xref: /freebsd/sys/netinet/tcp_output.c (revision af6a5351a1fdb1130f18be6c782c4d48916eb971)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. 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 #include <netipsec/ipsec_support.h>
94 
95 #include <machine/in_cksum.h>
96 
97 #include <security/mac/mac_framework.h>
98 
99 VNET_DEFINE(int, path_mtu_discovery) = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
101 	&VNET_NAME(path_mtu_discovery), 1,
102 	"Enable Path MTU Discovery");
103 
104 VNET_DEFINE(int, tcp_do_tso) = 1;
105 #define	V_tcp_do_tso		VNET(tcp_do_tso)
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
107 	&VNET_NAME(tcp_do_tso), 0,
108 	"Enable TCP Segmentation Offload");
109 
110 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
111 #define	V_tcp_sendspace	VNET(tcp_sendspace)
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
113 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114 
115 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
116 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
118 	&VNET_NAME(tcp_do_autosndbuf), 0,
119 	"Enable automatic send buffer sizing");
120 
121 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
122 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
124 	&VNET_NAME(tcp_autosndbuf_inc), 0,
125 	"Incrementor step size of automatic send buffer");
126 
127 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
128 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
130 	&VNET_NAME(tcp_autosndbuf_max), 0,
131 	"Max size of automatic send buffer");
132 
133 /*
134  * Make sure that either retransmit or persist timer is set for SYN, FIN and
135  * non-ACK.
136  */
137 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
138 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
139 	    tcp_timer_active((tp), TT_REXMT) ||				\
140 	    tcp_timer_active((tp), TT_PERSIST),				\
141 	    ("neither rexmt nor persist timer is set"))
142 
143 #ifdef TCP_HHOOK
144 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
145 			    struct tcphdr *th, struct tcpopt *to,
146 			    uint32_t len, int tso);
147 #endif
148 static void inline	cc_after_idle(struct tcpcb *tp);
149 
150 #ifdef TCP_HHOOK
151 /*
152  * Wrapper for the TCP established output helper hook.
153  */
154 static void inline
155 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
156     struct tcpopt *to, uint32_t len, int tso)
157 {
158 	struct tcp_hhook_data hhook_data;
159 
160 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
161 		hhook_data.tp = tp;
162 		hhook_data.th = th;
163 		hhook_data.to = to;
164 		hhook_data.len = len;
165 		hhook_data.tso = tso;
166 
167 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
168 		    tp->osd);
169 	}
170 }
171 #endif
172 
173 /*
174  * CC wrapper hook functions
175  */
176 static void inline
177 cc_after_idle(struct tcpcb *tp)
178 {
179 	INP_WLOCK_ASSERT(tp->t_inpcb);
180 
181 	if (CC_ALGO(tp)->after_idle != NULL)
182 		CC_ALGO(tp)->after_idle(tp->ccv);
183 }
184 
185 /*
186  * Tcp output routine: figure out what should be sent and send it.
187  */
188 int
189 tcp_output(struct tcpcb *tp)
190 {
191 	struct socket *so = tp->t_inpcb->inp_socket;
192 	int32_t len;
193 	uint32_t recwin, sendwin;
194 	int off, flags, error = 0;	/* Keep compiler happy */
195 	struct mbuf *m;
196 	struct ip *ip = NULL;
197 	struct ipovly *ipov = NULL;
198 	struct tcphdr *th;
199 	u_char opt[TCP_MAXOLEN];
200 	unsigned ipoptlen, optlen, hdrlen;
201 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
202 	unsigned ipsec_optlen = 0;
203 #endif
204 	int idle, sendalot;
205 	int sack_rxmit, sack_bytes_rxmt;
206 	struct sackhole *p;
207 	int tso, mtu;
208 	struct tcpopt to;
209 #if 0
210 	int maxburst = TCP_MAXBURST;
211 #endif
212 #ifdef INET6
213 	struct ip6_hdr *ip6 = NULL;
214 	int isipv6;
215 
216 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
217 #endif
218 
219 	INP_WLOCK_ASSERT(tp->t_inpcb);
220 
221 #ifdef TCP_OFFLOAD
222 	if (tp->t_flags & TF_TOE)
223 		return (tcp_offload_output(tp));
224 #endif
225 
226 #ifdef TCP_RFC7413
227 	/*
228 	 * For TFO connections in SYN_RECEIVED, only allow the initial
229 	 * SYN|ACK and those sent by the retransmit timer.
230 	 */
231 	if (IS_FASTOPEN(tp->t_flags) &&
232 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
233 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
234 	    (tp->snd_nxt != tp->snd_una))          /* not a retransmit */
235 		return (0);
236 #endif
237 	/*
238 	 * Determine length of data that should be transmitted,
239 	 * and flags that will be used.
240 	 * If there is some data or critical controls (SYN, RST)
241 	 * to send, then transmit; otherwise, investigate further.
242 	 */
243 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
244 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
245 		cc_after_idle(tp);
246 	tp->t_flags &= ~TF_LASTIDLE;
247 	if (idle) {
248 		if (tp->t_flags & TF_MORETOCOME) {
249 			tp->t_flags |= TF_LASTIDLE;
250 			idle = 0;
251 		}
252 	}
253 again:
254 	/*
255 	 * If we've recently taken a timeout, snd_max will be greater than
256 	 * snd_nxt.  There may be SACK information that allows us to avoid
257 	 * resending already delivered data.  Adjust snd_nxt accordingly.
258 	 */
259 	if ((tp->t_flags & TF_SACK_PERMIT) &&
260 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
261 		tcp_sack_adjust(tp);
262 	sendalot = 0;
263 	tso = 0;
264 	mtu = 0;
265 	off = tp->snd_nxt - tp->snd_una;
266 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
267 
268 	flags = tcp_outflags[tp->t_state];
269 	/*
270 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
271 	 * to send out new data (when sendalot is 1), bypass this function.
272 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
273 	 * we're replacing a (future) new transmission with a retransmission
274 	 * now, and we previously incremented snd_cwnd in tcp_input().
275 	 */
276 	/*
277 	 * Still in sack recovery , reset rxmit flag to zero.
278 	 */
279 	sack_rxmit = 0;
280 	sack_bytes_rxmt = 0;
281 	len = 0;
282 	p = NULL;
283 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
284 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
285 		uint32_t cwin;
286 
287 		cwin =
288 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
289 		/* Do not retransmit SACK segments beyond snd_recover */
290 		if (SEQ_GT(p->end, tp->snd_recover)) {
291 			/*
292 			 * (At least) part of sack hole extends beyond
293 			 * snd_recover. Check to see if we can rexmit data
294 			 * for this hole.
295 			 */
296 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
297 				/*
298 				 * Can't rexmit any more data for this hole.
299 				 * That data will be rexmitted in the next
300 				 * sack recovery episode, when snd_recover
301 				 * moves past p->rxmit.
302 				 */
303 				p = NULL;
304 				goto after_sack_rexmit;
305 			} else
306 				/* Can rexmit part of the current hole */
307 				len = ((int32_t)ulmin(cwin,
308 						   tp->snd_recover - p->rxmit));
309 		} else
310 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
311 		off = p->rxmit - tp->snd_una;
312 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
313 		    __func__, off));
314 		if (len > 0) {
315 			sack_rxmit = 1;
316 			sendalot = 1;
317 			TCPSTAT_INC(tcps_sack_rexmits);
318 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
319 			    min(len, tp->t_maxseg));
320 		}
321 	}
322 after_sack_rexmit:
323 	/*
324 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
325 	 * state flags.
