xref: /freebsd/sys/netinet/tcp_output.c (revision 5dae51da3da0cc94d17bd67b308fad304ebec7e0)
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  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #ifdef TCP_HHOOK
44 #include <sys/hhook.h>
45 #endif
46 #include <sys/kernel.h>
47 #include <sys/lock.h>
48 #include <sys/mbuf.h>
49 #include <sys/mutex.h>
50 #include <sys/protosw.h>
51 #include <sys/sdt.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/sysctl.h>
55 
56 #include <net/if.h>
57 #include <net/route.h>
58 #include <net/vnet.h>
59 
60 #include <netinet/in.h>
61 #include <netinet/in_kdtrace.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/ip_options.h>
67 #ifdef INET6
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/ip6_var.h>
71 #endif
72 #ifdef TCP_RFC7413
73 #include <netinet/tcp_fastopen.h>
74 #endif
75 #include <netinet/tcp.h>
76 #define	TCPOUTFLAGS
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcpip.h>
82 #include <netinet/cc/cc.h>
83 #ifdef TCPPCAP
84 #include <netinet/tcp_pcap.h>
85 #endif
86 #ifdef TCPDEBUG
87 #include <netinet/tcp_debug.h>
88 #endif
89 #ifdef TCP_OFFLOAD
90 #include <netinet/tcp_offload.h>
91 #endif
92 
93 #ifdef IPSEC
94 #include <netipsec/ipsec.h>
95 #endif /*IPSEC*/
96 
97 #include <machine/in_cksum.h>
98 
99 #include <security/mac/mac_framework.h>
100 
101 VNET_DEFINE(int, path_mtu_discovery) = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
103 	&VNET_NAME(path_mtu_discovery), 1,
104 	"Enable Path MTU Discovery");
105 
106 VNET_DEFINE(int, tcp_do_tso) = 1;
107 #define	V_tcp_do_tso		VNET(tcp_do_tso)
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
109 	&VNET_NAME(tcp_do_tso), 0,
110 	"Enable TCP Segmentation Offload");
111 
112 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
113 #define	V_tcp_sendspace	VNET(tcp_sendspace)
114 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
115 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
116 
117 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
118 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
120 	&VNET_NAME(tcp_do_autosndbuf), 0,
121 	"Enable automatic send buffer sizing");
122 
123 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
124 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
126 	&VNET_NAME(tcp_autosndbuf_inc), 0,
127 	"Incrementor step size of automatic send buffer");
128 
129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
130 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
132 	&VNET_NAME(tcp_autosndbuf_max), 0,
133 	"Max size of automatic send buffer");
134 
135 /*
136  * Make sure that either retransmit or persist timer is set for SYN, FIN and
137  * non-ACK.
138  */
139 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
140 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
141 	    tcp_timer_active((tp), TT_REXMT) ||				\
142 	    tcp_timer_active((tp), TT_PERSIST),				\
143 	    ("neither rexmt nor persist timer is set"))
144 
145 #ifdef TCP_HHOOK
146 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
147 			    struct tcphdr *th, struct tcpopt *to,
148 			    uint32_t len, int tso);
149 #endif
150 static void inline	cc_after_idle(struct tcpcb *tp);
151 
152 #ifdef TCP_HHOOK
153 /*
154  * Wrapper for the TCP established output helper hook.
155  */
156 static void inline
157 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
158     struct tcpopt *to, uint32_t len, int tso)
159 {
160 	struct tcp_hhook_data hhook_data;
161 
162 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
163 		hhook_data.tp = tp;
164 		hhook_data.th = th;
165 		hhook_data.to = to;
166 		hhook_data.len = len;
167 		hhook_data.tso = tso;
168 
169 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
170 		    tp->osd);
171 	}
172 }
173 #endif
174 
175 /*
176  * CC wrapper hook functions
177  */
178 static void inline
179 cc_after_idle(struct tcpcb *tp)
180 {
181 	INP_WLOCK_ASSERT(tp->t_inpcb);
182 
183 	if (CC_ALGO(tp)->after_idle != NULL)
184 		CC_ALGO(tp)->after_idle(tp->ccv);
185 }
186 
187 /*
188  * Tcp output routine: figure out what should be sent and send it.
189  */
190 int
191 tcp_output(struct tcpcb *tp)
192 {
193 	struct socket *so = tp->t_inpcb->inp_socket;
194 	int32_t len;
195 	uint32_t recwin, sendwin;
196 	int off, flags, error = 0;	/* Keep compiler happy */
197 	struct mbuf *m;
198 	struct ip *ip = NULL;
199 	struct ipovly *ipov = NULL;
200 	struct tcphdr *th;
201 	u_char opt[TCP_MAXOLEN];
202 	unsigned ipoptlen, optlen, hdrlen;
203 #ifdef IPSEC
204 	unsigned ipsec_optlen = 0;
205 #endif
206 	int idle, sendalot;
207 	int sack_rxmit, sack_bytes_rxmt;
208 	struct sackhole *p;
209 	int tso, mtu;
210 	struct tcpopt to;
211 #if 0
212 	int maxburst = TCP_MAXBURST;
213 #endif
214 #ifdef INET6
215 	struct ip6_hdr *ip6 = NULL;
216 	int isipv6;
217 
218 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
219 #endif
220 
221 	INP_WLOCK_ASSERT(tp->t_inpcb);
222 
223 #ifdef TCP_OFFLOAD
224 	if (tp->t_flags & TF_TOE)
225 		return (tcp_offload_output(tp));
226 #endif
227 
228 #ifdef TCP_RFC7413
229 	/*
230 	 * For TFO connections in SYN_RECEIVED, only allow the initial
231 	 * SYN|ACK and those sent by the retransmit timer.
232 	 */
233 	if (IS_FASTOPEN(tp->t_flags) &&
234 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
235 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
236 	    (tp->snd_nxt != tp->snd_una))          /* not a retransmit */
237 		return (0);
238 #endif
239 	/*
240 	 * Determine length of data that should be transmitted,
241 	 * and flags that will be used.
242 	 * If there is some data or critical controls (SYN, RST)
243 	 * to send, then transmit; otherwise, investigate further.
244 	 */
245 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
246 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
247 		cc_after_idle(tp);
248 	tp->t_flags &= ~TF_LASTIDLE;
249 	if (idle) {
250 		if (tp->t_flags & TF_MORETOCOME) {
251 			tp->t_flags |= TF_LASTIDLE;
252 			idle = 0;
253 		}
254 	}
255 again:
256 	/*
257 	 * If we've recently taken a timeout, snd_max will be greater than
258 	 * snd_nxt.  There may be SACK information that allows us to avoid
259 	 * resending already delivered data.  Adjust snd_nxt accordingly.
260 	 */
261 	if ((tp->t_flags & TF_SACK_PERMIT) &&
262 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
263 		tcp_sack_adjust(tp);
264 	sendalot = 0;
265 	tso = 0;
266 	mtu = 0;
267 	off = tp->snd_nxt - tp->snd_una;
268 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
269 
270 	flags = tcp_outflags[tp->t_state];
271 	/*
272 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
273 	 * to send out new data (when sendalot is 1), bypass this function.
274 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
275 	 * we're replacing a (future) new transmission with a retransmission
276 	 * now, and we previously incremented snd_cwnd in tcp_input().
277 	 */
278 	/*
279 	 * Still in sack recovery , reset rxmit flag to zero.
