xref: /dflybsd-src/sys/netinet/tcp_output.c (revision c7c26905e41d70fe0ea9d02c5af1fc0a688408d4)
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  */
69 
70 #include "opt_inet.h"
71 #include "opt_inet6.h"
72 #include "opt_ipsec.h"
73 #include "opt_tcpdebug.h"
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/sysctl.h>
79 #include <sys/mbuf.h>
80 #include <sys/domain.h>
81 #include <sys/protosw.h>
82 #include <sys/socket.h>
83 #include <sys/socketvar.h>
84 #include <sys/in_cksum.h>
85 #include <sys/thread.h>
86 #include <sys/globaldata.h>
87 
88 #include <net/if_var.h>
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 int tcp_do_tso = 1;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
155     &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
156 
157 static void	tcp_idle_cwnd_validate(struct tcpcb *);
158 
159 static int	tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
160 
161 /*
162  * Tcp output routine: figure out what should be sent and send it.
163  */
164 int
165 tcp_output(struct tcpcb *tp)
166 {
167 	struct inpcb * const inp = tp->t_inpcb;
168 	struct socket *so = inp->inp_socket;
169 	long len, recvwin, sendwin;
170 	int nsacked = 0;
171 	int off, flags, error = 0;
172 #ifdef TCP_SIGNATURE
173 	int sigoff = 0;
174 #endif
175 	struct mbuf *m;
176 	struct ip *ip;
177 	struct ipovly *ipov;
178 	struct tcphdr *th;
179 	u_char opt[TCP_MAXOLEN];
180 	unsigned int ipoptlen, optlen, hdrlen;
181 	int idle;
182 	boolean_t sendalot;
183 	struct ip6_hdr *ip6;
184 #ifdef INET6
185 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
186 #else
187 	const boolean_t isipv6 = FALSE;
188 #endif
189 	boolean_t can_tso = FALSE, use_tso;
190 	boolean_t report_sack, idle_cwv = FALSE;
191 	u_int segsz, tso_hlen;
192 
193 	KKASSERT(so->so_port == &curthread->td_msgport);
194 
195 	/*
196 	 * Determine length of data that should be transmitted,
197 	 * and flags that will be used.
198 	 * If there is some data or critical controls (SYN, RST)
199 	 * to send, then transmit; otherwise, investigate further.
200 	 */
201 
202 	/*
203 	 * If we have been idle for a while, the send congestion window
204 	 * could be no longer representative of the current state of the
205 	 * link; need to validate congestion window.  However, we should
206 	 * not perform congestion window validation here, since we could
207 	 * be asked to send pure ACK.
208 	 */
209 	if (tp->snd_max == tp->snd_una &&
210 	    (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
211 		idle_cwv = TRUE;
212 
213 	/*
214 	 * Calculate whether the transmit stream was previously idle
215 	 * and adjust TF_LASTIDLE for the next time.
216 	 */
217 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
218 	if (idle && (tp->t_flags & TF_MORETOCOME))
219 		tp->t_flags |= TF_LASTIDLE;
220 	else
221 		tp->t_flags &= ~TF_LASTIDLE;
222 
223 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
224 	    !IN_FASTRECOVERY(tp))
225 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
226 
227 	/*
228 	 * Find out whether TSO could be used or not
229 	 *
230 	 * For TSO capable devices, the following assumptions apply to
231 	 * the processing of TCP flags:
232 	 * - If FIN is set on the large TCP segment, the device must set
233 	 *   FIN on the last segment that it creates from the large TCP
234 	 *   segment.
235 	 * - If PUSH is set on the large TCP segment, the device must set
236 	 *   PUSH on the last segment that it creates from the large TCP
237 	 *   segment.
238 	 */
239 #if !defined(IPSEC) && !defined(FAST_IPSEC)
240 	if (tcp_do_tso
241 #ifdef TCP_SIGNATURE
242 	    && (tp->t_flags & TF_SIGNATURE) == 0
243 #endif
244 	) {
245 		if (!isipv6) {
246 			struct rtentry *rt = inp->inp_route.ro_rt;
247 
248 			if (rt != NULL && (rt->rt_flags & RTF_UP) &&
249 			    (rt->rt_ifp->if_hwassist & CSUM_TSO))
250 				can_tso = TRUE;
251 		}
252 	}
253 #endif	/* !IPSEC && !FAST_IPSEC */
254 
255 again:
256 	m = NULL;
257 	ip = NULL;
258 	ipov = NULL;
259 	th = NULL;
260 	ip6 = NULL;
261 
262 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
263 		TF_SACK_PERMITTED &&
264 	    (!TAILQ_EMPTY(&tp->t_segq) ||
265 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
266 		report_sack = TRUE;
267 	else
268 		report_sack = FALSE;
269 
270 	/* Make use of SACK information when slow-starting after a RTO. */
271 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
272 	    !IN_FASTRECOVERY(tp)) {
273 		tcp_seq old_snd_nxt = tp->snd_nxt;
274 
275 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
276 		nsacked += tp->snd_nxt - old_snd_nxt;
277 	}
278 
279 	sendalot = FALSE;
280 	off = tp->snd_nxt - tp->snd_una;
281 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
282 	sendwin = min(sendwin, tp->snd_bwnd);
283 
284 	flags = tcp_outflags[tp->t_state];
285 	/*
286 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
287 	 * state flags.
288 	 */
289 	if (tp->t_flags & TF_NEEDFIN)
290 		flags |= TH_FIN;
291 	if (tp->t_flags & TF_NEEDSYN)
292 		flags |= TH_SYN;
293 
294 	/*
295 	 * If in persist timeout with window of 0, send 1 byte.