326 	 */
327 	if (tp->t_flags & TF_NEEDFIN)
328 		flags |= TH_FIN;
329 	if (tp->t_flags & TF_NEEDSYN)
330 		flags |= TH_SYN;
331 
332 	SOCKBUF_LOCK(&so->so_snd);
333 	/*
334 	 * If in persist timeout with window of 0, send 1 byte.
335 	 * Otherwise, if window is small but nonzero
336 	 * and timer expired, we will send what we can
337 	 * and go to transmit state.
338 	 */
339 	if (tp->t_flags & TF_FORCEDATA) {
340 		if (sendwin == 0) {
341 			/*
342 			 * If we still have some data to send, then
343 			 * clear the FIN bit.  Usually this would
344 			 * happen below when it realizes that we
345 			 * aren't sending all the data.  However,
346 			 * if we have exactly 1 byte of unsent data,
347 			 * then it won't clear the FIN bit below,
348 			 * and if we are in persist state, we wind
349 			 * up sending the packet without recording
350 			 * that we sent the FIN bit.
351 			 *
352 			 * We can't just blindly clear the FIN bit,
353 			 * because if we don't have any more data
354 			 * to send then the probe will be the FIN
355 			 * itself.
356 			 */
357 			if (off < sbused(&so->so_snd))
358 				flags &= ~TH_FIN;
359 			sendwin = 1;
360 		} else {
361 			tcp_timer_activate(tp, TT_PERSIST, 0);
362 			tp->t_rxtshift = 0;
363 		}
364 	}
365 
366 	/*
367 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
368 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
369 	 * a negative length.  This can also occur when TCP opens up
370 	 * its congestion window while receiving additional duplicate
371 	 * acks after fast-retransmit because TCP will reset snd_nxt
372 	 * to snd_max after the fast-retransmit.
373 	 *
374 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
375 	 * be set to snd_una, the offset will be 0, and the length may
376 	 * wind up 0.
377 	 *
378 	 * If sack_rxmit is true we are retransmitting from the scoreboard
379 	 * in which case len is already set.
380 	 */
381 	if (sack_rxmit == 0) {
382 		if (sack_bytes_rxmt == 0)
383 			len = ((int32_t)ulmin(sbavail(&so->so_snd), sendwin) -
384 			    off);
385 		else {
386 			int32_t cwin;
387 
388                         /*
389 			 * We are inside of a SACK recovery episode and are
390 			 * sending new data, having retransmitted all the
391 			 * data possible in the scoreboard.
392 			 */
393 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
394 			    off);
395 			/*
396 			 * Don't remove this (len > 0) check !
397 			 * We explicitly check for len > 0 here (although it
398 			 * isn't really necessary), to work around a gcc
399 			 * optimization issue - to force gcc to compute
400 			 * len above. Without this check, the computation
401 			 * of len is bungled by the optimizer.
402 			 */
403 			if (len > 0) {
404 				cwin = tp->snd_cwnd -
405 					(tp->snd_nxt - tp->sack_newdata) -
406 					sack_bytes_rxmt;
407 				if (cwin < 0)
408 					cwin = 0;
409 				len = imin(len, cwin);
410 			}
411 		}
412 	}
413 
414 	/*
415 	 * Lop off SYN bit if it has already been sent.  However, if this
416 	 * is SYN-SENT state and if segment contains data and if we don't
417 	 * know that foreign host supports TAO, suppress sending segment.
418 	 */
419 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
420 		if (tp->t_state != TCPS_SYN_RECEIVED)
421 			flags &= ~TH_SYN;
422 #ifdef TCP_RFC7413
423 		/*
424 		 * When sending additional segments following a TFO SYN|ACK,
425 		 * do not include the SYN bit.
426 		 */
427 		if (IS_FASTOPEN(tp->t_flags) &&
428 		    (tp->t_state == TCPS_SYN_RECEIVED))
429 			flags &= ~TH_SYN;
430 #endif
431 		off--, len++;
432 	}
433 
434 	/*
435 	 * Be careful not to send data and/or FIN on SYN segments.
436 	 * This measure is needed to prevent interoperability problems
437 	 * with not fully conformant TCP implementations.
438 	 */
439 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
440 		len = 0;
441 		flags &= ~TH_FIN;
442 	}
443 
444 #ifdef TCP_RFC7413
445 	/*
446 	 * When retransmitting SYN|ACK on a passively-created TFO socket,
447 	 * don't include data, as the presence of data may have caused the
448 	 * original SYN|ACK to have been dropped by a middlebox.
449 	 */
450 	if (IS_FASTOPEN(tp->t_flags) &&
451 	    (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
452 	     (flags & TH_RST)))
453 		len = 0;
454 #endif
455 	if (len <= 0) {
456 		/*
457 		 * If FIN has been sent but not acked,
458 		 * but we haven't been called to retransmit,
459 		 * len will be < 0.  Otherwise, window shrank
460 		 * after we sent into it.  If window shrank to 0,
461 		 * cancel pending retransmit, pull snd_nxt back
462 		 * to (closed) window, and set the persist timer
463 		 * if it isn't already going.  If the window didn't
464 		 * close completely, just wait for an ACK.
465 		 *
466 		 * We also do a general check here to ensure that
467 		 * we will set the persist timer when we have data
468 		 * to send, but a 0-byte window. This makes sure
469 		 * the persist timer is set even if the packet
470 		 * hits one of the "goto send" lines below.
471 		 */
472 		len = 0;
473 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
474 			(off < (int) sbavail(&so->so_snd))) {
475 			tcp_timer_activate(tp, TT_REXMT, 0);
476 			tp->t_rxtshift = 0;
477 			tp->snd_nxt = tp->snd_una;
478 			if (!tcp_timer_active(tp, TT_PERSIST))
479 				tcp_setpersist(tp);
480 		}
481 	}
482 
483 	/* len will be >= 0 after this point. */
484 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
485 
486 	/*
487 	 * Automatic sizing of send socket buffer.  Often the send buffer
488 	 * size is not optimally adjusted to the actual network conditions
489 	 * at hand (delay bandwidth product).  Setting the buffer size too
490 	 * small limits throughput on links with high bandwidth and high
491 	 * delay (eg. trans-continental/oceanic links).  Setting the
492 	 * buffer size too big consumes too much real kernel memory,
493 	 * especially with many connections on busy servers.
494 	 *
495 	 * The criteria to step up the send buffer one notch are:
496 	 *  1. receive window of remote host is larger than send buffer
497 	 *     (with a fudge factor of 5/4th);
498 	 *  2. send buffer is filled to 7/8th with data (so we actually
499 	 *     have data to make use of it);
500 	 *  3. send buffer fill has not hit maximal automatic size;
501 	 *  4. our send window (slow start and cogestion controlled) is
502 	 *     larger than sent but unacknowledged data in send buffer.
503 	 *
504 	 * The remote host receive window scaling factor may limit the
505 	 * growing of the send buffer before it reaches its allowed
506 	 * maximum.
507 	 *
508 	 * It scales directly with slow start or congestion window
509 	 * and does at most one step per received ACK.  This fast
510 	 * scaling has the drawback of growing the send buffer beyond
511 	 * what is strictly necessary to make full use of a given
512 	 * delay*bandwidth product.  However testing has shown this not
513 	 * to be much of an problem.  At worst we are trading wasting
514 	 * of available bandwidth (the non-use of it) for wasting some
515 	 * socket buffer memory.