280 	 */
281 	sack_rxmit = 0;
282 	sack_bytes_rxmt = 0;
283 	len = 0;
284 	p = NULL;
285 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
286 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
287 		uint32_t cwin;
288 
289 		cwin =
290 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
291 		/* Do not retransmit SACK segments beyond snd_recover */
292 		if (SEQ_GT(p->end, tp->snd_recover)) {
293 			/*
294 			 * (At least) part of sack hole extends beyond
295 			 * snd_recover. Check to see if we can rexmit data
296 			 * for this hole.
297 			 */
298 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
299 				/*
300 				 * Can't rexmit any more data for this hole.
301 				 * That data will be rexmitted in the next
302 				 * sack recovery episode, when snd_recover
303 				 * moves past p->rxmit.
304 				 */
305 				p = NULL;
306 				goto after_sack_rexmit;
307 			} else
308 				/* Can rexmit part of the current hole */
309 				len = ((int32_t)ulmin(cwin,
310 						   tp->snd_recover - p->rxmit));
311 		} else
312 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
313 		off = p->rxmit - tp->snd_una;
314 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
315 		    __func__, off));
316 		if (len > 0) {
317 			sack_rxmit = 1;
318 			sendalot = 1;
319 			TCPSTAT_INC(tcps_sack_rexmits);
320 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
321 			    min(len, tp->t_maxseg));
322 		}
323 	}
324 after_sack_rexmit:
325 	/*
326 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
327 	 * state flags.
328 	 */
329 	if (tp->t_flags & TF_NEEDFIN)
330 		flags |= TH_FIN;
331 	if (tp->t_flags & TF_NEEDSYN)
332 		flags |= TH_SYN;
333 
334 	SOCKBUF_LOCK(&so->so_snd);
335 	/*
336 	 * If in persist timeout with window of 0, send 1 byte.
337 	 * Otherwise, if window is small but nonzero
338 	 * and timer expired, we will send what we can
339 	 * and go to transmit state.
340 	 */
341 	if (tp->t_flags & TF_FORCEDATA) {
342 		if (sendwin == 0) {
343 			/*
344 			 * If we still have some data to send, then
345 			 * clear the FIN bit.  Usually this would
346 			 * happen below when it realizes that we
347 			 * aren't sending all the data.  However,
348 			 * if we have exactly 1 byte of unsent data,
349 			 * then it won't clear the FIN bit below,
350 			 * and if we are in persist state, we wind
351 			 * up sending the packet without recording
352 			 * that we sent the FIN bit.
353 			 *
354 			 * We can't just blindly clear the FIN bit,
355 			 * because if we don't have any more data
356 			 * to send then the probe will be the FIN
357 			 * itself.
358 			 */
359 			if (off < sbused(&so->so_snd))
360 				flags &= ~TH_FIN;
361 			sendwin = 1;
362 		} else {
363 			tcp_timer_activate(tp, TT_PERSIST, 0);
364 			tp->t_rxtshift = 0;
365 		}
366 	}
367 
368 	/*
369 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
370 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
371 	 * a negative length.  This can also occur when TCP opens up
372 	 * its congestion window while receiving additional duplicate
373 	 * acks after fast-retransmit because TCP will reset snd_nxt
374 	 * to snd_max after the fast-retransmit.
375 	 *
376 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
377 	 * be set to snd_una, the offset will be 0, and the length may
378 	 * wind up 0.
379 	 *
380 	 * If sack_rxmit is true we are retransmitting from the scoreboard
381 	 * in which case len is already set.
382 	 */
383 	if (sack_rxmit == 0) {
384 		if (sack_bytes_rxmt == 0)
385 			len = ((int32_t)ulmin(sbavail(&so->so_snd), sendwin) -
386 			    off);
387 		else {
388 			int32_t cwin;
389 
390                         /*
391 			 * We are inside of a SACK recovery episode and are
392 			 * sending new data, having retransmitted all the
393 			 * data possible in the scoreboard.
394 			 */
395 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
396 			    off);
397 			/*
398 			 * Don't remove this (len > 0) check !
399 			 * We explicitly check for len > 0 here (although it
400 			 * isn't really necessary), to work around a gcc
401 			 * optimization issue - to force gcc to compute
402 			 * len above. Without this check, the computation
403 			 * of len is bungled by the optimizer.
404 			 */
405 			if (len > 0) {
406 				cwin = tp->snd_cwnd -
407 					(tp->snd_nxt - tp->sack_newdata) -
408 					sack_bytes_rxmt;
409 				if (cwin < 0)
410 					cwin = 0;
411 				len = imin(len, cwin);
412 			}
413 		}
414 	}
415 
416 	/*
417 	 * Lop off SYN bit if it has already been sent.  However, if this
418 	 * is SYN-SENT state and if segment contains data and if we don't
419 	 * know that foreign host supports TAO, suppress sending segment.
420 	 */
421 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
422 		if (tp->t_state != TCPS_SYN_RECEIVED)
423 			flags &= ~TH_SYN;
424 #ifdef TCP_RFC7413
425 		/*
426 		 * When sending additional segments following a TFO SYN|ACK,
427 		 * do not include the SYN bit.
428 		 */
429 		if (IS_FASTOPEN(tp->t_flags) &&
430 		    (tp->t_state == TCPS_SYN_RECEIVED))
431 			flags &= ~TH_SYN;
432 #endif
433 		off--, len++;
434 	}
435 
436 	/*
437 	 * Be careful not to send data and/or FIN on SYN segments.
438 	 * This measure is needed to prevent interoperability problems
439 	 * with not fully conformant TCP implementations.
440 	 */
441 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
442 		len = 0;
443 		flags &= ~TH_FIN;
444 	}
445 
446 #ifdef TCP_RFC7413
447 	/*
448 	 * When retransmitting SYN|ACK on a passively-created TFO socket,
449 	 * don't include data, as the presence of data may have caused the
450 	 * original SYN|ACK to have been dropped by a middlebox.
451 	 */
452 	if (IS_FASTOPEN(tp->t_flags) &&
453 	    (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
454 	     (flags & TH_RST)))
455 		len = 0;
456 #endif
457 	if (len <= 0) {
458 		/*
459 		 * If FIN has been sent but not acked,
460 		 * but we haven't been called to retransmit,
461 		 * len will be < 0.  Otherwise, window shrank
462 		 * after we sent into it.  If window shrank to 0,
463 		 * cancel pending retransmit, pull snd_nxt back
464 		 * to (closed) window, and set the persist timer
465 		 * if it isn't already going.  If the window didn't
466 		 * close completely, just wait for an ACK.
467 		 *
468 		 * We also do a general check here to ensure that
469 		 * we will set the persist timer when we have data
470 		 * to send, but a 0-byte window. This makes sure
471 		 * the persist timer is set even if the packet
472 		 * hits one of the "goto send" lines below.
473 		 */
474 		len = 0;
475 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
476 			(off < (int) sbavail(&so->so_snd))) {
477 			tcp_timer_activate(tp, TT_REXMT, 0);
478 			tp->t_rxtshift = 0;
479 			tp->snd_nxt = tp->snd_una;
480 			if (!tcp_timer_active(tp, TT_PERSIST))
481 				tcp_setpersist(tp);
482 		}
483 	}
484 
485 	/* len will be >= 0 after this point. */
486 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
487 
488 	/*
489 	 * Automatic sizing of send socket buffer.  Often the send buffer
490 	 * size is not optimally adjusted to the actual network conditions
491 	 * at hand (delay bandwidth product).  Setting the buffer size too
492 	 * small limits throughput on links with high bandwidth and high
493 	 * delay (eg. trans-continental/oceanic links).  Setting the
494 	 * buffer size too big consumes too much real kernel memory,
495 	 * especially with many connections on busy servers.