296 	 * Otherwise, if window is small but nonzero
297 	 * and timer expired, we will send what we can
298 	 * and go to transmit state.
299 	 */
300 	if (tp->t_flags & TF_FORCE) {
301 		if (sendwin == 0) {
302 			/*
303 			 * If we still have some data to send, then
304 			 * clear the FIN bit.  Usually this would
305 			 * happen below when it realizes that we
306 			 * aren't sending all the data.  However,
307 			 * if we have exactly 1 byte of unsent data,
308 			 * then it won't clear the FIN bit below,
309 			 * and if we are in persist state, we wind
310 			 * up sending the packet without recording
311 			 * that we sent the FIN bit.
312 			 *
313 			 * We can't just blindly clear the FIN bit,
314 			 * because if we don't have any more data
315 			 * to send then the probe will be the FIN
316 			 * itself.
317 			 */
318 			if (off < so->so_snd.ssb_cc)
319 				flags &= ~TH_FIN;
320 			sendwin = 1;
321 		} else {
322 			tcp_callout_stop(tp, tp->tt_persist);
323 			tp->t_rxtshift = 0;
324 		}
325 	}
326 
327 	/*
328 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
329 	 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
330 	 * a negative length.  This can also occur when TCP opens up
331 	 * its congestion window while receiving additional duplicate
332 	 * acks after fast-retransmit because TCP will reset snd_nxt
333 	 * to snd_max after the fast-retransmit.
334 	 *
335 	 * A negative length can also occur when we are in the
336 	 * TCPS_SYN_RECEIVED state due to a simultanious connect where
337 	 * our SYN has not been acked yet.
338 	 *
339 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
340 	 * be set to snd_una, the offset will be 0, and the length may
341 	 * wind up 0.
342 	 */
343 	len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
344 
345 	/*
346 	 * Lop off SYN bit if it has already been sent.  However, if this
347 	 * is SYN-SENT state and if segment contains data, suppress sending
348 	 * segment (sending the segment would be an option if we still
349 	 * did TAO and the remote host supported it).
350 	 */
351 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
352 		flags &= ~TH_SYN;
353 		off--, len++;
354 		if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
355 			tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
356 			return 0;
357 		}
358 	}
359 
360 	/*
361 	 * Be careful not to send data and/or FIN on SYN segments.
362 	 * This measure is needed to prevent interoperability problems
363 	 * with not fully conformant TCP implementations.
364 	 */
365 	if (flags & TH_SYN) {
366 		len = 0;
367 		flags &= ~TH_FIN;
368 	}
369 
370 	if (len < 0) {
371 		/*
372 		 * A negative len can occur if our FIN has been sent but not
373 		 * acked, or if we are in a simultanious connect in the
374 		 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
375 		 * acked.
376 		 *
377 		 * If our window has contracted to 0 in the FIN case
378 		 * (which can only occur if we have NOT been called to
379 		 * retransmit as per code a few paragraphs up) then we
380 		 * want to shift the retransmit timer over to the
381 		 * persist timer.
382 		 *
383 		 * However, if we are in the TCPS_SYN_RECEIVED state
384 		 * (the SYN case) we will be in a simultanious connect and
385 		 * the window may be zero degeneratively.  In this case we
386 		 * do not want to shift to the persist timer after the SYN
387 		 * or the SYN+ACK transmission.
388 		 */
389 		len = 0;
390 		if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
391 			tcp_callout_stop(tp, tp->tt_rexmt);
392 			tp->t_rxtshift = 0;
393 			tp->snd_nxt = tp->snd_una;
394 			if (!tcp_callout_active(tp, tp->tt_persist))
395 				tcp_setpersist(tp);
396 		}
397 	}
398 
399 	KASSERT(len >= 0, ("%s: len < 0", __func__));
400 	/*
401 	 * Automatic sizing of send socket buffer.  Often the send buffer
402 	 * size is not optimally adjusted to the actual network conditions
403 	 * at hand (delay bandwidth product).  Setting the buffer size too
404 	 * small limits throughput on links with high bandwidth and high
405 	 * delay (eg. trans-continental/oceanic links).  Setting the
406 	 * buffer size too big consumes too much real kernel memory,
407 	 * especially with many connections on busy servers.
408 	 *
409 	 * The criteria to step up the send buffer one notch are:
410 	 *  1. receive window of remote host is larger than send buffer
411 	 *     (with a fudge factor of 5/4th);
412 	 *  2. send buffer is filled to 7/8th with data (so we actually
413 	 *     have data to make use of it);
414 	 *  3. send buffer fill has not hit maximal automatic size;
415 	 *  4. our send window (slow start and cogestion controlled) is
416 	 *     larger than sent but unacknowledged data in send buffer.
417 	 *
418 	 * The remote host receive window scaling factor may limit the
419 	 * growing of the send buffer before it reaches its allowed
420 	 * maximum.
421 	 *
422 	 * It scales directly with slow start or congestion window
423 	 * and does at most one step per received ACK.  This fast
424 	 * scaling has the drawback of growing the send buffer beyond
425 	 * what is strictly necessary to make full use of a given
426 	 * delay*bandwith product.  However testing has shown this not
427 	 * to be much of an problem.  At worst we are trading wasting
428 	 * of available bandwith (the non-use of it) for wasting some
429 	 * socket buffer memory.
430 	 *
431 	 * TODO: Shrink send buffer during idle periods together
432 	 * with congestion window.  Requires another timer.  Has to
433 	 * wait for upcoming tcp timer rewrite.
