xref: /freebsd/sys/netinet/tcp_output.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
45 #ifdef TCP_HHOOK
46 #include <sys/hhook.h>
47 #endif
48 #include <sys/kernel.h>
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/sdt.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
57 
58 #include <net/if.h>
59 #include <net/route.h>
60 #include <net/vnet.h>
61 
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
69 #ifdef INET6
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
73 #endif
74 #include <netinet/tcp.h>
75 #define	TCPOUTFLAGS
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/tcp_seq.h>
78 #include <netinet/tcp_timer.h>
79 #include <netinet/tcp_var.h>
80 #include <netinet/tcpip.h>
81 #include <netinet/cc/cc.h>
82 #include <netinet/tcp_fastopen.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 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
106 	&VNET_NAME(tcp_do_tso), 0,
107 	"Enable TCP Segmentation Offload");
108 
109 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
110 #define	V_tcp_sendspace	VNET(tcp_sendspace)
111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
112 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
113 
114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
116 	&VNET_NAME(tcp_do_autosndbuf), 0,
117 	"Enable automatic send buffer sizing");
118 
119 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
121 	&VNET_NAME(tcp_autosndbuf_inc), 0,
122 	"Incrementor step size of automatic send buffer");
123 
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
126 	&VNET_NAME(tcp_autosndbuf_max), 0,
127 	"Max size of automatic send buffer");
128 
129 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
130 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
132 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
133 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
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 #ifdef TCPDEBUG
200 	struct ipovly *ipov = NULL;
201 #endif
202 	struct tcphdr *th;
203 	u_char opt[TCP_MAXOLEN];
204 	unsigned ipoptlen, optlen, hdrlen;
205 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
206 	unsigned ipsec_optlen = 0;
207 #endif
208 	int idle, sendalot;
209 	int sack_rxmit, sack_bytes_rxmt;
210 	struct sackhole *p;
211 	int tso, mtu;
212 	struct tcpopt to;
213 	unsigned int wanted_cookie = 0;
214 	unsigned int dont_sendalot = 0;
215 #if 0
216 	int maxburst = TCP_MAXBURST;
217 #endif
218 #ifdef INET6
219 	struct ip6_hdr *ip6 = NULL;
220 	int isipv6;
221 
222 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
223 #endif
224 
225 	INP_WLOCK_ASSERT(tp->t_inpcb);
226 
227 #ifdef TCP_OFFLOAD
228 	if (tp->t_flags & TF_TOE)
229 		return (tcp_offload_output(tp));
230 #endif
231 
232 	/*
233 	 * For TFO connections in SYN_RECEIVED, only allow the initial
234 	 * SYN|ACK and those sent by the retransmit timer.
235 	 */
236 	if (IS_FASTOPEN(tp->t_flags) &&
237 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
238 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
239 	    (tp->snd_nxt != tp->snd_una))         /* not a retransmit */
240 		return (0);
241 
242 	/*
243 	 * Determine length of data that should be transmitted,
244 	 * and flags that will be used.
245 	 * If there is some data or critical controls (SYN, RST)
246 	 * to send, then transmit; otherwise, investigate further.
247 	 */
248 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
249 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
250 		cc_after_idle(tp);
251 	tp->t_flags &= ~TF_LASTIDLE;
252 	if (idle) {
253 		if (tp->t_flags & TF_MORETOCOME) {
254 			tp->t_flags |= TF_LASTIDLE;
255 			idle = 0;
256 		}
257 	}
258 again:
259 	/*
260 	 * If we've recently taken a timeout, snd_max will be greater than
261 	 * snd_nxt.  There may be SACK information that allows us to avoid
262 	 * resending already delivered data.  Adjust snd_nxt accordingly.
263 	 */
264 	if ((tp->t_flags & TF_SACK_PERMIT) &&
265 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
266 		tcp_sack_adjust(tp);
267 	sendalot = 0;
268 	tso = 0;
269 	mtu = 0;
270 	off = tp->snd_nxt - tp->snd_una;
271 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
272 
273 	flags = tcp_outflags[tp->t_state];
274 	/*
275 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
276 	 * to send out new data (when sendalot is 1), bypass this function.
277 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
278 	 * we're replacing a (future) new transmission with a retransmission
279 	 * now, and we previously incremented snd_cwnd in tcp_input().
280 	 */
281 	/*
282 	 * Still in sack recovery , reset rxmit flag to zero.
283 	 */
284 	sack_rxmit = 0;
285 	sack_bytes_rxmt = 0;
286 	len = 0;
287 	p = NULL;
288 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
289 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
290 		uint32_t cwin;
291 
292 		cwin =
293 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
294 		/* Do not retransmit SACK segments beyond snd_recover */
295 		if (SEQ_GT(p->end, tp->snd_recover)) {
296 			/*
297 			 * (At least) part of sack hole extends beyond
298 			 * snd_recover. Check to see if we can rexmit data
299 			 * for this hole.
300 			 */
301 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
302 				/*
303 				 * Can't rexmit any more data for this hole.
304 				 * That data will be rexmitted in the next
305 				 * sack recovery episode, when snd_recover
306 				 * moves past p->rxmit.
307 				 */
308 				p = NULL;
309 				goto after_sack_rexmit;
310 			} else
311 				/* Can rexmit part of the current hole */
312 				len = ((int32_t)ulmin(cwin,
313 						   tp->snd_recover - p->rxmit));
314 		} else
315 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
316 		off = p->rxmit - tp->snd_una;
317 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
318 		    __func__, off));
319 		if (len > 0) {
320 			sack_rxmit = 1;
321 			sendalot = 1;
322 			TCPSTAT_INC(tcps_sack_rexmits);
323 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
324 			    min(len, tp->t_maxseg));
325 		}
326 	}
327 after_sack_rexmit:
328 	/*
329 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
330 	 * state flags.
331 	 */
332 	if (tp->t_flags & TF_NEEDFIN)
333 		flags |= TH_FIN;
334 	if (tp->t_flags & TF_NEEDSYN)
335 		flags |= TH_SYN;
336 
337 	SOCKBUF_LOCK(&so->so_snd);
338 	/*
339 	 * If in persist timeout with window of 0, send 1 byte.
340 	 * Otherwise, if window is small but nonzero
341 	 * and timer expired, we will send what we can
342 	 * and go to transmit state.
343 	 */
344 	if (tp->t_flags & TF_FORCEDATA) {
345 		if (sendwin == 0) {
346 			/*
347 			 * If we still have some data to send, then
348 			 * clear the FIN bit.  Usually this would
349 			 * happen below when it realizes that we
350 			 * aren't sending all the data.  However,
351 			 * if we have exactly 1 byte of unsent data,
352 			 * then it won't clear the FIN bit below,
353 			 * and if we are in persist state, we wind
354 			 * up sending the packet without recording
355 			 * that we sent the FIN bit.
356 			 *
357 			 * We can't just blindly clear the FIN bit,
358 			 * because if we don't have any more data
359 			 * to send then the probe will be the FIN
360 			 * itself.
361 			 */
362 			if (off < sbused(&so->so_snd))
363 				flags &= ~TH_FIN;
364 			sendwin = 1;
365 		} else {
366 			tcp_timer_activate(tp, TT_PERSIST, 0);
367 			tp->t_rxtshift = 0;
368 		}
369 	}
370 
371 	/*
372 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
373 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
374 	 * a negative length.  This can also occur when TCP opens up
375 	 * its congestion window while receiving additional duplicate
376 	 * acks after fast-retransmit because TCP will reset snd_nxt
377 	 * to snd_max after the fast-retransmit.