516 	 *
517 	 * TODO: Shrink send buffer during idle periods together
518 	 * with congestion window.  Requires another timer.  Has to
519 	 * wait for upcoming tcp timer rewrite.
520 	 *
521 	 * XXXGL: should there be used sbused() or sbavail()?
522 	 */
523 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
524 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
525 		    sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) &&
526 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
527 		    sendwin >= (sbused(&so->so_snd) -
528 		    (tp->snd_nxt - tp->snd_una))) {
529 			if (!sbreserve_locked(&so->so_snd,
530 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
531 			     V_tcp_autosndbuf_max), so, curthread))
532 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
533 		}
534 	}
535 
536 	/*
537 	 * Decide if we can use TCP Segmentation Offloading (if supported by
538 	 * hardware).
539 	 *
540 	 * TSO may only be used if we are in a pure bulk sending state.  The
541 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
542 	 * IP options prevent using TSO.  With TSO the TCP header is the same
543 	 * (except for the sequence number) for all generated packets.  This
544 	 * makes it impossible to transmit any options which vary per generated
545 	 * segment or packet.
546 	 *
547 	 * IPv4 handling has a clear separation of ip options and ip header
548 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
549 	 * the right thing below to provide length of just ip options and thus
550 	 * checking for ipoptlen is enough to decide if ip options are present.
551 	 */
552 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
553 	/*
554 	 * Pre-calculate here as we save another lookup into the darknesses
555 	 * of IPsec that way and can actually decide if TSO is ok.
556 	 */
557 #ifdef INET6
558 	if (isipv6 && IPSEC_ENABLED(ipv6))
559 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
560 #ifdef INET
561 	else
562 #endif
563 #endif /* INET6 */
564 #ifdef INET
565 	if (IPSEC_ENABLED(ipv4))
566 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
567 #endif /* INET */
568 #endif /* IPSEC */
569 #ifdef INET6
570 	if (isipv6)
571 		ipoptlen = ip6_optlen(tp->t_inpcb);
572 	else
573 #endif
574 	if (tp->t_inpcb->inp_options)
575 		ipoptlen = tp->t_inpcb->inp_options->m_len -
576 				offsetof(struct ipoption, ipopt_list);
577 	else
578 		ipoptlen = 0;
579 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
580 	ipoptlen += ipsec_optlen;
581 #endif
582 
583 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
584 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
585 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
586 	    ipoptlen == 0)
587 		tso = 1;
588 
589 	if (sack_rxmit) {
590 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
591 			flags &= ~TH_FIN;
592 	} else {
593 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
594 		    sbused(&so->so_snd)))
595 			flags &= ~TH_FIN;
596 	}
597 
598 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
599 	    (long)TCP_MAXWIN << tp->rcv_scale);
600 
601 	/*
602 	 * Sender silly window avoidance.   We transmit under the following
603 	 * conditions when len is non-zero:
604 	 *
605 	 *	- We have a full segment (or more with TSO)
606 	 *	- This is the last buffer in a write()/send() and we are
607 	 *	  either idle or running NODELAY
608 	 *	- we've timed out (e.g. persist timer)
609 	 *	- we have more then 1/2 the maximum send window's worth of
610 	 *	  data (receiver may be limited the window size)
611 	 *	- we need to retransmit
612 	 */
613 	if (len) {
614 		if (len >= tp->t_maxseg)
615 			goto send;
616 		/*
617 		 * NOTE! on localhost connections an 'ack' from the remote
618 		 * end may occur synchronously with the output and cause
619 		 * us to flush a buffer queued with moretocome.  XXX
620 		 *
621 		 * note: the len + off check is almost certainly unnecessary.
622 		 */
623 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
624 		    (idle || (tp->t_flags & TF_NODELAY)) &&
625 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
626 		    (tp->t_flags & TF_NOPUSH) == 0) {
627 			goto send;
628 		}
629 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
630 			goto send;
631 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
632 			goto send;
633 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
634 			goto send;
635 		if (sack_rxmit)
636 			goto send;
637 	}
638 
639 	/*
640 	 * Sending of standalone window updates.
641 	 *
642 	 * Window updates are important when we close our window due to a
643 	 * full socket buffer and are opening it again after the application
644 	 * reads data from it.  Once the window has opened again and the
645 	 * remote end starts to send again the ACK clock takes over and
646 	 * provides the most current window information.
647 	 *
648 	 * We must avoid the silly window syndrome whereas every read
649 	 * from the receive buffer, no matter how small, causes a window
650 	 * update to be sent.  We also should avoid sending a flurry of
651 	 * window updates when the socket buffer had queued a lot of data
652 	 * and the application is doing small reads.
653 	 *
654 	 * Prevent a flurry of pointless window updates by only sending
655 	 * an update when we can increase the advertized window by more
656 	 * than 1/4th of the socket buffer capacity.  When the buffer is
657 	 * getting full or is very small be more aggressive and send an
658 	 * update whenever we can increase by two mss sized segments.
659 	 * In all other situations the ACK's to new incoming data will
660 	 * carry further window increases.
661 	 *
662 	 * Don't send an independent window update if a delayed
663 	 * ACK is pending (it will get piggy-backed on it) or the
664 	 * remote side already has done a half-close and won't send
665 	 * more data.  Skip this if the connection is in T/TCP
666 	 * half-open state.
667 	 */
668 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
669 	    !(tp->t_flags & TF_DELACK) &&
670 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
671 		/*
672 		 * "adv" is the amount we could increase the window,
673 		 * taking into account that we are limited by
674 		 * TCP_MAXWIN << tp->rcv_scale.
675 		 */
676 		int32_t adv;
677 		int oldwin;
678 
679 		adv = recwin;
680 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
681 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
682 			adv -= oldwin;
683 		} else
684 			oldwin = 0;
685 
686 		/*
687 		 * If the new window size ends up being the same as or less
688 		 * than the old size when it is scaled, then don't force
689 		 * a window update.
690 		 */
691 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
692 			goto dontupdate;
693 
694 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
695 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
696 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
697 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
698 			goto send;
699 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
700 			goto send;
701 	}
702 dontupdate:
703 
704 	/*
705 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
706 	 * is also a catch-all for the retransmit timer timeout case.
707 	 */
708 	if (tp->t_flags & TF_ACKNOW)
709 		goto send;
710 	if ((flags & TH_RST) ||
711 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
712 		goto send;
713 	if (SEQ_GT(tp->snd_up, tp->snd_una))
714 		goto send;
715 	/*
716 	 * If our state indicates that FIN should be sent
717 	 * and we have not yet done so, then we need to send.
718 	 */
719 	if (flags & TH_FIN &&
720 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
721 		goto send;
722 	/*
723 	 * In SACK, it is possible for tcp_output to fail to send a segment
724 	 * after the retransmission timer has been turned off.  Make sure
725 	 * that the retransmission timer is set.
726 	 */
727 	if ((tp->t_flags & TF_SACK_PERMIT) &&
728 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
729 	    !tcp_timer_active(tp, TT_REXMT) &&
730 	    !tcp_timer_active(tp, TT_PERSIST)) {
731 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
732 		goto just_return;
733 	}
734 	/*
735 	 * TCP window updates are not reliable, rather a polling protocol
736 	 * using ``persist'' packets is used to insure receipt of window
737 	 * updates.  The three ``states'' for the output side are:
738 	 *	idle			not doing retransmits or persists
739 	 *	persisting		to move a small or zero window
740 	 *	(re)transmitting	and thereby not persisting
741 	 *
742 	 * tcp_timer_active(tp, TT_PERSIST)
743 	 *	is true when we are in persist state.