434 	 */
435 	if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
436 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
437 		    so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
438 		    so->so_snd.ssb_cc < tcp_autosndbuf_max &&
439 		    sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
440 			u_long newsize;
441 
442 			newsize = ulmin(so->so_snd.ssb_hiwat +
443 					 tcp_autosndbuf_inc,
444 					tcp_autosndbuf_max);
445 			if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
446 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
447 			if (newsize >= (TCP_MAXWIN << tp->snd_scale))
448 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
449 		}
450 	}
451 
452 	/*
453 	 * Don't use TSO, if:
454 	 * - Congestion window needs validation
455 	 * - There are SACK blocks to report
456 	 * - RST or SYN flags is set
457 	 * - URG will be set
458 	 *
459 	 * XXX
460 	 * Checking for SYN|RST looks overkill, just to be safe than sorry
461 	 */
462 	use_tso = can_tso;
463 	if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
464 		use_tso = FALSE;
465 	if (use_tso) {
466 		tcp_seq ugr_nxt = tp->snd_nxt;
467 
468 		if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
469 		    tp->snd_nxt == tp->snd_max)
470 			--ugr_nxt;
471 
472 		if (SEQ_GT(tp->snd_up, ugr_nxt))
473 			use_tso = FALSE;
474 	}
475 
476 	if (use_tso) {
477 		/*
478 		 * Find out segment size and header length for TSO
479 		 */
480 		error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
481 		if (error)
482 			use_tso = FALSE;
483 	}
484 	if (!use_tso) {
485 		segsz = tp->t_maxseg;
486 		tso_hlen = 0; /* not used */
487 	}
488 
489 	/*
490 	 * Truncate to the maximum segment length if not TSO, and ensure that
491 	 * FIN is removed if the length no longer contains the last data byte.
492 	 */
493 	if (len > segsz) {
494 		if (!use_tso) {
495 			len = segsz;
496 		} else {
497 			/*
498 			 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
499 			 * thoff), and make sure that we send equal size
500 			 * transfers down the stack (rather than big-small-
501 			 * big-small-...).
502 			 */
503 			len = (min(len, (IP_MAXPACKET - tso_hlen)) / segsz) *
504 			    segsz;
505 			if (len <= segsz)
506 				use_tso = FALSE;
507 		}
508 		sendalot = TRUE;
509 	} else {
510 		use_tso = FALSE;
511 	}
512 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
513 		flags &= ~TH_FIN;
514 
515 	recvwin = ssb_space(&so->so_rcv);
516 
517 	/*
518 	 * Sender silly window avoidance.   We transmit under the following
519 	 * conditions when len is non-zero:
520 	 *
521 	 *	- We have a full segment
522 	 *	- This is the last buffer in a write()/send() and we are
523 	 *	  either idle or running NODELAY
524 	 *	- we've timed out (e.g. persist timer)
525 	 *	- we have more then 1/2 the maximum send window's worth of
526 	 *	  data (receiver may be limiting the window size)
527 	 *	- we need to retransmit
528 	 */
529 	if (len) {
530 		if (len >= segsz)
531 			goto send;
532 		/*
533 		 * NOTE! on localhost connections an 'ack' from the remote
534 		 * end may occur synchronously with the output and cause
535 		 * us to flush a buffer queued with moretocome.  XXX
536 		 *
537 		 * note: the len + off check is almost certainly unnecessary.
538 		 */
539 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
540 		    (idle || (tp->t_flags & TF_NODELAY)) &&
541 		    len + off >= so->so_snd.ssb_cc &&
542 		    !(tp->t_flags & TF_NOPUSH)) {
543 			goto send;
544 		}
545 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
546 			goto send;
547 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
548 			goto send;
549 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
550 			goto send;
551 		if (tp->t_flags & TF_XMITNOW)
552 			goto send;
553 	}
554 
555 	/*
556 	 * Compare available window to amount of window
557 	 * known to peer (as advertised window less
558 	 * next expected input).  If the difference is at least two
559 	 * max size segments, or at least 50% of the maximum possible
560 	 * window, then want to send a window update to peer.
561 	 */
562 	if (recvwin > 0) {
563 		/*
564 		 * "adv" is the amount we can increase the window,
565 		 * taking into account that we are limited by
566 		 * TCP_MAXWIN << tp->rcv_scale.
567 		 */
568 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
569 			(tp->rcv_adv - tp->rcv_nxt);
570 		long hiwat;
571 
572 		/*
573 		 * This ack case typically occurs when the user has drained
574 		 * the TCP socket buffer sufficiently to warrent an ack
575 		 * containing a 'pure window update'... that is, an ack that
576 		 * ONLY updates the tcp window.
577 		 *
578 		 * It is unclear why we would need to do a pure window update
579 		 * past 2 segments if we are going to do one at 1/2 the high
580 		 * water mark anyway, especially since under normal conditions
581 		 * the user program will drain the socket buffer quickly.
582 		 * The 2-segment pure window update will often add a large
583 		 * number of extra, unnecessary acks to the stream.
584 		 *
585 		 * avoid_pure_win_update now defaults to 1.
586 		 */
587 		if (avoid_pure_win_update == 0 ||
588 		    (tp->t_flags & TF_RXRESIZED)) {
589 			if (adv >= (long) (2 * segsz)) {
590 				goto send;
591 			}
592 		}
593 		hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
594 		if (hiwat > (long)so->so_rcv.ssb_hiwat)
595 			hiwat = (long)so->so_rcv.ssb_hiwat;
596 		if (adv >= hiwat / 2)
597 			goto send;
598 	}
599 
600 	/*
601 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
602 	 * is also a catch-all for the retransmit timer timeout case.
603 	 */
604 	if (tp->t_flags & TF_ACKNOW)
605 		goto send;
606 	if ((flags & TH_RST) ||
607 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
608 		goto send;
609 	if (SEQ_GT(tp->snd_up, tp->snd_una))
610 		goto send;
611 	/*
612 	 * If our state indicates that FIN should be sent
613 	 * and we have not yet done so, then we need to send.
