xref: /freebsd/sys/netinet/tcp_output.c (revision b197d4b893974c9eb4d7b38704c6d5c486235d6f)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_kern_tls.h"
41 #include "opt_tcpdebug.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/arb.h>
46 #include <sys/domain.h>
47 #ifdef TCP_HHOOK
48 #include <sys/hhook.h>
49 #endif
50 #include <sys/kernel.h>
51 #ifdef KERN_TLS
52 #include <sys/ktls.h>
53 #endif
54 #include <sys/lock.h>
55 #include <sys/mbuf.h>
56 #include <sys/mutex.h>
57 #include <sys/protosw.h>
58 #include <sys/qmath.h>
59 #include <sys/sdt.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/sysctl.h>
63 #include <sys/stats.h>
64 
65 #include <net/if.h>
66 #include <net/route.h>
67 #include <net/route/nhop.h>
68 #include <net/vnet.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
77 #ifdef INET6
78 #include <netinet6/in6_pcb.h>
79 #include <netinet/ip6.h>
80 #include <netinet6/ip6_var.h>
81 #endif
82 #include <netinet/tcp.h>
83 #define	TCPOUTFLAGS
84 #include <netinet/tcp_fsm.h>
85 #include <netinet/tcp_log_buf.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_var.h>
88 #include <netinet/tcp_syncache.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcpip.h>
91 #include <netinet/cc/cc.h>
92 #include <netinet/tcp_fastopen.h>
93 #ifdef TCPPCAP
94 #include <netinet/tcp_pcap.h>
95 #endif
96 #ifdef TCPDEBUG
97 #include <netinet/tcp_debug.h>
98 #endif
99 #ifdef TCP_OFFLOAD
100 #include <netinet/tcp_offload.h>
101 #endif
102 #include <netinet/tcp_ecn.h>
103 
104 #include <netipsec/ipsec_support.h>
105 
106 #include <netinet/udp.h>
107 #include <netinet/udp_var.h>
108 #include <machine/in_cksum.h>
109 
110 #include <security/mac/mac_framework.h>
111 
112 VNET_DEFINE(int, path_mtu_discovery) = 1;
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
114 	&VNET_NAME(path_mtu_discovery), 1,
115 	"Enable Path MTU Discovery");
116 
117 VNET_DEFINE(int, tcp_do_tso) = 1;
118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
119 	&VNET_NAME(tcp_do_tso), 0,
120 	"Enable TCP Segmentation Offload");
121 
122 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
123 #define	V_tcp_sendspace	VNET(tcp_sendspace)
124 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
125 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
126 
127 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
129 	&VNET_NAME(tcp_do_autosndbuf), 0,
130 	"Enable automatic send buffer sizing");
131 
132 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
134 	&VNET_NAME(tcp_autosndbuf_inc), 0,
135 	"Incrementor step size of automatic send buffer");
136 
137 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
139 	&VNET_NAME(tcp_autosndbuf_max), 0,
140 	"Max size of automatic send buffer");
141 
142 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
143 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
145 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
146 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
147 
148 /*
149  * Make sure that either retransmit or persist timer is set for SYN, FIN and
150  * non-ACK.
151  */
152 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
153 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
154 	    tcp_timer_active((tp), TT_REXMT) ||				\
155 	    tcp_timer_active((tp), TT_PERSIST),				\
156 	    ("neither rexmt nor persist timer is set"))
157 
158 static void inline	cc_after_idle(struct tcpcb *tp);
159 
160 #ifdef TCP_HHOOK
161 /*
162  * Wrapper for the TCP established output helper hook.
