xref: /freebsd/sys/netinet/tcp_output.c (revision 6597ea4481bb29b6ac72b9c55038b1e98634bf92)
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 			TCPSTAT_INC(tcps_sack_rexmits);
334 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
335 			    min(len, tcp_maxseg(tp)));
336 		}
337 	}
338 after_sack_rexmit:
339 	/*
340 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
341 	 * state flags.
342 	 */
343 	if (tp->t_flags & TF_NEEDFIN)
344 		flags |= TH_FIN;
345 	if (tp->t_flags & TF_NEEDSYN)
346 		flags |= TH_SYN;
347 
348 	SOCKBUF_LOCK(&so->so_snd);
349 	/*
350 	 * If in persist timeout with window of 0, send 1 byte.
351 	 * Otherwise, if window is small but nonzero
352 	 * and timer expired, we will send what we can
353 	 * and go to transmit state.
354 	 */
355 	if (tp->t_flags & TF_FORCEDATA) {
356 		if (sendwin == 0) {
357 			/*
358 			 * If we still have some data to send, then
359 			 * clear the FIN bit.  Usually this would
360 			 * happen below when it realizes that we
361 			 * aren't sending all the data.  However,
362 			 * if we have exactly 1 byte of unsent data,
363 			 * then it won't clear the FIN bit below,
364 			 * and if we are in persist state, we wind
365 			 * up sending the packet without recording
366 			 * that we sent the FIN bit.
367 			 *
368 			 * We can't just blindly clear the FIN bit,
369 			 * because if we don't have any more data
370 			 * to send then the probe will be the FIN
371 			 * itself.
372 			 */
373 			if (off < sbused(&so->so_snd))
374 				flags &= ~TH_FIN;
375 			sendwin = 1;
376 		} else {
377 			tcp_timer_activate(tp, TT_PERSIST, 0);
378 			tp->t_rxtshift = 0;
379 		}
380 	}
381 
382 	/*
383 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
384 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
385 	 * a negative length.  This can also occur when TCP opens up
386 	 * its congestion window while receiving additional duplicate
387 	 * acks after fast-retransmit because TCP will reset snd_nxt
388 	 * to snd_max after the fast-retransmit.
389 	 *
390 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
391 	 * be set to snd_una, the offset will be 0, and the length may
392 	 * wind up 0.
393 	 *
394 	 * If sack_rxmit is true we are retransmitting from the scoreboard
395 	 * in which case len is already set.
396 	 */
397 	if (sack_rxmit == 0) {
398 		if (sack_bytes_rxmt == 0)
399 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
400 			    off);
401 		else {
402 			int32_t cwin;
403 
404 			/*
405 			 * We are inside of a SACK recovery episode and are
406 			 * sending new data, having retransmitted all the
407 			 * data possible in the scoreboard.
408 			 */
409 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
410 			    off);
411 			/*
412 			 * Don't remove this (len > 0) check !
413 			 * We explicitly check for len > 0 here (although it
414 			 * isn't really necessary), to work around a gcc
415 			 * optimization issue - to force gcc to compute
416 			 * len above. Without this check, the computation
417 			 * of len is bungled by the optimizer.
418 			 */
419 			if (len > 0) {
420 				cwin = tp->snd_cwnd - imax(0, (int32_t)
421 					(tp->snd_nxt - tp->snd_recover)) -
422 					sack_bytes_rxmt;
423 				if (cwin < 0)
424 					cwin = 0;
425 				len = imin(len, cwin);
426 			}
427 		}
428 	}
429 
430 	/*
431 	 * Lop off SYN bit if it has already been sent.  However, if this
432 	 * is SYN-SENT state and if segment contains data and if we don't
433 	 * know that foreign host supports TAO, suppress sending segment.
434 	 */
435 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
436 		if (tp->t_state != TCPS_SYN_RECEIVED)
437 			flags &= ~TH_SYN;
438 		/*
439 		 * When sending additional segments following a TFO SYN|ACK,
440 		 * do not include the SYN bit.
441 		 */
442 		if (IS_FASTOPEN(tp->t_flags) &&
443 		    (tp->t_state == TCPS_SYN_RECEIVED))
444 			flags &= ~TH_SYN;
445 		off--, len++;
446 	}
447 
448 	/*
449 	 * Be careful not to send data and/or FIN on SYN segments.
450 	 * This measure is needed to prevent interoperability problems
451 	 * with not fully conformant TCP implementations.
452 	 */
453 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
454 		len = 0;
455 		flags &= ~TH_FIN;
456 	}
457 
458 	/*
459 	 * On TFO sockets, ensure no data is sent in the following cases:
460 	 *
461 	 *  - When retransmitting SYN|ACK on a passively-created socket
462 	 *
463 	 *  - When retransmitting SYN on an actively created socket
464 	 *
465 	 *  - When sending a zero-length cookie (cookie request) on an
466 	 *    actively created socket
467 	 *
468 	 *  - When the socket is in the CLOSED state (RST is being sent)
469 	 */
470 	if (IS_FASTOPEN(tp->t_flags) &&
471 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
472 	     ((tp->t_state == TCPS_SYN_SENT) &&
473 	      (tp->t_tfo_client_cookie_len == 0)) ||
474 	     (flags & TH_RST)))
475 		len = 0;
476 	if (len <= 0) {
477 		/*
478 		 * If FIN has been sent but not acked,
479 		 * but we haven't been called to retransmit,
480 		 * len will be < 0.  Otherwise, window shrank
481 		 * after we sent into it.  If window shrank to 0,
482 		 * cancel pending retransmit, pull snd_nxt back
483 		 * to (closed) window, and set the persist timer
484 		 * if it isn't already going.  If the window didn't
485 		 * close completely, just wait for an ACK.
486 		 *
487 		 * We also do a general check here to ensure that
488 		 * we will set the persist timer when we have data
489 		 * to send, but a 0-byte window. This makes sure
490 		 * the persist timer is set even if the packet
491 		 * hits one of the "goto send" lines below.
492 		 */
493 		len = 0;
494 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
495 		    (off < (int) sbavail(&so->so_snd)) &&
496 		    !tcp_timer_active(tp, TT_PERSIST)) {
497 			tcp_timer_activate(tp, TT_REXMT, 0);
498 			tp->t_rxtshift = 0;
499 			tp->snd_nxt = tp->snd_una;
500 			if (!tcp_timer_active(tp, TT_PERSIST))
501 				tcp_setpersist(tp);
502 		}
503 	}
504 
505 	/* len will be >= 0 after this point. */
506 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
507 
508 	tcp_sndbuf_autoscale(tp, so, sendwin);
509 
510 	/*
511 	 * Decide if we can use TCP Segmentation Offloading (if supported by
512 	 * hardware).
513 	 *
514 	 * TSO may only be used if we are in a pure bulk sending state.  The
515 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
516 	 * IP options prevent using TSO.  With TSO the TCP header is the same
517 	 * (except for the sequence number) for all generated packets.  This
518 	 * makes it impossible to transmit any options which vary per generated
519 	 * segment or packet.
520 	 *
521 	 * IPv4 handling has a clear separation of ip options and ip header
522 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
523 	 * the right thing below to provide length of just ip options and thus
524 	 * checking for ipoptlen is enough to decide if ip options are present.
525 	 */
526 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
527 	/*
528 	 * Pre-calculate here as we save another lookup into the darknesses
529 	 * of IPsec that way and can actually decide if TSO is ok.