744 	 * (tp->t_flags & TF_FORCEDATA)
745 	 *	is set when we are called to send a persist packet.
746 	 * tcp_timer_active(tp, TT_REXMT)
747 	 *	is set when we are retransmitting
748 	 * The output side is idle when both timers are zero.
749 	 *
750 	 * If send window is too small, there is data to transmit, and no
751 	 * retransmit or persist is pending, then go to persist state.
752 	 * If nothing happens soon, send when timer expires:
753 	 * if window is nonzero, transmit what we can,
754 	 * otherwise force out a byte.
755 	 */
756 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
757 	    !tcp_timer_active(tp, TT_PERSIST)) {
758 		tp->t_rxtshift = 0;
759 		tcp_setpersist(tp);
760 	}
761 
762 	/*
763 	 * No reason to send a segment, just return.
764 	 */
765 just_return:
766 	SOCKBUF_UNLOCK(&so->so_snd);
767 	return (0);
768 
769 send:
770 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
771 	if (len > 0) {
772 		if (len >= tp->t_maxseg)
773 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
774 		else
775 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
776 	}
777 	/*
778 	 * Before ESTABLISHED, force sending of initial options
779 	 * unless TCP set not to do any options.
780 	 * NOTE: we assume that the IP/TCP header plus TCP options
781 	 * always fit in a single mbuf, leaving room for a maximum
782 	 * link header, i.e.
783 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
784 	 */
785 	optlen = 0;
786 #ifdef INET6
787 	if (isipv6)
788 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
789 	else
790 #endif
791 		hdrlen = sizeof (struct tcpiphdr);
792 
793 	/*
794 	 * Compute options for segment.
795 	 * We only have to care about SYN and established connection
796 	 * segments.  Options for SYN-ACK segments are handled in TCP
797 	 * syncache.
798 	 */
799 	to.to_flags = 0;
800 	if ((tp->t_flags & TF_NOOPT) == 0) {
801 		/* Maximum segment size. */
802 		if (flags & TH_SYN) {
803 			tp->snd_nxt = tp->iss;
804 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
805 			to.to_flags |= TOF_MSS;
806 #ifdef TCP_RFC7413
807 			/*
808 			 * Only include the TFO option on the first
809 			 * transmission of the SYN|ACK on a
810 			 * passively-created TFO socket, as the presence of
811 			 * the TFO option may have caused the original
812 			 * SYN|ACK to have been dropped by a middlebox.
813 			 */
814 			if (IS_FASTOPEN(tp->t_flags) &&
815 			    (tp->t_state == TCPS_SYN_RECEIVED) &&
816 			    (tp->t_rxtshift == 0)) {
817 				to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
818 				to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
819 				to.to_flags |= TOF_FASTOPEN;
820 			}
821 #endif
822 		}
823 		/* Window scaling. */
824 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
825 			to.to_wscale = tp->request_r_scale;
826 			to.to_flags |= TOF_SCALE;
827 		}
828 		/* Timestamps. */
829 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
830 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
831 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
832 			to.to_tsecr = tp->ts_recent;
833 			to.to_flags |= TOF_TS;
834 			/* Set receive buffer autosizing timestamp. */
835 			if (tp->rfbuf_ts == 0 &&
836 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
837 				tp->rfbuf_ts = tcp_ts_getticks();
838 		}
839 		/* Selective ACK's. */
840 		if (tp->t_flags & TF_SACK_PERMIT) {
841 			if (flags & TH_SYN)
842 				to.to_flags |= TOF_SACKPERM;
843 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
844 			    (tp->t_flags & TF_SACK_PERMIT) &&
845 			    tp->rcv_numsacks > 0) {
846 				to.to_flags |= TOF_SACK;
847 				to.to_nsacks = tp->rcv_numsacks;
848 				to.to_sacks = (u_char *)tp->sackblks;
849 			}
850 		}
851 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
852 		/* TCP-MD5 (RFC2385). */
853 		/*
854 		 * Check that TCP_MD5SIG is enabled in tcpcb to
855 		 * account the size needed to set this TCP option.
856 		 */
857 		if (tp->t_flags & TF_SIGNATURE)
858 			to.to_flags |= TOF_SIGNATURE;
859 #endif /* TCP_SIGNATURE */
860 
861 		/* Processing the options. */
862 		hdrlen += optlen = tcp_addoptions(&to, opt);
863 	}
864 
865 	/*
866 	 * Adjust data length if insertion of options will
867 	 * bump the packet length beyond the t_maxseg length.
868 	 * Clear the FIN bit because we cut off the tail of
869 	 * the segment.
870 	 */
871 	if (len + optlen + ipoptlen > tp->t_maxseg) {
872 		flags &= ~TH_FIN;
873 
874 		if (tso) {
875 			u_int if_hw_tsomax;
876 			u_int if_hw_tsomaxsegcount;
877 			u_int if_hw_tsomaxsegsize;
878 			struct mbuf *mb;
879 			u_int moff;
880 			int max_len;
881 
882 			/* extract TSO information */
883 			if_hw_tsomax = tp->t_tsomax;
884 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
885 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
886 
887 			/*
888 			 * Limit a TSO burst to prevent it from
889 			 * overflowing or exceeding the maximum length
890 			 * allowed by the network interface:
891 			 */
892 			KASSERT(ipoptlen == 0,
893 			    ("%s: TSO can't do IP options", __func__));
894 
895 			/*
896 			 * Check if we should limit by maximum payload
897 			 * length:
898 			 */
899 			if (if_hw_tsomax != 0) {
900 				/* compute maximum TSO length */
901 				max_len = (if_hw_tsomax - hdrlen -
902 				    max_linkhdr);
903 				if (max_len <= 0) {
904 					len = 0;
905 				} else if (len > max_len) {
906 					sendalot = 1;
907 					len = max_len;
908 				}
909 			}
910 
911 			/*
912 			 * Check if we should limit by maximum segment
913 			 * size and count:
914 			 */
915 			if (if_hw_tsomaxsegcount != 0 &&
916 			    if_hw_tsomaxsegsize != 0) {
917 				/*
918 				 * Subtract one segment for the LINK
919 				 * and TCP/IP headers mbuf that will
920 				 * be prepended to this mbuf chain
921 				 * after the code in this section
922 				 * limits the number of mbufs in the
923 				 * chain to if_hw_tsomaxsegcount.
924 				 */
925 				if_hw_tsomaxsegcount -= 1;
926 				max_len = 0;
927 				mb = sbsndmbuf(&so->so_snd, off, &moff);
928 
929 				while (mb != NULL && max_len < len) {
930 					u_int mlen;
931 					u_int frags;
932 
933 					/*
934 					 * Get length of mbuf fragment
935 					 * and how many hardware frags,
936 					 * rounded up, it would use:
937 					 */
938 					mlen = (mb->m_len - moff);
939 					frags = howmany(mlen,
940 					    if_hw_tsomaxsegsize);
941 
942 					/* Handle special case: Zero Length Mbuf */
943 					if (frags == 0)
944 						frags = 1;
945 
946 					/*
947 					 * Check if the fragment limit
948 					 * will be reached or exceeded:
949 					 */
950 					if (frags >= if_hw_tsomaxsegcount) {
951 						max_len += min(mlen,
952 						    if_hw_tsomaxsegcount *
953 						    if_hw_tsomaxsegsize);
954 						break;
955 					}
956 					max_len += mlen;
957 					if_hw_tsomaxsegcount -= frags;
958 					moff = 0;
959 					mb = mb->m_next;
960 				}
961 				if (max_len <= 0) {
962 					len = 0;
963 				} else if (len > max_len) {
964 					sendalot = 1;
965 					len = max_len;
966 				}
967 			}
968 
969 			/*
970 			 * Prevent the last segment from being
971 			 * fractional unless the send sockbuf can be
972 			 * emptied:
973 			 */
974 			max_len = (tp->t_maxseg - optlen);
975 			if (((uint32_t)off + (uint32_t)len) <
976 			    sbavail(&so->so_snd)) {
977 				moff = len % max_len;
978 				if (moff != 0) {
979 					len -= moff;
980 					sendalot = 1;
981 				}
982 			}
983 
984 			/*
985 			 * In case there are too many small fragments
986 			 * don't use TSO:
987 			 */
988 			if (len <= max_len) {
989 				len = max_len;
990 				sendalot = 1;
991 				tso = 0;
992 			}
993 
994 			/*
995 			 * Send the FIN in a separate segment
996 			 * after the bulk sending is done.