496 	 *
497 	 * The criteria to step up the send buffer one notch are:
498 	 *  1. receive window of remote host is larger than send buffer
499 	 *     (with a fudge factor of 5/4th);
500 	 *  2. send buffer is filled to 7/8th with data (so we actually
501 	 *     have data to make use of it);
502 	 *  3. send buffer fill has not hit maximal automatic size;
503 	 *  4. our send window (slow start and cogestion controlled) is
504 	 *     larger than sent but unacknowledged data in send buffer.
505 	 *
506 	 * The remote host receive window scaling factor may limit the
507 	 * growing of the send buffer before it reaches its allowed
508 	 * maximum.
509 	 *
510 	 * It scales directly with slow start or congestion window
511 	 * and does at most one step per received ACK.  This fast
512 	 * scaling has the drawback of growing the send buffer beyond
513 	 * what is strictly necessary to make full use of a given
514 	 * delay*bandwidth product.  However testing has shown this not
515 	 * to be much of an problem.  At worst we are trading wasting
516 	 * of available bandwidth (the non-use of it) for wasting some
517 	 * socket buffer memory.
518 	 *
519 	 * TODO: Shrink send buffer during idle periods together
520 	 * with congestion window.  Requires another timer.  Has to
521 	 * wait for upcoming tcp timer rewrite.
522 	 *
523 	 * XXXGL: should there be used sbused() or sbavail()?
524 	 */
525 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
526 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
527 		    sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) &&
528 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
529 		    sendwin >= (sbused(&so->so_snd) -
530 		    (tp->snd_nxt - tp->snd_una))) {
531 			if (!sbreserve_locked(&so->so_snd,
532 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
533 			     V_tcp_autosndbuf_max), so, curthread))
534 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
535 		}
536 	}
537 
538 	/*
539 	 * Decide if we can use TCP Segmentation Offloading (if supported by
540 	 * hardware).
541 	 *
542 	 * TSO may only be used if we are in a pure bulk sending state.  The
543 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
544 	 * IP options prevent using TSO.  With TSO the TCP header is the same
545 	 * (except for the sequence number) for all generated packets.  This
546 	 * makes it impossible to transmit any options which vary per generated
547 	 * segment or packet.
548 	 */
549 #ifdef IPSEC
550 	/*
551 	 * Pre-calculate here as we save another lookup into the darknesses
552 	 * of IPsec that way and can actually decide if TSO is ok.
553 	 */
554 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
555 #endif
556 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
557 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
558 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
559 #ifdef IPSEC
560 	    ipsec_optlen == 0 &&
561 #endif
562 	    tp->t_inpcb->inp_options == NULL &&
563 	    tp->t_inpcb->in6p_options == NULL)
564 		tso = 1;
565 
566 	if (sack_rxmit) {
567 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
568 			flags &= ~TH_FIN;
569 	} else {
570 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
571 		    sbused(&so->so_snd)))
572 			flags &= ~TH_FIN;
573 	}
574 
575 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
576 	    (long)TCP_MAXWIN << tp->rcv_scale);
577 
578 	/*
579 	 * Sender silly window avoidance.   We transmit under the following
580 	 * conditions when len is non-zero:
581 	 *
582 	 *	- We have a full segment (or more with TSO)
583 	 *	- This is the last buffer in a write()/send() and we are
584 	 *	  either idle or running NODELAY
585 	 *	- we've timed out (e.g. persist timer)
586 	 *	- we have more then 1/2 the maximum send window's worth of
587 	 *	  data (receiver may be limited the window size)
588 	 *	- we need to retransmit
589 	 */
590 	if (len) {
591 		if (len >= tp->t_maxseg)
592 			goto send;
593 		/*
594 		 * NOTE! on localhost connections an 'ack' from the remote
595 		 * end may occur synchronously with the output and cause
596 		 * us to flush a buffer queued with moretocome.  XXX
597 		 *
598 		 * note: the len + off check is almost certainly unnecessary.
599 		 */
600 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
601 		    (idle || (tp->t_flags & TF_NODELAY)) &&
602 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
603 		    (tp->t_flags & TF_NOPUSH) == 0) {
604 			goto send;
605 		}
606 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
607 			goto send;
608 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
609 			goto send;
610 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
611 			goto send;
612 		if (sack_rxmit)
613 			goto send;
614 	}
615 
616 	/*
617 	 * Sending of standalone window updates.
618 	 *
619 	 * Window updates are important when we close our window due to a
620 	 * full socket buffer and are opening it again after the application
621 	 * reads data from it.  Once the window has opened again and the
622 	 * remote end starts to send again the ACK clock takes over and
623 	 * provides the most current window information.
624 	 *
625 	 * We must avoid the silly window syndrome whereas every read
626 	 * from the receive buffer, no matter how small, causes a window
627 	 * update to be sent.  We also should avoid sending a flurry of
628 	 * window updates when the socket buffer had queued a lot of data
629 	 * and the application is doing small reads.
630 	 *
631 	 * Prevent a flurry of pointless window updates by only sending
632 	 * an update when we can increase the advertized window by more
633 	 * than 1/4th of the socket buffer capacity.  When the buffer is
634 	 * getting full or is very small be more aggressive and send an
635 	 * update whenever we can increase by two mss sized segments.
636 	 * In all other situations the ACK's to new incoming data will
637 	 * carry further window increases.
638 	 *
639 	 * Don't send an independent window update if a delayed
640 	 * ACK is pending (it will get piggy-backed on it) or the
641 	 * remote side already has done a half-close and won't send
642 	 * more data.  Skip this if the connection is in T/TCP
643 	 * half-open state.
644 	 */
645 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
646 	    !(tp->t_flags & TF_DELACK) &&
647 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
648 		/*
649 		 * "adv" is the amount we could increase the window,
650 		 * taking into account that we are limited by
651 		 * TCP_MAXWIN << tp->rcv_scale.
652 		 */
653 		int32_t adv;
654 		int oldwin;
655 
656 		adv = recwin;
657 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
658 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
659 			adv -= oldwin;
660 		} else
661 			oldwin = 0;
662 
663 		/*
664 		 * If the new window size ends up being the same as or less
665 		 * than the old size when it is scaled, then don't force
666 		 * a window update.
667 		 */
668 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
669 			goto dontupdate;
670 
671 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
672 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
673 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
674 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
675 			goto send;
676 	}
677 dontupdate:
678 
679 	/*
680 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
681 	 * is also a catch-all for the retransmit timer timeout case.
682 	 */
683 	if (tp->t_flags & TF_ACKNOW)
684 		goto send;
685 	if ((flags & TH_RST) ||
686 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
687 		goto send;
688 	if (SEQ_GT(tp->snd_up, tp->snd_una))
689 		goto send;
690 	/*
691 	 * If our state indicates that FIN should be sent
692 	 * and we have not yet done so, then we need to send.
693 	 */
694 	if (flags & TH_FIN &&
695 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
696 		goto send;
697 	/*
698 	 * In SACK, it is possible for tcp_output to fail to send a segment
699 	 * after the retransmission timer has been turned off.  Make sure
700 	 * that the retransmission timer is set.
701 	 */
702 	if ((tp->t_flags & TF_SACK_PERMIT) &&
703 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
704 	    !tcp_timer_active(tp, TT_REXMT) &&
705 	    !tcp_timer_active(tp, TT_PERSIST)) {
706 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
707 		goto just_return;
708 	}
709 	/*
710 	 * TCP window updates are not reliable, rather a polling protocol
711 	 * using ``persist'' packets is used to insure receipt of window
712 	 * updates.  The three ``states'' for the output side are:
713 	 *	idle			not doing retransmits or persists
714 	 *	persisting		to move a small or zero window
715 	 *	(re)transmitting	and thereby not persisting
716 	 *
717 	 * tcp_timer_active(tp, TT_PERSIST)
718 	 *	is true when we are in persist state.