530 	 */
531 #ifdef INET6
532 	if (isipv6 && IPSEC_ENABLED(ipv6))
533 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
534 #ifdef INET
535 	else
536 #endif
537 #endif /* INET6 */
538 #ifdef INET
539 	if (IPSEC_ENABLED(ipv4))
540 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
541 #endif /* INET */
542 #endif /* IPSEC */
543 #ifdef INET6
544 	if (isipv6)
545 		ipoptlen = ip6_optlen(inp);
546 	else
547 #endif
548 	if (inp->inp_options)
549 		ipoptlen = inp->inp_options->m_len -
550 				offsetof(struct ipoption, ipopt_list);
551 	else
552 		ipoptlen = 0;
553 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
554 	ipoptlen += ipsec_optlen;
555 #endif
556 
557 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
558 	    (tp->t_port == 0) &&
559 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
560 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
561 	    ipoptlen == 0 && !(flags & TH_SYN))
562 		tso = 1;
563 
564 	if (sack_rxmit) {
565 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
566 			flags &= ~TH_FIN;
567 	} else {
568 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
569 		    sbused(&so->so_snd)))
570 			flags &= ~TH_FIN;
571 	}
572 
573 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
574 	    (long)TCP_MAXWIN << tp->rcv_scale);
575 
576 	/*
577 	 * Sender silly window avoidance.   We transmit under the following
578 	 * conditions when len is non-zero:
579 	 *
580 	 *	- We have a full segment (or more with TSO)
581 	 *	- This is the last buffer in a write()/send() and we are
582 	 *	  either idle or running NODELAY
583 	 *	- we've timed out (e.g. persist timer)
584 	 *	- we have more then 1/2 the maximum send window's worth of
585 	 *	  data (receiver may be limited the window size)
586 	 *	- we need to retransmit
587 	 */
588 	if (len) {
589 		if (len >= tp->t_maxseg)
590 			goto send;
591 		/*
592 		 * As the TCP header options are now
593 		 * considered when setting up the initial
594 		 * window, we would not send the last segment
595 		 * if we skip considering the option length here.
596 		 * Note: this may not work when tcp headers change
597 		 * very dynamically in the future.
598 		 */
599 		if ((((tp->t_flags & TF_SIGNATURE) ?
600 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
601 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
602 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
603 		    len) >= tp->t_maxseg)
604 			goto send;
605 		/*
606 		 * NOTE! on localhost connections an 'ack' from the remote
607 		 * end may occur synchronously with the output and cause
608 		 * us to flush a buffer queued with moretocome.  XXX
609 		 *
610 		 * note: the len + off check is almost certainly unnecessary.
611 		 */
612 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
613 		    (idle || (tp->t_flags & TF_NODELAY)) &&
614 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
615 		    (tp->t_flags & TF_NOPUSH) == 0) {
616 			goto send;
617 		}
618 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
619 			goto send;
620 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
621 			goto send;
622 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
623 			goto send;
624 		if (sack_rxmit)
625 			goto send;
626 	}
627 
628 	/*
629 	 * Sending of standalone window updates.
630 	 *
631 	 * Window updates are important when we close our window due to a
632 	 * full socket buffer and are opening it again after the application
633 	 * reads data from it.  Once the window has opened again and the
634 	 * remote end starts to send again the ACK clock takes over and
635 	 * provides the most current window information.
636 	 *
637 	 * We must avoid the silly window syndrome whereas every read
638 	 * from the receive buffer, no matter how small, causes a window
639 	 * update to be sent.  We also should avoid sending a flurry of
640 	 * window updates when the socket buffer had queued a lot of data
641 	 * and the application is doing small reads.
642 	 *
643 	 * Prevent a flurry of pointless window updates by only sending
644 	 * an update when we can increase the advertized window by more
645 	 * than 1/4th of the socket buffer capacity.  When the buffer is
646 	 * getting full or is very small be more aggressive and send an
647 	 * update whenever we can increase by two mss sized segments.
648 	 * In all other situations the ACK's to new incoming data will
649 	 * carry further window increases.
650 	 *
651 	 * Don't send an independent window update if a delayed
652 	 * ACK is pending (it will get piggy-backed on it) or the
653 	 * remote side already has done a half-close and won't send
654 	 * more data.  Skip this if the connection is in T/TCP
655 	 * half-open state.
656 	 */
657 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
658 	    !(tp->t_flags & TF_DELACK) &&
659 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
660 		/*
661 		 * "adv" is the amount we could increase the window,
662 		 * taking into account that we are limited by
663 		 * TCP_MAXWIN << tp->rcv_scale.
664 		 */
665 		int32_t adv;
666 		int oldwin;
667 
668 		adv = recwin;
669 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
670 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
671 			if (adv > oldwin)
672 				adv -= oldwin;
673 			else
674 				adv = 0;
675 		} else
676 			oldwin = 0;
677 
678 		/*
679 		 * If the new window size ends up being the same as or less
680 		 * than the old size when it is scaled, then don't force
681 		 * a window update.
682 		 */
683 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
684 			goto dontupdate;
685 
686 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
687 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
688 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
689 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
690 		     adv >= TCP_MAXWIN << tp->rcv_scale))
691 			goto send;
692 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
693 			goto send;
694 	}
695 dontupdate:
696 
697 	/*
698 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
699 	 * is also a catch-all for the retransmit timer timeout case.
700 	 */
701 	if (tp->t_flags & TF_ACKNOW)
702 		goto send;
703 	if ((flags & TH_RST) ||
704 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
705 		goto send;
706 	if (SEQ_GT(tp->snd_up, tp->snd_una))
707 		goto send;
708 	/*
709 	 * If our state indicates that FIN should be sent
710 	 * and we have not yet done so, then we need to send.
711 	 */
712 	if (flags & TH_FIN &&
713 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
714 		goto send;
715 	/*
716 	 * In SACK, it is possible for tcp_output to fail to send a segment
717 	 * after the retransmission timer has been turned off.  Make sure
718 	 * that the retransmission timer is set.
719 	 */
720 	if ((tp->t_flags & TF_SACK_PERMIT) &&
721 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
722 	    !tcp_timer_active(tp, TT_REXMT) &&
723 	    !tcp_timer_active(tp, TT_PERSIST)) {
724 		tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
725 		goto just_return;
726 	}
727 	/*
728 	 * TCP window updates are not reliable, rather a polling protocol
729 	 * using ``persist'' packets is used to insure receipt of window
730 	 * updates.  The three ``states'' for the output side are:
731 	 *	idle			not doing retransmits or persists
732 	 *	persisting		to move a small or zero window
733 	 *	(re)transmitting	and thereby not persisting
734 	 *
735 	 * tcp_timer_active(tp, TT_PERSIST)
736 	 *	is true when we are in persist state.
737 	 * (tp->t_flags & TF_FORCEDATA)
738 	 *	is set when we are called to send a persist packet.
739 	 * tcp_timer_active(tp, TT_REXMT)
740 	 *	is set when we are retransmitting
741 	 * The output side is idle when both timers are zero.
742 	 *
743 	 * If send window is too small, there is data to transmit, and no
744 	 * retransmit or persist is pending, then go to persist state.
745 	 * If nothing happens soon, send when timer expires:
746 	 * if window is nonzero, transmit what we can,
747 	 * otherwise force out a byte.
748 	 */
749 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
750 	    !tcp_timer_active(tp, TT_PERSIST)) {
751 		tp->t_rxtshift = 0;
752 		tcp_setpersist(tp);
753 	}
754 
755 	/*
756 	 * No reason to send a segment, just return.
757 	 */
758 just_return:
759 	SOCKBUF_UNLOCK(&so->so_snd);
760 	return (0);
761 
762 send:
763 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
764 	if (len > 0) {
765 		if (len >= tp->t_maxseg)
766 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
767 		else
768 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
769 	}
770 	/*
771 	 * Before ESTABLISHED, force sending of initial options
772 	 * unless TCP set not to do any options.
773 	 * NOTE: we assume that the IP/TCP header plus TCP options
774 	 * always fit in a single mbuf, leaving room for a maximum
775 	 * link header, i.e.
776 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
777 	 */
778 	optlen = 0;
779 #ifdef INET6
780 	if (isipv6)
781 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
782 	else
783 #endif
784 		hdrlen = sizeof (struct tcpiphdr);
785 
786 	if (flags & TH_SYN) {
787 		tp->snd_nxt = tp->iss;
788 	}
789 
790 	/*
791 	 * Compute options for segment.
792 	 * We only have to care about SYN and established connection
793 	 * segments.  Options for SYN-ACK segments are handled in TCP
794 	 * syncache.