997 			 * We don't trust the TSO implementations
998 			 * to clear the FIN flag on all but the
999 			 * last segment.
1000 			 */
1001 			if (tp->t_flags & TF_NEEDFIN)
1002 				sendalot = 1;
1003 
1004 		} else {
1005 			len = tp->t_maxseg - optlen - ipoptlen;
1006 			sendalot = 1;
1007 		}
1008 	} else
1009 		tso = 0;
1010 
1011 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1012 	    ("%s: len > IP_MAXPACKET", __func__));
1013 
1014 /*#ifdef DIAGNOSTIC*/
1015 #ifdef INET6
1016 	if (max_linkhdr + hdrlen > MCLBYTES)
1017 #else
1018 	if (max_linkhdr + hdrlen > MHLEN)
1019 #endif
1020 		panic("tcphdr too big");
1021 /*#endif*/
1022 
1023 	/*
1024 	 * This KASSERT is here to catch edge cases at a well defined place.
1025 	 * Before, those had triggered (random) panic conditions further down.
1026 	 */
1027 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1028 
1029 	/*
1030 	 * Grab a header mbuf, attaching a copy of data to
1031 	 * be transmitted, and initialize the header from
1032 	 * the template for sends on this connection.
1033 	 */
1034 	if (len) {
1035 		struct mbuf *mb;
1036 		u_int moff;
1037 
1038 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1039 			TCPSTAT_INC(tcps_sndprobe);
1040 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1041 			tp->t_sndrexmitpack++;
1042 			TCPSTAT_INC(tcps_sndrexmitpack);
1043 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1044 		} else {
1045 			TCPSTAT_INC(tcps_sndpack);
1046 			TCPSTAT_ADD(tcps_sndbyte, len);
1047 		}
1048 #ifdef INET6
1049 		if (MHLEN < hdrlen + max_linkhdr)
1050 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1051 		else
1052 #endif
1053 			m = m_gethdr(M_NOWAIT, MT_DATA);
1054 
1055 		if (m == NULL) {
1056 			SOCKBUF_UNLOCK(&so->so_snd);
1057 			error = ENOBUFS;
1058 			sack_rxmit = 0;
1059 			goto out;
1060 		}
1061 
1062 		m->m_data += max_linkhdr;
1063 		m->m_len = hdrlen;
1064 
1065 		/*
1066 		 * Start the m_copy functions from the closest mbuf
1067 		 * to the offset in the socket buffer chain.
1068 		 */
1069 		mb = sbsndptr(&so->so_snd, off, len, &moff);
1070 
1071 		if (len <= MHLEN - hdrlen - max_linkhdr) {
1072 			m_copydata(mb, moff, len,
1073 			    mtod(m, caddr_t) + hdrlen);
1074 			m->m_len += len;
1075 		} else {
1076 			m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1077 			if (m->m_next == NULL) {
1078 				SOCKBUF_UNLOCK(&so->so_snd);
1079 				(void) m_free(m);
1080 				error = ENOBUFS;
1081 				sack_rxmit = 0;
1082 				goto out;
1083 			}
1084 		}
1085 
1086 		/*
1087 		 * If we're sending everything we've got, set PUSH.
1088 		 * (This will keep happy those implementations which only
1089 		 * give data to the user when a buffer fills or
1090 		 * a PUSH comes in.)
1091 		 */
1092 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1093 		    !(flags & TH_SYN))
1094 			flags |= TH_PUSH;
1095 		SOCKBUF_UNLOCK(&so->so_snd);
1096 	} else {
1097 		SOCKBUF_UNLOCK(&so->so_snd);
1098 		if (tp->t_flags & TF_ACKNOW)
1099 			TCPSTAT_INC(tcps_sndacks);
1100 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1101 			TCPSTAT_INC(tcps_sndctrl);
1102 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1103 			TCPSTAT_INC(tcps_sndurg);
1104 		else
1105 			TCPSTAT_INC(tcps_sndwinup);
1106 
1107 		m = m_gethdr(M_NOWAIT, MT_DATA);
1108 		if (m == NULL) {
1109 			error = ENOBUFS;
1110 			sack_rxmit = 0;
1111 			goto out;
1112 		}
1113 #ifdef INET6
1114 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1115 		    MHLEN >= hdrlen) {
1116 			M_ALIGN(m, hdrlen);
1117 		} else
1118 #endif
1119 		m->m_data += max_linkhdr;
1120 		m->m_len = hdrlen;
1121 	}
1122 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1123 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1124 #ifdef MAC
1125 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1126 #endif
1127 #ifdef INET6
1128 	if (isipv6) {
1129 		ip6 = mtod(m, struct ip6_hdr *);
1130 		th = (struct tcphdr *)(ip6 + 1);
1131 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1132 	} else
1133 #endif /* INET6 */
1134 	{
1135 		ip = mtod(m, struct ip *);
1136 		ipov = (struct ipovly *)ip;
1137 		th = (struct tcphdr *)(ip + 1);
1138 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1139 	}
1140 
1141 	/*
1142 	 * Fill in fields, remembering maximum advertised
1143 	 * window for use in delaying messages about window sizes.
1144 	 * If resending a FIN, be sure not to use a new sequence number.
1145 	 */
1146 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1147 	    tp->snd_nxt == tp->snd_max)
1148 		tp->snd_nxt--;
1149 	/*
1150 	 * If we are starting a connection, send ECN setup
1151 	 * SYN packet. If we are on a retransmit, we may
1152 	 * resend those bits a number of times as per
1153 	 * RFC 3168.
1154 	 */
1155 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1156 		if (tp->t_rxtshift >= 1) {
1157 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1158 				flags |= TH_ECE|TH_CWR;
1159 		} else
1160 			flags |= TH_ECE|TH_CWR;
1161 	}
1162 
1163 	if (tp->t_state == TCPS_ESTABLISHED &&
1164 	    (tp->t_flags & TF_ECN_PERMIT)) {
1165 		/*
1166 		 * If the peer has ECN, mark data packets with
1167 		 * ECN capable transmission (ECT).
1168 		 * Ignore pure ack packets, retransmissions and window probes.
1169 		 */
1170 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1171 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1172 #ifdef INET6
1173 			if (isipv6)
1174 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1175 			else
1176 #endif
1177 				ip->ip_tos |= IPTOS_ECN_ECT0;
1178 			TCPSTAT_INC(tcps_ecn_ect0);
1179 		}
1180 
1181 		/*
1182 		 * Reply with proper ECN notifications.
