xref: /freebsd/sys/netinet/tcp_output.c (revision f5463265955b829775bbb32e1fd0bc11dafc36ce)
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 
32 #include <sys/cdefs.h>
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_kern_tls.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/arb.h>
41 #include <sys/domain.h>
42 #ifdef TCP_HHOOK
43 #include <sys/hhook.h>
44 #endif
45 #include <sys/kernel.h>
46 #ifdef KERN_TLS
47 #include <sys/ktls.h>
48 #endif
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/qmath.h>
54 #include <sys/sdt.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/stats.h>
59 
60 #include <net/if.h>
61 #include <net/route.h>
62 #include <net/route/nhop.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_kdtrace.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_pcb.h>
70 #include <netinet/ip_var.h>
71 #include <netinet/ip_options.h>
72 #ifdef INET6
73 #include <netinet6/in6_pcb.h>
74 #include <netinet/ip6.h>
75 #include <netinet6/ip6_var.h>
76 #endif
77 #include <netinet/tcp.h>
78 #define	TCPOUTFLAGS
79 #include <netinet/tcp_fsm.h>
80 #include <netinet/tcp_seq.h>
81 #include <netinet/tcp_var.h>
82 #include <netinet/tcp_log_buf.h>
83 #include <netinet/tcp_syncache.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcpip.h>
86 #include <netinet/cc/cc.h>
87 #include <netinet/tcp_fastopen.h>
88 #ifdef TCPPCAP
89 #include <netinet/tcp_pcap.h>
90 #endif
91 #ifdef TCP_OFFLOAD
92 #include <netinet/tcp_offload.h>
93 #endif
94 #include <netinet/tcp_ecn.h>
95 
96 #include <netipsec/ipsec_support.h>
97 
98 #include <netinet/udp.h>
99 #include <netinet/udp_var.h>
100 #include <machine/in_cksum.h>
101 
102 #include <security/mac/mac_framework.h>
103 
104 VNET_DEFINE(int, path_mtu_discovery) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
106 	&VNET_NAME(path_mtu_discovery), 1,
107 	"Enable Path MTU Discovery");
108 
109 VNET_DEFINE(int, tcp_do_tso) = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
111 	&VNET_NAME(tcp_do_tso), 0,
112 	"Enable TCP Segmentation Offload");
113 
114 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
115 #define	V_tcp_sendspace	VNET(tcp_sendspace)
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
117 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
118 
119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
121 	&VNET_NAME(tcp_do_autosndbuf), 0,
122 	"Enable automatic send buffer sizing");
123 
124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
126 	&VNET_NAME(tcp_autosndbuf_inc), 0,
127 	"Incrementor step size of automatic send buffer");
128 
129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
131 	&VNET_NAME(tcp_autosndbuf_max), 0,
132 	"Max size of automatic send buffer");
133 
134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
135 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
137 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
138 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
139 
140 /*
141  * Make sure that either retransmit or persist timer is set for SYN, FIN and
142  * non-ACK.
143  */
144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
145 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
146 	    tcp_timer_active((tp), TT_REXMT) ||				\
147 	    tcp_timer_active((tp), TT_PERSIST),				\
148 	    ("neither rexmt nor persist timer is set"))
149 
150 #ifdef TCP_HHOOK
151 /*
152  * Wrapper for the TCP established output helper hook.
153  */
154 void
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->t_osd);
169 	}
170 }
171 #endif
172 
173 /*
174  * CC wrapper hook functions
175  */
176 void
177 cc_after_idle(struct tcpcb *tp)
178 {
179 	INP_WLOCK_ASSERT(tptoinpcb(tp));
180 
181 	if (CC_ALGO(tp)->after_idle != NULL)
182 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
183 }
184 
185 /*
186  * Tcp output routine: figure out what should be sent and send it.
187  */
188 int
189 tcp_default_output(struct tcpcb *tp)
190 {
191 	struct socket *so = tptosocket(tp);
192 	struct inpcb *inp = tptoinpcb(tp);
193 	int32_t len;
194 	uint32_t recwin, sendwin;
195 	uint16_t flags;
196 	int off, error = 0;	/* Keep compiler happy */
197 	u_int if_hw_tsomaxsegcount = 0;
198 	u_int if_hw_tsomaxsegsize = 0;
199 	struct mbuf *m;
200 	struct ip *ip = NULL;
201 	struct tcphdr *th;
202 	u_char opt[TCP_MAXOLEN];
203 	unsigned ipoptlen, optlen, hdrlen, ulen;
204 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
205 	unsigned ipsec_optlen = 0;
206 #endif
207 	int idle, sendalot, curticks;
208 	int sack_rxmit, sack_bytes_rxmt;
209 	struct sackhole *p;
210 	int tso, mtu;
211 	struct tcpopt to;
212 	struct udphdr *udp = NULL;
213 	struct tcp_log_buffer *lgb;
214 	unsigned int wanted_cookie = 0;
215 	unsigned int dont_sendalot = 0;
216 #if 0
217 	int maxburst = TCP_MAXBURST;
218 #endif
219 #ifdef INET6
220 	struct ip6_hdr *ip6 = NULL;
221 	const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
222 #endif
223 #ifdef KERN_TLS
224 	const bool hw_tls = tp->t_nic_ktls_xmit != 0;
225 #else
226 	const bool hw_tls = false;
227 #endif
228 
229 	NET_EPOCH_ASSERT();
230 	INP_WLOCK_ASSERT(inp);
231 
232 #ifdef TCP_OFFLOAD
233 	if (tp->t_flags & TF_TOE)
234 		return (tcp_offload_output(tp));
235 #endif
236 
237 	/*
238 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
239 	 * only allow the initial SYN or SYN|ACK and those sent
240 	 * by the retransmit timer.
241 	 */
242 	if (IS_FASTOPEN(tp->t_flags) &&
243 	    ((tp->t_state == TCPS_SYN_SENT) ||
244 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
245 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
246 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
247 		return (0);
248 
249 	/*
250 	 * Determine length of data that should be transmitted,
251 	 * and flags that will be used.
252 	 * If there is some data or critical controls (SYN, RST)
253 	 * to send, then transmit; otherwise, investigate further.
254 	 */
255 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
256 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
257 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
258 		cc_after_idle(tp);
259 	tp->t_flags &= ~TF_LASTIDLE;
260 	if (idle) {
261 		if (tp->t_flags & TF_MORETOCOME) {
262 			tp->t_flags |= TF_LASTIDLE;
263 			idle = 0;
264 		}
265 	}
266 again:
267 	/*
268 	 * If we've recently taken a timeout, snd_max will be greater than
269 	 * snd_nxt.  There may be SACK information that allows us to avoid
270 	 * resending already delivered data.  Adjust snd_nxt accordingly.
271 	 */
272 	if ((tp->t_flags & TF_SACK_PERMIT) &&
273 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
274 		tcp_sack_adjust(tp);
275 	sendalot = 0;
276 	tso = 0;
277 	mtu = 0;
278 	off = tp->snd_nxt - tp->snd_una;
279 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
280 
281 	flags = tcp_outflags[tp->t_state];
282 	/*
283 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
284 	 * to send out new data (when sendalot is 1), bypass this function.
285 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
286 	 * we're replacing a (future) new transmission with a retransmission
287 	 * now, and we previously incremented snd_cwnd in tcp_input().
288 	 */
289 	/*
290 	 * Still in sack recovery , reset rxmit flag to zero.
291 	 */
292 	sack_rxmit = 0;
293 	sack_bytes_rxmt = 0;
294 	len = 0;
295 	p = NULL;
296 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
297 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
298 		uint32_t cwin;
299 
300 		cwin =
301 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
302 		/* Do not retransmit SACK segments beyond snd_recover */
303 		if (SEQ_GT(p->end, tp->snd_recover)) {
304 			/*
305 			 * (At least) part of sack hole extends beyond
306 			 * snd_recover. Check to see if we can rexmit data
307 			 * for this hole.
308 			 */
309 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
310 				/*
311 				 * Can't rexmit any more data for this hole.
312 				 * That data will be rexmitted in the next
313 				 * sack recovery episode, when snd_recover
314 				 * moves past p->rxmit.
315 				 */
316 				p = NULL;
317 				goto after_sack_rexmit;
318 			} else {
319 				/* Can rexmit part of the current hole */
320 				len = ((int32_t)ulmin(cwin,
321 				    SEQ_SUB(tp->snd_recover, p->rxmit)));
322 			}
323 		} else {
324 			len = ((int32_t)ulmin(cwin,
325 			    SEQ_SUB(p->end, p->rxmit)));
326 		}
327 		if (len > 0) {
328 			off = SEQ_SUB(p->rxmit, tp->snd_una);
329 			KASSERT(off >= 0,("%s: sack block to the left of una : %d",
330 			    __func__, off));
331 			sack_rxmit = 1;
332 			sendalot = 1;
333 		}
334 	}
335 after_sack_rexmit:
336 	/*
337 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
338 	 * state flags.