795 	 */
796 	to.to_flags = 0;
797 	if ((tp->t_flags & TF_NOOPT) == 0) {
798 		/* Maximum segment size. */
799 		if (flags & TH_SYN) {
800 			to.to_mss = tcp_mssopt(&inp->inp_inc);
801 			if (tp->t_port)
802 				to.to_mss -= V_tcp_udp_tunneling_overhead;
803 			to.to_flags |= TOF_MSS;
804 
805 			/*
806 			 * On SYN or SYN|ACK transmits on TFO connections,
807 			 * only include the TFO option if it is not a
808 			 * retransmit, as the presence of the TFO option may
809 			 * have caused the original SYN or SYN|ACK to have
810 			 * been dropped by a middlebox.
811 			 */
812 			if (IS_FASTOPEN(tp->t_flags) &&
813 			    (tp->t_rxtshift == 0)) {
814 				if (tp->t_state == TCPS_SYN_RECEIVED) {
815 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
816 					to.to_tfo_cookie =
817 					    (u_int8_t *)&tp->t_tfo_cookie.server;
818 					to.to_flags |= TOF_FASTOPEN;
819 					wanted_cookie = 1;
820 				} else if (tp->t_state == TCPS_SYN_SENT) {
821 					to.to_tfo_len =
822 					    tp->t_tfo_client_cookie_len;
823 					to.to_tfo_cookie =
824 					    tp->t_tfo_cookie.client;
825 					to.to_flags |= TOF_FASTOPEN;
826 					wanted_cookie = 1;
827 					/*
828 					 * If we wind up having more data to
829 					 * send with the SYN than can fit in
830 					 * one segment, don't send any more
831 					 * until the SYN|ACK comes back from
832 					 * the other end.
833 					 */
834 					dont_sendalot = 1;
835 				}
836 			}
837 		}
838 		/* Window scaling. */
839 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
840 			to.to_wscale = tp->request_r_scale;
841 			to.to_flags |= TOF_SCALE;
842 		}
843 		/* Timestamps. */
844 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
845 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
846 			curticks = tcp_ts_getticks();
847 			to.to_tsval = curticks + tp->ts_offset;
848 			to.to_tsecr = tp->ts_recent;
849 			to.to_flags |= TOF_TS;
850 			if (tp->t_rxtshift == 1)
851 				tp->t_badrxtwin = curticks;
852 		}
853 
854 		/* Set receive buffer autosizing timestamp. */
855 		if (tp->rfbuf_ts == 0 &&
856 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
857 			tp->rfbuf_ts = tcp_ts_getticks();
858 
859 		/* Selective ACK's. */
860 		if (tp->t_flags & TF_SACK_PERMIT) {
861 			if (flags & TH_SYN)
862 				to.to_flags |= TOF_SACKPERM;
863 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
864 			    tp->rcv_numsacks > 0) {
865 				to.to_flags |= TOF_SACK;
866 				to.to_nsacks = tp->rcv_numsacks;
867 				to.to_sacks = (u_char *)tp->sackblks;
868 			}
869 		}
870 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
871 		/* TCP-MD5 (RFC2385). */
872 		/*
873 		 * Check that TCP_MD5SIG is enabled in tcpcb to
874 		 * account the size needed to set this TCP option.
875 		 */
876 		if (tp->t_flags & TF_SIGNATURE)
877 			to.to_flags |= TOF_SIGNATURE;
878 #endif /* TCP_SIGNATURE */
879 
880 		/* Processing the options. */
881 		hdrlen += optlen = tcp_addoptions(&to, opt);
882 		/*
883 		 * If we wanted a TFO option to be added, but it was unable
884 		 * to fit, ensure no data is sent.
885 		 */
886 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
887 		    !(to.to_flags & TOF_FASTOPEN))
888 			len = 0;
889 	}
890 	if (tp->t_port) {
891 		if (V_tcp_udp_tunneling_port == 0) {
892 			/* The port was removed?? */
893 			SOCKBUF_UNLOCK(&so->so_snd);
894 			return (EHOSTUNREACH);
895 		}
896 		hdrlen += sizeof(struct udphdr);
897 	}
898 	/*
899 	 * Adjust data length if insertion of options will
900 	 * bump the packet length beyond the t_maxseg length.
901 	 * Clear the FIN bit because we cut off the tail of
902 	 * the segment.
903 	 */
904 	if (len + optlen + ipoptlen > tp->t_maxseg) {
905 		flags &= ~TH_FIN;
906 
907 		if (tso) {
908 			u_int if_hw_tsomax;
909 			u_int moff;
910 			int max_len;
911 
912 			/* extract TSO information */
913 			if_hw_tsomax = tp->t_tsomax;
914 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
915 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
916 
917 			/*
918 			 * Limit a TSO burst to prevent it from
919 			 * overflowing or exceeding the maximum length
920 			 * allowed by the network interface:
921 			 */
922 			KASSERT(ipoptlen == 0,
923 			    ("%s: TSO can't do IP options", __func__));
924 
925 			/*
926 			 * Check if we should limit by maximum payload
927 			 * length:
928 			 */
929 			if (if_hw_tsomax != 0) {
930 				/* compute maximum TSO length */
931 				max_len = (if_hw_tsomax - hdrlen -
932 				    max_linkhdr);
933 				if (max_len <= 0) {
934 					len = 0;
935 				} else if (len > max_len) {
936 					sendalot = 1;
937 					len = max_len;
938 				}
939 			}
940 
941 			/*
942 			 * Prevent the last segment from being
943 			 * fractional unless the send sockbuf can be
944 			 * emptied:
945 			 */
946 			max_len = (tp->t_maxseg - optlen);
947 			if (((uint32_t)off + (uint32_t)len) <
948 			    sbavail(&so->so_snd)) {
949 				moff = len % max_len;
950 				if (moff != 0) {
951 					len -= moff;
952 					sendalot = 1;
953 				}
954 			}
955 
956 			/*
957 			 * In case there are too many small fragments
958 			 * don't use TSO:
959 			 */
960 			if (len <= max_len) {
961 				len = max_len;
962 				sendalot = 1;
963 				tso = 0;
964 			}
965 
966 			/*
967 			 * Send the FIN in a separate segment
968 			 * after the bulk sending is done.
969 			 * We don't trust the TSO implementations
970 			 * to clear the FIN flag on all but the
971 			 * last segment.
972 			 */
973 			if (tp->t_flags & TF_NEEDFIN)
974 				sendalot = 1;
975 		} else {
976 			if (optlen + ipoptlen >= tp->t_maxseg) {
977 				/*
978 				 * Since we don't have enough space to put
979 				 * the IP header chain and the TCP header in
980 				 * one packet as required by RFC 7112, don't
981 				 * send it. Also ensure that at least one
982 				 * byte of the payload can be put into the
983 				 * TCP segment.
984 				 */
985 				SOCKBUF_UNLOCK(&so->so_snd);
986 				error = EMSGSIZE;
987 				sack_rxmit = 0;
988 				goto out;
989 			}
990 			len = tp->t_maxseg - optlen - ipoptlen;
991 			sendalot = 1;
992 			if (dont_sendalot)
993 				sendalot = 0;
994 		}
995 	} else
996 		tso = 0;
997 
998 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
999 	    ("%s: len > IP_MAXPACKET", __func__));
1000 
1001 /*#ifdef DIAGNOSTIC*/
1002 #ifdef INET6
1003 	if (max_linkhdr + hdrlen > MCLBYTES)
1004 #else
1005 	if (max_linkhdr + hdrlen > MHLEN)
1006 #endif
1007 		panic("tcphdr too big");
1008 /*#endif*/
1009 
1010 	/*
1011 	 * This KASSERT is here to catch edge cases at a well defined place.
1012 	 * Before, those had triggered (random) panic conditions further down.
1013 	 */
1014 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1015 
1016 	/*
1017 	 * Grab a header mbuf, attaching a copy of data to
1018 	 * be transmitted, and initialize the header from
1019 	 * the template for sends on this connection.