378 	 *
379 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
380 	 * be set to snd_una, the offset will be 0, and the length may
381 	 * wind up 0.
382 	 *
383 	 * If sack_rxmit is true we are retransmitting from the scoreboard
384 	 * in which case len is already set.
385 	 */
386 	if (sack_rxmit == 0) {
387 		if (sack_bytes_rxmt == 0)
388 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
389 			    off);
390 		else {
391 			int32_t cwin;
392 
393                         /*
394 			 * We are inside of a SACK recovery episode and are
395 			 * sending new data, having retransmitted all the
396 			 * data possible in the scoreboard.
397 			 */
398 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
399 			    off);
400 			/*
401 			 * Don't remove this (len > 0) check !
402 			 * We explicitly check for len > 0 here (although it
403 			 * isn't really necessary), to work around a gcc
404 			 * optimization issue - to force gcc to compute
405 			 * len above. Without this check, the computation
406 			 * of len is bungled by the optimizer.
407 			 */
408 			if (len > 0) {
409 				cwin = tp->snd_cwnd -
410 					(tp->snd_nxt - tp->sack_newdata) -
411 					sack_bytes_rxmt;
412 				if (cwin < 0)
413 					cwin = 0;
414 				len = imin(len, cwin);
415 			}
416 		}
417 	}
418 
419 	/*
420 	 * Lop off SYN bit if it has already been sent.  However, if this
421 	 * is SYN-SENT state and if segment contains data and if we don't
422 	 * know that foreign host supports TAO, suppress sending segment.
423 	 */
424 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
425 		if (tp->t_state != TCPS_SYN_RECEIVED)
426 			flags &= ~TH_SYN;
427 		/*
428 		 * When sending additional segments following a TFO SYN|ACK,
429 		 * do not include the SYN bit.
430 		 */
431 		if (IS_FASTOPEN(tp->t_flags) &&
432 		    (tp->t_state == TCPS_SYN_RECEIVED))
433 			flags &= ~TH_SYN;
434 		off--, len++;
435 	}
436 
437 	/*
438 	 * Be careful not to send data and/or FIN on SYN segments.
439 	 * This measure is needed to prevent interoperability problems
440 	 * with not fully conformant TCP implementations.
441 	 */
442 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
443 		len = 0;
444 		flags &= ~TH_FIN;
445 	}
446 
447 	/*
448 	 * On TFO sockets, ensure no data is sent in the following cases:
449 	 *
450 	 *  - When retransmitting SYN|ACK on a passively-created socket
451 	 *
452 	 *  - When retransmitting SYN on an actively created socket
453 	 *
454 	 *  - When sending a zero-length cookie (cookie request) on an
455 	 *    actively created socket
456 	 *
457 	 *  - When the socket is in the CLOSED state (RST is being sent)
458 	 */
459 	if (IS_FASTOPEN(tp->t_flags) &&
460 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
461 	     ((tp->t_state == TCPS_SYN_SENT) &&
462 	      (tp->t_tfo_client_cookie_len == 0)) ||
463 	     (flags & TH_RST)))
464 		len = 0;
465 	if (len <= 0) {
466 		/*
467 		 * If FIN has been sent but not acked,
468 		 * but we haven't been called to retransmit,
469 		 * len will be < 0.  Otherwise, window shrank
470 		 * after we sent into it.  If window shrank to 0,
471 		 * cancel pending retransmit, pull snd_nxt back
472 		 * to (closed) window, and set the persist timer
473 		 * if it isn't already going.  If the window didn't
474 		 * close completely, just wait for an ACK.
475 		 *
476 		 * We also do a general check here to ensure that
477 		 * we will set the persist timer when we have data
478 		 * to send, but a 0-byte window. This makes sure
479 		 * the persist timer is set even if the packet
480 		 * hits one of the "goto send" lines below.
481 		 */
482 		len = 0;
483 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
484 			(off < (int) sbavail(&so->so_snd))) {
485 			tcp_timer_activate(tp, TT_REXMT, 0);
486 			tp->t_rxtshift = 0;
487 			tp->snd_nxt = tp->snd_una;
488 			if (!tcp_timer_active(tp, TT_PERSIST))
489 				tcp_setpersist(tp);
490 		}
491 	}
492 
493 	/* len will be >= 0 after this point. */
494 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
495 
496 	tcp_sndbuf_autoscale(tp, so, sendwin);
497 
498 	/*
499 	 * Decide if we can use TCP Segmentation Offloading (if supported by
500 	 * hardware).
501 	 *
502 	 * TSO may only be used if we are in a pure bulk sending state.  The
503 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
504 	 * IP options prevent using TSO.  With TSO the TCP header is the same
505 	 * (except for the sequence number) for all generated packets.  This
506 	 * makes it impossible to transmit any options which vary per generated
507 	 * segment or packet.
508 	 *
509 	 * IPv4 handling has a clear separation of ip options and ip header
510 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
511 	 * the right thing below to provide length of just ip options and thus
512 	 * checking for ipoptlen is enough to decide if ip options are present.
513 	 */
514 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
515 	/*
516 	 * Pre-calculate here as we save another lookup into the darknesses
517 	 * of IPsec that way and can actually decide if TSO is ok.
518 	 */
519 #ifdef INET6
520 	if (isipv6 && IPSEC_ENABLED(ipv6))
521 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
522 #ifdef INET
523 	else
524 #endif
525 #endif /* INET6 */
526 #ifdef INET
527 	if (IPSEC_ENABLED(ipv4))
528 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
529 #endif /* INET */
530 #endif /* IPSEC */
531 #ifdef INET6
532 	if (isipv6)
533 		ipoptlen = ip6_optlen(tp->t_inpcb);
534 	else
535 #endif
536 	if (tp->t_inpcb->inp_options)
537 		ipoptlen = tp->t_inpcb->inp_options->m_len -
538 				offsetof(struct ipoption, ipopt_list);
539 	else
540 		ipoptlen = 0;
541 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
542 	ipoptlen += ipsec_optlen;
543 #endif
544 
545 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
546 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
547 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
548 	    ipoptlen == 0 && !(flags & TH_SYN))
549 		tso = 1;
550 
551 	if (sack_rxmit) {
552 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
553 			flags &= ~TH_FIN;
554 	} else {
555 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
556 		    sbused(&so->so_snd)))
557 			flags &= ~TH_FIN;
558 	}
559 
560 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
561 	    (long)TCP_MAXWIN << tp->rcv_scale);
562 
563 	/*
564 	 * Sender silly window avoidance.   We transmit under the following
565 	 * conditions when len is non-zero:
566 	 *
567 	 *	- We have a full segment (or more with TSO)
568 	 *	- This is the last buffer in a write()/send() and we are
569 	 *	  either idle or running NODELAY
570 	 *	- we've timed out (e.g. persist timer)
571 	 *	- we have more then 1/2 the maximum send window's worth of
572 	 *	  data (receiver may be limited the window size)
573 	 *	- we need to retransmit
574 	 */
575 	if (len) {
576 		if (len >= tp->t_maxseg)
577 			goto send;
578 		/*
579 		 * NOTE! on localhost connections an 'ack' from the remote
580 		 * end may occur synchronously with the output and cause
581 		 * us to flush a buffer queued with moretocome.  XXX
582 		 *
583 		 * note: the len + off check is almost certainly unnecessary.