719 	 * (tp->t_flags & TF_FORCEDATA)
720 	 *	is set when we are called to send a persist packet.
721 	 * tcp_timer_active(tp, TT_REXMT)
722 	 *	is set when we are retransmitting
723 	 * The output side is idle when both timers are zero.
724 	 *
725 	 * If send window is too small, there is data to transmit, and no
726 	 * retransmit or persist is pending, then go to persist state.
727 	 * If nothing happens soon, send when timer expires:
728 	 * if window is nonzero, transmit what we can,
729 	 * otherwise force out a byte.
730 	 */
731 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
732 	    !tcp_timer_active(tp, TT_PERSIST)) {
733 		tp->t_rxtshift = 0;
734 		tcp_setpersist(tp);
735 	}
736 
737 	/*
738 	 * No reason to send a segment, just return.
739 	 */
740 just_return:
741 	SOCKBUF_UNLOCK(&so->so_snd);
742 	return (0);
743 
744 send:
745 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
746 	if (len > 0) {
747 		if (len >= tp->t_maxseg)
748 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
749 		else
750 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
751 	}
752 	/*
753 	 * Before ESTABLISHED, force sending of initial options
754 	 * unless TCP set not to do any options.
755 	 * NOTE: we assume that the IP/TCP header plus TCP options
756 	 * always fit in a single mbuf, leaving room for a maximum
757 	 * link header, i.e.
758 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
759 	 */
760 	optlen = 0;
761 #ifdef INET6
762 	if (isipv6)
763 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
764 	else
765 #endif
766 		hdrlen = sizeof (struct tcpiphdr);
767 
768 	/*
769 	 * Compute options for segment.
770 	 * We only have to care about SYN and established connection
771 	 * segments.  Options for SYN-ACK segments are handled in TCP
772 	 * syncache.
773 	 */
774 	to.to_flags = 0;
775 	if ((tp->t_flags & TF_NOOPT) == 0) {
776 		/* Maximum segment size. */
777 		if (flags & TH_SYN) {
778 			tp->snd_nxt = tp->iss;
779 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
780 			to.to_flags |= TOF_MSS;
781 #ifdef TCP_RFC7413
782 			/*
783 			 * Only include the TFO option on the first
784 			 * transmission of the SYN|ACK on a
785 			 * passively-created TFO socket, as the presence of
786 			 * the TFO option may have caused the original
787 			 * SYN|ACK to have been dropped by a middlebox.
788 			 */
789 			if (IS_FASTOPEN(tp->t_flags) &&
790 			    (tp->t_state == TCPS_SYN_RECEIVED) &&
791 			    (tp->t_rxtshift == 0)) {
792 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
793 				to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
794 				to.to_flags |= TOF_FASTOPEN;
795 			}
796 #endif
797 		}
798 		/* Window scaling. */
799 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
800 			to.to_wscale = tp->request_r_scale;
801 			to.to_flags |= TOF_SCALE;
802 		}
803 		/* Timestamps. */
804 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
805 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
806 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
807 			to.to_tsecr = tp->ts_recent;
808 			to.to_flags |= TOF_TS;
809 			/* Set receive buffer autosizing timestamp. */
810 			if (tp->rfbuf_ts == 0 &&
811 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
812 				tp->rfbuf_ts = tcp_ts_getticks();
813 		}
814 		/* Selective ACK's. */
815 		if (tp->t_flags & TF_SACK_PERMIT) {
816 			if (flags & TH_SYN)
817 				to.to_flags |= TOF_SACKPERM;
818 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
819 			    (tp->t_flags & TF_SACK_PERMIT) &&
820 			    tp->rcv_numsacks > 0) {
821 				to.to_flags |= TOF_SACK;
822 				to.to_nsacks = tp->rcv_numsacks;
823 				to.to_sacks = (u_char *)tp->sackblks;
824 			}
825 		}
826 #ifdef TCP_SIGNATURE
827 		/* TCP-MD5 (RFC2385). */
828 		if (tp->t_flags & TF_SIGNATURE)
829 			to.to_flags |= TOF_SIGNATURE;
830 #endif /* TCP_SIGNATURE */
831 
832 		/* Processing the options. */
833 		hdrlen += optlen = tcp_addoptions(&to, opt);
834 	}
835 
836 #ifdef INET6
837 	if (isipv6)
838 		ipoptlen = ip6_optlen(tp->t_inpcb);
839 	else
840 #endif
841 	if (tp->t_inpcb->inp_options)
842 		ipoptlen = tp->t_inpcb->inp_options->m_len -
843 				offsetof(struct ipoption, ipopt_list);
844 	else
845 		ipoptlen = 0;
846 #ifdef IPSEC
847 	ipoptlen += ipsec_optlen;
848 #endif
849 
850 	/*
851 	 * Adjust data length if insertion of options will
852 	 * bump the packet length beyond the t_maxseg length.
853 	 * Clear the FIN bit because we cut off the tail of
854 	 * the segment.
855 	 */
856 	if (len + optlen + ipoptlen > tp->t_maxseg) {
857 		flags &= ~TH_FIN;
858 
859 		if (tso) {
860 			u_int if_hw_tsomax;
861 			u_int if_hw_tsomaxsegcount;
862 			u_int if_hw_tsomaxsegsize;
863 			struct mbuf *mb;
864 			u_int moff;
865 			int max_len;
866 
867 			/* extract TSO information */
868 			if_hw_tsomax = tp->t_tsomax;
869 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
870 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
871 
872 			/*
873 			 * Limit a TSO burst to prevent it from
874 			 * overflowing or exceeding the maximum length
875 			 * allowed by the network interface:
876 			 */
877 			KASSERT(ipoptlen == 0,
878 			    ("%s: TSO can't do IP options", __func__));
879 
880 			/*
881 			 * Check if we should limit by maximum payload
882 			 * length:
883 			 */
884 			if (if_hw_tsomax != 0) {
885 				/* compute maximum TSO length */
886 				max_len = (if_hw_tsomax - hdrlen -
887 				    max_linkhdr);
888 				if (max_len <= 0) {
889 					len = 0;
890 				} else if (len > max_len) {
891 					sendalot = 1;
892 					len = max_len;
893 				}
894 			}
895 
896 			/*
897 			 * Check if we should limit by maximum segment
898 			 * size and count:
899 			 */
900 			if (if_hw_tsomaxsegcount != 0 &&
901 			    if_hw_tsomaxsegsize != 0) {
902 				/*
903 				 * Subtract one segment for the LINK
904 				 * and TCP/IP headers mbuf that will
905 				 * be prepended to this mbuf chain
906 				 * after the code in this section
907 				 * limits the number of mbufs in the
908 				 * chain to if_hw_tsomaxsegcount.