1020 	 */
1021 	if (len) {
1022 		struct mbuf *mb;
1023 		struct sockbuf *msb;
1024 		u_int moff;
1025 
1026 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1027 			TCPSTAT_INC(tcps_sndprobe);
1028 #ifdef STATS
1029 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1030 				stats_voi_update_abs_u32(tp->t_stats,
1031 				VOI_TCP_RETXPB, len);
1032 			else
1033 				stats_voi_update_abs_u64(tp->t_stats,
1034 				    VOI_TCP_TXPB, len);
1035 #endif /* STATS */
1036 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1037 			tp->t_sndrexmitpack++;
1038 			TCPSTAT_INC(tcps_sndrexmitpack);
1039 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1040 #ifdef STATS
1041 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1042 			    len);
1043 #endif /* STATS */
1044 		} else {
1045 			TCPSTAT_INC(tcps_sndpack);
1046 			TCPSTAT_ADD(tcps_sndbyte, len);
1047 #ifdef STATS
1048 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1049 			    len);
1050 #endif /* STATS */
1051 		}
1052 #ifdef INET6
1053 		if (MHLEN < hdrlen + max_linkhdr)
1054 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1055 		else
1056 #endif
1057 			m = m_gethdr(M_NOWAIT, MT_DATA);
1058 
1059 		if (m == NULL) {
1060 			SOCKBUF_UNLOCK(&so->so_snd);
1061 			error = ENOBUFS;
1062 			sack_rxmit = 0;
1063 			goto out;
1064 		}
1065 
1066 		m->m_data += max_linkhdr;
1067 		m->m_len = hdrlen;
1068 
1069 		/*
1070 		 * Start the m_copy functions from the closest mbuf
1071 		 * to the offset in the socket buffer chain.
1072 		 */
1073 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1074 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1075 			m_copydata(mb, moff, len,
1076 			    mtod(m, caddr_t) + hdrlen);
1077 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1078 				sbsndptr_adv(&so->so_snd, mb, len);
1079 			m->m_len += len;
1080 		} else {
1081 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1082 				msb = NULL;
1083 			else
1084 				msb = &so->so_snd;
1085 			m->m_next = tcp_m_copym(mb, moff,
1086 			    &len, if_hw_tsomaxsegcount,
1087 			    if_hw_tsomaxsegsize, msb, hw_tls);
1088 			if (len <= (tp->t_maxseg - optlen)) {
1089 				/*
1090 				 * Must have ran out of mbufs for the copy
1091 				 * shorten it to no longer need tso. Lets
1092 				 * not put on sendalot since we are low on
1093 				 * mbufs.
1094 				 */
1095 				tso = 0;
1096 			}
1097 			if (m->m_next == NULL) {
1098 				SOCKBUF_UNLOCK(&so->so_snd);
1099 				(void) m_free(m);
1100 				error = ENOBUFS;
1101 				sack_rxmit = 0;
1102 				goto out;
1103 			}
1104 		}
1105 
1106 		/*
1107 		 * If we're sending everything we've got, set PUSH.
1108 		 * (This will keep happy those implementations which only
1109 		 * give data to the user when a buffer fills or
1110 		 * a PUSH comes in.)
1111 		 */
1112 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1113 		    !(flags & TH_SYN))
1114 			flags |= TH_PUSH;
1115 		SOCKBUF_UNLOCK(&so->so_snd);
1116 	} else {
1117 		SOCKBUF_UNLOCK(&so->so_snd);
1118 		if (tp->t_flags & TF_ACKNOW)
1119 			TCPSTAT_INC(tcps_sndacks);
1120 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1121 			TCPSTAT_INC(tcps_sndctrl);
1122 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1123 			TCPSTAT_INC(tcps_sndurg);
1124 		else
1125 			TCPSTAT_INC(tcps_sndwinup);
1126 
1127 		m = m_gethdr(M_NOWAIT, MT_DATA);
1128 		if (m == NULL) {
1129 			error = ENOBUFS;
1130 			sack_rxmit = 0;
1131 			goto out;
1132 		}
1133 #ifdef INET6
1134 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1135 		    MHLEN >= hdrlen) {
1136 			M_ALIGN(m, hdrlen);
1137 		} else
1138 #endif
1139 		m->m_data += max_linkhdr;
1140 		m->m_len = hdrlen;
1141 	}
1142 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1143 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1144 #ifdef MAC
1145 	mac_inpcb_create_mbuf(inp, m);
1146 #endif
1147 #ifdef INET6
1148 	if (isipv6) {
1149 		ip6 = mtod(m, struct ip6_hdr *);
1150 		if (tp->t_port) {
1151 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1152 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1153 			udp->uh_dport = tp->t_port;
1154 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1155 			udp->uh_ulen = htons(ulen);
1156 			th = (struct tcphdr *)(udp + 1);
1157 		} else {
1158 			th = (struct tcphdr *)(ip6 + 1);
1159 		}
1160 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
1161 	} else
1162 #endif /* INET6 */
1163 	{
1164 		ip = mtod(m, struct ip *);
1165 		if (tp->t_port) {
1166 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1167 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1168 			udp->uh_dport = tp->t_port;
1169 			ulen = hdrlen + len - sizeof(struct ip);
1170 			udp->uh_ulen = htons(ulen);
1171 			th = (struct tcphdr *)(udp + 1);
1172 		} else
1173 			th = (struct tcphdr *)(ip + 1);
1174 		tcpip_fillheaders(inp, tp->t_port, ip, th);
1175 	}
1176 
1177 	/*
1178 	 * Fill in fields, remembering maximum advertised
1179 	 * window for use in delaying messages about window sizes.
1180 	 * If resending a FIN, be sure not to use a new sequence number.
1181 	 */
1182 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1183 	    tp->snd_nxt == tp->snd_max)
1184 		tp->snd_nxt--;
1185 	/*
1186 	 * If we are starting a connection, send ECN setup
1187 	 * SYN packet. If we are on a retransmit, we may
1188 	 * resend those bits a number of times as per
1189 	 * RFC 3168.
1190 	 */
1191 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1192 		flags |= tcp_ecn_output_syn_sent(tp);
1193 	}
1194 	/* Also handle parallel SYN for ECN */
1195 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1196 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1197 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1198 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1199 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1200 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1201 #ifdef INET6
1202 		if (isipv6) {
1203 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1204 			ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1205 		}
1206 		else
1207 #endif
1208 		{
1209 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1210 			ip->ip_tos |= ect;
1211 		}
1212 	}
1213 
1214 	/*
1215 	 * If we are doing retransmissions, then snd_nxt will
1216 	 * not reflect the first unsent octet.  For ACK only
1217 	 * packets, we do not want the sequence number of the
1218 	 * retransmitted packet, we want the sequence number
1219 	 * of the next unsent octet.  So, if there is no data
1220 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1221 	 * when filling in ti_seq.  But if we are in persist
1222 	 * state, snd_max might reflect one byte beyond the
1223 	 * right edge of the window, so use snd_nxt in that
1224 	 * case, since we know we aren't doing a retransmission.
1225 	 * (retransmit and persist are mutually exclusive...)
1226 	 */
1227 	if (sack_rxmit == 0) {
1228 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1229 		    tcp_timer_active(tp, TT_PERSIST))
1230 			th->th_seq = htonl(tp->snd_nxt);
1231 		else
1232 			th->th_seq = htonl(tp->snd_max);
1233 	} else {
1234 		th->th_seq = htonl(p->rxmit);
1235 		p->rxmit += len;
1236 		/*
1237 		 * Lost Retransmission Detection
1238 		 * trigger resending of a (then
1239 		 * still existing) hole, when
1240 		 * fack acks recoverypoint.