339 	 */
340 	if (tp->t_flags & TF_NEEDFIN)
341 		flags |= TH_FIN;
342 	if (tp->t_flags & TF_NEEDSYN)
343 		flags |= TH_SYN;
344 
345 	SOCKBUF_LOCK(&so->so_snd);
346 	/*
347 	 * If in persist timeout with window of 0, send 1 byte.
348 	 * Otherwise, if window is small but nonzero
349 	 * and timer expired, we will send what we can
350 	 * and go to transmit state.
351 	 */
352 	if (tp->t_flags & TF_FORCEDATA) {
353 		if (sendwin == 0) {
354 			/*
355 			 * If we still have some data to send, then
356 			 * clear the FIN bit.  Usually this would
357 			 * happen below when it realizes that we
358 			 * aren't sending all the data.  However,
359 			 * if we have exactly 1 byte of unsent data,
360 			 * then it won't clear the FIN bit below,
361 			 * and if we are in persist state, we wind
362 			 * up sending the packet without recording
363 			 * that we sent the FIN bit.
364 			 *
365 			 * We can't just blindly clear the FIN bit,
366 			 * because if we don't have any more data
367 			 * to send then the probe will be the FIN
368 			 * itself.
369 			 */
370 			if (off < sbused(&so->so_snd))
371 				flags &= ~TH_FIN;
372 			sendwin = 1;
373 		} else {
374 			tcp_timer_activate(tp, TT_PERSIST, 0);
375 			tp->t_rxtshift = 0;
376 		}
377 	}
378 
379 	/*
380 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
381 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
382 	 * a negative length.  This can also occur when TCP opens up
383 	 * its congestion window while receiving additional duplicate
384 	 * acks after fast-retransmit because TCP will reset snd_nxt
385 	 * to snd_max after the fast-retransmit.
386 	 *
387 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
388 	 * be set to snd_una, the offset will be 0, and the length may
389 	 * wind up 0.
390 	 *
391 	 * If sack_rxmit is true we are retransmitting from the scoreboard
392 	 * in which case len is already set.
393 	 */
394 	if (sack_rxmit == 0) {
395 		if (sack_bytes_rxmt == 0) {
396 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
397 			    off);
398 		} else {
399 			int32_t cwin;
400 
401 			/*
402 			 * We are inside of a SACK recovery episode and are
403 			 * sending new data, having retransmitted all the
404 			 * data possible in the scoreboard.
405 			 */
406 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
407 			    off);
408 			/*
409 			 * Don't remove this (len > 0) check !
410 			 * We explicitly check for len > 0 here (although it
411 			 * isn't really necessary), to work around a gcc
412 			 * optimization issue - to force gcc to compute
413 			 * len above. Without this check, the computation
414 			 * of len is bungled by the optimizer.
415 			 */
416 			if (len > 0) {
417 				cwin = tp->snd_cwnd - imax(0, (int32_t)
418 					(tp->snd_nxt - tp->snd_recover)) -
419 					sack_bytes_rxmt;
420 				if (cwin < 0)
421 					cwin = 0;
422 				len = imin(len, cwin);
423 			}
424 		}
425 	}
426 
427 	/*
428 	 * Lop off SYN bit if it has already been sent.  However, if this
429 	 * is SYN-SENT state and if segment contains data and if we don't
430 	 * know that foreign host supports TAO, suppress sending segment.
431 	 */
432 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
433 		if (tp->t_state != TCPS_SYN_RECEIVED)
434 			flags &= ~TH_SYN;
435 		/*
436 		 * When sending additional segments following a TFO SYN|ACK,
437 		 * do not include the SYN bit.
438 		 */
439 		if (IS_FASTOPEN(tp->t_flags) &&
440 		    (tp->t_state == TCPS_SYN_RECEIVED))
441 			flags &= ~TH_SYN;
442 		off--, len++;
443 	}
444 
445 	/*
446 	 * Be careful not to send data and/or FIN on SYN segments.
447 	 * This measure is needed to prevent interoperability problems
448 	 * with not fully conformant TCP implementations.
449 	 */
450 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
451 		len = 0;
452 		flags &= ~TH_FIN;
453 	}
454 
455 	/*
456 	 * On TFO sockets, ensure no data is sent in the following cases:
457 	 *
458 	 *  - When retransmitting SYN|ACK on a passively-created socket
459 	 *
460 	 *  - When retransmitting SYN on an actively created socket
461 	 *
462 	 *  - When sending a zero-length cookie (cookie request) on an
463 	 *    actively created socket
464 	 *
465 	 *  - When the socket is in the CLOSED state (RST is being sent)
466 	 */
467 	if (IS_FASTOPEN(tp->t_flags) &&
468 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
469 	     ((tp->t_state == TCPS_SYN_SENT) &&
470 	      (tp->t_tfo_client_cookie_len == 0)) ||
471 	     (flags & TH_RST)))
472 		len = 0;
473 	if (len <= 0) {
474 		/*
475 		 * If FIN has been sent but not acked,
476 		 * but we haven't been called to retransmit,
477 		 * len will be < 0.  Otherwise, window shrank
478 		 * after we sent into it.  If window shrank to 0,
479 		 * cancel pending retransmit, pull snd_nxt back
480 		 * to (closed) window, and set the persist timer
481 		 * if it isn't already going.  If the window didn't
482 		 * close completely, just wait for an ACK.
483 		 *
484 		 * We also do a general check here to ensure that
485 		 * we will set the persist timer when we have data
486 		 * to send, but a 0-byte window. This makes sure
487 		 * the persist timer is set even if the packet
488 		 * hits one of the "goto send" lines below.
489 		 */
490 		len = 0;
491 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
492 		    (off < (int) sbavail(&so->so_snd)) &&
493 		    !tcp_timer_active(tp, TT_PERSIST)) {
494 			tcp_timer_activate(tp, TT_REXMT, 0);
495 			tp->t_rxtshift = 0;
496 			tp->snd_nxt = tp->snd_una;
497 			if (!tcp_timer_active(tp, TT_PERSIST))
498 				tcp_setpersist(tp);
499 		}
500 	}
501 
502 	/* len will be >= 0 after this point. */
503 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
504 
505 	tcp_sndbuf_autoscale(tp, so, sendwin);
506 
507 	/*
508 	 * Decide if we can use TCP Segmentation Offloading (if supported by
509 	 * hardware).
510 	 *
511 	 * TSO may only be used if we are in a pure bulk sending state.  The
512 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
513 	 * IP options prevent using TSO.  With TSO the TCP header is the same
514 	 * (except for the sequence number) for all generated packets.  This
515 	 * makes it impossible to transmit any options which vary per generated
516 	 * segment or packet.
517 	 *
518 	 * IPv4 handling has a clear separation of ip options and ip header
519 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
520 	 * the right thing below to provide length of just ip options and thus
521 	 * checking for ipoptlen is enough to decide if ip options are present.
522 	 */
523 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
524 	/*
525 	 * Pre-calculate here as we save another lookup into the darknesses
526 	 * of IPsec that way and can actually decide if TSO is ok.
527 	 */
528 #ifdef INET6
529 	if (isipv6 && IPSEC_ENABLED(ipv6))
530 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
531 #ifdef INET
532 	else
533 #endif
534 #endif /* INET6 */
535 #ifdef INET
536 	if (IPSEC_ENABLED(ipv4))
537 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
538 #endif /* INET */
539 #endif /* IPSEC */
540 #ifdef INET6
541 	if (isipv6)
542 		ipoptlen = ip6_optlen(inp);
543 	else
544 #endif
545 	if (inp->inp_options)
546 		ipoptlen = inp->inp_options->m_len -
547 				offsetof(struct ipoption, ipopt_list);
548 	else
549 		ipoptlen = 0;
550 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
551 	ipoptlen += ipsec_optlen;
552 #endif
553 
554 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
555 	    (tp->t_port == 0) &&
556 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
557 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
558 	    ipoptlen == 0 && !(flags & TH_SYN))
559 		tso = 1;
560 
561 	if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
562 		    tp->snd_una + sbused(&so->so_snd))) {
563 			flags &= ~TH_FIN;
564 	}
565 
566 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
567 	    (long)TCP_MAXWIN << tp->rcv_scale);
568 
569 	/*
570 	 * Sender silly window avoidance.   We transmit under the following
571 	 * conditions when len is non-zero:
572 	 *
573 	 *	- We have a full segment (or more with TSO)
574 	 *	- This is the last buffer in a write()/send() and we are
575 	 *	  either idle or running NODELAY
576 	 *	- we've timed out (e.g. persist timer)
577 	 *	- we have more then 1/2 the maximum send window's worth of
578 	 *	  data (receiver may be limited the window size)
579 	 *	- we need to retransmit
580 	 */
581 	if (len) {
582 		if (len >= tp->t_maxseg)
583 			goto send;
584 		/*
585 		 * As the TCP header options are now
586 		 * considered when setting up the initial
587 		 * window, we would not send the last segment
588 		 * if we skip considering the option length here.