614 	 */
615 	if ((flags & TH_FIN) &&
616 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
617 		goto send;
618 
619 	/*
620 	 * TCP window updates are not reliable, rather a polling protocol
621 	 * using ``persist'' packets is used to insure receipt of window
622 	 * updates.  The three ``states'' for the output side are:
623 	 *	idle			not doing retransmits or persists
624 	 *	persisting		to move a small or zero window
625 	 *	(re)transmitting	and thereby not persisting
626 	 *
627 	 * tcp_callout_active(tp, tp->tt_persist)
628 	 *	is true when we are in persist state.
629 	 * The TF_FORCE flag in tp->t_flags
630 	 *	is set when we are called to send a persist packet.
631 	 * tcp_callout_active(tp, tp->tt_rexmt)
632 	 *	is set when we are retransmitting
633 	 * The output side is idle when both timers are zero.
634 	 *
635 	 * If send window is too small, there is data to transmit, and no
636 	 * retransmit or persist is pending, then go to persist state.
637 	 *
638 	 * If nothing happens soon, send when timer expires:
639 	 * if window is nonzero, transmit what we can, otherwise force out
640 	 * a byte.
641 	 *
642 	 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
643 	 * with data pending.  This situation can occur during a
644 	 * simultanious connect.
645 	 */
646 	if (so->so_snd.ssb_cc > 0 &&
647 	    tp->t_state != TCPS_SYN_RECEIVED &&
648 	    !tcp_callout_active(tp, tp->tt_rexmt) &&
649 	    !tcp_callout_active(tp, tp->tt_persist)) {
650 		tp->t_rxtshift = 0;
651 		tcp_setpersist(tp);
652 	}
653 
654 	/*
655 	 * No reason to send a segment, just return.
656 	 */
657 	tp->t_flags &= ~TF_XMITNOW;
658 	return (0);
659 
660 send:
661 	/*
662 	 * Before ESTABLISHED, force sending of initial options
663 	 * unless TCP set not to do any options.
664 	 * NOTE: we assume that the IP/TCP header plus TCP options
665 	 * always fit in a single mbuf, leaving room for a maximum
666 	 * link header, i.e.
667 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
668 	 */
669 	optlen = 0;
670 	if (isipv6)
671 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
672 	else
673 		hdrlen = sizeof(struct tcpiphdr);
674 	if (flags & TH_SYN) {
675 		tp->snd_nxt = tp->iss;
676 		if (!(tp->t_flags & TF_NOOPT)) {
677 			u_short mss;
678 
679 			opt[0] = TCPOPT_MAXSEG;
680 			opt[1] = TCPOLEN_MAXSEG;
681 			mss = htons((u_short) tcp_mssopt(tp));
682 			memcpy(opt + 2, &mss, sizeof mss);
683 			optlen = TCPOLEN_MAXSEG;
684 
685 			if ((tp->t_flags & TF_REQ_SCALE) &&
686 			    (!(flags & TH_ACK) ||
687 			     (tp->t_flags & TF_RCVD_SCALE))) {
688 				*((u_int32_t *)(opt + optlen)) = htonl(
689 					TCPOPT_NOP << 24 |
690 					TCPOPT_WINDOW << 16 |
691 					TCPOLEN_WINDOW << 8 |
692 					tp->request_r_scale);
693 				optlen += 4;
694 			}
695 
696 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
697 			    tp->t_flags & TF_SACK_PERMITTED) {
698 				uint32_t *lp = (uint32_t *)(opt + optlen);
699 
700 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
701 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
702 			}
703 		}
704 	}
705 
706 	/*
707 	 * Send a timestamp and echo-reply if this is a SYN and our side
708 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
709 	 * and our peer have sent timestamps in our SYN's.
710 	 */
711 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
712 	    !(flags & TH_RST) &&
713 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
714 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
715 
716 		/* Form timestamp option as shown in appendix A of RFC 1323. */
717 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
718 		*lp++ = htonl(ticks);
719 		*lp   = htonl(tp->ts_recent);
720 		optlen += TCPOLEN_TSTAMP_APPA;
721 	}
722 
723 	/* Set receive buffer autosizing timestamp. */
724 	if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
725 		tp->rfbuf_ts = ticks;
726 
727 	/*
728 	 * If this is a SACK connection and we have a block to report,
729 	 * fill in the SACK blocks in the TCP options.
730 	 */
731 	if (report_sack)
732 		tcp_sack_fill_report(tp, opt, &optlen);
733 
734 #ifdef TCP_SIGNATURE
735 	if (tp->t_flags & TF_SIGNATURE) {
736 		int i;
737 		u_char *bp;
738 		/*
739 		 * Initialize TCP-MD5 option (RFC2385)
740 		 */
741 		bp = (u_char *)opt + optlen;
742 		*bp++ = TCPOPT_SIGNATURE;
743 		*bp++ = TCPOLEN_SIGNATURE;
744 		sigoff = optlen + 2;
745 		for (i = 0; i < TCP_SIGLEN; i++)
746 			*bp++ = 0;
747 		optlen += TCPOLEN_SIGNATURE;
748 		/*
749 		 * Terminate options list and maintain 32-bit alignment.