1241 		 */
1242 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1243 			p->rxmit = tp->snd_recover;
1244 		tp->sackhint.sack_bytes_rexmit += len;
1245 	}
1246 	if (IN_RECOVERY(tp->t_flags)) {
1247 		/*
1248 		 * Account all bytes transmitted while
1249 		 * IN_RECOVERY, simplifying PRR and
1250 		 * Lost Retransmit Detection
1251 		 */
1252 		tp->sackhint.prr_out += len;
1253 	}
1254 	th->th_ack = htonl(tp->rcv_nxt);
1255 	if (optlen) {
1256 		bcopy(opt, th + 1, optlen);
1257 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1258 	}
1259 	tcp_set_flags(th, flags);
1260 	/*
1261 	 * Calculate receive window.  Don't shrink window,
1262 	 * but avoid silly window syndrome.
1263 	 * If a RST segment is sent, advertise a window of zero.
1264 	 */
1265 	if (flags & TH_RST) {
1266 		recwin = 0;
1267 	} else {
1268 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1269 		    recwin < tp->t_maxseg)
1270 			recwin = 0;
1271 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1272 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1273 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1274 	}
1275 	/*
1276 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1277 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1278 	 * case is handled in syncache.
1279 	 */
1280 	if (flags & TH_SYN)
1281 		th->th_win = htons((u_short)
1282 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1283 	else {
1284 		/* Avoid shrinking window with window scaling. */
1285 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1286 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1287 	}
1288 
1289 	/*
1290 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1291 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1292 	 * flag tells soreceive() to disable delayed acknowledgements when
1293 	 * draining the buffer.  This can occur if the receiver is attempting
1294 	 * to read more data than can be buffered prior to transmitting on
1295 	 * the connection.
1296 	 */
1297 	if (th->th_win == 0) {
1298 		tp->t_sndzerowin++;
1299 		tp->t_flags |= TF_RXWIN0SENT;
1300 	} else
1301 		tp->t_flags &= ~TF_RXWIN0SENT;
1302 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1303 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1304 		th->th_flags |= TH_URG;
1305 	} else
1306 		/*
1307 		 * If no urgent pointer to send, then we pull
1308 		 * the urgent pointer to the left edge of the send window
1309 		 * so that it doesn't drift into the send window on sequence
1310 		 * number wraparound.
1311 		 */
1312 		tp->snd_up = tp->snd_una;		/* drag it along */
1313 
1314 	/*
1315 	 * Put TCP length in extended header, and then
1316 	 * checksum extended header and data.
1317 	 */
1318 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1319 
1320 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1321 	if (to.to_flags & TOF_SIGNATURE) {
1322 		/*
1323 		 * Calculate MD5 signature and put it into the place
1324 		 * determined before.
1325 		 * NOTE: since TCP options buffer doesn't point into
1326 		 * mbuf's data, calculate offset and use it.
1327 		 */
1328 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1329 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1330 			/*
1331 			 * Do not send segment if the calculation of MD5
1332 			 * digest has failed.
1333 			 */
1334 			m_freem(m);
1335 			goto out;
1336 		}
1337 	}
1338 #endif
1339 #ifdef INET6
1340 	if (isipv6) {
1341 		/*
1342 		 * There is no need to fill in ip6_plen right now.
1343 		 * It will be filled later by ip6_output.
1344 		 */
1345 		if (tp->t_port) {
1346 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1347 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1348 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1349 			th->th_sum = htons(0);
1350 			UDPSTAT_INC(udps_opackets);
1351 		} else {
1352 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1353 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1354 			th->th_sum = in6_cksum_pseudo(ip6,
1355 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1356 			    0);
1357 		}
1358 	}
1359 #endif
1360 #if defined(INET6) && defined(INET)
1361 	else
1362 #endif
1363 #ifdef INET
1364 	{
1365 		if (tp->t_port) {
1366 			m->m_pkthdr.csum_flags = CSUM_UDP;
1367 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1368 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1369 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1370 			th->th_sum = htons(0);
1371 			UDPSTAT_INC(udps_opackets);
1372 		} else {
1373 			m->m_pkthdr.csum_flags = CSUM_TCP;
1374 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1375 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1376 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1377 			    IPPROTO_TCP + len + optlen));
1378 		}
1379 
1380 		/* IP version must be set here for ipv4/ipv6 checking later */
1381 		KASSERT(ip->ip_v == IPVERSION,
1382 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1383 	}
1384 #endif
1385 
1386 	/*
1387 	 * Enable TSO and specify the size of the segments.
1388 	 * The TCP pseudo header checksum is always provided.
1389 	 */
1390 	if (tso) {
1391 		KASSERT(len > tp->t_maxseg - optlen,
1392 		    ("%s: len <= tso_segsz", __func__));
1393 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1394 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1395 	}
1396 
1397 	KASSERT(len + hdrlen == m_length(m, NULL),
1398 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1399 	    __func__, len, hdrlen, m_length(m, NULL)));
1400 
1401 #ifdef TCP_HHOOK
1402 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1403 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1404 #endif
1405 
1406 	TCP_PROBE3(debug__output, tp, th, m);
1407 
1408 	/* We're getting ready to send; log now. */
1409 	/* XXXMT: We are not honoring verbose logging. */
1410 
1411 	if (tcp_bblogging_on(tp))
1412 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1413 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1414 		    NULL);
1415 	else
1416 		lgb = NULL;
1417 
1418 	/*
1419 	 * Fill in IP length and desired time to live and
1420 	 * send to IP level.  There should be a better way
1421 	 * to handle ttl and tos; we could keep them in
1422 	 * the template, but need a way to checksum without them.
1423 	 */
1424 	/*
1425 	 * m->m_pkthdr.len should have been set before checksum calculation,
1426 	 * because in6_cksum() need it.
1427 	 */
1428 #ifdef INET6
1429 	if (isipv6) {
1430 		/*
1431 		 * we separately set hoplimit for every segment, since the
1432 		 * user might want to change the value via setsockopt.
1433 		 * Also, desired default hop limit might be changed via
1434 		 * Neighbor Discovery.
1435 		 */
1436 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1437 
1438 		/*
1439 		 * Set the packet size here for the benefit of DTrace probes.
1440 		 * ip6_output() will set it properly; it's supposed to include
1441 		 * the option header lengths as well.
1442 		 */
1443 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1444 
1445 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1446 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1447 		else
1448 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1449 
1450 		if (tp->t_state == TCPS_SYN_SENT)
1451 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1452 
1453 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1454 
1455 #ifdef TCPPCAP
1456 		/* Save packet, if requested. */
1457 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1458 #endif
1459 
1460 		/* TODO: IPv6 IP6TOS_ECT bit on */
1461 		error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1462 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1463 		    NULL, NULL, inp);
1464 
1465 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1466 			mtu = inp->inp_route6.ro_nh->nh_mtu;
1467 	}
1468 #endif /* INET6 */
1469 #if defined(INET) && defined(INET6)
1470 	else
1471 #endif
1472 #ifdef INET
1473     {
1474 	ip->ip_len = htons(m->m_pkthdr.len);
1475 #ifdef INET6
1476 	if (inp->inp_vflag & INP_IPV6PROTO)
1477 		ip->ip_ttl = in6_selecthlim(inp, NULL);
1478 #endif /* INET6 */
1479 	/*
1480 	 * If we do path MTU discovery, then we set DF on every packet.
1481 	 * This might not be the best thing to do according to RFC3390
1482 	 * Section 2. However the tcp hostcache migitates the problem
1483 	 * so it affects only the first tcp connection with a host.
1484 	 *
1485 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1486 	 */
1487 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1488 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1489 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1490 			ip->ip_off |= htons(IP_DF);
1491 		}
1492 	} else {
1493 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1494 	}
1495 
1496 	if (tp->t_state == TCPS_SYN_SENT)
1497 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1498 
1499 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1500 
1501 #ifdef TCPPCAP
1502 	/* Save packet, if requested. */
1503 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1504 #endif
1505 
1506 	error = ip_output(m, inp->inp_options, &inp->inp_route,
1507 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1508 
1509 	if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1510 		mtu = inp->inp_route.ro_nh->nh_mtu;
1511     }
1512 #endif /* INET */
1513 
1514 	if (lgb != NULL) {
1515 		lgb->tlb_errno = error;
1516 		lgb = NULL;
1517 	}
1518 out:
1519 	if (error == 0)
1520 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1521 	/*
1522 	 * In transmit state, time the transmission and arrange for
1523 	 * the retransmit.  In persist state, just set snd_max.