589 		 * Note: this may not work when tcp headers change
590 		 * very dynamically in the future.
591 		 */
592 		if ((((tp->t_flags & TF_SIGNATURE) ?
593 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
594 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
595 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
596 		    len) >= tp->t_maxseg)
597 			goto send;
598 		/*
599 		 * NOTE! on localhost connections an 'ack' from the remote
600 		 * end may occur synchronously with the output and cause
601 		 * us to flush a buffer queued with moretocome.  XXX
602 		 *
603 		 * note: the len + off check is almost certainly unnecessary.
604 		 */
605 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
606 		    (idle || (tp->t_flags & TF_NODELAY)) &&
607 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
608 		    (tp->t_flags & TF_NOPUSH) == 0) {
609 			goto send;
610 		}
611 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
612 			goto send;
613 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
614 			goto send;
615 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
616 			goto send;
617 		if (sack_rxmit)
618 			goto send;
619 	}
620 
621 	/*
622 	 * Sending of standalone window updates.
623 	 *
624 	 * Window updates are important when we close our window due to a
625 	 * full socket buffer and are opening it again after the application
626 	 * reads data from it.  Once the window has opened again and the
627 	 * remote end starts to send again the ACK clock takes over and
628 	 * provides the most current window information.
629 	 *
630 	 * We must avoid the silly window syndrome whereas every read
631 	 * from the receive buffer, no matter how small, causes a window
632 	 * update to be sent.  We also should avoid sending a flurry of
633 	 * window updates when the socket buffer had queued a lot of data
634 	 * and the application is doing small reads.
635 	 *
636 	 * Prevent a flurry of pointless window updates by only sending
637 	 * an update when we can increase the advertized window by more
638 	 * than 1/4th of the socket buffer capacity.  When the buffer is
639 	 * getting full or is very small be more aggressive and send an
640 	 * update whenever we can increase by two mss sized segments.
641 	 * In all other situations the ACK's to new incoming data will
642 	 * carry further window increases.
643 	 *
644 	 * Don't send an independent window update if a delayed
645 	 * ACK is pending (it will get piggy-backed on it) or the
646 	 * remote side already has done a half-close and won't send
647 	 * more data.
648 	 */
649 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
650 	    !(tp->t_flags & TF_DELACK) &&
651 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
652 		/*
653 		 * "adv" is the amount we could increase the window,
654 		 * taking into account that we are limited by
655 		 * TCP_MAXWIN << tp->rcv_scale.
656 		 */
657 		int32_t adv;
658 		int oldwin;
659 
660 		adv = recwin;
661 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
662 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
663 			if (adv > oldwin)
664 				adv -= oldwin;
665 			else
666 				adv = 0;
667 		} else
668 			oldwin = 0;
669 
670 		/*
671 		 * If the new window size ends up being the same as or less
672 		 * than the old size when it is scaled, then don't force
673 		 * a window update.
674 		 */
675 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
676 			goto dontupdate;
677 
678 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
679 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
680 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
681 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
682 		     adv >= TCP_MAXWIN << tp->rcv_scale))
683 			goto send;
684 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
685 			goto send;
686 	}
687 dontupdate:
688 
689 	/*
690 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
691 	 * is also a catch-all for the retransmit timer timeout case.
692 	 */
693 	if (tp->t_flags & TF_ACKNOW)
694 		goto send;
695 	if ((flags & TH_RST) ||
696 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
697 		goto send;
698 	if (SEQ_GT(tp->snd_up, tp->snd_una))
699 		goto send;
700 	/*
701 	 * If our state indicates that FIN should be sent
702 	 * and we have not yet done so, then we need to send.
703 	 */
704 	if (flags & TH_FIN &&
705 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
706 		goto send;
707 	/*
708 	 * In SACK, it is possible for tcp_output to fail to send a segment
709 	 * after the retransmission timer has been turned off.  Make sure
710 	 * that the retransmission timer is set.
711 	 */
712 	if ((tp->t_flags & TF_SACK_PERMIT) &&
713 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
714 	    !tcp_timer_active(tp, TT_REXMT) &&
715 	    !tcp_timer_active(tp, TT_PERSIST)) {
716 		tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
717 		goto just_return;
718 	}
719 	/*
720 	 * TCP window updates are not reliable, rather a polling protocol
721 	 * using ``persist'' packets is used to insure receipt of window
722 	 * updates.  The three ``states'' for the output side are:
723 	 *	idle			not doing retransmits or persists
724 	 *	persisting		to move a small or zero window
725 	 *	(re)transmitting	and thereby not persisting
726 	 *
727 	 * tcp_timer_active(tp, TT_PERSIST)
728 	 *	is true when we are in persist state.
729 	 * (tp->t_flags & TF_FORCEDATA)
730 	 *	is set when we are called to send a persist packet.
731 	 * tcp_timer_active(tp, TT_REXMT)
732 	 *	is set when we are retransmitting
733 	 * The output side is idle when both timers are zero.
734 	 *
735 	 * If send window is too small, there is data to transmit, and no
736 	 * retransmit or persist is pending, then go to persist state.
737 	 * If nothing happens soon, send when timer expires:
738 	 * if window is nonzero, transmit what we can,
739 	 * otherwise force out a byte.
740 	 */
741 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
742 	    !tcp_timer_active(tp, TT_PERSIST)) {
743 		tp->t_rxtshift = 0;
744 		tcp_setpersist(tp);
745 	}
746 
747 	/*
748 	 * No reason to send a segment, just return.
749 	 */
750 just_return:
751 	SOCKBUF_UNLOCK(&so->so_snd);
752 	return (0);
753 
754 send:
755 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
756 	if (len > 0) {
757 		if (len >= tp->t_maxseg)
758 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
759 		else
760 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
761 	}
762 	/*
763 	 * Before ESTABLISHED, force sending of initial options
764 	 * unless TCP set not to do any options.
765 	 * NOTE: we assume that the IP/TCP header plus TCP options
766 	 * always fit in a single mbuf, leaving room for a maximum
767 	 * link header, i.e.
768 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
769 	 */
770 	optlen = 0;
771 #ifdef INET6
772 	if (isipv6)
773 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
774 	else
775 #endif
776 		hdrlen = sizeof (struct tcpiphdr);
777 
778 	if (flags & TH_SYN) {
779 		tp->snd_nxt = tp->iss;
780 	}
781 
782 	/*
783 	 * Compute options for segment.
784 	 * We only have to care about SYN and established connection
785 	 * segments.  Options for SYN-ACK segments are handled in TCP
786 	 * syncache.
787 	 */
788 	to.to_flags = 0;
789 	if ((tp->t_flags & TF_NOOPT) == 0) {
790 		/* Maximum segment size. */
791 		if (flags & TH_SYN) {
792 			to.to_mss = tcp_mssopt(&inp->inp_inc);
793 			if (tp->t_port)
794 				to.to_mss -= V_tcp_udp_tunneling_overhead;
795 			to.to_flags |= TOF_MSS;
796 
797 			/*
798 			 * On SYN or SYN|ACK transmits on TFO connections,
799 			 * only include the TFO option if it is not a
800 			 * retransmit, as the presence of the TFO option may
801 			 * have caused the original SYN or SYN|ACK to have
802 			 * been dropped by a middlebox.
803 			 */
804 			if (IS_FASTOPEN(tp->t_flags) &&
805 			    (tp->t_rxtshift == 0)) {
806 				if (tp->t_state == TCPS_SYN_RECEIVED) {
807 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
808 					to.to_tfo_cookie =
809 					    (u_int8_t *)&tp->t_tfo_cookie.server;
810 					to.to_flags |= TOF_FASTOPEN;
811 					wanted_cookie = 1;
812 				} else if (tp->t_state == TCPS_SYN_SENT) {
813 					to.to_tfo_len =
814 					    tp->t_tfo_client_cookie_len;
815 					to.to_tfo_cookie =
816 					    tp->t_tfo_cookie.client;
817 					to.to_flags |= TOF_FASTOPEN;
818 					wanted_cookie = 1;
819 					/*
820 					 * If we wind up having more data to
821 					 * send with the SYN than can fit in
822 					 * one segment, don't send any more
823 					 * until the SYN|ACK comes back from
824 					 * the other end.