584 		 */
585 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
586 		    (idle || (tp->t_flags & TF_NODELAY)) &&
587 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
588 		    (tp->t_flags & TF_NOPUSH) == 0) {
589 			goto send;
590 		}
591 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
592 			goto send;
593 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
594 			goto send;
595 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
596 			goto send;
597 		if (sack_rxmit)
598 			goto send;
599 	}
600 
601 	/*
602 	 * Sending of standalone window updates.
603 	 *
604 	 * Window updates are important when we close our window due to a
605 	 * full socket buffer and are opening it again after the application
606 	 * reads data from it.  Once the window has opened again and the
607 	 * remote end starts to send again the ACK clock takes over and
608 	 * provides the most current window information.
609 	 *
610 	 * We must avoid the silly window syndrome whereas every read
611 	 * from the receive buffer, no matter how small, causes a window
612 	 * update to be sent.  We also should avoid sending a flurry of
613 	 * window updates when the socket buffer had queued a lot of data
614 	 * and the application is doing small reads.
615 	 *
616 	 * Prevent a flurry of pointless window updates by only sending
617 	 * an update when we can increase the advertized window by more
618 	 * than 1/4th of the socket buffer capacity.  When the buffer is
619 	 * getting full or is very small be more aggressive and send an
620 	 * update whenever we can increase by two mss sized segments.
621 	 * In all other situations the ACK's to new incoming data will
622 	 * carry further window increases.
623 	 *
624 	 * Don't send an independent window update if a delayed
625 	 * ACK is pending (it will get piggy-backed on it) or the
626 	 * remote side already has done a half-close and won't send
627 	 * more data.  Skip this if the connection is in T/TCP
628 	 * half-open state.
629 	 */
630 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
631 	    !(tp->t_flags & TF_DELACK) &&
632 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
633 		/*
634 		 * "adv" is the amount we could increase the window,
635 		 * taking into account that we are limited by
636 		 * TCP_MAXWIN << tp->rcv_scale.
637 		 */
638 		int32_t adv;
639 		int oldwin;
640 
641 		adv = recwin;
642 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
643 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
644 			adv -= oldwin;
645 		} else
646 			oldwin = 0;
647 
648 		/*
649 		 * If the new window size ends up being the same as or less
650 		 * than the old size when it is scaled, then don't force
651 		 * a window update.
652 		 */
653 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
654 			goto dontupdate;
655 
656 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
657 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
658 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
659 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
660 			goto send;
661 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
662 			goto send;
663 	}
664 dontupdate:
665 
666 	/*
667 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
668 	 * is also a catch-all for the retransmit timer timeout case.
669 	 */
670 	if (tp->t_flags & TF_ACKNOW)
671 		goto send;
672 	if ((flags & TH_RST) ||
673 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
674 		goto send;
675 	if (SEQ_GT(tp->snd_up, tp->snd_una))
676 		goto send;
677 	/*
678 	 * If our state indicates that FIN should be sent
679 	 * and we have not yet done so, then we need to send.
680 	 */
681 	if (flags & TH_FIN &&
682 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
683 		goto send;
684 	/*
685 	 * In SACK, it is possible for tcp_output to fail to send a segment
686 	 * after the retransmission timer has been turned off.  Make sure
687 	 * that the retransmission timer is set.
688 	 */
689 	if ((tp->t_flags & TF_SACK_PERMIT) &&
690 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
691 	    !tcp_timer_active(tp, TT_REXMT) &&
692 	    !tcp_timer_active(tp, TT_PERSIST)) {
693 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
694 		goto just_return;
695 	}
696 	/*
697 	 * TCP window updates are not reliable, rather a polling protocol
698 	 * using ``persist'' packets is used to insure receipt of window
699 	 * updates.  The three ``states'' for the output side are:
700 	 *	idle			not doing retransmits or persists
701 	 *	persisting		to move a small or zero window
702 	 *	(re)transmitting	and thereby not persisting
703 	 *
704 	 * tcp_timer_active(tp, TT_PERSIST)
705 	 *	is true when we are in persist state.
706 	 * (tp->t_flags & TF_FORCEDATA)
707 	 *	is set when we are called to send a persist packet.
708 	 * tcp_timer_active(tp, TT_REXMT)
709 	 *	is set when we are retransmitting
710 	 * The output side is idle when both timers are zero.
711 	 *
712 	 * If send window is too small, there is data to transmit, and no
713 	 * retransmit or persist is pending, then go to persist state.
714 	 * If nothing happens soon, send when timer expires:
715 	 * if window is nonzero, transmit what we can,
716 	 * otherwise force out a byte.
717 	 */
718 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
719 	    !tcp_timer_active(tp, TT_PERSIST)) {
720 		tp->t_rxtshift = 0;
721 		tcp_setpersist(tp);
722 	}
723 
724 	/*
725 	 * No reason to send a segment, just return.
726 	 */
727 just_return:
728 	SOCKBUF_UNLOCK(&so->so_snd);
729 	return (0);
730 
731 send:
732 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
733 	if (len > 0) {
734 		if (len >= tp->t_maxseg)
735 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
736 		else
737 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
738 	}
739 	/*
740 	 * Before ESTABLISHED, force sending of initial options
741 	 * unless TCP set not to do any options.
742 	 * NOTE: we assume that the IP/TCP header plus TCP options
743 	 * always fit in a single mbuf, leaving room for a maximum
744 	 * link header, i.e.
745 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
746 	 */
747 	optlen = 0;
748 #ifdef INET6
749 	if (isipv6)
750 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
751 	else
752 #endif
753 		hdrlen = sizeof (struct tcpiphdr);
754 
755 	/*
756 	 * Compute options for segment.
757 	 * We only have to care about SYN and established connection
758 	 * segments.  Options for SYN-ACK segments are handled in TCP
759 	 * syncache.
760 	 */
761 	to.to_flags = 0;
762 	if ((tp->t_flags & TF_NOOPT) == 0) {
763 		/* Maximum segment size. */
764 		if (flags & TH_SYN) {
765 			tp->snd_nxt = tp->iss;
766 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
767 			to.to_flags |= TOF_MSS;
768 
769 			/*
770 			 * On SYN or SYN|ACK transmits on TFO connections,
771 			 * only include the TFO option if it is not a
772 			 * retransmit, as the presence of the TFO option may
773 			 * have caused the original SYN or SYN|ACK to have
774 			 * been dropped by a middlebox.
775 			 */
776 			if (IS_FASTOPEN(tp->t_flags) &&
777 			    (tp->t_rxtshift == 0)) {
778 				if (tp->t_state == TCPS_SYN_RECEIVED) {
779 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
780 					to.to_tfo_cookie =
781 					    (u_int8_t *)&tp->t_tfo_cookie.server;
782 					to.to_flags |= TOF_FASTOPEN;
783 					wanted_cookie = 1;
784 				} else if (tp->t_state == TCPS_SYN_SENT) {
785 					to.to_tfo_len =
786 					    tp->t_tfo_client_cookie_len;
787 					to.to_tfo_cookie =
788 					    tp->t_tfo_cookie.client;
789 					to.to_flags |= TOF_FASTOPEN;
790 					wanted_cookie = 1;
791 					/*
792 					 * If we wind up having more data to
793 					 * send with the SYN than can fit in
794 					 * one segment, don't send any more
795 					 * until the SYN|ACK comes back from
796 					 * the other end.