1524 	 */
1525 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1526 	    !tcp_timer_active(tp, TT_PERSIST)) {
1527 		tcp_seq startseq = tp->snd_nxt;
1528 
1529 		/*
1530 		 * Advance snd_nxt over sequence space of this segment.
1531 		 */
1532 		if (flags & (TH_SYN|TH_FIN)) {
1533 			if (flags & TH_SYN)
1534 				tp->snd_nxt++;
1535 			if (flags & TH_FIN) {
1536 				tp->snd_nxt++;
1537 				tp->t_flags |= TF_SENTFIN;
1538 			}
1539 		}
1540 		if (sack_rxmit)
1541 			goto timer;
1542 		tp->snd_nxt += len;
1543 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1544 			/*
1545 			 * Update "made progress" indication if we just
1546 			 * added new data to an empty socket buffer.
1547 			 */
1548 			if (tp->snd_una == tp->snd_max)
1549 				tp->t_acktime = ticks;
1550 			tp->snd_max = tp->snd_nxt;
1551 			/*
1552 			 * Time this transmission if not a retransmission and
1553 			 * not currently timing anything.
1554 			 */
1555 			tp->t_sndtime = ticks;
1556 			if (tp->t_rtttime == 0) {
1557 				tp->t_rtttime = ticks;
1558 				tp->t_rtseq = startseq;
1559 				TCPSTAT_INC(tcps_segstimed);
1560 			}
1561 #ifdef STATS
1562 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1563 				tp->t_flags |= TF_GPUTINPROG;
1564 				tp->gput_seq = startseq;
1565 				tp->gput_ack = startseq +
1566 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1567 				tp->gput_ts = tcp_ts_getticks();
1568 			}
1569 #endif /* STATS */
1570 		}
1571 
1572 		/*
1573 		 * Set retransmit timer if not currently set,
1574 		 * and not doing a pure ack or a keep-alive probe.
1575 		 * Initial value for retransmit timer is smoothed
1576 		 * round-trip time + 2 * round-trip time variance.
1577 		 * Initialize shift counter which is used for backoff
1578 		 * of retransmit time.
1579 		 */
1580 timer:
1581 		if (!tcp_timer_active(tp, TT_REXMT) &&
1582 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1583 		     (tp->snd_nxt != tp->snd_una))) {
1584 			if (tcp_timer_active(tp, TT_PERSIST)) {
1585 				tcp_timer_activate(tp, TT_PERSIST, 0);
1586 				tp->t_rxtshift = 0;
1587 			}
1588 			tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1589 		} else if (len == 0 && sbavail(&so->so_snd) &&
1590 		    !tcp_timer_active(tp, TT_REXMT) &&
1591 		    !tcp_timer_active(tp, TT_PERSIST)) {
1592 			/*
1593 			 * Avoid a situation where we do not set persist timer
1594 			 * after a zero window condition. For example:
1595 			 * 1) A -> B: packet with enough data to fill the window
1596 			 * 2) B -> A: ACK for #1 + new data (0 window
1597 			 *    advertisement)
1598 			 * 3) A -> B: ACK for #2, 0 len packet
1599 			 *
1600 			 * In this case, A will not activate the persist timer,
1601 			 * because it chose to send a packet. Unless tcp_output
1602 			 * is called for some other reason (delayed ack timer,
1603 			 * another input packet from B, socket syscall), A will
1604 			 * not send zero window probes.
1605 			 *
1606 			 * So, if you send a 0-length packet, but there is data
1607 			 * in the socket buffer, and neither the rexmt or
1608 			 * persist timer is already set, then activate the
1609 			 * persist timer.
1610 			 */
1611 			tp->t_rxtshift = 0;
1612 			tcp_setpersist(tp);
1613 		}
1614 	} else {
1615 		/*
1616 		 * Persist case, update snd_max but since we are in
1617 		 * persist mode (no window) we do not update snd_nxt.
1618 		 */
1619 		int xlen = len;
1620 		if (flags & TH_SYN)
1621 			++xlen;
1622 		if (flags & TH_FIN) {
1623 			++xlen;
1624 			tp->t_flags |= TF_SENTFIN;
1625 		}
1626 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1627 			tp->snd_max = tp->snd_nxt + xlen;
1628 	}
1629 	if ((error == 0) &&
1630 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1631 	     (tp->t_flags & TF_SACK_PERMIT) &&
1632 	     tp->rcv_numsacks > 0)) {
1633 		    /* Clean up any DSACK's sent */
1634 		    tcp_clean_dsack_blocks(tp);
1635 	}
1636 	if (error) {
1637 		/*
1638 		 * We know that the packet was lost, so back out the
1639 		 * sequence number advance, if any.
1640 		 *
1641 		 * If the error is EPERM the packet got blocked by the
1642 		 * local firewall.  Normally we should terminate the
1643 		 * connection but the blocking may have been spurious
1644 		 * due to a firewall reconfiguration cycle.  So we treat
1645 		 * it like a packet loss and let the retransmit timer and
1646 		 * timeouts do their work over time.
1647 		 * XXX: It is a POLA question whether calling tcp_drop right
1648 		 * away would be the really correct behavior instead.
1649 		 */
1650 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1651 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1652 		    ((flags & TH_SYN) == 0) &&
1653 		    (error != EPERM)) {
1654 			if (sack_rxmit) {
1655 				p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1656 				tp->sackhint.sack_bytes_rexmit -= len;
1657 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1658 				    ("sackhint bytes rtx >= 0"));
1659 				KASSERT((flags & TH_FIN) == 0,
1660 				    ("error while FIN with SACK rxmit"));
1661 			} else {
1662 				tp->snd_nxt -= len;
1663 				if (flags & TH_FIN)
1664 					tp->snd_nxt--;
1665 			}
1666 			if (IN_RECOVERY(tp->t_flags))
1667 				tp->sackhint.prr_out -= len;
1668 		}
1669 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1670 		switch (error) {
1671 		case EACCES:
1672 		case EPERM:
1673 			tp->t_softerror = error;
1674 			return (error);
1675 		case ENOBUFS:
1676 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1677 			tp->snd_cwnd = tp->t_maxseg;
1678 			return (0);
1679 		case EMSGSIZE:
1680 			/*
1681 			 * For some reason the interface we used initially
1682 			 * to send segments changed to another or lowered
1683 			 * its MTU.
1684 			 * If TSO was active we either got an interface
1685 			 * without TSO capabilits or TSO was turned off.
1686 			 * If we obtained mtu from ip_output() then update
1687 			 * it and try again.
1688 			 */
1689 			if (tso)
1690 				tp->t_flags &= ~TF_TSO;
1691 			if (mtu != 0) {
1692 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1693 				goto again;
1694 			}
1695 			return (error);
1696 		case EHOSTDOWN:
1697 		case EHOSTUNREACH:
1698 		case ENETDOWN:
1699 		case ENETUNREACH:
1700 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1701 				tp->t_softerror = error;
1702 				return (0);
1703 			}
1704 			/* FALLTHROUGH */
1705 		default:
1706 			return (error);
1707 		}
1708 	}
1709 	TCPSTAT_INC(tcps_sndtotal);
1710 
1711 	/*
1712 	 * Data sent (as far as we can tell).
1713 	 * If this advertises a larger window than any other segment,
1714 	 * then remember the size of the advertised window.
1715 	 * Any pending ACK has now been sent.
1716 	 */
1717 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1718 		tp->rcv_adv = tp->rcv_nxt + recwin;
1719 	tp->last_ack_sent = tp->rcv_nxt;
1720 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1721 	if (tcp_timer_active(tp, TT_DELACK))
1722 		tcp_timer_activate(tp, TT_DELACK, 0);
1723 #if 0
1724 	/*
1725 	 * This completely breaks TCP if newreno is turned on.  What happens
1726 	 * is that if delayed-acks are turned on on the receiver, this code
1727 	 * on the transmitter effectively destroys the TCP window, forcing
1728 	 * it to four packets (1.5Kx4 = 6K window).