825 					 */
826 					dont_sendalot = 1;
827 				}
828 			}
829 		}
830 		/* Window scaling. */
831 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
832 			to.to_wscale = tp->request_r_scale;
833 			to.to_flags |= TOF_SCALE;
834 		}
835 		/* Timestamps. */
836 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
837 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
838 			curticks = tcp_ts_getticks();
839 			to.to_tsval = curticks + tp->ts_offset;
840 			to.to_tsecr = tp->ts_recent;
841 			to.to_flags |= TOF_TS;
842 			if (tp->t_rxtshift == 1)
843 				tp->t_badrxtwin = curticks;
844 		}
845 
846 		/* Set receive buffer autosizing timestamp. */
847 		if (tp->rfbuf_ts == 0 &&
848 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
849 			tp->rfbuf_ts = tcp_ts_getticks();
850 
851 		/* Selective ACK's. */
852 		if (tp->t_flags & TF_SACK_PERMIT) {
853 			if (flags & TH_SYN)
854 				to.to_flags |= TOF_SACKPERM;
855 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
856 			    tp->rcv_numsacks > 0) {
857 				to.to_flags |= TOF_SACK;
858 				to.to_nsacks = tp->rcv_numsacks;
859 				to.to_sacks = (u_char *)tp->sackblks;
860 			}
861 		}
862 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
863 		/* TCP-MD5 (RFC2385). */
864 		/*
865 		 * Check that TCP_MD5SIG is enabled in tcpcb to
866 		 * account the size needed to set this TCP option.
867 		 */
868 		if (tp->t_flags & TF_SIGNATURE)
869 			to.to_flags |= TOF_SIGNATURE;
870 #endif /* TCP_SIGNATURE */
871 
872 		/* Processing the options. */
873 		hdrlen += optlen = tcp_addoptions(&to, opt);
874 		/*
875 		 * If we wanted a TFO option to be added, but it was unable
876 		 * to fit, ensure no data is sent.
877 		 */
878 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
879 		    !(to.to_flags & TOF_FASTOPEN))
880 			len = 0;
881 	}
882 	if (tp->t_port) {
883 		if (V_tcp_udp_tunneling_port == 0) {
884 			/* The port was removed?? */
885 			SOCKBUF_UNLOCK(&so->so_snd);
886 			return (EHOSTUNREACH);
887 		}
888 		hdrlen += sizeof(struct udphdr);
889 	}
890 	/*
891 	 * Adjust data length if insertion of options will
892 	 * bump the packet length beyond the t_maxseg length.
893 	 * Clear the FIN bit because we cut off the tail of
894 	 * the segment.
895 	 */
896 	if (len + optlen + ipoptlen > tp->t_maxseg) {
897 		flags &= ~TH_FIN;
898 
899 		if (tso) {
900 			u_int if_hw_tsomax;
901 			u_int moff;
902 			int max_len;
903 
904 			/* extract TSO information */
905 			if_hw_tsomax = tp->t_tsomax;
906 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
907 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
908 
909 			/*
910 			 * Limit a TSO burst to prevent it from
911 			 * overflowing or exceeding the maximum length
912 			 * allowed by the network interface:
913 			 */
914 			KASSERT(ipoptlen == 0,
915 			    ("%s: TSO can't do IP options", __func__));
916 
917 			/*
918 			 * Check if we should limit by maximum payload
919 			 * length:
920 			 */
921 			if (if_hw_tsomax != 0) {
922 				/* compute maximum TSO length */
923 				max_len = (if_hw_tsomax - hdrlen -
924 				    max_linkhdr);
925 				if (max_len <= 0) {
926 					len = 0;
927 				} else if (len > max_len) {
928 					sendalot = 1;
929 					len = max_len;
930 				}
931 			}
932 
933 			/*
934 			 * Prevent the last segment from being
935 			 * fractional unless the send sockbuf can be
936 			 * emptied:
937 			 */
938 			max_len = (tp->t_maxseg - optlen);
939 			if (((uint32_t)off + (uint32_t)len) <
940 			    sbavail(&so->so_snd)) {
941 				moff = len % max_len;
942 				if (moff != 0) {
943 					len -= moff;
944 					sendalot = 1;
945 				}
946 			}
947 
948 			/*
949 			 * In case there are too many small fragments
950 			 * don't use TSO:
951 			 */
952 			if (len <= max_len) {
953 				len = max_len;
954 				sendalot = 1;
955 				tso = 0;
956 			}
957 
958 			/*
959 			 * Send the FIN in a separate segment
960 			 * after the bulk sending is done.
961 			 * We don't trust the TSO implementations
962 			 * to clear the FIN flag on all but the
963 			 * last segment.
964 			 */
965 			if (tp->t_flags & TF_NEEDFIN)
966 				sendalot = 1;
967 		} else {
968 			if (optlen + ipoptlen >= tp->t_maxseg) {
969 				/*
970 				 * Since we don't have enough space to put
971 				 * the IP header chain and the TCP header in
972 				 * one packet as required by RFC 7112, don't
973 				 * send it. Also ensure that at least one
974 				 * byte of the payload can be put into the
975 				 * TCP segment.
976 				 */
977 				SOCKBUF_UNLOCK(&so->so_snd);
978 				error = EMSGSIZE;
979 				sack_rxmit = 0;
980 				goto out;
981 			}
982 			len = tp->t_maxseg - optlen - ipoptlen;
983 			sendalot = 1;
984 			if (dont_sendalot)
985 				sendalot = 0;
986 		}
987 	} else
988 		tso = 0;
989 
990 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
991 	    ("%s: len > IP_MAXPACKET", __func__));
992 
993 /*#ifdef DIAGNOSTIC*/
994 #ifdef INET6
995 	if (max_linkhdr + hdrlen > MCLBYTES)
996 #else
997 	if (max_linkhdr + hdrlen > MHLEN)
998 #endif
999 		panic("tcphdr too big");
1000 /*#endif*/
1001 
1002 	/*
1003 	 * This KASSERT is here to catch edge cases at a well defined place.
1004 	 * Before, those had triggered (random) panic conditions further down.
1005 	 */
1006 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1007 
1008 	/*
1009 	 * Grab a header mbuf, attaching a copy of data to
1010 	 * be transmitted, and initialize the header from
1011 	 * the template for sends on this connection.
1012 	 */
1013 	if (len) {
1014 		struct mbuf *mb;
1015 		struct sockbuf *msb;
1016 		u_int moff;
1017 
1018 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1019 			TCPSTAT_INC(tcps_sndprobe);
1020 #ifdef STATS
1021 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1022 				stats_voi_update_abs_u32(tp->t_stats,
1023 				VOI_TCP_RETXPB, len);
1024 			else
1025 				stats_voi_update_abs_u64(tp->t_stats,
1026 				    VOI_TCP_TXPB, len);
1027 #endif /* STATS */
1028 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1029 			tp->t_sndrexmitpack++;
1030 			TCPSTAT_INC(tcps_sndrexmitpack);
1031 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1032 			if (sack_rxmit) {
1033 				TCPSTAT_INC(tcps_sack_rexmits);
1034 				TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1035 			}
1036 #ifdef STATS
1037 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1038 			    len);
1039 #endif /* STATS */
1040 		} else {
1041 			TCPSTAT_INC(tcps_sndpack);
1042 			TCPSTAT_ADD(tcps_sndbyte, len);
1043 #ifdef STATS
1044 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1045 			    len);
1046 #endif /* STATS */
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_noadv(&so->so_snd, off, &moff);
1070 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1071 			m_copydata(mb, moff, len,
1072 			    mtod(m, caddr_t) + hdrlen);
1073 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1074 				sbsndptr_adv(&so->so_snd, mb, len);
1075 			m->m_len += len;
1076 		} else {
1077 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1078 				msb = NULL;
1079 			else
1080 				msb = &so->so_snd;
1081 			m->m_next = tcp_m_copym(mb, moff,
1082 			    &len, if_hw_tsomaxsegcount,
1083 			    if_hw_tsomaxsegsize, msb, hw_tls);
1084 			if (len <= (tp->t_maxseg - optlen)) {
1085 				/*
1086 				 * Must have ran out of mbufs for the copy
1087 				 * shorten it to no longer need tso. Lets
1088 				 * not put on sendalot since we are low on
1089 				 * mbufs.
1090 				 */
1091 				tso = 0;
1092 			}
1093 			if (m->m_next == NULL) {
1094 				SOCKBUF_UNLOCK(&so->so_snd);
1095 				(void) m_free(m);
1096 				error = ENOBUFS;
1097 				sack_rxmit = 0;
1098 				goto out;
1099 			}
1100 		}
1101 
1102 		/*
1103 		 * If we're sending everything we've got, set PUSH.
1104 		 * (This will keep happy those implementations which only
1105 		 * give data to the user when a buffer fills or
1106 		 * a PUSH comes in.)