909 				 */
910 				if_hw_tsomaxsegcount -= 1;
911 				max_len = 0;
912 				mb = sbsndmbuf(&so->so_snd, off, &moff);
913 
914 				while (mb != NULL && max_len < len) {
915 					u_int mlen;
916 					u_int frags;
917 
918 					/*
919 					 * Get length of mbuf fragment
920 					 * and how many hardware frags,
921 					 * rounded up, it would use:
922 					 */
923 					mlen = (mb->m_len - moff);
924 					frags = howmany(mlen,
925 					    if_hw_tsomaxsegsize);
926 
927 					/* Handle special case: Zero Length Mbuf */
928 					if (frags == 0)
929 						frags = 1;
930 
931 					/*
932 					 * Check if the fragment limit
933 					 * will be reached or exceeded:
934 					 */
935 					if (frags >= if_hw_tsomaxsegcount) {
936 						max_len += min(mlen,
937 						    if_hw_tsomaxsegcount *
938 						    if_hw_tsomaxsegsize);
939 						break;
940 					}
941 					max_len += mlen;
942 					if_hw_tsomaxsegcount -= frags;
943 					moff = 0;
944 					mb = mb->m_next;
945 				}
946 				if (max_len <= 0) {
947 					len = 0;
948 				} else if (len > max_len) {
949 					sendalot = 1;
950 					len = max_len;
951 				}
952 			}
953 
954 			/*
955 			 * Prevent the last segment from being
956 			 * fractional unless the send sockbuf can be
957 			 * emptied:
958 			 */
959 			max_len = (tp->t_maxseg - optlen);
960 			if (((uint32_t)off + (uint32_t)len) <
961 			    sbavail(&so->so_snd)) {
962 				moff = len % max_len;
963 				if (moff != 0) {
964 					len -= moff;
965 					sendalot = 1;
966 				}
967 			}
968 
969 			/*
970 			 * In case there are too many small fragments
971 			 * don't use TSO:
972 			 */
973 			if (len <= max_len) {
974 				len = max_len;
975 				sendalot = 1;
976 				tso = 0;
977 			}
978 
979 			/*
980 			 * Send the FIN in a separate segment
981 			 * after the bulk sending is done.
982 			 * We don't trust the TSO implementations
983 			 * to clear the FIN flag on all but the
984 			 * last segment.
985 			 */
986 			if (tp->t_flags & TF_NEEDFIN)
987 				sendalot = 1;
988 
989 		} else {
990 			len = tp->t_maxseg - optlen - ipoptlen;
991 			sendalot = 1;
992 		}
993 	} else
994 		tso = 0;
995 
996 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
997 	    ("%s: len > IP_MAXPACKET", __func__));
998 
999 /*#ifdef DIAGNOSTIC*/
1000 #ifdef INET6
1001 	if (max_linkhdr + hdrlen > MCLBYTES)
1002 #else
1003 	if (max_linkhdr + hdrlen > MHLEN)
1004 #endif
1005 		panic("tcphdr too big");
1006 /*#endif*/
1007 
1008 	/*
1009 	 * This KASSERT is here to catch edge cases at a well defined place.
1010 	 * Before, those had triggered (random) panic conditions further down.
1011 	 */
1012 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1013 
1014 	/*
1015 	 * Grab a header mbuf, attaching a copy of data to
1016 	 * be transmitted, and initialize the header from
1017 	 * the template for sends on this connection.
1018 	 */
1019 	if (len) {
1020 		struct mbuf *mb;
1021 		u_int moff;
1022 
1023 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1024 			TCPSTAT_INC(tcps_sndprobe);
1025 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1026 			tp->t_sndrexmitpack++;
1027 			TCPSTAT_INC(tcps_sndrexmitpack);
1028 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1029 		} else {
1030 			TCPSTAT_INC(tcps_sndpack);
1031 			TCPSTAT_ADD(tcps_sndbyte, len);
1032 		}
1033 #ifdef INET6
1034 		if (MHLEN < hdrlen + max_linkhdr)
1035 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1036 		else
1037 #endif
1038 			m = m_gethdr(M_NOWAIT, MT_DATA);
1039 
1040 		if (m == NULL) {
1041 			SOCKBUF_UNLOCK(&so->so_snd);
1042 			error = ENOBUFS;
1043 			sack_rxmit = 0;
1044 			goto out;
1045 		}
1046 
1047 		m->m_data += max_linkhdr;
1048 		m->m_len = hdrlen;
1049 
1050 		/*
1051 		 * Start the m_copy functions from the closest mbuf
1052 		 * to the offset in the socket buffer chain.
1053 		 */
1054 		mb = sbsndptr(&so->so_snd, off, len, &moff);
1055 
1056 		if (len <= MHLEN - hdrlen - max_linkhdr) {
1057 			m_copydata(mb, moff, len,
1058 			    mtod(m, caddr_t) + hdrlen);
1059 			m->m_len += len;
1060 		} else {
1061 			m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1062 			if (m->m_next == NULL) {
1063 				SOCKBUF_UNLOCK(&so->so_snd);
1064 				(void) m_free(m);
1065 				error = ENOBUFS;
1066 				sack_rxmit = 0;
1067 				goto out;
1068 			}
1069 		}
1070 
1071 		/*
1072 		 * If we're sending everything we've got, set PUSH.
1073 		 * (This will keep happy those implementations which only
1074 		 * give data to the user when a buffer fills or
1075 		 * a PUSH comes in.)
1076 		 */
1077 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1078 		    !(flags & TH_SYN))
1079 			flags |= TH_PUSH;
1080 		SOCKBUF_UNLOCK(&so->so_snd);
1081 	} else {
1082 		SOCKBUF_UNLOCK(&so->so_snd);
1083 		if (tp->t_flags & TF_ACKNOW)
1084 			TCPSTAT_INC(tcps_sndacks);
1085 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1086 			TCPSTAT_INC(tcps_sndctrl);
1087 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1088 			TCPSTAT_INC(tcps_sndurg);
1089 		else
1090 			TCPSTAT_INC(tcps_sndwinup);
1091 
1092 		m = m_gethdr(M_NOWAIT, MT_DATA);
1093 		if (m == NULL) {
1094 			error = ENOBUFS;
1095 			sack_rxmit = 0;
1096 			goto out;
1097 		}
1098 #ifdef INET6
1099 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1100 		    MHLEN >= hdrlen) {
1101 			M_ALIGN(m, hdrlen);
1102 		} else
1103 #endif
1104 		m->m_data += max_linkhdr;
1105 		m->m_len = hdrlen;
1106 	}
1107 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1108 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1109 #ifdef MAC
1110 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1111 #endif
1112 #ifdef INET6
1113 	if (isipv6) {
1114 		ip6 = mtod(m, struct ip6_hdr *);
1115 		th = (struct tcphdr *)(ip6 + 1);
1116 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1117 	} else
1118 #endif /* INET6 */
1119 	{
1120 		ip = mtod(m, struct ip *);
1121 		ipov = (struct ipovly *)ip;
1122 		th = (struct tcphdr *)(ip + 1);
1123 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1124 	}
1125 
1126 	/*
1127 	 * Fill in fields, remembering maximum advertised
1128 	 * window for use in delaying messages about window sizes.
1129 	 * If resending a FIN, be sure not to use a new sequence number.
1130 	 */
1131 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1132 	    tp->snd_nxt == tp->snd_max)
1133 		tp->snd_nxt--;
1134 	/*
1135 	 * If we are starting a connection, send ECN setup
1136 	 * SYN packet. If we are on a retransmit, we may
1137 	 * resend those bits a number of times as per
1138 	 * RFC 3168.
1139 	 */
1140 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1141 		if (tp->t_rxtshift >= 1) {
1142 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1143 				flags |= TH_ECE|TH_CWR;
1144 		} else
1145 			flags |= TH_ECE|TH_CWR;
1146 	}
1147 
1148 	if (tp->t_state == TCPS_ESTABLISHED &&
1149 	    (tp->t_flags & TF_ECN_PERMIT)) {
1150 		/*
1151 		 * If the peer has ECN, mark data packets with
1152 		 * ECN capable transmission (ECT).