1183 		 */
1184 		if (tp->t_flags & TF_ECN_SND_CWR) {
1185 			flags |= TH_CWR;
1186 			tp->t_flags &= ~TF_ECN_SND_CWR;
1187 		}
1188 		if (tp->t_flags & TF_ECN_SND_ECE)
1189 			flags |= TH_ECE;
1190 	}
1191 
1192 	/*
1193 	 * If we are doing retransmissions, then snd_nxt will
1194 	 * not reflect the first unsent octet.  For ACK only
1195 	 * packets, we do not want the sequence number of the
1196 	 * retransmitted packet, we want the sequence number
1197 	 * of the next unsent octet.  So, if there is no data
1198 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1199 	 * when filling in ti_seq.  But if we are in persist
1200 	 * state, snd_max might reflect one byte beyond the
1201 	 * right edge of the window, so use snd_nxt in that
1202 	 * case, since we know we aren't doing a retransmission.
1203 	 * (retransmit and persist are mutually exclusive...)
1204 	 */
1205 	if (sack_rxmit == 0) {
1206 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1207 		    tcp_timer_active(tp, TT_PERSIST))
1208 			th->th_seq = htonl(tp->snd_nxt);
1209 		else
1210 			th->th_seq = htonl(tp->snd_max);
1211 	} else {
1212 		th->th_seq = htonl(p->rxmit);
1213 		p->rxmit += len;
1214 		tp->sackhint.sack_bytes_rexmit += len;
1215 	}
1216 	th->th_ack = htonl(tp->rcv_nxt);
1217 	if (optlen) {
1218 		bcopy(opt, th + 1, optlen);
1219 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1220 	}
1221 	th->th_flags = flags;
1222 	/*
1223 	 * Calculate receive window.  Don't shrink window,
1224 	 * but avoid silly window syndrome.
1225 	 */
1226 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1227 	    recwin < tp->t_maxseg)
1228 		recwin = 0;
1229 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1230 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
1231 		recwin = (tp->rcv_adv - tp->rcv_nxt);
1232 
1233 	/*
1234 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1235 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1236 	 * case is handled in syncache.
1237 	 */
1238 	if (flags & TH_SYN)
1239 		th->th_win = htons((u_short)
1240 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1241 	else
1242 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1243 
1244 	/*
1245 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1246 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1247 	 * flag tells soreceive() to disable delayed acknowledgements when
1248 	 * draining the buffer.  This can occur if the receiver is attempting
1249 	 * to read more data than can be buffered prior to transmitting on
1250 	 * the connection.
1251 	 */
1252 	if (th->th_win == 0) {
1253 		tp->t_sndzerowin++;
1254 		tp->t_flags |= TF_RXWIN0SENT;
1255 	} else
1256 		tp->t_flags &= ~TF_RXWIN0SENT;
1257 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1258 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1259 		th->th_flags |= TH_URG;
1260 	} else
1261 		/*
1262 		 * If no urgent pointer to send, then we pull
1263 		 * the urgent pointer to the left edge of the send window
1264 		 * so that it doesn't drift into the send window on sequence
1265 		 * number wraparound.
1266 		 */
1267 		tp->snd_up = tp->snd_una;		/* drag it along */
1268 
1269 	/*
1270 	 * Put TCP length in extended header, and then
1271 	 * checksum extended header and data.
1272 	 */
1273 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1274 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1275 
1276 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1277 	if (to.to_flags & TOF_SIGNATURE) {
1278 		/*
1279 		 * Calculate MD5 signature and put it into the place
1280 		 * determined before.
1281 		 * NOTE: since TCP options buffer doesn't point into
1282 		 * mbuf's data, calculate offset and use it.
1283 		 */
1284 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
1285 		    (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
1286 			/*
1287 			 * Do not send segment if the calculation of MD5
1288 			 * digest has failed.
1289 			 */
1290 			goto out;
1291 		}
1292 	}
1293 #endif
1294 #ifdef INET6
1295 	if (isipv6) {
1296 		/*
1297 		 * There is no need to fill in ip6_plen right now.
1298 		 * It will be filled later by ip6_output.
1299 		 */
1300 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1301 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1302 		    optlen + len, IPPROTO_TCP, 0);
1303 	}
1304 #endif
1305 #if defined(INET6) && defined(INET)
1306 	else
1307 #endif
1308 #ifdef INET
1309 	{
1310 		m->m_pkthdr.csum_flags = CSUM_TCP;
1311 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1312 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1313 
1314 		/* IP version must be set here for ipv4/ipv6 checking later */
1315 		KASSERT(ip->ip_v == IPVERSION,
1316 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1317 	}
1318 #endif
1319 
1320 	/*
1321 	 * Enable TSO and specify the size of the segments.
1322 	 * The TCP pseudo header checksum is always provided.
1323 	 */
1324 	if (tso) {
1325 		KASSERT(len > tp->t_maxseg - optlen,
1326 		    ("%s: len <= tso_segsz", __func__));
1327 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1328 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1329 	}
1330 
1331 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1332 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1333 	    ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1334 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1335 #else
1336 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1337 	    ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1338 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1339 #endif
1340 
1341 #ifdef TCP_HHOOK
1342 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1343 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1344 #endif
1345 
1346 #ifdef TCPDEBUG
1347 	/*
1348 	 * Trace.
1349 	 */
1350 	if (so->so_options & SO_DEBUG) {
1351 		u_short save = 0;
1352 #ifdef INET6
1353 		if (!isipv6)
1354 #endif
1355 		{
1356 			save = ipov->ih_len;
1357 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1358 		}
1359 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1360 #ifdef INET6
1361 		if (!isipv6)
1362 #endif
1363 		ipov->ih_len = save;
1364 	}
1365 #endif /* TCPDEBUG */
1366 	TCP_PROBE3(debug__output, tp, th, m);
1367 
1368 	/*
1369 	 * Fill in IP length and desired time to live and
1370 	 * send to IP level.  There should be a better way
1371 	 * to handle ttl and tos; we could keep them in
1372 	 * the template, but need a way to checksum without them.
1373 	 */
1374 	/*
1375 	 * m->m_pkthdr.len should have been set before checksum calculation,
1376 	 * because in6_cksum() need it.
1377 	 */
1378 #ifdef INET6
1379 	if (isipv6) {
1380 		struct route_in6 ro;
1381 
1382 		bzero(&ro, sizeof(ro));
1383 		/*
1384 		 * we separately set hoplimit for every segment, since the
1385 		 * user might want to change the value via setsockopt.
1386 		 * Also, desired default hop limit might be changed via
1387 		 * Neighbor Discovery.
1388 		 */
1389 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1390 
1391 		/*
1392 		 * Set the packet size here for the benefit of DTrace probes.
1393 		 * ip6_output() will set it properly; it's supposed to include
1394 		 * the option header lengths as well.
1395 		 */
1396 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1397 
1398 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1399 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1400 		else
1401 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1402 
1403 		if (tp->t_state == TCPS_SYN_SENT)
1404 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1405 
1406 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1407 
1408 #ifdef TCPPCAP
1409 		/* Save packet, if requested. */
1410 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1411 #endif
1412 
1413 		/* TODO: IPv6 IP6TOS_ECT bit on */
1414 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1415 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1416 		    NULL, NULL, tp->t_inpcb);
1417 
1418 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1419 			mtu = ro.ro_rt->rt_mtu;
1420 		RO_RTFREE(&ro);
1421 	}
1422 #endif /* INET6 */
1423 #if defined(INET) && defined(INET6)
1424 	else
1425 #endif
1426 #ifdef INET
1427     {
1428 	ip->ip_len = htons(m->m_pkthdr.len);
1429 #ifdef INET6
1430 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1431 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1432 #endif /* INET6 */
1433 	/*
1434 	 * If we do path MTU discovery, then we set DF on every packet.