797 					 */
798 					dont_sendalot = 1;
799 				}
800 			}
801 		}
802 		/* Window scaling. */
803 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
804 			to.to_wscale = tp->request_r_scale;
805 			to.to_flags |= TOF_SCALE;
806 		}
807 		/* Timestamps. */
808 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
809 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
810 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
811 			to.to_tsecr = tp->ts_recent;
812 			to.to_flags |= TOF_TS;
813 		}
814 
815 		/* Set receive buffer autosizing timestamp. */
816 		if (tp->rfbuf_ts == 0 &&
817 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
818 			tp->rfbuf_ts = tcp_ts_getticks();
819 
820 		/* Selective ACK's. */
821 		if (tp->t_flags & TF_SACK_PERMIT) {
822 			if (flags & TH_SYN)
823 				to.to_flags |= TOF_SACKPERM;
824 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
825 			    (tp->t_flags & TF_SACK_PERMIT) &&
826 			    tp->rcv_numsacks > 0) {
827 				to.to_flags |= TOF_SACK;
828 				to.to_nsacks = tp->rcv_numsacks;
829 				to.to_sacks = (u_char *)tp->sackblks;
830 			}
831 		}
832 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
833 		/* TCP-MD5 (RFC2385). */
834 		/*
835 		 * Check that TCP_MD5SIG is enabled in tcpcb to
836 		 * account the size needed to set this TCP option.
837 		 */
838 		if (tp->t_flags & TF_SIGNATURE)
839 			to.to_flags |= TOF_SIGNATURE;
840 #endif /* TCP_SIGNATURE */
841 
842 		/* Processing the options. */
843 		hdrlen += optlen = tcp_addoptions(&to, opt);
844 		/*
845 		 * If we wanted a TFO option to be added, but it was unable
846 		 * to fit, ensure no data is sent.
847 		 */
848 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
849 		    !(to.to_flags & TOF_FASTOPEN))
850 			len = 0;
851 	}
852 
853 	/*
854 	 * Adjust data length if insertion of options will
855 	 * bump the packet length beyond the t_maxseg length.
856 	 * Clear the FIN bit because we cut off the tail of
857 	 * the segment.
858 	 */
859 	if (len + optlen + ipoptlen > tp->t_maxseg) {
860 		flags &= ~TH_FIN;
861 
862 		if (tso) {
863 			u_int if_hw_tsomax;
864 			u_int if_hw_tsomaxsegcount;
865 			u_int if_hw_tsomaxsegsize;
866 			struct mbuf *mb;
867 			u_int moff;
868 			int max_len;
869 
870 			/* extract TSO information */
871 			if_hw_tsomax = tp->t_tsomax;
872 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
873 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
874 
875 			/*
876 			 * Limit a TSO burst to prevent it from
877 			 * overflowing or exceeding the maximum length
878 			 * allowed by the network interface:
879 			 */
880 			KASSERT(ipoptlen == 0,
881 			    ("%s: TSO can't do IP options", __func__));
882 
883 			/*
884 			 * Check if we should limit by maximum payload
885 			 * length:
886 			 */
887 			if (if_hw_tsomax != 0) {
888 				/* compute maximum TSO length */
889 				max_len = (if_hw_tsomax - hdrlen -
890 				    max_linkhdr);
891 				if (max_len <= 0) {
892 					len = 0;
893 				} else if (len > max_len) {
894 					sendalot = 1;
895 					len = max_len;
896 				}
897 			}
898 
899 			/*
900 			 * Check if we should limit by maximum segment
901 			 * size and count:
902 			 */
903 			if (if_hw_tsomaxsegcount != 0 &&
904 			    if_hw_tsomaxsegsize != 0) {
905 				/*
906 				 * Subtract one segment for the LINK
907 				 * and TCP/IP headers mbuf that will
908 				 * be prepended to this mbuf chain
909 				 * after the code in this section
910 				 * limits the number of mbufs in the
911 				 * chain to if_hw_tsomaxsegcount.
912 				 */
913 				if_hw_tsomaxsegcount -= 1;
914 				max_len = 0;
915 				mb = sbsndmbuf(&so->so_snd, off, &moff);
916 
917 				while (mb != NULL && max_len < len) {
918 					u_int mlen;
919 					u_int frags;
920 
921 					/*
922 					 * Get length of mbuf fragment
923 					 * and how many hardware frags,
924 					 * rounded up, it would use:
925 					 */
926 					mlen = (mb->m_len - moff);
927 					frags = howmany(mlen,
928 					    if_hw_tsomaxsegsize);
929 
930 					/* Handle special case: Zero Length Mbuf */
931 					if (frags == 0)
932 						frags = 1;
933 
934 					/*
935 					 * Check if the fragment limit
936 					 * will be reached or exceeded:
937 					 */
938 					if (frags >= if_hw_tsomaxsegcount) {
939 						max_len += min(mlen,
940 						    if_hw_tsomaxsegcount *
941 						    if_hw_tsomaxsegsize);
942 						break;
943 					}
944 					max_len += mlen;
945 					if_hw_tsomaxsegcount -= frags;
946 					moff = 0;
947 					mb = mb->m_next;
948 				}
949 				if (max_len <= 0) {
950 					len = 0;
951 				} else if (len > max_len) {
952 					sendalot = 1;
953 					len = max_len;
954 				}
955 			}
956 
957 			/*
958 			 * Prevent the last segment from being
959 			 * fractional unless the send sockbuf can be
960 			 * emptied:
961 			 */
962 			max_len = (tp->t_maxseg - optlen);
963 			if (((uint32_t)off + (uint32_t)len) <
964 			    sbavail(&so->so_snd)) {
965 				moff = len % max_len;
966 				if (moff != 0) {
967 					len -= moff;
968 					sendalot = 1;
969 				}
970 			}
971 
972 			/*
973 			 * In case there are too many small fragments
974 			 * don't use TSO:
975 			 */
976 			if (len <= max_len) {
977 				len = max_len;
978 				sendalot = 1;
979 				tso = 0;
980 			}
981 
982 			/*
983 			 * Send the FIN in a separate segment
984 			 * after the bulk sending is done.
985 			 * We don't trust the TSO implementations
986 			 * to clear the FIN flag on all but the
987 			 * last segment.
988 			 */
989 			if (tp->t_flags & TF_NEEDFIN)
990 				sendalot = 1;
991 
992 		} else {
993 			len = tp->t_maxseg - optlen - ipoptlen;
994 			sendalot = 1;
995 			if (dont_sendalot)
996 				sendalot = 0;
997 		}
998 	} else
999 		tso = 0;
1000 
1001 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1002 	    ("%s: len > IP_MAXPACKET", __func__));
1003 
1004 /*#ifdef DIAGNOSTIC*/
1005 #ifdef INET6
1006 	if (max_linkhdr + hdrlen > MCLBYTES)
1007 #else
1008 	if (max_linkhdr + hdrlen > MHLEN)
1009 #endif
1010 		panic("tcphdr too big");
1011 /*#endif*/
1012 
1013 	/*
1014 	 * This KASSERT is here to catch edge cases at a well defined place.
1015 	 * Before, those had triggered (random) panic conditions further down.
1016 	 */
1017 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1018 
1019 	/*
1020 	 * Grab a header mbuf, attaching a copy of data to
1021 	 * be transmitted, and initialize the header from
1022 	 * the template for sends on this connection.