1153 		 * Ignore pure ack packets, retransmissions and window probes.
1154 		 */
1155 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1156 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1157 #ifdef INET6
1158 			if (isipv6)
1159 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1160 			else
1161 #endif
1162 				ip->ip_tos |= IPTOS_ECN_ECT0;
1163 			TCPSTAT_INC(tcps_ecn_ect0);
1164 		}
1165 
1166 		/*
1167 		 * Reply with proper ECN notifications.
1168 		 */
1169 		if (tp->t_flags & TF_ECN_SND_CWR) {
1170 			flags |= TH_CWR;
1171 			tp->t_flags &= ~TF_ECN_SND_CWR;
1172 		}
1173 		if (tp->t_flags & TF_ECN_SND_ECE)
1174 			flags |= TH_ECE;
1175 	}
1176 
1177 	/*
1178 	 * If we are doing retransmissions, then snd_nxt will
1179 	 * not reflect the first unsent octet.  For ACK only
1180 	 * packets, we do not want the sequence number of the
1181 	 * retransmitted packet, we want the sequence number
1182 	 * of the next unsent octet.  So, if there is no data
1183 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1184 	 * when filling in ti_seq.  But if we are in persist
1185 	 * state, snd_max might reflect one byte beyond the
1186 	 * right edge of the window, so use snd_nxt in that
1187 	 * case, since we know we aren't doing a retransmission.
1188 	 * (retransmit and persist are mutually exclusive...)
1189 	 */
1190 	if (sack_rxmit == 0) {
1191 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1192 		    tcp_timer_active(tp, TT_PERSIST))
1193 			th->th_seq = htonl(tp->snd_nxt);
1194 		else
1195 			th->th_seq = htonl(tp->snd_max);
1196 	} else {
1197 		th->th_seq = htonl(p->rxmit);
1198 		p->rxmit += len;
1199 		tp->sackhint.sack_bytes_rexmit += len;
1200 	}
1201 	th->th_ack = htonl(tp->rcv_nxt);
1202 	if (optlen) {
1203 		bcopy(opt, th + 1, optlen);
1204 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1205 	}
1206 	th->th_flags = flags;
1207 	/*
1208 	 * Calculate receive window.  Don't shrink window,
1209 	 * but avoid silly window syndrome.
1210 	 */
1211 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1212 	    recwin < tp->t_maxseg)
1213 		recwin = 0;
1214 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1215 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
1216 		recwin = (tp->rcv_adv - tp->rcv_nxt);
1217 
1218 	/*
1219 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1220 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1221 	 * case is handled in syncache.
1222 	 */
1223 	if (flags & TH_SYN)
1224 		th->th_win = htons((u_short)
1225 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1226 	else
1227 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1228 
1229 	/*
1230 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1231 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1232 	 * flag tells soreceive() to disable delayed acknowledgements when
1233 	 * draining the buffer.  This can occur if the receiver is attempting
1234 	 * to read more data than can be buffered prior to transmitting on
1235 	 * the connection.
1236 	 */
1237 	if (th->th_win == 0) {
1238 		tp->t_sndzerowin++;
1239 		tp->t_flags |= TF_RXWIN0SENT;
1240 	} else
1241 		tp->t_flags &= ~TF_RXWIN0SENT;
1242 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1243 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1244 		th->th_flags |= TH_URG;
1245 	} else
1246 		/*
1247 		 * If no urgent pointer to send, then we pull
1248 		 * the urgent pointer to the left edge of the send window
1249 		 * so that it doesn't drift into the send window on sequence
1250 		 * number wraparound.
1251 		 */
1252 		tp->snd_up = tp->snd_una;		/* drag it along */
1253 
1254 #ifdef TCP_SIGNATURE
1255 	if (to.to_flags & TOF_SIGNATURE) {
1256 		int sigoff = to.to_signature - opt;
1257 		tcp_signature_compute(m, 0, len, optlen,
1258 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1259 	}
1260 #endif
1261 
1262 	/*
1263 	 * Put TCP length in extended header, and then
1264 	 * checksum extended header and data.
1265 	 */
1266 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1267 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1268 #ifdef INET6
1269 	if (isipv6) {
1270 		/*
1271 		 * ip6_plen is not need to be filled now, and will be filled
1272 		 * in ip6_output.
1273 		 */
1274 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1275 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1276 		    optlen + len, IPPROTO_TCP, 0);
1277 	}
1278 #endif
1279 #if defined(INET6) && defined(INET)
1280 	else
1281 #endif
1282 #ifdef INET
1283 	{
1284 		m->m_pkthdr.csum_flags = CSUM_TCP;
1285 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1286 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1287 
1288 		/* IP version must be set here for ipv4/ipv6 checking later */
1289 		KASSERT(ip->ip_v == IPVERSION,
1290 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1291 	}
1292 #endif
1293 
1294 	/*
1295 	 * Enable TSO and specify the size of the segments.
1296 	 * The TCP pseudo header checksum is always provided.
1297 	 */
1298 	if (tso) {
1299 		KASSERT(len > tp->t_maxseg - optlen,
1300 		    ("%s: len <= tso_segsz", __func__));
1301 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1302 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1303 	}
1304 
1305 #ifdef IPSEC
1306 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1307 	    ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1308 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1309 #else
1310 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1311 	    ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1312 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1313 #endif
1314 
1315 #ifdef TCP_HHOOK
1316 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1317 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1318 #endif
1319 
1320 #ifdef TCPDEBUG
1321 	/*
1322 	 * Trace.
1323 	 */
1324 	if (so->so_options & SO_DEBUG) {
1325 		u_short save = 0;
1326 #ifdef INET6
1327 		if (!isipv6)
1328 #endif
1329 		{
1330 			save = ipov->ih_len;
1331 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1332 		}
1333 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1334 #ifdef INET6
1335 		if (!isipv6)
1336 #endif
1337 		ipov->ih_len = save;
1338 	}
1339 #endif /* TCPDEBUG */
1340 	TCP_PROBE3(debug__output, tp, th, mtod(m, const char *));
1341 
1342 	/*
1343 	 * Fill in IP length and desired time to live and
1344 	 * send to IP level.  There should be a better way
1345 	 * to handle ttl and tos; we could keep them in
1346 	 * the template, but need a way to checksum without them.
1347 	 */
1348 	/*
1349 	 * m->m_pkthdr.len should have been set before checksum calculation,
1350 	 * because in6_cksum() need it.
1351 	 */
1352 #ifdef INET6
1353 	if (isipv6) {
1354 		struct route_in6 ro;
1355 
1356 		bzero(&ro, sizeof(ro));
1357 		/*
1358 		 * we separately set hoplimit for every segment, since the
1359 		 * user might want to change the value via setsockopt.
1360 		 * Also, desired default hop limit might be changed via
1361 		 * Neighbor Discovery.
1362 		 */
1363 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1364 
1365 		/*
1366 		 * Set the packet size here for the benefit of DTrace probes.
1367 		 * ip6_output() will set it properly; it's supposed to include
1368 		 * the option header lengths as well.
1369 		 */
1370 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1371 
1372 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1373 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1374 		else
1375 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1376 
1377 		if (tp->t_state == TCPS_SYN_SENT)
1378 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1379 
1380 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1381 
1382 #ifdef TCPPCAP
1383 		/* Save packet, if requested. */
1384 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1385 #endif
1386 
1387 		/* TODO: IPv6 IP6TOS_ECT bit on */
1388 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1389 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1390 		    NULL, NULL, tp->t_inpcb);
1391 
1392 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1393 			mtu = ro.ro_rt->rt_mtu;
1394 		RO_RTFREE(&ro);
1395 	}
1396 #endif /* INET6 */
1397 #if defined(INET) && defined(INET6)
1398 	else
1399 #endif
1400 #ifdef INET
1401     {
1402 	ip->ip_len = htons(m->m_pkthdr.len);
1403 #ifdef INET6
1404 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1405 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1406 #endif /* INET6 */
1407 	/*
1408 	 * If we do path MTU discovery, then we set DF on every packet.