750 		 */
751 		*bp++ = TCPOPT_NOP;
752 		*bp++ = TCPOPT_EOL;
753 		optlen += 2;
754 	}
755 #endif /* TCP_SIGNATURE */
756 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
757 	hdrlen += optlen;
758 
759 	if (isipv6) {
760 		ipoptlen = ip6_optlen(inp);
761 	} else {
762 		if (inp->inp_options) {
763 			ipoptlen = inp->inp_options->m_len -
764 			    offsetof(struct ipoption, ipopt_list);
765 		} else {
766 			ipoptlen = 0;
767 		}
768 	}
769 #ifdef IPSEC
770 	ipoptlen += ipsec_hdrsiz_tcp(tp);
771 #endif
772 
773 	if (use_tso) {
774 		KASSERT(len > segsz,
775 		    ("invalid TSO len %ld, segsz %u", len, segsz));
776 	} else {
777 		KASSERT(len <= segsz,
778 		    ("invalid len %ld, segsz %u", len, segsz));
779 
780 		/*
781 		 * Adjust data length if insertion of options will bump
782 		 * the packet length beyond the t_maxopd length.  Clear
783 		 * FIN to prevent premature closure since there is still
784 		 * more data to send after this (now truncated) packet.
785 		 *
786 		 * If just the options do not fit we are in a no-win
787 		 * situation and we treat it as an unreachable host.
788 		 */
789 		if (len + optlen + ipoptlen > tp->t_maxopd) {
790 			if (tp->t_maxopd <= optlen + ipoptlen) {
791 				static time_t last_optlen_report;
792 
793 				if (last_optlen_report != time_second) {
794 					last_optlen_report = time_second;
795 					kprintf("tcpcb %p: MSS (%d) too "
796 					    "small to hold options!\n",
797 					    tp, tp->t_maxopd);
798 				}
799 				error = EHOSTUNREACH;
800 				goto out;
801 			} else {
802 				flags &= ~TH_FIN;
803 				len = tp->t_maxopd - optlen - ipoptlen;
804 				sendalot = TRUE;
805 			}
806 		}
807 	}
808 
809 #ifdef INET6
810 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
811 #else
812 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
813 #endif
814 
815 	/*
816 	 * Grab a header mbuf, attaching a copy of data to
817 	 * be transmitted, and initialize the header from
818 	 * the template for sends on this connection.
819 	 */
820 	if (len) {
821 		if ((tp->t_flags & TF_FORCE) && len == 1)
822 			tcpstat.tcps_sndprobe++;
823 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
824 			if (tp->snd_nxt == tp->snd_una)
825 				tp->snd_max_rexmt = tp->snd_max;
826 			if (nsacked) {
827 				tcpstat.tcps_sndsackrtopack++;
828 				tcpstat.tcps_sndsackrtobyte += len;
829 			}
830 			tcpstat.tcps_sndrexmitpack++;
831 			tcpstat.tcps_sndrexmitbyte += len;
832 		} else {
833 			tcpstat.tcps_sndpack++;
834 			tcpstat.tcps_sndbyte += len;
835 		}
836 		if (idle_cwv) {
837 			idle_cwv = FALSE;
838 			tcp_idle_cwnd_validate(tp);
839 		}
840 		/* Update last send time after CWV */
841 		tp->snd_last = ticks;
842 #ifdef notyet
843 		if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
844 		    max_linkhdr + hdrlen)) == NULL) {
845 			error = ENOBUFS;
846 			goto after_th;
847 		}
848 		/*
849 		 * m_copypack left space for our hdr; use it.
850 		 */
851 		m->m_len += hdrlen;
852 		m->m_data -= hdrlen;
853 #else
854 #ifndef INET6
855 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
856 #else
857 		m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
858 			   M_PKTHDR, NULL);
859 #endif
860 		if (m == NULL) {
861 			error = ENOBUFS;
862 			goto after_th;
863 		}
864 		m->m_data += max_linkhdr;
865 		m->m_len = hdrlen;
866 		if (len <= MHLEN - hdrlen - max_linkhdr) {
867 			m_copydata(so->so_snd.ssb_mb, off, (int) len,
868 			    mtod(m, caddr_t) + hdrlen);
869 			m->m_len += len;
870 		} else {
871 			m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
872 			if (m->m_next == NULL) {
873 				m_free(m);
874 				m = NULL;
875 				error = ENOBUFS;
876 				goto after_th;
877 			}
878 		}
879 #endif
880 		/*
881 		 * If we're sending everything we've got, set PUSH.
882 		 * (This will keep happy those implementations which only
883 		 * give data to the user when a buffer fills or
884 		 * a PUSH comes in.)
885 		 */
886 		if (off + len == so->so_snd.ssb_cc)
887 			flags |= TH_PUSH;
888 	} else {
889 		if (tp->t_flags & TF_ACKNOW)
890 			tcpstat.tcps_sndacks++;
891 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
892 			tcpstat.tcps_sndctrl++;
893 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
894 			tcpstat.tcps_sndurg++;
895 		else
896 			tcpstat.tcps_sndwinup++;
897 
898 		MGETHDR(m, MB_DONTWAIT, MT_HEADER);
899 		if (m == NULL) {
900 			error = ENOBUFS;
901 			goto after_th;
902 		}
903 		if (isipv6 &&
904 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
905 			MH_ALIGN(m, hdrlen);
906 		else
907 			m->m_data += max_linkhdr;
908 		m->m_len = hdrlen;
909 	}
910 	m->m_pkthdr.rcvif = NULL;
911 	if (isipv6) {
912 		ip6 = mtod(m, struct ip6_hdr *);
913 		th = (struct tcphdr *)(ip6 + 1);
914 		tcp_fillheaders(tp, ip6, th, use_tso);
915 	} else {
916 		ip = mtod(m, struct ip *);
917 		ipov = (struct ipovly *)ip;
918 		th = (struct tcphdr *)(ip + 1);
919 		/* this picks up the pseudo header (w/o the length) */
920 		tcp_fillheaders(tp, ip, th, use_tso);
921 	}
922 after_th:
923 	/*
924 	 * Fill in fields, remembering maximum advertised
925 	 * window for use in delaying messages about window sizes.