1729 	 */
1730 	if (sendalot && --maxburst)
1731 		goto again;
1732 #endif
1733 	if (sendalot)
1734 		goto again;
1735 	return (0);
1736 }
1737 
1738 void
1739 tcp_setpersist(struct tcpcb *tp)
1740 {
1741 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1742 	int tt;
1743 	int maxunacktime;
1744 
1745 	tp->t_flags &= ~TF_PREVVALID;
1746 	if (tcp_timer_active(tp, TT_REXMT))
1747 		panic("tcp_setpersist: retransmit pending");
1748 	/*
1749 	 * If the state is already closed, don't bother.
1750 	 */
1751 	if (tp->t_state == TCPS_CLOSED)
1752 		return;
1753 
1754 	/*
1755 	 * Start/restart persistence timer.
1756 	 */
1757 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1758 		      tcp_persmin, tcp_persmax);
1759 	if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1760 		maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1761 		if (maxunacktime < 1)
1762 			maxunacktime = 1;
1763 		if (maxunacktime < tt)
1764 			tt = maxunacktime;
1765 	}
1766 	tcp_timer_activate(tp, TT_PERSIST, tt);
1767 	if (tp->t_rxtshift < V_tcp_retries)
1768 		tp->t_rxtshift++;
1769 }
1770 
1771 /*
1772  * Insert TCP options according to the supplied parameters to the place
1773  * optp in a consistent way.  Can handle unaligned destinations.
1774  *
1775  * The order of the option processing is crucial for optimal packing and
1776  * alignment for the scarce option space.
1777  *
1778  * The optimal order for a SYN/SYN-ACK segment is:
1779  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1780  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1781  *
1782  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1783  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1784  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1785  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1786  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1787  */
1788 int
1789 tcp_addoptions(struct tcpopt *to, u_char *optp)
1790 {
1791 	u_int32_t mask, optlen = 0;
1792 
1793 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1794 		if ((to->to_flags & mask) != mask)
1795 			continue;
1796 		if (optlen == TCP_MAXOLEN)
1797 			break;
1798 		switch (to->to_flags & mask) {
1799 		case TOF_MSS:
1800 			while (optlen % 4) {
1801 				optlen += TCPOLEN_NOP;
1802 				*optp++ = TCPOPT_NOP;
1803 			}
1804 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1805 				continue;
1806 			optlen += TCPOLEN_MAXSEG;
1807 			*optp++ = TCPOPT_MAXSEG;
1808 			*optp++ = TCPOLEN_MAXSEG;
1809 			to->to_mss = htons(to->to_mss);
1810 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1811 			optp += sizeof(to->to_mss);
1812 			break;
1813 		case TOF_SCALE:
1814 			while (!optlen || optlen % 2 != 1) {
1815 				optlen += TCPOLEN_NOP;
1816 				*optp++ = TCPOPT_NOP;
1817 			}
1818 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1819 				continue;
1820 			optlen += TCPOLEN_WINDOW;
1821 			*optp++ = TCPOPT_WINDOW;
1822 			*optp++ = TCPOLEN_WINDOW;
1823 			*optp++ = to->to_wscale;
1824 			break;
1825 		case TOF_SACKPERM:
1826 			while (optlen % 2) {
1827 				optlen += TCPOLEN_NOP;
1828 				*optp++ = TCPOPT_NOP;
1829 			}
1830 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1831 				continue;
1832 			optlen += TCPOLEN_SACK_PERMITTED;
1833 			*optp++ = TCPOPT_SACK_PERMITTED;
1834 			*optp++ = TCPOLEN_SACK_PERMITTED;
1835 			break;
1836 		case TOF_TS:
1837 			while (!optlen || optlen % 4 != 2) {
1838 				optlen += TCPOLEN_NOP;
1839 				*optp++ = TCPOPT_NOP;
1840 			}
1841 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1842 				continue;
1843 			optlen += TCPOLEN_TIMESTAMP;
1844 			*optp++ = TCPOPT_TIMESTAMP;
1845 			*optp++ = TCPOLEN_TIMESTAMP;
1846 			to->to_tsval = htonl(to->to_tsval);
1847 			to->to_tsecr = htonl(to->to_tsecr);
1848 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1849 			optp += sizeof(to->to_tsval);
1850 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1851 			optp += sizeof(to->to_tsecr);
1852 			break;
1853 		case TOF_SIGNATURE:
1854 			{
1855 			int siglen = TCPOLEN_SIGNATURE - 2;
1856 
1857 			while (!optlen || optlen % 4 != 2) {
1858 				optlen += TCPOLEN_NOP;
1859 				*optp++ = TCPOPT_NOP;
1860 			}
1861 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1862 				to->to_flags &= ~TOF_SIGNATURE;
1863 				continue;
1864 			}
1865 			optlen += TCPOLEN_SIGNATURE;
1866 			*optp++ = TCPOPT_SIGNATURE;
1867 			*optp++ = TCPOLEN_SIGNATURE;
1868 			to->to_signature = optp;
1869 			while (siglen--)
1870 				 *optp++ = 0;
1871 			break;
1872 			}
1873 		case TOF_SACK:
1874 			{
1875 			int sackblks = 0;
1876 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1877 			tcp_seq sack_seq;
1878 
1879 			while (!optlen || optlen % 4 != 2) {
1880 				optlen += TCPOLEN_NOP;
1881 				*optp++ = TCPOPT_NOP;
1882 			}
1883 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1884 				continue;
1885 			optlen += TCPOLEN_SACKHDR;
1886 			*optp++ = TCPOPT_SACK;
1887 			sackblks = min(to->to_nsacks,
1888 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1889 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1890 			while (sackblks--) {
1891 				sack_seq = htonl(sack->start);
1892 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1893 				optp += sizeof(sack_seq);
1894 				sack_seq = htonl(sack->end);
1895 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1896 				optp += sizeof(sack_seq);
1897 				optlen += TCPOLEN_SACK;
1898 				sack++;
1899 			}
1900 			TCPSTAT_INC(tcps_sack_send_blocks);
1901 			break;
1902 			}
1903 		case TOF_FASTOPEN:
1904 			{
1905 			int total_len;
1906 
1907 			/* XXX is there any point to aligning this option? */
1908 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1909 			if (TCP_MAXOLEN - optlen < total_len) {
1910 				to->to_flags &= ~TOF_FASTOPEN;
1911 				continue;
1912 			}
1913 			*optp++ = TCPOPT_FAST_OPEN;
1914 			*optp++ = total_len;
1915 			if (to->to_tfo_len > 0) {
1916 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1917 				optp += to->to_tfo_len;
1918 			}
1919 			optlen += total_len;
1920 			break;
1921 			}
1922 		default:
1923 			panic("%s: unknown TCP option type", __func__);
1924 			break;
1925 		}
1926 	}
1927 
1928 	/* Terminate and pad TCP options to a 4 byte boundary. */
1929 	if (optlen % 4) {
1930 		optlen += TCPOLEN_EOL;
1931 		*optp++ = TCPOPT_EOL;
1932 	}
1933 	/*
1934 	 * According to RFC 793 (STD0007):
1935 	 *   "The content of the header beyond the End-of-Option option
1936 	 *    must be header padding (i.e., zero)."
1937 	 *   and later: "The padding is composed of zeros."
1938 	 */
1939 	while (optlen % 4) {
1940 		optlen += TCPOLEN_PAD;
1941 		*optp++ = TCPOPT_PAD;
1942 	}
1943 
1944 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1945 	return (optlen);
1946 }
1947 
1948 /*
1949  * This is a copy of m_copym(), taking the TSO segment size/limit
1950  * constraints into account, and advancing the sndptr as it goes.
