109 			 * not do so here.
1110 			 */
1111 			tcp_mtudisc(inp, 0);
1112 			return 0;
1113 		}
1114 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1115 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1116 			tp->t_softerror = error;
1117 			return (0);
1118 		}
1119 		return (error);
1120 	}
1121 	tcpstat.tcps_sndtotal++;
1122 
1123 	/*
1124 	 * Data sent (as far as we can tell).
1125 	 *
1126 	 * If this advertises a larger window than any other segment,
1127 	 * then remember the size of the advertised window.
1128 	 *
1129 	 * Any pending ACK has now been sent.
1130 	 */
1131 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1132 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1133 		tp->t_flags &= ~TF_RXRESIZED;
1134 	}
1135 	tp->last_ack_sent = tp->rcv_nxt;
1136 	tp->t_flags &= ~TF_ACKNOW;
1137 	if (tcp_delack_enabled)
1138 		tcp_callout_stop(tp, tp->tt_delack);
1139 	if (sendalot)
1140 		goto again;
1141 	return (0);
1142 }
1143 
1144 void
1145 tcp_setpersist(struct tcpcb *tp)
1146 {
1147 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1148 	int tt;
1149 
1150 	if (tp->t_state == TCPS_SYN_SENT ||
1151 	    tp->t_state == TCPS_SYN_RECEIVED) {
1152 		panic("tcp_setpersist: not established yet, current %s\n",
1153 		      tp->t_state == TCPS_SYN_SENT ?
1154 		      "SYN_SENT" : "SYN_RECEIVED");
1155 	}
1156 
1157 	if (tcp_callout_active(tp, tp->tt_rexmt))
1158 		panic("tcp_setpersist: retransmit pending");
1159 	/*
1160 	 * Start/restart persistance timer.
1161 	 */
1162 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1163 		      TCPTV_PERSMAX);
1164 	tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1165 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1166 		tp->t_rxtshift++;
1167 }
1168 
1169 static void
1170 tcp_idle_cwnd_validate(struct tcpcb *tp)
1171 {
1172 	u_long initial_cwnd = tcp_initial_window(tp);
1173 	u_long min_cwnd;
1174 
1175 	tcpstat.tcps_sndidle++;
1176 
1177 	/* According to RFC5681: RW=min(IW,cwnd) */
1178 	min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1179 
1180 	if (tcp_idle_cwv) {
1181 		u_long idle_time, decay_cwnd;
1182 
1183 		/*
1184 		 * RFC2861, but only after idle period.
1185 		 */
1186 
1187 		/*
1188 		 * Before the congestion window is reduced, ssthresh
1189 		 * is set to the maximum of its current value and 3/4
1190 		 * cwnd.  If the sender then has more data to send
1191 		 * than the decayed cwnd allows, the TCP will slow-
1192 		 * start (perform exponential increase) at least
1193 		 * half-way back up to the old value of cwnd.
1194 		 */
1195 		tp->snd_ssthresh = max(tp->snd_ssthresh,
1196 		    (3 * tp->snd_cwnd) / 4);
1197 
1198 		/*
1199 		 * Decay the congestion window by half for every RTT
1200 		 * that the flow remains inactive.
1201 		 *
1202 		 * The difference between our implementation and
1203 		 * RFC2861 is that we don't allow cwnd to go below
1204 		 * the value allowed by RFC5681 (min_cwnd).
1205 		 */
1206 		idle_time = ticks - tp->snd_last;
1207 		decay_cwnd = tp->snd_cwnd;
1208 		while (idle_time >= tp->t_rxtcur &&
1209 		    decay_cwnd > min_cwnd) {
1210 			decay_cwnd >>= 1;
1211 			idle_time -= tp->t_rxtcur;
1212 		}
1213 		tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1214 	} else {
1215 		/*
1216 		 * Slow-start from scratch to re-determine the send
1217 		 * congestion window.
1218 		 */
1219 		tp->snd_cwnd = min_cwnd;
1220 	}
1221 
1222 	/* Restart ABC counting during congestion avoidance */
1223 	tp->snd_wacked = 0;
1224 }
1225