1435 	 * This might not be the best thing to do according to RFC3390
1436 	 * Section 2. However the tcp hostcache migitates the problem
1437 	 * so it affects only the first tcp connection with a host.
1438 	 *
1439 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1440 	 */
1441 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1442 		ip->ip_off |= htons(IP_DF);
1443 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1444 	} else {
1445 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1446 	}
1447 
1448 	if (tp->t_state == TCPS_SYN_SENT)
1449 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1450 
1451 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1452 
1453 #ifdef TCPPCAP
1454 	/* Save packet, if requested. */
1455 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1456 #endif
1457 
1458 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1459 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1460 	    tp->t_inpcb);
1461 
1462 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1463 		mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1464     }
1465 #endif /* INET */
1466 
1467 out:
1468 	/*
1469 	 * In transmit state, time the transmission and arrange for
1470 	 * the retransmit.  In persist state, just set snd_max.
1471 	 */
1472 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1473 	    !tcp_timer_active(tp, TT_PERSIST)) {
1474 		tcp_seq startseq = tp->snd_nxt;
1475 
1476 		/*
1477 		 * Advance snd_nxt over sequence space of this segment.
1478 		 */
1479 		if (flags & (TH_SYN|TH_FIN)) {
1480 			if (flags & TH_SYN)
1481 				tp->snd_nxt++;
1482 			if (flags & TH_FIN) {
1483 				tp->snd_nxt++;
1484 				tp->t_flags |= TF_SENTFIN;
1485 			}
1486 		}
1487 		if (sack_rxmit)
1488 			goto timer;
1489 		tp->snd_nxt += len;
1490 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1491 			tp->snd_max = tp->snd_nxt;
1492 			/*
1493 			 * Time this transmission if not a retransmission and
1494 			 * not currently timing anything.
1495 			 */
1496 			if (tp->t_rtttime == 0) {
1497 				tp->t_rtttime = ticks;
1498 				tp->t_rtseq = startseq;
1499 				TCPSTAT_INC(tcps_segstimed);
1500 			}
1501 		}
1502 
1503 		/*
1504 		 * Set retransmit timer if not currently set,
1505 		 * and not doing a pure ack or a keep-alive probe.
1506 		 * Initial value for retransmit timer is smoothed
1507 		 * round-trip time + 2 * round-trip time variance.
1508 		 * Initialize shift counter which is used for backoff
1509 		 * of retransmit time.
1510 		 */
1511 timer:
1512 		if (!tcp_timer_active(tp, TT_REXMT) &&
1513 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1514 		     (tp->snd_nxt != tp->snd_una))) {
1515 			if (tcp_timer_active(tp, TT_PERSIST)) {
1516 				tcp_timer_activate(tp, TT_PERSIST, 0);
1517 				tp->t_rxtshift = 0;
1518 			}
1519 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1520 		} else if (len == 0 && sbavail(&so->so_snd) &&
1521 		    !tcp_timer_active(tp, TT_REXMT) &&
1522 		    !tcp_timer_active(tp, TT_PERSIST)) {
1523 			/*
1524 			 * Avoid a situation where we do not set persist timer
1525 			 * after a zero window condition. For example:
1526 			 * 1) A -> B: packet with enough data to fill the window
1527 			 * 2) B -> A: ACK for #1 + new data (0 window
1528 			 *    advertisement)
1529 			 * 3) A -> B: ACK for #2, 0 len packet
1530 			 *
1531 			 * In this case, A will not activate the persist timer,
1532 			 * because it chose to send a packet. Unless tcp_output
1533 			 * is called for some other reason (delayed ack timer,
1534 			 * another input packet from B, socket syscall), A will
1535 			 * not send zero window probes.
1536 			 *
1537 			 * So, if you send a 0-length packet, but there is data
1538 			 * in the socket buffer, and neither the rexmt or
1539 			 * persist timer is already set, then activate the
1540 			 * persist timer.
1541 			 */
1542 			tp->t_rxtshift = 0;
1543 			tcp_setpersist(tp);
1544 		}
1545 	} else {
1546 		/*
1547 		 * Persist case, update snd_max but since we are in
1548 		 * persist mode (no window) we do not update snd_nxt.
1549 		 */
1550 		int xlen = len;
1551 		if (flags & TH_SYN)
1552 			++xlen;
1553 		if (flags & TH_FIN) {
1554 			++xlen;
1555 			tp->t_flags |= TF_SENTFIN;
1556 		}
1557 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1558 			tp->snd_max = tp->snd_nxt + xlen;
1559 	}
1560 
1561 	if (error) {
1562 
1563 		/*
1564 		 * We know that the packet was lost, so back out the
1565 		 * sequence number advance, if any.
1566 		 *
1567 		 * If the error is EPERM the packet got blocked by the
1568 		 * local firewall.  Normally we should terminate the
1569 		 * connection but the blocking may have been spurious
1570 		 * due to a firewall reconfiguration cycle.  So we treat
1571 		 * it like a packet loss and let the retransmit timer and
1572 		 * timeouts do their work over time.
1573 		 * XXX: It is a POLA question whether calling tcp_drop right
1574 		 * away would be the really correct behavior instead.
1575 		 */
1576 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1577 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1578 		    ((flags & TH_SYN) == 0) &&
1579 		    (error != EPERM)) {
1580 			if (sack_rxmit) {
1581 				p->rxmit -= len;
1582 				tp->sackhint.sack_bytes_rexmit -= len;
1583 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1584 				    ("sackhint bytes rtx >= 0"));
1585 			} else
1586 				tp->snd_nxt -= len;
1587 		}
1588 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1589 		switch (error) {
1590 		case EACCES:
1591 			tp->t_softerror = error;
1592 			return (0);
1593 		case EPERM:
1594 			tp->t_softerror = error;
1595 			return (error);
1596 		case ENOBUFS:
1597 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1598 			tp->snd_cwnd = tp->t_maxseg;
1599 			return (0);
1600 		case EMSGSIZE:
1601 			/*
1602 			 * For some reason the interface we used initially
1603 			 * to send segments changed to another or lowered
1604 			 * its MTU.
1605 			 * If TSO was active we either got an interface
1606 			 * without TSO capabilits or TSO was turned off.
1607 			 * If we obtained mtu from ip_output() then update
1608 			 * it and try again.
1609 			 */
1610 			if (tso)
1611 				tp->t_flags &= ~TF_TSO;
1612 			if (mtu != 0) {
1613 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1614 				goto again;
1615 			}
1616 			return (error);
1617 		case EHOSTDOWN:
1618 		case EHOSTUNREACH:
1619 		case ENETDOWN:
1620 		case ENETUNREACH:
1621 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1622 				tp->t_softerror = error;
1623 				return (0);
1624 			}
1625 			/* FALLTHROUGH */
1626 		default:
1627 			return (error);
1628 		}
1629 	}
1630 	TCPSTAT_INC(tcps_sndtotal);
1631 
1632 	/*
1633 	 * Data sent (as far as we can tell).
1634 	 * If this advertises a larger window than any other segment,
1635 	 * then remember the size of the advertised window.
1636 	 * Any pending ACK has now been sent.