163  */
164 void
165 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
166     struct tcpopt *to, uint32_t len, int tso)
167 {
168 	struct tcp_hhook_data hhook_data;
169 
170 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
171 		hhook_data.tp = tp;
172 		hhook_data.th = th;
173 		hhook_data.to = to;
174 		hhook_data.len = len;
175 		hhook_data.tso = tso;
176 
177 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
178 		    tp->osd);
179 	}
180 }
181 #endif
182 
183 /*
184  * CC wrapper hook functions
185  */
186 static void inline
187 cc_after_idle(struct tcpcb *tp)
188 {
189 	INP_WLOCK_ASSERT(tp->t_inpcb);
190 
191 	if (CC_ALGO(tp)->after_idle != NULL)
192 		CC_ALGO(tp)->after_idle(tp->ccv);
193 }
194 
195 /*
196  * Tcp output routine: figure out what should be sent and send it.
197  */
198 int
199 tcp_default_output(struct tcpcb *tp)
200 {
201 	struct socket *so = tp->t_inpcb->inp_socket;
202 	int32_t len;
203 	uint32_t recwin, sendwin;
204 	uint16_t flags;
205 	int off, error = 0;	/* Keep compiler happy */
206 	u_int if_hw_tsomaxsegcount = 0;
207 	u_int if_hw_tsomaxsegsize = 0;
208 	struct mbuf *m;
209 	struct ip *ip = NULL;
210 #ifdef TCPDEBUG
211 	struct ipovly *ipov = NULL;
212 #endif
213 	struct tcphdr *th;
214 	u_char opt[TCP_MAXOLEN];
215 	unsigned ipoptlen, optlen, hdrlen, ulen;
216 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
217 	unsigned ipsec_optlen = 0;
218 #endif
219 	int idle, sendalot, curticks;
220 	int sack_rxmit, sack_bytes_rxmt;
221 	struct sackhole *p;
222 	int tso, mtu;
223 	struct tcpopt to;
224 	struct udphdr *udp = NULL;
225 	struct tcp_log_buffer *lgb;
226 	unsigned int wanted_cookie = 0;
227 	unsigned int dont_sendalot = 0;
228 #if 0
229 	int maxburst = TCP_MAXBURST;
230 #endif
231 #ifdef INET6
232 	struct ip6_hdr *ip6 = NULL;
233 	int isipv6;
234 
235 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
236 #endif
237 #ifdef KERN_TLS
238 	const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
239 #else
240 	const bool hw_tls = false;
241 #endif
242 
243 	NET_EPOCH_ASSERT();
244 	INP_WLOCK_ASSERT(tp->t_inpcb);
245 
246 #ifdef TCP_OFFLOAD
247 	if (tp->t_flags & TF_TOE)
248 		return (tcp_offload_output(tp));
249 #endif
250 
251 	/*
252 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
253 	 * only allow the initial SYN or SYN|ACK and those sent
254 	 * by the retransmit timer.
255 	 */
256 	if (IS_FASTOPEN(tp->t_flags) &&
257 	    ((tp->t_state == TCPS_SYN_SENT) ||
258 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
259 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
260 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
261 		return (0);
262 
263 	/*
264 	 * Determine length of data that should be transmitted,
265 	 * and flags that will be used.
266 	 * If there is some data or critical controls (SYN, RST)
267 	 * to send, then transmit; otherwise, investigate further.
268 	 */
269 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
270 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
271 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
272 		cc_after_idle(tp);
273 	tp->t_flags &= ~TF_LASTIDLE;
274 	if (idle) {
275 		if (tp->t_flags & TF_MORETOCOME) {
276 			tp->t_flags |= TF_LASTIDLE;
277 			idle = 0;
278 		}
279 	}
280 again:
281 	/*
282 	 * If we've recently taken a timeout, snd_max will be greater than
283 	 * snd_nxt.  There may be SACK information that allows us to avoid
284 	 * resending already delivered data.  Adjust snd_nxt accordingly.
285 	 */
286 	if ((tp->t_flags & TF_SACK_PERMIT) &&
287 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
288 		tcp_sack_adjust(tp);
289 	sendalot = 0;
290 	tso = 0;
291 	mtu = 0;
292 	off = tp->snd_nxt - tp->snd_una;
293 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
294 
295 	flags = tcp_outflags[tp->t_state];
296 	/*
297 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
298 	 * to send out new data (when sendalot is 1), bypass this function.