1409 	 * This might not be the best thing to do according to RFC3390
1410 	 * Section 2. However the tcp hostcache migitates the problem
1411 	 * so it affects only the first tcp connection with a host.
1412 	 *
1413 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1414 	 */
1415 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1416 		ip->ip_off |= htons(IP_DF);
1417 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1418 	} else {
1419 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1420 	}
1421 
1422 	if (tp->t_state == TCPS_SYN_SENT)
1423 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1424 
1425 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1426 
1427 #ifdef TCPPCAP
1428 	/* Save packet, if requested. */
1429 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1430 #endif
1431 
1432 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1433 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1434 	    tp->t_inpcb);
1435 
1436 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1437 		mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1438     }
1439 #endif /* INET */
1440 
1441 out:
1442 	/*
1443 	 * In transmit state, time the transmission and arrange for
1444 	 * the retransmit.  In persist state, just set snd_max.
1445 	 */
1446 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1447 	    !tcp_timer_active(tp, TT_PERSIST)) {
1448 		tcp_seq startseq = tp->snd_nxt;
1449 
1450 		/*
1451 		 * Advance snd_nxt over sequence space of this segment.
1452 		 */
1453 		if (flags & (TH_SYN|TH_FIN)) {
1454 			if (flags & TH_SYN)
1455 				tp->snd_nxt++;
1456 			if (flags & TH_FIN) {
1457 				tp->snd_nxt++;
1458 				tp->t_flags |= TF_SENTFIN;
1459 			}
1460 		}
1461 		if (sack_rxmit)
1462 			goto timer;
1463 		tp->snd_nxt += len;
1464 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1465 			tp->snd_max = tp->snd_nxt;
1466 			/*
1467 			 * Time this transmission if not a retransmission and
1468 			 * not currently timing anything.
1469 			 */
1470 			if (tp->t_rtttime == 0) {
1471 				tp->t_rtttime = ticks;
1472 				tp->t_rtseq = startseq;
1473 				TCPSTAT_INC(tcps_segstimed);
1474 			}
1475 		}
1476 
1477 		/*
1478 		 * Set retransmit timer if not currently set,
1479 		 * and not doing a pure ack or a keep-alive probe.
1480 		 * Initial value for retransmit timer is smoothed
1481 		 * round-trip time + 2 * round-trip time variance.
1482 		 * Initialize shift counter which is used for backoff
1483 		 * of retransmit time.
1484 		 */
1485 timer:
1486 		if (!tcp_timer_active(tp, TT_REXMT) &&
1487 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1488 		     (tp->snd_nxt != tp->snd_una))) {
1489 			if (tcp_timer_active(tp, TT_PERSIST)) {
1490 				tcp_timer_activate(tp, TT_PERSIST, 0);
1491 				tp->t_rxtshift = 0;
1492 			}
1493 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1494 		} else if (len == 0 && sbavail(&so->so_snd) &&
1495 		    !tcp_timer_active(tp, TT_REXMT) &&
1496 		    !tcp_timer_active(tp, TT_PERSIST)) {
1497 			/*
1498 			 * Avoid a situation where we do not set persist timer
1499 			 * after a zero window condition. For example:
1500 			 * 1) A -> B: packet with enough data to fill the window
1501 			 * 2) B -> A: ACK for #1 + new data (0 window
1502 			 *    advertisement)
1503 			 * 3) A -> B: ACK for #2, 0 len packet
1504 			 *
1505 			 * In this case, A will not activate the persist timer,
1506 			 * because it chose to send a packet. Unless tcp_output
1507 			 * is called for some other reason (delayed ack timer,
1508 			 * another input packet from B, socket syscall), A will
1509 			 * not send zero window probes.
1510 			 *
1511 			 * So, if you send a 0-length packet, but there is data
1512 			 * in the socket buffer, and neither the rexmt or
1513 			 * persist timer is already set, then activate the
1514 			 * persist timer.
1515 			 */
1516 			tp->t_rxtshift = 0;
1517 			tcp_setpersist(tp);
1518 		}
1519 	} else {
1520 		/*
1521 		 * Persist case, update snd_max but since we are in
1522 		 * persist mode (no window) we do not update snd_nxt.
1523 		 */
1524 		int xlen = len;
1525 		if (flags & TH_SYN)
1526 			++xlen;
1527 		if (flags & TH_FIN) {
1528 			++xlen;
1529 			tp->t_flags |= TF_SENTFIN;
1530 		}
1531 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1532 			tp->snd_max = tp->snd_nxt + xlen;
1533 	}
1534 
1535 	if (error) {
1536 
1537 		/*
1538 		 * We know that the packet was lost, so back out the
1539 		 * sequence number advance, if any.
1540 		 *
1541 		 * If the error is EPERM the packet got blocked by the
1542 		 * local firewall.  Normally we should terminate the
1543 		 * connection but the blocking may have been spurious
1544 		 * due to a firewall reconfiguration cycle.  So we treat
1545 		 * it like a packet loss and let the retransmit timer and
1546 		 * timeouts do their work over time.
1547 		 * XXX: It is a POLA question whether calling tcp_drop right
1548 		 * away would be the really correct behavior instead.
1549 		 */
1550 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1551 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1552 		    ((flags & TH_SYN) == 0) &&
1553 		    (error != EPERM)) {
1554 			if (sack_rxmit) {
1555 				p->rxmit -= len;
1556 				tp->sackhint.sack_bytes_rexmit -= len;
1557 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1558 				    ("sackhint bytes rtx >= 0"));
1559 			} else
1560 				tp->snd_nxt -= len;
1561 		}
1562 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1563 		switch (error) {
1564 		case EPERM:
1565 			tp->t_softerror = error;
1566 			return (error);
1567 		case ENOBUFS:
1568 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1569 			tp->snd_cwnd = tp->t_maxseg;
1570 			return (0);
1571 		case EMSGSIZE:
1572 			/*
1573 			 * For some reason the interface we used initially
1574 			 * to send segments changed to another or lowered
1575 			 * its MTU.
1576 			 * If TSO was active we either got an interface
1577 			 * without TSO capabilits or TSO was turned off.
1578 			 * If we obtained mtu from ip_output() then update
1579 			 * it and try again.
1580 			 */
1581 			if (tso)
1582 				tp->t_flags &= ~TF_TSO;
1583 			if (mtu != 0) {
1584 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1585 				goto again;
1586 			}
1587 			return (error);
1588 		case EHOSTDOWN:
1589 		case EHOSTUNREACH:
1590 		case ENETDOWN:
1591 		case ENETUNREACH:
1592 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1593 				tp->t_softerror = error;
1594 				return (0);
1595 			}
1596 			/* FALLTHROUGH */
1597 		default:
1598 			return (error);
1599 		}
1600 	}
1601 	TCPSTAT_INC(tcps_sndtotal);
1602 
1603 	/*
1604 	 * Data sent (as far as we can tell).
1605 	 * If this advertises a larger window than any other segment,
1606 	 * then remember the size of the advertised window.
1607 	 * Any pending ACK has now been sent.