1023 	 */
1024 	if (len) {
1025 		struct mbuf *mb;
1026 		u_int moff;
1027 
1028 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1029 			TCPSTAT_INC(tcps_sndprobe);
1030 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1031 			tp->t_sndrexmitpack++;
1032 			TCPSTAT_INC(tcps_sndrexmitpack);
1033 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1034 		} else {
1035 			TCPSTAT_INC(tcps_sndpack);
1036 			TCPSTAT_ADD(tcps_sndbyte, len);
1037 		}
1038 #ifdef INET6
1039 		if (MHLEN < hdrlen + max_linkhdr)
1040 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1041 		else
1042 #endif
1043 			m = m_gethdr(M_NOWAIT, MT_DATA);
1044 
1045 		if (m == NULL) {
1046 			SOCKBUF_UNLOCK(&so->so_snd);
1047 			error = ENOBUFS;
1048 			sack_rxmit = 0;
1049 			goto out;
1050 		}
1051 
1052 		m->m_data += max_linkhdr;
1053 		m->m_len = hdrlen;
1054 
1055 		/*
1056 		 * Start the m_copy functions from the closest mbuf
1057 		 * to the offset in the socket buffer chain.
1058 		 */
1059 		mb = sbsndptr(&so->so_snd, off, len, &moff);
1060 
1061 		if (len <= MHLEN - hdrlen - max_linkhdr) {
1062 			m_copydata(mb, moff, len,
1063 			    mtod(m, caddr_t) + hdrlen);
1064 			m->m_len += len;
1065 		} else {
1066 			m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1067 			if (m->m_next == NULL) {
1068 				SOCKBUF_UNLOCK(&so->so_snd);
1069 				(void) m_free(m);
1070 				error = ENOBUFS;
1071 				sack_rxmit = 0;
1072 				goto out;
1073 			}
1074 		}
1075 
1076 		/*
1077 		 * If we're sending everything we've got, set PUSH.
1078 		 * (This will keep happy those implementations which only
1079 		 * give data to the user when a buffer fills or
1080 		 * a PUSH comes in.)
1081 		 */
1082 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1083 		    !(flags & TH_SYN))
1084 			flags |= TH_PUSH;
1085 		SOCKBUF_UNLOCK(&so->so_snd);
1086 	} else {
1087 		SOCKBUF_UNLOCK(&so->so_snd);
1088 		if (tp->t_flags & TF_ACKNOW)
1089 			TCPSTAT_INC(tcps_sndacks);
1090 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1091 			TCPSTAT_INC(tcps_sndctrl);
1092 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1093 			TCPSTAT_INC(tcps_sndurg);
1094 		else
1095 			TCPSTAT_INC(tcps_sndwinup);
1096 
1097 		m = m_gethdr(M_NOWAIT, MT_DATA);
1098 		if (m == NULL) {
1099 			error = ENOBUFS;
1100 			sack_rxmit = 0;
1101 			goto out;
1102 		}
1103 #ifdef INET6
1104 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1105 		    MHLEN >= hdrlen) {
1106 			M_ALIGN(m, hdrlen);
1107 		} else
1108 #endif
1109 		m->m_data += max_linkhdr;
1110 		m->m_len = hdrlen;
1111 	}
1112 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1113 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1114 #ifdef MAC
1115 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1116 #endif
1117 #ifdef INET6
1118 	if (isipv6) {
1119 		ip6 = mtod(m, struct ip6_hdr *);
1120 		th = (struct tcphdr *)(ip6 + 1);
1121 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1122 	} else
1123 #endif /* INET6 */
1124 	{
1125 		ip = mtod(m, struct ip *);
1126 #ifdef TCPDEBUG
1127 		ipov = (struct ipovly *)ip;
1128 #endif
1129 		th = (struct tcphdr *)(ip + 1);
1130 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1131 	}
1132 
1133 	/*
1134 	 * Fill in fields, remembering maximum advertised
1135 	 * window for use in delaying messages about window sizes.
1136 	 * If resending a FIN, be sure not to use a new sequence number.
1137 	 */
1138 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1139 	    tp->snd_nxt == tp->snd_max)
1140 		tp->snd_nxt--;
1141 	/*
1142 	 * If we are starting a connection, send ECN setup
1143 	 * SYN packet. If we are on a retransmit, we may
1144 	 * resend those bits a number of times as per
1145 	 * RFC 3168.
1146 	 */
1147 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1148 		if (tp->t_rxtshift >= 1) {
1149 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1150 				flags |= TH_ECE|TH_CWR;
1151 		} else
1152 			flags |= TH_ECE|TH_CWR;
1153 	}
1154 
1155 	if (tp->t_state == TCPS_ESTABLISHED &&
1156 	    (tp->t_flags & TF_ECN_PERMIT)) {
1157 		/*
1158 		 * If the peer has ECN, mark data packets with
1159 		 * ECN capable transmission (ECT).
1160 		 * Ignore pure ack packets, retransmissions and window probes.
1161 		 */
1162 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1163 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1164 #ifdef INET6
1165 			if (isipv6)
1166 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1167 			else
1168 #endif
1169 				ip->ip_tos |= IPTOS_ECN_ECT0;
1170 			TCPSTAT_INC(tcps_ecn_ect0);
1171 		}
1172 
1173 		/*
1174 		 * Reply with proper ECN notifications.
1175 		 */
1176 		if (tp->t_flags & TF_ECN_SND_CWR) {
1177 			flags |= TH_CWR;
1178 			tp->t_flags &= ~TF_ECN_SND_CWR;
1179 		}
1180 		if (tp->t_flags & TF_ECN_SND_ECE)
1181 			flags |= TH_ECE;
1182 	}
1183 
1184 	/*
1185 	 * If we are doing retransmissions, then snd_nxt will
1186 	 * not reflect the first unsent octet.  For ACK only
1187 	 * packets, we do not want the sequence number of the
1188 	 * retransmitted packet, we want the sequence number
1189 	 * of the next unsent octet.  So, if there is no data
1190 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1191 	 * when filling in ti_seq.  But if we are in persist
1192 	 * state, snd_max might reflect one byte beyond the
1193 	 * right edge of the window, so use snd_nxt in that
1194 	 * case, since we know we aren't doing a retransmission.
1195 	 * (retransmit and persist are mutually exclusive...)
1196 	 */
1197 	if (sack_rxmit == 0) {
1198 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1199 		    tcp_timer_active(tp, TT_PERSIST))
1200 			th->th_seq = htonl(tp->snd_nxt);
1201 		else
1202 			th->th_seq = htonl(tp->snd_max);
1203 	} else {
1204 		th->th_seq = htonl(p->rxmit);
1205 		p->rxmit += len;
1206 		tp->sackhint.sack_bytes_rexmit += len;
1207 	}
1208 	th->th_ack = htonl(tp->rcv_nxt);
1209 	if (optlen) {
1210 		bcopy(opt, th + 1, optlen);
1211 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1212 	}
1213 	th->th_flags = flags;
1214 	/*
1215 	 * Calculate receive window.  Don't shrink window,
1216 	 * but avoid silly window syndrome.
1217 	 */
1218 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1219 	    recwin < tp->t_maxseg)
1220 		recwin = 0;
1221 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1222 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
1223 		recwin = (tp->rcv_adv - tp->rcv_nxt);
1224 
1225 	/*
1226 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1227 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1228 	 * case is handled in syncache.
1229 	 */
1230 	if (flags & TH_SYN)
1231 		th->th_win = htons((u_short)
1232 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1233 	else
1234 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1235 
1236 	/*
1237 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1238 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1239 	 * flag tells soreceive() to disable delayed acknowledgements when
1240 	 * draining the buffer.  This can occur if the receiver is attempting
1241 	 * to read more data than can be buffered prior to transmitting on
1242 	 * the connection.