926 	 * If resending a FIN, be sure not to use a new sequence number.
927 	 */
928 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
929 	    tp->snd_nxt == tp->snd_max)
930 		tp->snd_nxt--;
931 
932 	if (th != NULL) {
933 		/*
934 		 * If we are doing retransmissions, then snd_nxt will
935 		 * not reflect the first unsent octet.  For ACK only
936 		 * packets, we do not want the sequence number of the
937 		 * retransmitted packet, we want the sequence number
938 		 * of the next unsent octet.  So, if there is no data
939 		 * (and no SYN or FIN), use snd_max instead of snd_nxt
940 		 * when filling in ti_seq.  But if we are in persist
941 		 * state, snd_max might reflect one byte beyond the
942 		 * right edge of the window, so use snd_nxt in that
943 		 * case, since we know we aren't doing a retransmission.
944 		 * (retransmit and persist are mutually exclusive...)
945 		 */
946 		if (len || (flags & (TH_SYN|TH_FIN)) ||
947 		    tcp_callout_active(tp, tp->tt_persist))
948 			th->th_seq = htonl(tp->snd_nxt);
949 		else
950 			th->th_seq = htonl(tp->snd_max);
951 		th->th_ack = htonl(tp->rcv_nxt);
952 		if (optlen) {
953 			bcopy(opt, th + 1, optlen);
954 			th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
955 		}
956 		th->th_flags = flags;
957 	}
958 
959 	/*
960 	 * Calculate receive window.  Don't shrink window, but avoid
961 	 * silly window syndrome by sending a 0 window if the actual
962 	 * window is less then one segment.
963 	 */
964 	if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
965 	    recvwin < (long)segsz)
966 		recvwin = 0;
967 	if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
968 		recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
969 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
970 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
971 
972 	/*
973 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
974 	 * a 0 window.  This may cause the remote transmitter to stall.  This
975 	 * flag tells soreceive() to disable delayed acknowledgements when
976 	 * draining the buffer.  This can occur if the receiver is attempting
977 	 * to read more data then can be buffered prior to transmitting on
978 	 * the connection.
979 	 */
980 	if (recvwin == 0)
981 		tp->t_flags |= TF_RXWIN0SENT;
982 	else
983 		tp->t_flags &= ~TF_RXWIN0SENT;
984 
985 	if (th != NULL)
986 		th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
987 
988 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
989 		KASSERT(!use_tso, ("URG with TSO"));
990 		if (th != NULL) {
991 			th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
992 			th->th_flags |= TH_URG;
993 		}
994 	} else {
995 		/*
996 		 * If no urgent pointer to send, then we pull
997 		 * the urgent pointer to the left edge of the send window
998 		 * so that it doesn't drift into the send window on sequence
999 		 * number wraparound.
1000 		 */
1001 		tp->snd_up = tp->snd_una;		/* drag it along */
1002 	}
1003 
1004 	if (th != NULL) {
1005 #ifdef TCP_SIGNATURE
1006 		if (tp->t_flags & TF_SIGNATURE) {
1007 			tcpsignature_compute(m, len, optlen,
1008 			    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1009 		}
1010 #endif /* TCP_SIGNATURE */
1011 
1012 		/*
1013 		 * Put TCP length in extended header, and then
1014 		 * checksum extended header and data.
1015 		 */
1016 		m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1017 		if (isipv6) {
1018 			/*
1019 			 * ip6_plen is not need to be filled now, and will be
1020 			 * filled in ip6_output().
1021 			 */
1022 			th->th_sum = in6_cksum(m, IPPROTO_TCP,
1023 			    sizeof(struct ip6_hdr),
1024 			    sizeof(struct tcphdr) + optlen + len);
1025 		} else {
1026 			m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1027 			if (use_tso) {
1028 				m->m_pkthdr.csum_flags = CSUM_TSO;
1029 				m->m_pkthdr.tso_segsz = segsz;
1030 			} else {
1031 				m->m_pkthdr.csum_flags = CSUM_TCP;
1032 				m->m_pkthdr.csum_data =
1033 				    offsetof(struct tcphdr, th_sum);
1034 				if (len + optlen) {
1035 					th->th_sum = in_addword(th->th_sum,
1036 					    htons((u_short)(optlen + len)));
1037 				}
1038 			}
1039 
1040 			/*
1041 			 * IP version must be set here for ipv4/ipv6 checking
1042 			 * later
1043 			 */
1044 			KASSERT(ip->ip_v == IPVERSION,
1045 			    ("%s: IP version incorrect: %d",
1046 			     __func__, ip->ip_v));
1047 		}
1048 	}
1049 
1050 	/*
1051 	 * In transmit state, time the transmission and arrange for
1052 	 * the retransmit.  In persist state, just set snd_max.
1053 	 */
1054 	if (!(tp->t_flags & TF_FORCE) ||
1055 	    !tcp_callout_active(tp, tp->tt_persist)) {
1056 		tcp_seq startseq = tp->snd_nxt;
1057 
1058 		/*
1059 		 * Advance snd_nxt over sequence space of this segment.
1060 		 */
1061 		if (flags & (TH_SYN | TH_FIN)) {
1062 			if (flags & TH_SYN)
1063 				tp->snd_nxt++;
1064 			if (flags & TH_FIN) {
1065 				tp->snd_nxt++;
1066 				tp->t_flags |= TF_SENTFIN;
1067 			}
1068 		}
1069 		tp->snd_nxt += len;
1070 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1071 			tp->snd_max = tp->snd_nxt;
1072 			/*
1073 			 * Time this transmission if not a retransmission and
1074 			 * not currently timing anything.