1107 		 */
1108 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1109 		    !(flags & TH_SYN))
1110 			flags |= TH_PUSH;
1111 		SOCKBUF_UNLOCK(&so->so_snd);
1112 	} else {
1113 		SOCKBUF_UNLOCK(&so->so_snd);
1114 		if (tp->t_flags & TF_ACKNOW)
1115 			TCPSTAT_INC(tcps_sndacks);
1116 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1117 			TCPSTAT_INC(tcps_sndctrl);
1118 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1119 			TCPSTAT_INC(tcps_sndurg);
1120 		else
1121 			TCPSTAT_INC(tcps_sndwinup);
1122 
1123 		m = m_gethdr(M_NOWAIT, MT_DATA);
1124 		if (m == NULL) {
1125 			error = ENOBUFS;
1126 			sack_rxmit = 0;
1127 			goto out;
1128 		}
1129 #ifdef INET6
1130 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1131 		    MHLEN >= hdrlen) {
1132 			M_ALIGN(m, hdrlen);
1133 		} else
1134 #endif
1135 		m->m_data += max_linkhdr;
1136 		m->m_len = hdrlen;
1137 	}
1138 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1139 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1140 #ifdef MAC
1141 	mac_inpcb_create_mbuf(inp, m);
1142 #endif
1143 #ifdef INET6
1144 	if (isipv6) {
1145 		ip6 = mtod(m, struct ip6_hdr *);
1146 		if (tp->t_port) {
1147 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1148 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1149 			udp->uh_dport = tp->t_port;
1150 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1151 			udp->uh_ulen = htons(ulen);
1152 			th = (struct tcphdr *)(udp + 1);
1153 		} else {
1154 			th = (struct tcphdr *)(ip6 + 1);
1155 		}
1156 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
1157 	} else
1158 #endif /* INET6 */
1159 	{
1160 		ip = mtod(m, struct ip *);
1161 		if (tp->t_port) {
1162 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1163 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1164 			udp->uh_dport = tp->t_port;
1165 			ulen = hdrlen + len - sizeof(struct ip);
1166 			udp->uh_ulen = htons(ulen);
1167 			th = (struct tcphdr *)(udp + 1);
1168 		} else
1169 			th = (struct tcphdr *)(ip + 1);
1170 		tcpip_fillheaders(inp, tp->t_port, ip, th);
1171 	}
1172 
1173 	/*
1174 	 * Fill in fields, remembering maximum advertised
1175 	 * window for use in delaying messages about window sizes.
1176 	 * If resending a FIN, be sure not to use a new sequence number.
1177 	 */
1178 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1179 	    tp->snd_nxt == tp->snd_max)
1180 		tp->snd_nxt--;
1181 	/*
1182 	 * If we are starting a connection, send ECN setup
1183 	 * SYN packet. If we are on a retransmit, we may
1184 	 * resend those bits a number of times as per
1185 	 * RFC 3168.
1186 	 */
1187 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1188 		flags |= tcp_ecn_output_syn_sent(tp);
1189 	}
1190 	/* Also handle parallel SYN for ECN */
1191 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1192 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1193 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1194 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1195 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1196 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1197 #ifdef INET6
1198 		if (isipv6) {
1199 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1200 			ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1201 		}
1202 		else
1203 #endif
1204 		{
1205 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1206 			ip->ip_tos |= ect;
1207 		}
1208 	}
1209 
1210 	/*
1211 	 * If we are doing retransmissions, then snd_nxt will
1212 	 * not reflect the first unsent octet.  For ACK only
1213 	 * packets, we do not want the sequence number of the
1214 	 * retransmitted packet, we want the sequence number
1215 	 * of the next unsent octet.  So, if there is no data
1216 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1217 	 * when filling in ti_seq.  But if we are in persist
1218 	 * state, snd_max might reflect one byte beyond the
1219 	 * right edge of the window, so use snd_nxt in that
1220 	 * case, since we know we aren't doing a retransmission.
1221 	 * (retransmit and persist are mutually exclusive...)
1222 	 */
1223 	if (sack_rxmit == 0) {
1224 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1225 		    tcp_timer_active(tp, TT_PERSIST))
1226 			th->th_seq = htonl(tp->snd_nxt);
1227 		else
1228 			th->th_seq = htonl(tp->snd_max);
1229 	} else {
1230 		th->th_seq = htonl(p->rxmit);
1231 		p->rxmit += len;
1232 		/*
1233 		 * Lost Retransmission Detection
1234 		 * trigger resending of a (then
1235 		 * still existing) hole, when
1236 		 * fack acks recoverypoint.
1237 		 */
1238 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1239 			p->rxmit = tp->snd_recover;
1240 		tp->sackhint.sack_bytes_rexmit += len;
1241 	}
1242 	if (IN_RECOVERY(tp->t_flags)) {
1243 		/*
1244 		 * Account all bytes transmitted while
1245 		 * IN_RECOVERY, simplifying PRR and
1246 		 * Lost Retransmit Detection
1247 		 */
1248 		tp->sackhint.prr_out += len;
1249 	}
1250 	th->th_ack = htonl(tp->rcv_nxt);
1251 	if (optlen) {
1252 		bcopy(opt, th + 1, optlen);
1253 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1254 	}
1255 	tcp_set_flags(th, flags);
1256 	/*
1257 	 * Calculate receive window.  Don't shrink window,
1258 	 * but avoid silly window syndrome.
1259 	 * If a RST segment is sent, advertise a window of zero.
1260 	 */
1261 	if (flags & TH_RST) {
1262 		recwin = 0;
1263 	} else {
1264 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1265 		    recwin < tp->t_maxseg)
1266 			recwin = 0;
1267 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1268 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1269 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1270 	}
1271 	/*
1272 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1273 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1274 	 * case is handled in syncache.
1275 	 */
1276 	if (flags & TH_SYN)
1277 		th->th_win = htons((u_short)
1278 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1279 	else {
1280 		/* Avoid shrinking window with window scaling. */
1281 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1282 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1283 	}
1284 
1285 	/*
1286 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1287 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1288 	 * flag tells soreceive() to disable delayed acknowledgements when
1289 	 * draining the buffer.  This can occur if the receiver is attempting
1290 	 * to read more data than can be buffered prior to transmitting on
1291 	 * the connection.
1292 	 */
1293 	if (th->th_win == 0) {
1294 		tp->t_sndzerowin++;
1295 		tp->t_flags |= TF_RXWIN0SENT;
1296 	} else
1297 		tp->t_flags &= ~TF_RXWIN0SENT;
1298 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1299 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1300 		th->th_flags |= TH_URG;
1301 	} else
1302 		/*
1303 		 * If no urgent pointer to send, then we pull
1304 		 * the urgent pointer to the left edge of the send window
1305 		 * so that it doesn't drift into the send window on sequence
1306 		 * number wraparound.
1307 		 */
1308 		tp->snd_up = tp->snd_una;		/* drag it along */
1309 
1310 	/*
1311 	 * Put TCP length in extended header, and then
1312 	 * checksum extended header and data.
1313 	 */
1314 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1315 
1316 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1317 	if (to.to_flags & TOF_SIGNATURE) {
1318 		/*
1319 		 * Calculate MD5 signature and put it into the place
1320 		 * determined before.
1321 		 * NOTE: since TCP options buffer doesn't point into
1322 		 * mbuf's data, calculate offset and use it.
1323 		 */
1324 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1325 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1326 			/*
1327 			 * Do not send segment if the calculation of MD5
1328 			 * digest has failed.
1329 			 */
1330 			m_freem(m);
1331 			goto out;
1332 		}
1333 	}
1334 #endif
1335 #ifdef INET6
1336 	if (isipv6) {
1337 		/*
1338 		 * There is no need to fill in ip6_plen right now.
1339 		 * It will be filled later by ip6_output.
1340 		 */
1341 		if (tp->t_port) {
1342 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1343 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1344 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1345 			th->th_sum = htons(0);
1346 			UDPSTAT_INC(udps_opackets);
1347 		} else {
1348 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1349 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1350 			th->th_sum = in6_cksum_pseudo(ip6,
1351 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1352 			    0);
1353 		}
1354 	}
1355 #endif
1356 #if defined(INET6) && defined(INET)
1357 	else
1358 #endif
1359 #ifdef INET
1360 	{
1361 		if (tp->t_port) {
1362 			m->m_pkthdr.csum_flags = CSUM_UDP;
1363 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1364 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1365 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1366 			th->th_sum = htons(0);
1367 			UDPSTAT_INC(udps_opackets);
1368 		} else {
1369 			m->m_pkthdr.csum_flags = CSUM_TCP;
1370 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1371 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1372 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1373 			    IPPROTO_TCP + len + optlen));
1374 		}
1375 
1376 		/* IP version must be set here for ipv4/ipv6 checking later */
1377 		KASSERT(ip->ip_v == IPVERSION,
1378 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1379 	}
1380 #endif
1381 
1382 	/*
1383 	 * Enable TSO and specify the size of the segments.
1384 	 * The TCP pseudo header checksum is always provided.