1243 	 */
1244 	if (th->th_win == 0) {
1245 		tp->t_sndzerowin++;
1246 		tp->t_flags |= TF_RXWIN0SENT;
1247 	} else
1248 		tp->t_flags &= ~TF_RXWIN0SENT;
1249 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1250 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1251 		th->th_flags |= TH_URG;
1252 	} else
1253 		/*
1254 		 * If no urgent pointer to send, then we pull
1255 		 * the urgent pointer to the left edge of the send window
1256 		 * so that it doesn't drift into the send window on sequence
1257 		 * number wraparound.
1258 		 */
1259 		tp->snd_up = tp->snd_una;		/* drag it along */
1260 
1261 	/*
1262 	 * Put TCP length in extended header, and then
1263 	 * checksum extended header and data.
1264 	 */
1265 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1266 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1267 
1268 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1269 	if (to.to_flags & TOF_SIGNATURE) {
1270 		/*
1271 		 * Calculate MD5 signature and put it into the place
1272 		 * determined before.
1273 		 * NOTE: since TCP options buffer doesn't point into
1274 		 * mbuf's data, calculate offset and use it.
1275 		 */
1276 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1277 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1278 			/*
1279 			 * Do not send segment if the calculation of MD5
1280 			 * digest has failed.
1281 			 */
1282 			m_freem(m);
1283 			goto out;
1284 		}
1285 	}
1286 #endif
1287 #ifdef INET6
1288 	if (isipv6) {
1289 		/*
1290 		 * There is no need to fill in ip6_plen right now.
1291 		 * It will be filled later by ip6_output.
1292 		 */
1293 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1294 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1295 		    optlen + len, IPPROTO_TCP, 0);
1296 	}
1297 #endif
1298 #if defined(INET6) && defined(INET)
1299 	else
1300 #endif
1301 #ifdef INET
1302 	{
1303 		m->m_pkthdr.csum_flags = CSUM_TCP;
1304 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1305 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1306 
1307 		/* IP version must be set here for ipv4/ipv6 checking later */
1308 		KASSERT(ip->ip_v == IPVERSION,
1309 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1310 	}
1311 #endif
1312 
1313 	/*
1314 	 * Enable TSO and specify the size of the segments.
1315 	 * The TCP pseudo header checksum is always provided.
1316 	 */
1317 	if (tso) {
1318 		KASSERT(len > tp->t_maxseg - optlen,
1319 		    ("%s: len <= tso_segsz", __func__));
1320 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1321 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1322 	}
1323 
1324 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1325 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1326 	    ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1327 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1328 #else
1329 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1330 	    ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1331 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1332 #endif
1333 
1334 #ifdef TCP_HHOOK
1335 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1336 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1337 #endif
1338 
1339 #ifdef TCPDEBUG
1340 	/*
1341 	 * Trace.
1342 	 */
1343 	if (so->so_options & SO_DEBUG) {
1344 		u_short save = 0;
1345 #ifdef INET6
1346 		if (!isipv6)
1347 #endif
1348 		{
1349 			save = ipov->ih_len;
1350 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1351 		}
1352 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1353 #ifdef INET6
1354 		if (!isipv6)
1355 #endif
1356 		ipov->ih_len = save;
1357 	}
1358 #endif /* TCPDEBUG */
1359 	TCP_PROBE3(debug__output, tp, th, m);
1360 
1361 	/*
1362 	 * Fill in IP length and desired time to live and
1363 	 * send to IP level.  There should be a better way
1364 	 * to handle ttl and tos; we could keep them in
1365 	 * the template, but need a way to checksum without them.
1366 	 */
1367 	/*
1368 	 * m->m_pkthdr.len should have been set before checksum calculation,
1369 	 * because in6_cksum() need it.
1370 	 */
1371 #ifdef INET6
1372 	if (isipv6) {
1373 		/*
1374 		 * we separately set hoplimit for every segment, since the
1375 		 * user might want to change the value via setsockopt.
1376 		 * Also, desired default hop limit might be changed via
1377 		 * Neighbor Discovery.
1378 		 */
1379 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1380 
1381 		/*
1382 		 * Set the packet size here for the benefit of DTrace probes.
1383 		 * ip6_output() will set it properly; it's supposed to include
1384 		 * the option header lengths as well.
1385 		 */
1386 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1387 
1388 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1389 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1390 		else
1391 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1392 
1393 		if (tp->t_state == TCPS_SYN_SENT)
1394 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1395 
1396 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1397 
1398 #ifdef TCPPCAP
1399 		/* Save packet, if requested. */
1400 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1401 #endif
1402 
1403 		/* TODO: IPv6 IP6TOS_ECT bit on */
1404 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1405 		    &tp->t_inpcb->inp_route6,
1406 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1407 		    NULL, NULL, tp->t_inpcb);
1408 
1409 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1410 			mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1411 	}
1412 #endif /* INET6 */
1413 #if defined(INET) && defined(INET6)
1414 	else
1415 #endif
1416 #ifdef INET
1417     {
1418 	ip->ip_len = htons(m->m_pkthdr.len);
1419 #ifdef INET6
1420 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1421 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1422 #endif /* INET6 */
1423 	/*
1424 	 * If we do path MTU discovery, then we set DF on every packet.
1425 	 * This might not be the best thing to do according to RFC3390
1426 	 * Section 2. However the tcp hostcache migitates the problem
1427 	 * so it affects only the first tcp connection with a host.
1428 	 *
1429 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1430 	 */
1431 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1432 		ip->ip_off |= htons(IP_DF);
1433 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1434 	} else {
1435 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1436 	}
1437 
1438 	if (tp->t_state == TCPS_SYN_SENT)
1439 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1440 
1441 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1442 
1443 #ifdef TCPPCAP
1444 	/* Save packet, if requested. */
1445 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1446 #endif
1447 
1448 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1449 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1450 	    tp->t_inpcb);
1451 
1452 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1453 		mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1454     }
1455 #endif /* INET */
1456 
1457 out:
1458 	/*
1459 	 * In transmit state, time the transmission and arrange for
1460 	 * the retransmit.  In persist state, just set snd_max.
1461 	 */
1462 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1463 	    !tcp_timer_active(tp, TT_PERSIST)) {
1464 		tcp_seq startseq = tp->snd_nxt;
1465 
1466 		/*
1467 		 * Advance snd_nxt over sequence space of this segment.
1468 		 */
1469 		if (flags & (TH_SYN|TH_FIN)) {
1470 			if (flags & TH_SYN)
1471 				tp->snd_nxt++;
1472 			if (flags & TH_FIN) {
1473 				tp->snd_nxt++;
1474 				tp->t_flags |= TF_SENTFIN;
1475 			}
1476 		}
1477 		if (sack_rxmit)
1478 			goto timer;
1479 		tp->snd_nxt += len;
1480 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1481 			tp->snd_max = tp->snd_nxt;
1482 			/*
1483 			 * Time this transmission if not a retransmission and
1484 			 * not currently timing anything.
1485 			 */
1486 			if (tp->t_rtttime == 0) {
1487 				tp->t_rtttime = ticks;
1488 				tp->t_rtseq = startseq;
1489 				TCPSTAT_INC(tcps_segstimed);
1490 			}
1491 		}
1492 
1493 		/*
1494 		 * Set retransmit timer if not currently set,
1495 		 * and not doing a pure ack or a keep-alive probe.
1496 		 * Initial value for retransmit timer is smoothed
1497 		 * round-trip time + 2 * round-trip time variance.