1951  */
1952 struct mbuf *
1953 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1954     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1955 {
1956 #ifdef KERN_TLS
1957 	struct ktls_session *tls, *ntls;
1958 	struct mbuf *start __diagused;
1959 #endif
1960 	struct mbuf *n, **np;
1961 	struct mbuf *top;
1962 	int32_t off = off0;
1963 	int32_t len = *plen;
1964 	int32_t fragsize;
1965 	int32_t len_cp = 0;
1966 	int32_t *pkthdrlen;
1967 	uint32_t mlen, frags;
1968 	bool copyhdr;
1969 
1970 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1971 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1972 	if (off == 0 && m->m_flags & M_PKTHDR)
1973 		copyhdr = true;
1974 	else
1975 		copyhdr = false;
1976 	while (off > 0) {
1977 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1978 		if (off < m->m_len)
1979 			break;
1980 		off -= m->m_len;
1981 		if ((sb) && (m == sb->sb_sndptr)) {
1982 			sb->sb_sndptroff += m->m_len;
1983 			sb->sb_sndptr = m->m_next;
1984 		}
1985 		m = m->m_next;
1986 	}
1987 	np = &top;
1988 	top = NULL;
1989 	pkthdrlen = NULL;
1990 #ifdef KERN_TLS
1991 	if (hw_tls && (m->m_flags & M_EXTPG))
1992 		tls = m->m_epg_tls;
1993 	else
1994 		tls = NULL;
1995 	start = m;
1996 #endif
1997 	while (len > 0) {
1998 		if (m == NULL) {
1999 			KASSERT(len == M_COPYALL,
2000 			    ("tcp_m_copym, length > size of mbuf chain"));
2001 			*plen = len_cp;
2002 			if (pkthdrlen != NULL)
2003 				*pkthdrlen = len_cp;
2004 			break;
2005 		}
2006 #ifdef KERN_TLS
2007 		if (hw_tls) {
2008 			if (m->m_flags & M_EXTPG)
2009 				ntls = m->m_epg_tls;
2010 			else
2011 				ntls = NULL;
2012 
2013 			/*
2014 			 * Avoid mixing TLS records with handshake
2015 			 * data or TLS records from different
2016 			 * sessions.
2017 			 */
2018 			if (tls != ntls) {
2019 				MPASS(m != start);
2020 				*plen = len_cp;
2021 				if (pkthdrlen != NULL)
2022 					*pkthdrlen = len_cp;
2023 				break;
2024 			}
2025 		}
2026 #endif
2027 		mlen = min(len, m->m_len - off);
2028 		if (seglimit) {
2029 			/*
2030 			 * For M_EXTPG mbufs, add 3 segments
2031 			 * + 1 in case we are crossing page boundaries
2032 			 * + 2 in case the TLS hdr/trailer are used
2033 			 * It is cheaper to just add the segments
2034 			 * than it is to take the cache miss to look
2035 			 * at the mbuf ext_pgs state in detail.
2036 			 */
2037 			if (m->m_flags & M_EXTPG) {
2038 				fragsize = min(segsize, PAGE_SIZE);
2039 				frags = 3;
2040 			} else {
2041 				fragsize = segsize;
2042 				frags = 0;
2043 			}
2044 
2045 			/* Break if we really can't fit anymore. */
2046 			if ((frags + 1) >= seglimit) {
2047 				*plen =	len_cp;
2048 				if (pkthdrlen != NULL)
2049 					*pkthdrlen = len_cp;
2050 				break;
2051 			}
2052 
2053 			/*
2054 			 * Reduce size if you can't copy the whole
2055 			 * mbuf. If we can't copy the whole mbuf, also
2056 			 * adjust len so the loop will end after this
2057 			 * mbuf.
2058 			 */
2059 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2060 				mlen = (seglimit - frags - 1) * fragsize;
2061 				len = mlen;
2062 				*plen = len_cp + len;
2063 				if (pkthdrlen != NULL)
2064 					*pkthdrlen = *plen;
2065 			}
2066 			frags += howmany(mlen, fragsize);
2067 			if (frags == 0)
2068 				frags++;
2069 			seglimit -= frags;
2070 			KASSERT(seglimit > 0,
2071 			    ("%s: seglimit went too low", __func__));
2072 		}
2073 		if (copyhdr)
2074 			n = m_gethdr(M_NOWAIT, m->m_type);
2075 		else
2076 			n = m_get(M_NOWAIT, m->m_type);
2077 		*np = n;
2078 		if (n == NULL)
2079 			goto nospace;
2080 		if (copyhdr) {
2081 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2082 				goto nospace;
2083 			if (len == M_COPYALL)
2084 				n->m_pkthdr.len -= off0;
2085 			else
2086 				n->m_pkthdr.len = len;
2087 			pkthdrlen = &n->m_pkthdr.len;
2088 			copyhdr = false;
2089 		}
2090 		n->m_len = mlen;
2091 		len_cp += n->m_len;
2092 		if (m->m_flags & (M_EXT | M_EXTPG)) {
2093 			n->m_data = m->m_data + off;
2094 			mb_dupcl(n, m);
2095 		} else
2096 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2097 			    (u_int)n->m_len);
2098 
2099 		if (sb && (sb->sb_sndptr == m) &&
2100 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2101 			sb->sb_sndptroff += m->m_len;
2102 			sb->sb_sndptr = m->m_next;
2103 		}
2104 		off = 0;
2105 		if (len != M_COPYALL) {
2106 			len -= n->m_len;
2107 		}
2108 		m = m->m_next;
2109 		np = &n->m_next;
2110 	}
2111 	return (top);
2112 nospace:
2113 	m_freem(top);
2114 	return (NULL);
2115 }
2116 
2117 void
2118 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2119 {
2120 
2121 	/*
2122 	 * Automatic sizing of send socket buffer.  Often the send buffer
2123 	 * size is not optimally adjusted to the actual network conditions
2124 	 * at hand (delay bandwidth product).  Setting the buffer size too
2125 	 * small limits throughput on links with high bandwidth and high
2126 	 * delay (eg. trans-continental/oceanic links).  Setting the
2127 	 * buffer size too big consumes too much real kernel memory,
2128 	 * especially with many connections on busy servers.
2129 	 *
2130 	 * The criteria to step up the send buffer one notch are:
2131 	 *  1. receive window of remote host is larger than send buffer
2132 	 *     (with a fudge factor of 5/4th);
2133 	 *  2. send buffer is filled to 7/8th with data (so we actually
2134 	 *     have data to make use of it);
2135 	 *  3. send buffer fill has not hit maximal automatic size;
2136 	 *  4. our send window (slow start and cogestion controlled) is
2137 	 *     larger than sent but unacknowledged data in send buffer.
2138 	 *
2139 	 * The remote host receive window scaling factor may limit the
2140 	 * growing of the send buffer before it reaches its allowed
2141 	 * maximum.
2142 	 *
2143 	 * It scales directly with slow start or congestion window
2144 	 * and does at most one step per received ACK.  This fast
2145 	 * scaling has the drawback of growing the send buffer beyond
2146 	 * what is strictly necessary to make full use of a given
2147 	 * delay*bandwidth product.  However testing has shown this not
2148 	 * to be much of an problem.  At worst we are trading wasting
2149 	 * of available bandwidth (the non-use of it) for wasting some
2150 	 * socket buffer memory.
2151 	 *
2152 	 * TODO: Shrink send buffer during idle periods together
2153 	 * with congestion window.  Requires another timer.  Has to
2154 	 * wait for upcoming tcp timer rewrite.
2155 	 *
2156 	 * XXXGL: should there be used sbused() or sbavail()?
2157 	 */
2158 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2159 		int lowat;
2160 
2161 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2162 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2163 		    sbused(&so->so_snd) >=
2164 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2165 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2166 		    sendwin >= (sbused(&so->so_snd) -
2167 		    (tp->snd_nxt - tp->snd_una))) {
2168 			if (!sbreserve_locked(so, SO_SND,
2169 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2170 			     V_tcp_autosndbuf_max), curthread))
2171 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2172 		}
2173 	}
2174 }
2175