1385 	 */
1386 	if (tso) {
1387 		KASSERT(len > tp->t_maxseg - optlen,
1388 		    ("%s: len <= tso_segsz", __func__));
1389 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1390 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1391 	}
1392 
1393 	KASSERT(len + hdrlen == m_length(m, NULL),
1394 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1395 	    __func__, len, hdrlen, m_length(m, NULL)));
1396 
1397 #ifdef TCP_HHOOK
1398 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1399 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1400 #endif
1401 
1402 	TCP_PROBE3(debug__output, tp, th, m);
1403 
1404 	/* We're getting ready to send; log now. */
1405 	/* XXXMT: We are not honoring verbose logging. */
1406 
1407 	if (tcp_bblogging_on(tp))
1408 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1409 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1410 		    NULL);
1411 	else
1412 		lgb = NULL;
1413 
1414 	/*
1415 	 * Fill in IP length and desired time to live and
1416 	 * send to IP level.  There should be a better way
1417 	 * to handle ttl and tos; we could keep them in
1418 	 * the template, but need a way to checksum without them.
1419 	 */
1420 	/*
1421 	 * m->m_pkthdr.len should have been set before checksum calculation,
1422 	 * because in6_cksum() need it.
1423 	 */
1424 #ifdef INET6
1425 	if (isipv6) {
1426 		/*
1427 		 * we separately set hoplimit for every segment, since the
1428 		 * user might want to change the value via setsockopt.
1429 		 * Also, desired default hop limit might be changed via
1430 		 * Neighbor Discovery.
1431 		 */
1432 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1433 
1434 		/*
1435 		 * Set the packet size here for the benefit of DTrace probes.
1436 		 * ip6_output() will set it properly; it's supposed to include
1437 		 * the option header lengths as well.
1438 		 */
1439 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1440 
1441 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1442 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1443 		else
1444 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1445 
1446 		if (tp->t_state == TCPS_SYN_SENT)
1447 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1448 
1449 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1450 
1451 #ifdef TCPPCAP
1452 		/* Save packet, if requested. */
1453 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1454 #endif
1455 
1456 		/* TODO: IPv6 IP6TOS_ECT bit on */
1457 		error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1458 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1459 		    NULL, NULL, inp);
1460 
1461 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1462 			mtu = inp->inp_route6.ro_nh->nh_mtu;
1463 	}
1464 #endif /* INET6 */
1465 #if defined(INET) && defined(INET6)
1466 	else
1467 #endif
1468 #ifdef INET
1469     {
1470 	ip->ip_len = htons(m->m_pkthdr.len);
1471 #ifdef INET6
1472 	if (inp->inp_vflag & INP_IPV6PROTO)
1473 		ip->ip_ttl = in6_selecthlim(inp, NULL);
1474 #endif /* INET6 */
1475 	/*
1476 	 * If we do path MTU discovery, then we set DF on every packet.
1477 	 * This might not be the best thing to do according to RFC3390
1478 	 * Section 2. However the tcp hostcache migitates the problem
1479 	 * so it affects only the first tcp connection with a host.
1480 	 *
1481 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1482 	 */
1483 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1484 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1485 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1486 			ip->ip_off |= htons(IP_DF);
1487 		}
1488 	} else {
1489 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1490 	}
1491 
1492 	if (tp->t_state == TCPS_SYN_SENT)
1493 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1494 
1495 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1496 
1497 #ifdef TCPPCAP
1498 	/* Save packet, if requested. */
1499 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1500 #endif
1501 
1502 	error = ip_output(m, inp->inp_options, &inp->inp_route,
1503 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1504 
1505 	if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1506 		mtu = inp->inp_route.ro_nh->nh_mtu;
1507     }
1508 #endif /* INET */
1509 
1510 	if (lgb != NULL) {
1511 		lgb->tlb_errno = error;
1512 		lgb = NULL;
1513 	}
1514 out:
1515 	if (error == 0)
1516 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1517 	/*
1518 	 * In transmit state, time the transmission and arrange for
1519 	 * the retransmit.  In persist state, just set snd_max.
1520 	 */
1521 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1522 	    !tcp_timer_active(tp, TT_PERSIST)) {
1523 		tcp_seq startseq = tp->snd_nxt;
1524 
1525 		/*
1526 		 * Advance snd_nxt over sequence space of this segment.
1527 		 */
1528 		if (flags & (TH_SYN|TH_FIN)) {
1529 			if (flags & TH_SYN)
1530 				tp->snd_nxt++;
1531 			if (flags & TH_FIN) {
1532 				tp->snd_nxt++;
1533 				tp->t_flags |= TF_SENTFIN;
1534 			}
1535 		}
1536 		if (sack_rxmit)
1537 			goto timer;
1538 		tp->snd_nxt += len;
1539 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1540 			/*
1541 			 * Update "made progress" indication if we just
1542 			 * added new data to an empty socket buffer.
1543 			 */
1544 			if (tp->snd_una == tp->snd_max)
1545 				tp->t_acktime = ticks;
1546 			tp->snd_max = tp->snd_nxt;
1547 			/*
1548 			 * Time this transmission if not a retransmission and
1549 			 * not currently timing anything.
1550 			 */
1551 			tp->t_sndtime = ticks;
1552 			if (tp->t_rtttime == 0) {
1553 				tp->t_rtttime = ticks;
1554 				tp->t_rtseq = startseq;
1555 				TCPSTAT_INC(tcps_segstimed);
1556 			}
1557 #ifdef STATS
1558 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1559 				tp->t_flags |= TF_GPUTINPROG;
1560 				tp->gput_seq = startseq;
1561 				tp->gput_ack = startseq +
1562 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1563 				tp->gput_ts = tcp_ts_getticks();
1564 			}
1565 #endif /* STATS */
1566 		}
1567 
1568 		/*
1569 		 * Set retransmit timer if not currently set,
1570 		 * and not doing a pure ack or a keep-alive probe.
1571 		 * Initial value for retransmit timer is smoothed
1572 		 * round-trip time + 2 * round-trip time variance.
1573 		 * Initialize shift counter which is used for backoff
1574 		 * of retransmit time.
1575 		 */
1576 timer:
1577 		if (!tcp_timer_active(tp, TT_REXMT) &&
1578 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1579 		     (tp->snd_nxt != tp->snd_una))) {
1580 			if (tcp_timer_active(tp, TT_PERSIST)) {
1581 				tcp_timer_activate(tp, TT_PERSIST, 0);
1582 				tp->t_rxtshift = 0;
1583 			}
1584 			tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1585 		} else if (len == 0 && sbavail(&so->so_snd) &&
1586 		    !tcp_timer_active(tp, TT_REXMT) &&
1587 		    !tcp_timer_active(tp, TT_PERSIST)) {
1588 			/*
1589 			 * Avoid a situation where we do not set persist timer
1590 			 * after a zero window condition. For example:
1591 			 * 1) A -> B: packet with enough data to fill the window
1592 			 * 2) B -> A: ACK for #1 + new data (0 window
1593 			 *    advertisement)
1594 			 * 3) A -> B: ACK for #2, 0 len packet
1595 			 *
1596 			 * In this case, A will not activate the persist timer,
1597 			 * because it chose to send a packet. Unless tcp_output
1598 			 * is called for some other reason (delayed ack timer,
1599 			 * another input packet from B, socket syscall), A will
1600 			 * not send zero window probes.
1601 			 *
1602 			 * So, if you send a 0-length packet, but there is data
1603 			 * in the socket buffer, and neither the rexmt or
1604 			 * persist timer is already set, then activate the
1605 			 * persist timer.
1606 			 */
1607 			tp->t_rxtshift = 0;
1608 			tcp_setpersist(tp);
1609 		}
1610 	} else {
1611 		/*
1612 		 * Persist case, update snd_max but since we are in
1613 		 * persist mode (no window) we do not update snd_nxt.
1614 		 */
1615 		int xlen = len;
1616 		if (flags & TH_SYN)
1617 			++xlen;
1618 		if (flags & TH_FIN) {
1619 			++xlen;
1620 			tp->t_flags |= TF_SENTFIN;
1621 		}
1622 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1623 			tp->snd_max = tp->snd_nxt + xlen;
1624 	}
1625 	if ((error == 0) &&
1626 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1627 	     (tp->t_flags & TF_SACK_PERMIT) &&
1628 	     tp->rcv_numsacks > 0)) {
1629 		    /* Clean up any DSACK's sent */
1630 		    tcp_clean_dsack_blocks(tp);
1631 	}
1632 	if (error) {
1633 		/*
1634 		 * We know that the packet was lost, so back out the
1635 		 * sequence number advance, if any.
1636 		 *
1637 		 * If the error is EPERM the packet got blocked by the
1638 		 * local firewall.  Normally we should terminate the
1639 		 * connection but the blocking may have been spurious
1640 		 * due to a firewall reconfiguration cycle.  So we treat
1641 		 * it like a packet loss and let the retransmit timer and
1642 		 * timeouts do their work over time.
1643 		 * XXX: It is a POLA question whether calling tcp_drop right
1644 		 * away would be the really correct behavior instead.