1637 	 */
1638 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1639 		tp->rcv_adv = tp->rcv_nxt + recwin;
1640 	tp->last_ack_sent = tp->rcv_nxt;
1641 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1642 	if (tcp_timer_active(tp, TT_DELACK))
1643 		tcp_timer_activate(tp, TT_DELACK, 0);
1644 #if 0
1645 	/*
1646 	 * This completely breaks TCP if newreno is turned on.  What happens
1647 	 * is that if delayed-acks are turned on on the receiver, this code
1648 	 * on the transmitter effectively destroys the TCP window, forcing
1649 	 * it to four packets (1.5Kx4 = 6K window).
1650 	 */
1651 	if (sendalot && --maxburst)
1652 		goto again;
1653 #endif
1654 	if (sendalot)
1655 		goto again;
1656 	return (0);
1657 }
1658 
1659 void
1660 tcp_setpersist(struct tcpcb *tp)
1661 {
1662 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1663 	int tt;
1664 
1665 	tp->t_flags &= ~TF_PREVVALID;
1666 	if (tcp_timer_active(tp, TT_REXMT))
1667 		panic("tcp_setpersist: retransmit pending");
1668 	/*
1669 	 * Start/restart persistence timer.
1670 	 */
1671 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1672 		      tcp_persmin, tcp_persmax);
1673 	tcp_timer_activate(tp, TT_PERSIST, tt);
1674 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1675 		tp->t_rxtshift++;
1676 }
1677 
1678 /*
1679  * Insert TCP options according to the supplied parameters to the place
1680  * optp in a consistent way.  Can handle unaligned destinations.
1681  *
1682  * The order of the option processing is crucial for optimal packing and
1683  * alignment for the scarce option space.
1684  *
1685  * The optimal order for a SYN/SYN-ACK segment is:
1686  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1687  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1688  *
1689  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1690  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1691  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1692  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1693  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1694  */
1695 int
1696 tcp_addoptions(struct tcpopt *to, u_char *optp)
1697 {
1698 	u_int32_t mask, optlen = 0;
1699 
1700 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1701 		if ((to->to_flags & mask) != mask)
1702 			continue;
1703 		if (optlen == TCP_MAXOLEN)
1704 			break;
1705 		switch (to->to_flags & mask) {
1706 		case TOF_MSS:
1707 			while (optlen % 4) {
1708 				optlen += TCPOLEN_NOP;
1709 				*optp++ = TCPOPT_NOP;
1710 			}
1711 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1712 				continue;
1713 			optlen += TCPOLEN_MAXSEG;
1714 			*optp++ = TCPOPT_MAXSEG;
1715 			*optp++ = TCPOLEN_MAXSEG;
1716 			to->to_mss = htons(to->to_mss);
1717 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1718 			optp += sizeof(to->to_mss);
1719 			break;
1720 		case TOF_SCALE:
1721 			while (!optlen || optlen % 2 != 1) {
1722 				optlen += TCPOLEN_NOP;
1723 				*optp++ = TCPOPT_NOP;
1724 			}
1725 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1726 				continue;
1727 			optlen += TCPOLEN_WINDOW;
1728 			*optp++ = TCPOPT_WINDOW;
1729 			*optp++ = TCPOLEN_WINDOW;
1730 			*optp++ = to->to_wscale;
1731 			break;
1732 		case TOF_SACKPERM:
1733 			while (optlen % 2) {
1734 				optlen += TCPOLEN_NOP;
1735 				*optp++ = TCPOPT_NOP;
1736 			}
1737 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1738 				continue;
1739 			optlen += TCPOLEN_SACK_PERMITTED;
1740 			*optp++ = TCPOPT_SACK_PERMITTED;
1741 			*optp++ = TCPOLEN_SACK_PERMITTED;
1742 			break;
1743 		case TOF_TS:
1744 			while (!optlen || optlen % 4 != 2) {
1745 				optlen += TCPOLEN_NOP;
1746 				*optp++ = TCPOPT_NOP;
1747 			}
1748 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1749 				continue;
1750 			optlen += TCPOLEN_TIMESTAMP;
1751 			*optp++ = TCPOPT_TIMESTAMP;
1752 			*optp++ = TCPOLEN_TIMESTAMP;
1753 			to->to_tsval = htonl(to->to_tsval);
1754 			to->to_tsecr = htonl(to->to_tsecr);
1755 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1756 			optp += sizeof(to->to_tsval);
1757 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1758 			optp += sizeof(to->to_tsecr);
1759 			break;
1760 		case TOF_SIGNATURE:
1761 			{
1762 			int siglen = TCPOLEN_SIGNATURE - 2;
1763 
1764 			while (!optlen || optlen % 4 != 2) {
1765 				optlen += TCPOLEN_NOP;
1766 				*optp++ = TCPOPT_NOP;
1767 			}
1768 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1769 				to->to_flags &= ~TOF_SIGNATURE;
1770 				continue;
1771 			}
1772 			optlen += TCPOLEN_SIGNATURE;
1773 			*optp++ = TCPOPT_SIGNATURE;
1774 			*optp++ = TCPOLEN_SIGNATURE;
1775 			to->to_signature = optp;
1776 			while (siglen--)
1777 				 *optp++ = 0;
1778 			break;
1779 			}
1780 		case TOF_SACK:
1781 			{
1782 			int sackblks = 0;
1783 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1784 			tcp_seq sack_seq;
1785 
1786 			while (!optlen || optlen % 4 != 2) {
1787 				optlen += TCPOLEN_NOP;
1788 				*optp++ = TCPOPT_NOP;
1789 			}
1790 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1791 				continue;
1792 			optlen += TCPOLEN_SACKHDR;
1793 			*optp++ = TCPOPT_SACK;
1794 			sackblks = min(to->to_nsacks,
1795 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1796 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1797 			while (sackblks--) {
1798 				sack_seq = htonl(sack->start);
1799 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1800 				optp += sizeof(sack_seq);
1801 				sack_seq = htonl(sack->end);
1802 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1803 				optp += sizeof(sack_seq);
1804 				optlen += TCPOLEN_SACK;
1805 				sack++;
1806 			}
1807 			TCPSTAT_INC(tcps_sack_send_blocks);
1808 			break;
1809 			}
1810 #ifdef TCP_RFC7413
1811 		case TOF_FASTOPEN:
1812 			{
1813 			int total_len;
1814 
1815 			/* XXX is there any point to aligning this option? */
1816 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1817 			if (TCP_MAXOLEN - optlen < total_len)
1818 				continue;
1819 			*optp++ = TCPOPT_FAST_OPEN;
1820 			*optp++ = total_len;
1821 			if (to->to_tfo_len > 0) {
1822 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1823 				optp += to->to_tfo_len;
1824 			}
1825 			optlen += total_len;
1826 			break;
1827 			}
1828 #endif
1829 		default:
1830 			panic("%s: unknown TCP option type", __func__);
1831 			break;
1832 		}
1833 	}
1834 
1835 	/* Terminate and pad TCP options to a 4 byte boundary. */
1836 	if (optlen % 4) {
1837 		optlen += TCPOLEN_EOL;
1838 		*optp++ = TCPOPT_EOL;
1839 	}
1840 	/*
1841 	 * According to RFC 793 (STD0007):
1842 	 *   "The content of the header beyond the End-of-Option option
1843 	 *    must be header padding (i.e., zero)."
1844 	 *   and later: "The padding is composed of zeros."
1845 	 */
1846 	while (optlen % 4) {
1847 		optlen += TCPOLEN_PAD;
1848 		*optp++ = TCPOPT_PAD;
1849 	}
1850 
1851 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1852 	return (optlen);
1853 }
1854