1608 	 */
1609 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1610 		tp->rcv_adv = tp->rcv_nxt + recwin;
1611 	tp->last_ack_sent = tp->rcv_nxt;
1612 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1613 	if (tcp_timer_active(tp, TT_DELACK))
1614 		tcp_timer_activate(tp, TT_DELACK, 0);
1615 #if 0
1616 	/*
1617 	 * This completely breaks TCP if newreno is turned on.  What happens
1618 	 * is that if delayed-acks are turned on on the receiver, this code
1619 	 * on the transmitter effectively destroys the TCP window, forcing
1620 	 * it to four packets (1.5Kx4 = 6K window).
1621 	 */
1622 	if (sendalot && --maxburst)
1623 		goto again;
1624 #endif
1625 	if (sendalot)
1626 		goto again;
1627 	return (0);
1628 }
1629 
1630 void
1631 tcp_setpersist(struct tcpcb *tp)
1632 {
1633 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1634 	int tt;
1635 
1636 	tp->t_flags &= ~TF_PREVVALID;
1637 	if (tcp_timer_active(tp, TT_REXMT))
1638 		panic("tcp_setpersist: retransmit pending");
1639 	/*
1640 	 * Start/restart persistence timer.
1641 	 */
1642 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1643 		      tcp_persmin, tcp_persmax);
1644 	tcp_timer_activate(tp, TT_PERSIST, tt);
1645 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1646 		tp->t_rxtshift++;
1647 }
1648 
1649 /*
1650  * Insert TCP options according to the supplied parameters to the place
1651  * optp in a consistent way.  Can handle unaligned destinations.
1652  *
1653  * The order of the option processing is crucial for optimal packing and
1654  * alignment for the scarce option space.
1655  *
1656  * The optimal order for a SYN/SYN-ACK segment is:
1657  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1658  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1659  *
1660  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1661  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1662  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1663  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1664  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1665  */
1666 int
1667 tcp_addoptions(struct tcpopt *to, u_char *optp)
1668 {
1669 	u_int32_t mask, optlen = 0;
1670 
1671 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1672 		if ((to->to_flags & mask) != mask)
1673 			continue;
1674 		if (optlen == TCP_MAXOLEN)
1675 			break;
1676 		switch (to->to_flags & mask) {
1677 		case TOF_MSS:
1678 			while (optlen % 4) {
1679 				optlen += TCPOLEN_NOP;
1680 				*optp++ = TCPOPT_NOP;
1681 			}
1682 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1683 				continue;
1684 			optlen += TCPOLEN_MAXSEG;
1685 			*optp++ = TCPOPT_MAXSEG;
1686 			*optp++ = TCPOLEN_MAXSEG;
1687 			to->to_mss = htons(to->to_mss);
1688 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1689 			optp += sizeof(to->to_mss);
1690 			break;
1691 		case TOF_SCALE:
1692 			while (!optlen || optlen % 2 != 1) {
1693 				optlen += TCPOLEN_NOP;
1694 				*optp++ = TCPOPT_NOP;
1695 			}
1696 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1697 				continue;
1698 			optlen += TCPOLEN_WINDOW;
1699 			*optp++ = TCPOPT_WINDOW;
1700 			*optp++ = TCPOLEN_WINDOW;
1701 			*optp++ = to->to_wscale;
1702 			break;
1703 		case TOF_SACKPERM:
1704 			while (optlen % 2) {
1705 				optlen += TCPOLEN_NOP;
1706 				*optp++ = TCPOPT_NOP;
1707 			}
1708 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1709 				continue;
1710 			optlen += TCPOLEN_SACK_PERMITTED;
1711 			*optp++ = TCPOPT_SACK_PERMITTED;
1712 			*optp++ = TCPOLEN_SACK_PERMITTED;
1713 			break;
1714 		case TOF_TS:
1715 			while (!optlen || optlen % 4 != 2) {
1716 				optlen += TCPOLEN_NOP;
1717 				*optp++ = TCPOPT_NOP;
1718 			}
1719 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1720 				continue;
1721 			optlen += TCPOLEN_TIMESTAMP;
1722 			*optp++ = TCPOPT_TIMESTAMP;
1723 			*optp++ = TCPOLEN_TIMESTAMP;
1724 			to->to_tsval = htonl(to->to_tsval);
1725 			to->to_tsecr = htonl(to->to_tsecr);
1726 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1727 			optp += sizeof(to->to_tsval);
1728 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1729 			optp += sizeof(to->to_tsecr);
1730 			break;
1731 #ifdef TCP_SIGNATURE
1732 		case TOF_SIGNATURE:
1733 			{
1734 			int siglen = TCPOLEN_SIGNATURE - 2;
1735 
1736 			while (!optlen || optlen % 4 != 2) {
1737 				optlen += TCPOLEN_NOP;
1738 				*optp++ = TCPOPT_NOP;
1739 			}
1740 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1741 				continue;
1742 			optlen += TCPOLEN_SIGNATURE;
1743 			*optp++ = TCPOPT_SIGNATURE;
1744 			*optp++ = TCPOLEN_SIGNATURE;
1745 			to->to_signature = optp;
1746 			while (siglen--)
1747 				 *optp++ = 0;
1748 			break;
1749 			}
1750 #endif
1751 		case TOF_SACK:
1752 			{
1753 			int sackblks = 0;
1754 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1755 			tcp_seq sack_seq;
1756 
1757 			while (!optlen || optlen % 4 != 2) {
1758 				optlen += TCPOLEN_NOP;
1759 				*optp++ = TCPOPT_NOP;
1760 			}
1761 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1762 				continue;
1763 			optlen += TCPOLEN_SACKHDR;
1764 			*optp++ = TCPOPT_SACK;
1765 			sackblks = min(to->to_nsacks,
1766 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1767 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1768 			while (sackblks--) {
1769 				sack_seq = htonl(sack->start);
1770 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1771 				optp += sizeof(sack_seq);
1772 				sack_seq = htonl(sack->end);
1773 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1774 				optp += sizeof(sack_seq);
1775 				optlen += TCPOLEN_SACK;
1776 				sack++;
1777 			}
1778 			TCPSTAT_INC(tcps_sack_send_blocks);
1779 			break;
1780 			}
1781 #ifdef TCP_RFC7413
1782 		case TOF_FASTOPEN:
1783 			{
1784 			int total_len;
1785 
1786 			/* XXX is there any point to aligning this option? */
1787 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1788 			if (TCP_MAXOLEN - optlen < total_len)
1789 				continue;
1790 			*optp++ = TCPOPT_FAST_OPEN;
1791 			*optp++ = total_len;
1792 			if (to->to_tfo_len > 0) {
1793 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1794 				optp += to->to_tfo_len;
1795 			}
1796 			optlen += total_len;
1797 			break;
1798 			}
1799 #endif
1800 		default:
1801 			panic("%s: unknown TCP option type", __func__);
1802 			break;
1803 		}
1804 	}
1805 
1806 	/* Terminate and pad TCP options to a 4 byte boundary. */
1807 	if (optlen % 4) {
1808 		optlen += TCPOLEN_EOL;
1809 		*optp++ = TCPOPT_EOL;
1810 	}
1811 	/*
1812 	 * According to RFC 793 (STD0007):
1813 	 *   "The content of the header beyond the End-of-Option option
1814 	 *    must be header padding (i.e., zero)."
1815 	 *   and later: "The padding is composed of zeros."
1816 	 */
1817 	while (optlen % 4) {
1818 		optlen += TCPOLEN_PAD;
1819 		*optp++ = TCPOPT_PAD;
1820 	}
1821 
1822 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1823 	return (optlen);
1824 }
1825