1645 		 */
1646 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1647 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1648 		    ((flags & TH_SYN) == 0) &&
1649 		    (error != EPERM)) {
1650 			if (sack_rxmit) {
1651 				p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1652 				tp->sackhint.sack_bytes_rexmit -= len;
1653 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1654 				    ("sackhint bytes rtx >= 0"));
1655 				KASSERT((flags & TH_FIN) == 0,
1656 				    ("error while FIN with SACK rxmit"));
1657 			} else {
1658 				tp->snd_nxt -= len;
1659 				if (flags & TH_FIN)
1660 					tp->snd_nxt--;
1661 			}
1662 			if (IN_RECOVERY(tp->t_flags))
1663 				tp->sackhint.prr_out -= len;
1664 		}
1665 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1666 		switch (error) {
1667 		case EACCES:
1668 		case EPERM:
1669 			tp->t_softerror = error;
1670 			return (error);
1671 		case ENOBUFS:
1672 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1673 			tp->snd_cwnd = tp->t_maxseg;
1674 			return (0);
1675 		case EMSGSIZE:
1676 			/*
1677 			 * For some reason the interface we used initially
1678 			 * to send segments changed to another or lowered
1679 			 * its MTU.
1680 			 * If TSO was active we either got an interface
1681 			 * without TSO capabilits or TSO was turned off.
1682 			 * If we obtained mtu from ip_output() then update
1683 			 * it and try again.
1684 			 */
1685 			if (tso)
1686 				tp->t_flags &= ~TF_TSO;
1687 			if (mtu != 0) {
1688 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1689 				goto again;
1690 			}
1691 			return (error);
1692 		case EHOSTDOWN:
1693 		case EHOSTUNREACH:
1694 		case ENETDOWN:
1695 		case ENETUNREACH:
1696 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1697 				tp->t_softerror = error;
1698 				return (0);
1699 			}
1700 			/* FALLTHROUGH */
1701 		default:
1702 			return (error);
1703 		}
1704 	}
1705 	TCPSTAT_INC(tcps_sndtotal);
1706 
1707 	/*
1708 	 * Data sent (as far as we can tell).
1709 	 * If this advertises a larger window than any other segment,
1710 	 * then remember the size of the advertised window.
1711 	 * Any pending ACK has now been sent.
1712 	 */
1713 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1714 		tp->rcv_adv = tp->rcv_nxt + recwin;
1715 	tp->last_ack_sent = tp->rcv_nxt;
1716 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1717 	if (tcp_timer_active(tp, TT_DELACK))
1718 		tcp_timer_activate(tp, TT_DELACK, 0);
1719 	if (sendalot)
1720 		goto again;
1721 	return (0);
1722 }
1723 
1724 void
1725 tcp_setpersist(struct tcpcb *tp)
1726 {
1727 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1728 	int tt;
1729 	int maxunacktime;
1730 
1731 	tp->t_flags &= ~TF_PREVVALID;
1732 	if (tcp_timer_active(tp, TT_REXMT))
1733 		panic("tcp_setpersist: retransmit pending");
1734 	/*
1735 	 * If the state is already closed, don't bother.
1736 	 */
1737 	if (tp->t_state == TCPS_CLOSED)
1738 		return;
1739 
1740 	/*
1741 	 * Start/restart persistence timer.
1742 	 */
1743 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1744 		      tcp_persmin, tcp_persmax);
1745 	if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1746 		maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1747 		if (maxunacktime < 1)
1748 			maxunacktime = 1;
1749 		if (maxunacktime < tt)
1750 			tt = maxunacktime;
1751 	}
1752 	tcp_timer_activate(tp, TT_PERSIST, tt);
1753 	if (tp->t_rxtshift < V_tcp_retries)
1754 		tp->t_rxtshift++;
1755 }
1756 
1757 /*
1758  * Insert TCP options according to the supplied parameters to the place
1759  * optp in a consistent way.  Can handle unaligned destinations.
1760  *
1761  * The order of the option processing is crucial for optimal packing and
1762  * alignment for the scarce option space.
1763  *
1764  * The optimal order for a SYN/SYN-ACK segment is:
1765  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1766  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1767  *
1768  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1769  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1770  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1771  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1772  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1773  */
1774 int
1775 tcp_addoptions(struct tcpopt *to, u_char *optp)
1776 {
1777 	u_int32_t mask, optlen = 0;
1778 
1779 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1780 		if ((to->to_flags & mask) != mask)
1781 			continue;
1782 		if (optlen == TCP_MAXOLEN)
1783 			break;
1784 		switch (to->to_flags & mask) {
1785 		case TOF_MSS:
1786 			while (optlen % 4) {
1787 				optlen += TCPOLEN_NOP;
1788 				*optp++ = TCPOPT_NOP;
1789 			}
1790 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1791 				continue;
1792 			optlen += TCPOLEN_MAXSEG;
1793 			*optp++ = TCPOPT_MAXSEG;
1794 			*optp++ = TCPOLEN_MAXSEG;
1795 			to->to_mss = htons(to->to_mss);
1796 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1797 			optp += sizeof(to->to_mss);
1798 			break;
1799 		case TOF_SCALE:
1800 			while (!optlen || optlen % 2 != 1) {
1801 				optlen += TCPOLEN_NOP;
1802 				*optp++ = TCPOPT_NOP;
1803 			}
1804 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1805 				continue;
1806 			optlen += TCPOLEN_WINDOW;
1807 			*optp++ = TCPOPT_WINDOW;
1808 			*optp++ = TCPOLEN_WINDOW;
1809 			*optp++ = to->to_wscale;
1810 			break;
1811 		case TOF_SACKPERM:
1812 			while (optlen % 2) {
1813 				optlen += TCPOLEN_NOP;
1814 				*optp++ = TCPOPT_NOP;
1815 			}
1816 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1817 				continue;
1818 			optlen += TCPOLEN_SACK_PERMITTED;
1819 			*optp++ = TCPOPT_SACK_PERMITTED;
1820 			*optp++ = TCPOLEN_SACK_PERMITTED;
1821 			break;
1822 		case TOF_TS:
1823 			while (!optlen || optlen % 4 != 2) {
1824 				optlen += TCPOLEN_NOP;
1825 				*optp++ = TCPOPT_NOP;
1826 			}
1827 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1828 				continue;
1829 			optlen += TCPOLEN_TIMESTAMP;
1830 			*optp++ = TCPOPT_TIMESTAMP;
1831 			*optp++ = TCPOLEN_TIMESTAMP;
1832 			to->to_tsval = htonl(to->to_tsval);
1833 			to->to_tsecr = htonl(to->to_tsecr);
1834 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1835 			optp += sizeof(to->to_tsval);
1836 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1837 			optp += sizeof(to->to_tsecr);
1838 			break;
1839 		case TOF_SIGNATURE:
1840 			{
1841 			int siglen = TCPOLEN_SIGNATURE - 2;
1842 
1843 			while (!optlen || optlen % 4 != 2) {
1844 				optlen += TCPOLEN_NOP;
1845 				*optp++ = TCPOPT_NOP;
1846 			}
1847 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1848 				to->to_flags &= ~TOF_SIGNATURE;
1849 				continue;
1850 			}
1851 			optlen += TCPOLEN_SIGNATURE;
1852 			*optp++ = TCPOPT_SIGNATURE;
1853 			*optp++ = TCPOLEN_SIGNATURE;
1854 			to->to_signature = optp;
1855 			while (siglen--)
1856 				 *optp++ = 0;
1857 			break;
1858 			}
1859 		case TOF_SACK:
1860 			{
1861 			int sackblks = 0;
1862 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1863 			tcp_seq sack_seq;
1864 
1865 			while (!optlen || optlen % 4 != 2) {
1866 				optlen += TCPOLEN_NOP;
1867 				*optp++ = TCPOPT_NOP;
1868 			}
1869 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1870 				continue;
1871 			optlen += TCPOLEN_SACKHDR;
1872 			*optp++ = TCPOPT_SACK;
1873 			sackblks = min(to->to_nsacks,
1874 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1875 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1876 			while (sackblks--) {
1877 				sack_seq = htonl(sack->start);
1878 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1879 				optp += sizeof(sack_seq);
1880 				sack_seq = htonl(sack->end);
1881 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1882 				optp += sizeof(sack_seq);
1883 				optlen += TCPOLEN_SACK;
1884 				sack++;
1885 			}
1886 			TCPSTAT_INC(tcps_sack_send_blocks);
1887 			break;
1888 			}
1889 		case TOF_FASTOPEN:
1890 			{
1891 			int total_len;
1892 
1893 			/* XXX is there any point to aligning this option? */
1894 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1895 			if (TCP_MAXOLEN - optlen < total_len) {
1896 				to->to_flags &= ~TOF_FASTOPEN;
1897 				continue;
1898 			}
1899 			*optp++ = TCPOPT_FAST_OPEN;
1900 			*optp++ = total_len;
1901 			if (to->to_tfo_len > 0) {
1902 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1903 				optp += to->to_tfo_len;
1904 			}
1905 			optlen += total_len;
1906 			break;
1907 			}
1908 		default:
1909 			panic("%s: unknown TCP option type", __func__);
1910 			break;
1911 		}
1912 	}
1913 
1914 	/* Terminate and pad TCP options to a 4 byte boundary. */
1915 	if (optlen % 4) {
1916 		optlen += TCPOLEN_EOL;
1917 		*optp++ = TCPOPT_EOL;
1918 	}
1919 	/*
1920 	 * According to RFC 793 (STD0007):
1921 	 *   "The content of the header beyond the End-of-Option option
1922 	 *    must be header padding (i.e., zero)."