1498 		 * Initialize shift counter which is used for backoff
1499 		 * of retransmit time.
1500 		 */
1501 timer:
1502 		if (!tcp_timer_active(tp, TT_REXMT) &&
1503 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1504 		     (tp->snd_nxt != tp->snd_una))) {
1505 			if (tcp_timer_active(tp, TT_PERSIST)) {
1506 				tcp_timer_activate(tp, TT_PERSIST, 0);
1507 				tp->t_rxtshift = 0;
1508 			}
1509 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1510 		} else if (len == 0 && sbavail(&so->so_snd) &&
1511 		    !tcp_timer_active(tp, TT_REXMT) &&
1512 		    !tcp_timer_active(tp, TT_PERSIST)) {
1513 			/*
1514 			 * Avoid a situation where we do not set persist timer
1515 			 * after a zero window condition. For example:
1516 			 * 1) A -> B: packet with enough data to fill the window
1517 			 * 2) B -> A: ACK for #1 + new data (0 window
1518 			 *    advertisement)
1519 			 * 3) A -> B: ACK for #2, 0 len packet
1520 			 *
1521 			 * In this case, A will not activate the persist timer,
1522 			 * because it chose to send a packet. Unless tcp_output
1523 			 * is called for some other reason (delayed ack timer,
1524 			 * another input packet from B, socket syscall), A will
1525 			 * not send zero window probes.
1526 			 *
1527 			 * So, if you send a 0-length packet, but there is data
1528 			 * in the socket buffer, and neither the rexmt or
1529 			 * persist timer is already set, then activate the
1530 			 * persist timer.
1531 			 */
1532 			tp->t_rxtshift = 0;
1533 			tcp_setpersist(tp);
1534 		}
1535 	} else {
1536 		/*
1537 		 * Persist case, update snd_max but since we are in
1538 		 * persist mode (no window) we do not update snd_nxt.
1539 		 */
1540 		int xlen = len;
1541 		if (flags & TH_SYN)
1542 			++xlen;
1543 		if (flags & TH_FIN) {
1544 			++xlen;
1545 			tp->t_flags |= TF_SENTFIN;
1546 		}
1547 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1548 			tp->snd_max = tp->snd_nxt + xlen;
1549 	}
1550 
1551 	if (error) {
1552 
1553 		/*
1554 		 * We know that the packet was lost, so back out the
1555 		 * sequence number advance, if any.
1556 		 *
1557 		 * If the error is EPERM the packet got blocked by the
1558 		 * local firewall.  Normally we should terminate the
1559 		 * connection but the blocking may have been spurious
1560 		 * due to a firewall reconfiguration cycle.  So we treat
1561 		 * it like a packet loss and let the retransmit timer and
1562 		 * timeouts do their work over time.
1563 		 * XXX: It is a POLA question whether calling tcp_drop right
1564 		 * away would be the really correct behavior instead.
1565 		 */
1566 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1567 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1568 		    ((flags & TH_SYN) == 0) &&
1569 		    (error != EPERM)) {
1570 			if (sack_rxmit) {
1571 				p->rxmit -= len;
1572 				tp->sackhint.sack_bytes_rexmit -= len;
1573 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1574 				    ("sackhint bytes rtx >= 0"));
1575 			} else
1576 				tp->snd_nxt -= len;
1577 		}
1578 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1579 		switch (error) {
1580 		case EACCES:
1581 			tp->t_softerror = error;
1582 			return (0);
1583 		case EPERM:
1584 			tp->t_softerror = error;
1585 			return (error);
1586 		case ENOBUFS:
1587 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1588 			tp->snd_cwnd = tp->t_maxseg;
1589 			return (0);
1590 		case EMSGSIZE:
1591 			/*
1592 			 * For some reason the interface we used initially
1593 			 * to send segments changed to another or lowered
1594 			 * its MTU.
1595 			 * If TSO was active we either got an interface
1596 			 * without TSO capabilits or TSO was turned off.
1597 			 * If we obtained mtu from ip_output() then update
1598 			 * it and try again.
1599 			 */
1600 			if (tso)
1601 				tp->t_flags &= ~TF_TSO;
1602 			if (mtu != 0) {
1603 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1604 				goto again;
1605 			}
1606 			return (error);
1607 		case EHOSTDOWN:
1608 		case EHOSTUNREACH:
1609 		case ENETDOWN:
1610 		case ENETUNREACH:
1611 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1612 				tp->t_softerror = error;
1613 				return (0);
1614 			}
1615 			/* FALLTHROUGH */
1616 		default:
1617 			return (error);
1618 		}
1619 	}
1620 	TCPSTAT_INC(tcps_sndtotal);
1621 
1622 	/*
1623 	 * Data sent (as far as we can tell).
1624 	 * If this advertises a larger window than any other segment,
1625 	 * then remember the size of the advertised window.
1626 	 * Any pending ACK has now been sent.
1627 	 */
1628 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1629 		tp->rcv_adv = tp->rcv_nxt + recwin;
1630 	tp->last_ack_sent = tp->rcv_nxt;
1631 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1632 	if (tcp_timer_active(tp, TT_DELACK))
1633 		tcp_timer_activate(tp, TT_DELACK, 0);
1634 #if 0
1635 	/*
1636 	 * This completely breaks TCP if newreno is turned on.  What happens
1637 	 * is that if delayed-acks are turned on on the receiver, this code
1638 	 * on the transmitter effectively destroys the TCP window, forcing
1639 	 * it to four packets (1.5Kx4 = 6K window).
1640 	 */
1641 	if (sendalot && --maxburst)
1642 		goto again;
1643 #endif
1644 	if (sendalot)
1645 		goto again;
1646 	return (0);
1647 }
1648 
1649 void
1650 tcp_setpersist(struct tcpcb *tp)
1651 {
1652 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1653 	int tt;
1654 
1655 	tp->t_flags &= ~TF_PREVVALID;
1656 	if (tcp_timer_active(tp, TT_REXMT))
1657 		panic("tcp_setpersist: retransmit pending");
1658 	/*
1659 	 * Start/restart persistence timer.
1660 	 */
1661 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1662 		      tcp_persmin, tcp_persmax);
1663 	tcp_timer_activate(tp, TT_PERSIST, tt);
1664 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1665 		tp->t_rxtshift++;
1666 }
1667 
1668 /*
1669  * Insert TCP options according to the supplied parameters to the place
1670  * optp in a consistent way.  Can handle unaligned destinations.
1671  *
1672  * The order of the option processing is crucial for optimal packing and
1673  * alignment for the scarce option space.