1075 			 */
1076 			if (tp->t_rtttime == 0) {
1077 				tp->t_rtttime = ticks;
1078 				tp->t_rtseq = startseq;
1079 				tcpstat.tcps_segstimed++;
1080 			}
1081 		}
1082 
1083 		/*
1084 		 * Set retransmit timer if not currently set,
1085 		 * and not doing a pure ack or a keep-alive probe.
1086 		 * Initial value for retransmit timer is smoothed
1087 		 * round-trip time + 2 * round-trip time variance.
1088 		 * Initialize shift counter which is used for backoff
1089 		 * of retransmit time.
1090 		 */
1091 		if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1092 		    tp->snd_nxt != tp->snd_una) {
1093 			if (tcp_callout_active(tp, tp->tt_persist)) {
1094 				tcp_callout_stop(tp, tp->tt_persist);
1095 				tp->t_rxtshift = 0;
1096 			}
1097 			tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1098 			    tcp_timer_rexmt);
1099 		}
1100 	} else {
1101 		/*
1102 		 * Persist case, update snd_max but since we are in
1103 		 * persist mode (no window) we do not update snd_nxt.
1104 		 */
1105 		int xlen = len;
1106 		if (flags & TH_SYN)
1107 			panic("tcp_output: persist timer to send SYN");
1108 		if (flags & TH_FIN) {
1109 			++xlen;
1110 			tp->t_flags |= TF_SENTFIN;
1111 		}
1112 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1113 			tp->snd_max = tp->snd_nxt + xlen;
1114 	}
1115 
1116 	if (th != NULL) {
1117 #ifdef TCPDEBUG
1118 		/* Trace. */
1119 		if (so->so_options & SO_DEBUG) {
1120 			tcp_trace(TA_OUTPUT, tp->t_state, tp,
1121 			    mtod(m, void *), th, 0);
1122 		}
1123 #endif
1124 
1125 		/*
1126 		 * Fill in IP length and desired time to live and
1127 		 * send to IP level.  There should be a better way
1128 		 * to handle ttl and tos; we could keep them in
1129 		 * the template, but need a way to checksum without them.
1130 		 */
1131 		/*
1132 		 * m->m_pkthdr.len should have been set before cksum
1133 		 * calcuration, because in6_cksum() need it.
1134 		 */
1135 		if (isipv6) {
1136 			/*
1137 			 * we separately set hoplimit for every segment,
1138 			 * since the user might want to change the value
1139 			 * via setsockopt.  Also, desired default hop
1140 			 * limit might be changed via Neighbor Discovery.
1141 			 */
1142 			ip6->ip6_hlim = in6_selecthlim(inp,
1143 			    (inp->in6p_route.ro_rt ?
1144 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
1145 
1146 			/* TODO: IPv6 IP6TOS_ECT bit on */
1147 			error = ip6_output(m, inp->in6p_outputopts,
1148 			    &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1149 			    NULL, NULL, inp);
1150 		} else {
1151 			struct rtentry *rt;
1152 			ip->ip_len = m->m_pkthdr.len;
1153 #ifdef INET6
1154 			if (INP_CHECK_SOCKAF(so, AF_INET6))
1155 				ip->ip_ttl = in6_selecthlim(inp,
1156 				    (inp->in6p_route.ro_rt ?
1157 				     inp->in6p_route.ro_rt->rt_ifp : NULL));
1158 			else
1159 #endif
1160 				ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
1161 
1162 			ip->ip_tos = inp->inp_ip_tos;	/* XXX */
1163 			/*
1164 			 * See if we should do MTU discovery.
1165 			 * We do it only if the following are true:
1166 			 *	1) we have a valid route to the destination
1167 			 *	2) the MTU is not locked (if it is,
1168 			 *	   then discovery has been disabled)
1169 			 */
1170 			if (path_mtu_discovery &&
1171 			    (rt = inp->inp_route.ro_rt) &&
1172 			    (rt->rt_flags & RTF_UP) &&
1173 			    !(rt->rt_rmx.rmx_locks & RTV_MTU))
1174 				ip->ip_off |= IP_DF;
1175 
1176 			error = ip_output(m, inp->inp_options, &inp->inp_route,
1177 					  (so->so_options & SO_DONTROUTE) |
1178 					  IP_DEBUGROUTE, NULL, inp);
1179 		}
1180 	} else {
1181 		KASSERT(error != 0, ("no error, but th not set"));
1182 	}
1183 	if (error) {
1184 		tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1185 
1186 		/*
1187 		 * We know that the packet was lost, so back out the
1188 		 * sequence number advance, if any.
1189 		 */
1190 		if (!(tp->t_flags & TF_FORCE) ||
1191 		    !tcp_callout_active(tp, tp->tt_persist)) {
1192 			/*
1193 			 * No need to check for TH_FIN here because
1194 			 * the TF_SENTFIN flag handles that case.
1195 			 */
1196 			if (!(flags & TH_SYN))
1197 				tp->snd_nxt -= len;
1198 		}
1199 
1200 out:
1201 		if (error == ENOBUFS) {
1202 			/*
1203 			 * If we can't send, make sure there is something
1204 			 * to get us going again later.
1205 			 *
1206 			 * The persist timer isn't necessarily allowed in all
1207 			 * states, use the rexmt timer.
1208 			 */
1209 			if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1210 			    !tcp_callout_active(tp, tp->tt_persist)) {
1211 				tcp_callout_reset(tp, tp->tt_rexmt,
1212 						  tp->t_rxtcur,
1213 						  tcp_timer_rexmt);
1214 #if 0
1215 				tp->t_rxtshift = 0;
1216 				tcp_setpersist(tp);
1217 #endif
1218 			}
1219 			tcp_quench(inp, 0);
1220 			return (0);
1221 		}
1222 		if (error == EMSGSIZE) {
1223 			/*
1224 			 * ip_output() will have already fixed the route
1225 			 * for us.  tcp_mtudisc() will, as its last action,
1226 			 * initiate retransmission, so it is important to
1227 			 * not do so here.