1923 	 *   and later: "The padding is composed of zeros."
1924 	 */
1925 	while (optlen % 4) {
1926 		optlen += TCPOLEN_PAD;
1927 		*optp++ = TCPOPT_PAD;
1928 	}
1929 
1930 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1931 	return (optlen);
1932 }
1933 
1934 /*
1935  * This is a copy of m_copym(), taking the TSO segment size/limit
1936  * constraints into account, and advancing the sndptr as it goes.
1937  */
1938 struct mbuf *
1939 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1940     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1941 {
1942 #ifdef KERN_TLS
1943 	struct ktls_session *tls, *ntls;
1944 	struct mbuf *start __diagused;
1945 #endif
1946 	struct mbuf *n, **np;
1947 	struct mbuf *top;
1948 	int32_t off = off0;
1949 	int32_t len = *plen;
1950 	int32_t fragsize;
1951 	int32_t len_cp = 0;
1952 	int32_t *pkthdrlen;
1953 	uint32_t mlen, frags;
1954 	bool copyhdr;
1955 
1956 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1957 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1958 	if (off == 0 && m->m_flags & M_PKTHDR)
1959 		copyhdr = true;
1960 	else
1961 		copyhdr = false;
1962 	while (off > 0) {
1963 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1964 		if (off < m->m_len)
1965 			break;
1966 		off -= m->m_len;
1967 		if ((sb) && (m == sb->sb_sndptr)) {
1968 			sb->sb_sndptroff += m->m_len;
1969 			sb->sb_sndptr = m->m_next;
1970 		}
1971 		m = m->m_next;
1972 	}
1973 	np = &top;
1974 	top = NULL;
1975 	pkthdrlen = NULL;
1976 #ifdef KERN_TLS
1977 	if (hw_tls && (m->m_flags & M_EXTPG))
1978 		tls = m->m_epg_tls;
1979 	else
1980 		tls = NULL;
1981 	start = m;
1982 #endif
1983 	while (len > 0) {
1984 		if (m == NULL) {
1985 			KASSERT(len == M_COPYALL,
1986 			    ("tcp_m_copym, length > size of mbuf chain"));
1987 			*plen = len_cp;
1988 			if (pkthdrlen != NULL)
1989 				*pkthdrlen = len_cp;
1990 			break;
1991 		}
1992 #ifdef KERN_TLS
1993 		if (hw_tls) {
1994 			if (m->m_flags & M_EXTPG)
1995 				ntls = m->m_epg_tls;
1996 			else
1997 				ntls = NULL;
1998 
1999 			/*
2000 			 * Avoid mixing TLS records with handshake
2001 			 * data or TLS records from different
2002 			 * sessions.
2003 			 */
2004 			if (tls != ntls) {
2005 				MPASS(m != start);
2006 				*plen = len_cp;
2007 				if (pkthdrlen != NULL)
2008 					*pkthdrlen = len_cp;
2009 				break;
2010 			}
2011 		}
2012 #endif
2013 		mlen = min(len, m->m_len - off);
2014 		if (seglimit) {
2015 			/*
2016 			 * For M_EXTPG mbufs, add 3 segments
2017 			 * + 1 in case we are crossing page boundaries
2018 			 * + 2 in case the TLS hdr/trailer are used
2019 			 * It is cheaper to just add the segments
2020 			 * than it is to take the cache miss to look
2021 			 * at the mbuf ext_pgs state in detail.
2022 			 */
2023 			if (m->m_flags & M_EXTPG) {
2024 				fragsize = min(segsize, PAGE_SIZE);
2025 				frags = 3;
2026 			} else {
2027 				fragsize = segsize;
2028 				frags = 0;
2029 			}
2030 
2031 			/* Break if we really can't fit anymore. */
2032 			if ((frags + 1) >= seglimit) {
2033 				*plen =	len_cp;
2034 				if (pkthdrlen != NULL)
2035 					*pkthdrlen = len_cp;
2036 				break;
2037 			}
2038 
2039 			/*
2040 			 * Reduce size if you can't copy the whole
2041 			 * mbuf. If we can't copy the whole mbuf, also
2042 			 * adjust len so the loop will end after this
2043 			 * mbuf.
2044 			 */
2045 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2046 				mlen = (seglimit - frags - 1) * fragsize;
2047 				len = mlen;
2048 				*plen = len_cp + len;
2049 				if (pkthdrlen != NULL)
2050 					*pkthdrlen = *plen;
2051 			}
2052 			frags += howmany(mlen, fragsize);
2053 			if (frags == 0)
2054 				frags++;
2055 			seglimit -= frags;
2056 			KASSERT(seglimit > 0,
2057 			    ("%s: seglimit went too low", __func__));
2058 		}
2059 		if (copyhdr)
2060 			n = m_gethdr(M_NOWAIT, m->m_type);
2061 		else
2062 			n = m_get(M_NOWAIT, m->m_type);
2063 		*np = n;
2064 		if (n == NULL)
2065 			goto nospace;
2066 		if (copyhdr) {
2067 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2068 				goto nospace;
2069 			if (len == M_COPYALL)
2070 				n->m_pkthdr.len -= off0;
2071 			else
2072 				n->m_pkthdr.len = len;
2073 			pkthdrlen = &n->m_pkthdr.len;
2074 			copyhdr = false;
2075 		}
2076 		n->m_len = mlen;
2077 		len_cp += n->m_len;
2078 		if (m->m_flags & (M_EXT | M_EXTPG)) {
2079 			n->m_data = m->m_data + off;
2080 			mb_dupcl(n, m);
2081 		} else
2082 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2083 			    (u_int)n->m_len);
2084 
2085 		if (sb && (sb->sb_sndptr == m) &&
2086 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2087 			sb->sb_sndptroff += m->m_len;
2088 			sb->sb_sndptr = m->m_next;
2089 		}
2090 		off = 0;
2091 		if (len != M_COPYALL) {
2092 			len -= n->m_len;
2093 		}
2094 		m = m->m_next;
2095 		np = &n->m_next;
2096 	}
2097 	return (top);
2098 nospace:
2099 	m_freem(top);
2100 	return (NULL);
2101 }
2102 
2103 void
2104 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2105 {
2106 
2107 	/*
2108 	 * Automatic sizing of send socket buffer.  Often the send buffer
2109 	 * size is not optimally adjusted to the actual network conditions
2110 	 * at hand (delay bandwidth product).  Setting the buffer size too
2111 	 * small limits throughput on links with high bandwidth and high
2112 	 * delay (eg. trans-continental/oceanic links).  Setting the
2113 	 * buffer size too big consumes too much real kernel memory,
2114 	 * especially with many connections on busy servers.
2115 	 *
2116 	 * The criteria to step up the send buffer one notch are:
2117 	 *  1. receive window of remote host is larger than send buffer
2118 	 *     (with a fudge factor of 5/4th);
2119 	 *  2. send buffer is filled to 7/8th with data (so we actually
2120 	 *     have data to make use of it);
2121 	 *  3. send buffer fill has not hit maximal automatic size;
2122 	 *  4. our send window (slow start and cogestion controlled) is
2123 	 *     larger than sent but unacknowledged data in send buffer.
2124 	 *
2125 	 * The remote host receive window scaling factor may limit the
2126 	 * growing of the send buffer before it reaches its allowed
2127 	 * maximum.
2128 	 *
2129 	 * It scales directly with slow start or congestion window
2130 	 * and does at most one step per received ACK.  This fast
2131 	 * scaling has the drawback of growing the send buffer beyond
2132 	 * what is strictly necessary to make full use of a given
2133 	 * delay*bandwidth product.  However testing has shown this not
2134 	 * to be much of an problem.  At worst we are trading wasting
2135 	 * of available bandwidth (the non-use of it) for wasting some
2136 	 * socket buffer memory.
2137 	 *
2138 	 * TODO: Shrink send buffer during idle periods together
2139 	 * with congestion window.  Requires another timer.  Has to
2140 	 * wait for upcoming tcp timer rewrite.
2141 	 *
2142 	 * XXXGL: should there be used sbused() or sbavail()?
2143 	 */
2144 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2145 		int lowat;
2146 
2147 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2148 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2149 		    sbused(&so->so_snd) >=
2150 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2151 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2152 		    sendwin >= (sbused(&so->so_snd) -
2153 		    (tp->snd_nxt - tp->snd_una))) {
2154 			if (!sbreserve_locked(so, SO_SND,
2155 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2156 			     V_tcp_autosndbuf_max), curthread))
2157 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2158 		}
2159 	}
2160 }
2161