1674  *
1675  * The optimal order for a SYN/SYN-ACK segment is:
1676  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1677  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1678  *
1679  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1680  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1681  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1682  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1683  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1684  */
1685 int
1686 tcp_addoptions(struct tcpopt *to, u_char *optp)
1687 {
1688 	u_int32_t mask, optlen = 0;
1689 
1690 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1691 		if ((to->to_flags & mask) != mask)
1692 			continue;
1693 		if (optlen == TCP_MAXOLEN)
1694 			break;
1695 		switch (to->to_flags & mask) {
1696 		case TOF_MSS:
1697 			while (optlen % 4) {
1698 				optlen += TCPOLEN_NOP;
1699 				*optp++ = TCPOPT_NOP;
1700 			}
1701 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1702 				continue;
1703 			optlen += TCPOLEN_MAXSEG;
1704 			*optp++ = TCPOPT_MAXSEG;
1705 			*optp++ = TCPOLEN_MAXSEG;
1706 			to->to_mss = htons(to->to_mss);
1707 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1708 			optp += sizeof(to->to_mss);
1709 			break;
1710 		case TOF_SCALE:
1711 			while (!optlen || optlen % 2 != 1) {
1712 				optlen += TCPOLEN_NOP;
1713 				*optp++ = TCPOPT_NOP;
1714 			}
1715 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1716 				continue;
1717 			optlen += TCPOLEN_WINDOW;
1718 			*optp++ = TCPOPT_WINDOW;
1719 			*optp++ = TCPOLEN_WINDOW;
1720 			*optp++ = to->to_wscale;
1721 			break;
1722 		case TOF_SACKPERM:
1723 			while (optlen % 2) {
1724 				optlen += TCPOLEN_NOP;
1725 				*optp++ = TCPOPT_NOP;
1726 			}
1727 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1728 				continue;
1729 			optlen += TCPOLEN_SACK_PERMITTED;
1730 			*optp++ = TCPOPT_SACK_PERMITTED;
1731 			*optp++ = TCPOLEN_SACK_PERMITTED;
1732 			break;
1733 		case TOF_TS:
1734 			while (!optlen || optlen % 4 != 2) {
1735 				optlen += TCPOLEN_NOP;
1736 				*optp++ = TCPOPT_NOP;
1737 			}
1738 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1739 				continue;
1740 			optlen += TCPOLEN_TIMESTAMP;
1741 			*optp++ = TCPOPT_TIMESTAMP;
1742 			*optp++ = TCPOLEN_TIMESTAMP;
1743 			to->to_tsval = htonl(to->to_tsval);
1744 			to->to_tsecr = htonl(to->to_tsecr);
1745 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1746 			optp += sizeof(to->to_tsval);
1747 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1748 			optp += sizeof(to->to_tsecr);
1749 			break;
1750 		case TOF_SIGNATURE:
1751 			{
1752 			int siglen = TCPOLEN_SIGNATURE - 2;
1753 
1754 			while (!optlen || optlen % 4 != 2) {
1755 				optlen += TCPOLEN_NOP;
1756 				*optp++ = TCPOPT_NOP;
1757 			}
1758 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1759 				to->to_flags &= ~TOF_SIGNATURE;
1760 				continue;
1761 			}
1762 			optlen += TCPOLEN_SIGNATURE;
1763 			*optp++ = TCPOPT_SIGNATURE;
1764 			*optp++ = TCPOLEN_SIGNATURE;
1765 			to->to_signature = optp;
1766 			while (siglen--)
1767 				 *optp++ = 0;
1768 			break;
1769 			}
1770 		case TOF_SACK:
1771 			{
1772 			int sackblks = 0;
1773 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1774 			tcp_seq sack_seq;
1775 
1776 			while (!optlen || optlen % 4 != 2) {
1777 				optlen += TCPOLEN_NOP;
1778 				*optp++ = TCPOPT_NOP;
1779 			}
1780 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1781 				continue;
1782 			optlen += TCPOLEN_SACKHDR;
1783 			*optp++ = TCPOPT_SACK;
1784 			sackblks = min(to->to_nsacks,
1785 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1786 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1787 			while (sackblks--) {
1788 				sack_seq = htonl(sack->start);
1789 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1790 				optp += sizeof(sack_seq);
1791 				sack_seq = htonl(sack->end);
1792 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1793 				optp += sizeof(sack_seq);
1794 				optlen += TCPOLEN_SACK;
1795 				sack++;
1796 			}
1797 			TCPSTAT_INC(tcps_sack_send_blocks);
1798 			break;
1799 			}
1800 		case TOF_FASTOPEN:
1801 			{
1802 			int total_len;
1803 
1804 			/* XXX is there any point to aligning this option? */
1805 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1806 			if (TCP_MAXOLEN - optlen < total_len) {
1807 				to->to_flags &= ~TOF_FASTOPEN;
1808 				continue;
1809 			}
1810 			*optp++ = TCPOPT_FAST_OPEN;
1811 			*optp++ = total_len;
1812 			if (to->to_tfo_len > 0) {
1813 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1814 				optp += to->to_tfo_len;
1815 			}
1816 			optlen += total_len;
1817 			break;
1818 			}
1819 		default:
1820 			panic("%s: unknown TCP option type", __func__);
1821 			break;
1822 		}
1823 	}
1824 
1825 	/* Terminate and pad TCP options to a 4 byte boundary. */
1826 	if (optlen % 4) {
1827 		optlen += TCPOLEN_EOL;
1828 		*optp++ = TCPOPT_EOL;
1829 	}
1830 	/*
1831 	 * According to RFC 793 (STD0007):
1832 	 *   "The content of the header beyond the End-of-Option option
1833 	 *    must be header padding (i.e., zero)."
1834 	 *   and later: "The padding is composed of zeros."
1835 	 */
1836 	while (optlen % 4) {
1837 		optlen += TCPOLEN_PAD;
1838 		*optp++ = TCPOPT_PAD;
1839 	}
1840 
1841 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1842 	return (optlen);
1843 }
1844 
1845 void
1846 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1847 {
1848 
1849 	/*
1850 	 * Automatic sizing of send socket buffer.  Often the send buffer
1851 	 * size is not optimally adjusted to the actual network conditions
1852 	 * at hand (delay bandwidth product).  Setting the buffer size too
1853 	 * small limits throughput on links with high bandwidth and high
1854 	 * delay (eg. trans-continental/oceanic links).  Setting the
1855 	 * buffer size too big consumes too much real kernel memory,
1856 	 * especially with many connections on busy servers.
1857 	 *
1858 	 * The criteria to step up the send buffer one notch are:
1859 	 *  1. receive window of remote host is larger than send buffer
1860 	 *     (with a fudge factor of 5/4th);
1861 	 *  2. send buffer is filled to 7/8th with data (so we actually
1862 	 *     have data to make use of it);
1863 	 *  3. send buffer fill has not hit maximal automatic size;
1864 	 *  4. our send window (slow start and cogestion controlled) is
1865 	 *     larger than sent but unacknowledged data in send buffer.
1866 	 *
1867 	 * The remote host receive window scaling factor may limit the
1868 	 * growing of the send buffer before it reaches its allowed
1869 	 * maximum.
1870 	 *
1871 	 * It scales directly with slow start or congestion window
1872 	 * and does at most one step per received ACK.  This fast
1873 	 * scaling has the drawback of growing the send buffer beyond
1874 	 * what is strictly necessary to make full use of a given
1875 	 * delay*bandwidth product.  However testing has shown this not
1876 	 * to be much of an problem.  At worst we are trading wasting
1877 	 * of available bandwidth (the non-use of it) for wasting some
1878 	 * socket buffer memory.
1879 	 *
1880 	 * TODO: Shrink send buffer during idle periods together
1881 	 * with congestion window.  Requires another timer.  Has to
1882 	 * wait for upcoming tcp timer rewrite.
1883 	 *
1884 	 * XXXGL: should there be used sbused() or sbavail()?
1885 	 */
1886 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
1887 		int lowat;
1888 
1889 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
1890 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
1891 		    sbused(&so->so_snd) >=
1892 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
1893 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
1894 		    sendwin >= (sbused(&so->so_snd) -
1895 		    (tp->snd_nxt - tp->snd_una))) {
1896 			if (!sbreserve_locked(&so->so_snd,
1897 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
1898 			     V_tcp_autosndbuf_max), so, curthread))
1899 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
1900 		}
1901 	}
1902 }
1903