1228 			 */
1229 			tcp_mtudisc(inp, 0);
1230 			return 0;
1231 		}
1232 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1233 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1234 			tp->t_softerror = error;
1235 			return (0);
1236 		}
1237 		return (error);
1238 	}
1239 	tcpstat.tcps_sndtotal++;
1240 
1241 	/*
1242 	 * Data sent (as far as we can tell).
1243 	 *
1244 	 * If this advertises a larger window than any other segment,
1245 	 * then remember the size of the advertised window.
1246 	 *
1247 	 * Any pending ACK has now been sent.
1248 	 */
1249 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1250 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1251 		tp->t_flags &= ~TF_RXRESIZED;
1252 	}
1253 	tp->last_ack_sent = tp->rcv_nxt;
1254 	tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1255 	if (tcp_delack_enabled)
1256 		tcp_callout_stop(tp, tp->tt_delack);
1257 	if (sendalot)
1258 		goto again;
1259 	return (0);
1260 }
1261 
1262 void
1263 tcp_setpersist(struct tcpcb *tp)
1264 {
1265 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1266 	int tt;
1267 
1268 	if (tp->t_state == TCPS_SYN_SENT ||
1269 	    tp->t_state == TCPS_SYN_RECEIVED) {
1270 		panic("tcp_setpersist: not established yet, current %s",
1271 		      tp->t_state == TCPS_SYN_SENT ?
1272 		      "SYN_SENT" : "SYN_RECEIVED");
1273 	}
1274 
1275 	if (tcp_callout_active(tp, tp->tt_rexmt))
1276 		panic("tcp_setpersist: retransmit pending");
1277 	/*
1278 	 * Start/restart persistance timer.
1279 	 */
1280 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1281 		      TCPTV_PERSMAX);
1282 	tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1283 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1284 		tp->t_rxtshift++;
1285 }
1286 
1287 static void
1288 tcp_idle_cwnd_validate(struct tcpcb *tp)
1289 {
1290 	u_long initial_cwnd = tcp_initial_window(tp);
1291 	u_long min_cwnd;
1292 
1293 	tcpstat.tcps_sndidle++;
1294 
1295 	/* According to RFC5681: RW=min(IW,cwnd) */
1296 	min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1297 
1298 	if (tcp_idle_cwv) {
1299 		u_long idle_time, decay_cwnd;
1300 
1301 		/*
1302 		 * RFC2861, but only after idle period.
1303 		 */
1304 
1305 		/*
1306 		 * Before the congestion window is reduced, ssthresh
1307 		 * is set to the maximum of its current value and 3/4
1308 		 * cwnd.  If the sender then has more data to send
1309 		 * than the decayed cwnd allows, the TCP will slow-
1310 		 * start (perform exponential increase) at least
1311 		 * half-way back up to the old value of cwnd.
1312 		 */
1313 		tp->snd_ssthresh = max(tp->snd_ssthresh,
1314 		    (3 * tp->snd_cwnd) / 4);
1315 
1316 		/*
1317 		 * Decay the congestion window by half for every RTT
1318 		 * that the flow remains inactive.
1319 		 *
1320 		 * The difference between our implementation and
1321 		 * RFC2861 is that we don't allow cwnd to go below
1322 		 * the value allowed by RFC5681 (min_cwnd).
1323 		 */
1324 		idle_time = ticks - tp->snd_last;
1325 		decay_cwnd = tp->snd_cwnd;
1326 		while (idle_time >= tp->t_rxtcur &&
1327 		    decay_cwnd > min_cwnd) {
1328 			decay_cwnd >>= 1;
1329 			idle_time -= tp->t_rxtcur;
1330 		}
1331 		tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1332 	} else {
1333 		/*
1334 		 * Slow-start from scratch to re-determine the send
1335 		 * congestion window.
1336 		 */
1337 		tp->snd_cwnd = min_cwnd;
1338 	}
1339 
1340 	/* Restart ABC counting during congestion avoidance */
1341 	tp->snd_wacked = 0;
1342 }
1343 
1344 static int
1345 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1346 {
1347 	struct inpcb * const inp = tp->t_inpcb;
1348 #ifdef INET6
1349 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1350 #else
1351 	const boolean_t isipv6 = FALSE;
1352 #endif
1353 	unsigned int ipoptlen, optlen;
1354 	u_int hlen;
1355 
1356 	hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1357 
1358 	if (isipv6) {
1359 		ipoptlen = ip6_optlen(inp);
1360 	} else {
1361 		if (inp->inp_options) {
1362 			ipoptlen = inp->inp_options->m_len -
1363 			    offsetof(struct ipoption, ipopt_list);
1364 		} else {
1365 			ipoptlen = 0;
1366 		}
1367 	}
1368 #ifdef IPSEC
1369 	ipoptlen += ipsec_hdrsiz_tcp(tp);
1370 #endif
1371 	hlen += ipoptlen;
1372 
1373 	optlen = 0;
1374 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1375 	    (tp->t_flags & TF_RCVD_TSTMP))
1376 		optlen += TCPOLEN_TSTAMP_APPA;
1377 	hlen += optlen;
1378 
1379 	if (tp->t_maxopd <= optlen + ipoptlen)
1380 		return EHOSTUNREACH;
1381 
1382 	*segsz = tp->t_maxopd - optlen - ipoptlen;
1383 	*hlen0 = hlen;
1384 	return 0;
1385 }
1386