299 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
300 	 * we're replacing a (future) new transmission with a retransmission
301 	 * now, and we previously incremented snd_cwnd in tcp_input().
302 	 */
303 	/*
304 	 * Still in sack recovery , reset rxmit flag to zero.
305 	 */
306 	sack_rxmit = 0;
307 	sack_bytes_rxmt = 0;
308 	len = 0;
309 	p = NULL;
310 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
311 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
312 		uint32_t cwin;
313 
314 		cwin =
315 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
316 		/* Do not retransmit SACK segments beyond snd_recover */
317 		if (SEQ_GT(p->end, tp->snd_recover)) {
318 			/*
319 			 * (At least) part of sack hole extends beyond
320 			 * snd_recover. Check to see if we can rexmit data
321 			 * for this hole.
322 			 */
323 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
324 				/*
325 				 * Can't rexmit any more data for this hole.
326 				 * That data will be rexmitted in the next
327 				 * sack recovery episode, when snd_recover
328 				 * moves past p->rxmit.
329 				 */
330 				p = NULL;
331 				goto after_sack_rexmit;
332 			} else {
333 				/* Can rexmit part of the current hole */
334 				len = ((int32_t)ulmin(cwin,
335 				    SEQ_SUB(tp->snd_recover, p->rxmit)));
336 			}
337 		} else {
338 			len = ((int32_t)ulmin(cwin,
339 			    SEQ_SUB(p->end, p->rxmit)));
340 		}
341 		off = SEQ_SUB(p->rxmit, tp->snd_una);
342 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
343 		    __func__, off));
344 		if (len > 0) {
345 			sack_rxmit = 1;
346 			sendalot = 1;
347 			TCPSTAT_INC(tcps_sack_rexmits);
348 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
349 			    min(len, tcp_maxseg(tp)));
350 		}
351 	}
352 after_sack_rexmit:
353 	/*
354 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
355 	 * state flags.
356 	 */
357 	if (tp->t_flags & TF_NEEDFIN)
358 		flags |= TH_FIN;
359 	if (tp->t_flags & TF_NEEDSYN)
360 		flags |= TH_SYN;
361 
362 	SOCKBUF_LOCK(&so->so_snd);
363 	/*
364 	 * If in persist timeout with window of 0, send 1 byte.
365 	 * Otherwise, if window is small but nonzero
366 	 * and timer expired, we will send what we can
367 	 * and go to transmit state.
368 	 */
369 	if (tp->t_flags & TF_FORCEDATA) {
370 		if (sendwin == 0) {
371 			/*
372 			 * If we still have some data to send, then
373 			 * clear the FIN bit.  Usually this would
374 			 * happen below when it realizes that we
375 			 * aren't sending all the data.  However,
376 			 * if we have exactly 1 byte of unsent data,
377 			 * then it won't clear the FIN bit below,
378 			 * and if we are in persist state, we wind
379 			 * up sending the packet without recording
380 			 * that we sent the FIN bit.
381 			 *
382 			 * We can't just blindly clear the FIN bit,
383 			 * because if we don't have any more data
384 			 * to send then the probe will be the FIN
385 			 * itself.
386 			 */
387 			if (off < sbused(&so->so_snd))
388 				flags &= ~TH_FIN;
389 			sendwin = 1;
390 		} else {
391 			tcp_timer_activate(tp, TT_PERSIST, 0);
392 			tp->t_rxtshift = 0;
393 		}
394 	}
395 
396 	/*
397 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
398 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
399 	 * a negative length.  This can also occur when TCP opens up
400 	 * its congestion window while receiving additional duplicate
401 	 * acks after fast-retransmit because TCP will reset snd_nxt
402 	 * to snd_max after the fast-retransmit.
403 	 *
404 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
405 	 * be set to snd_una, the offset will be 0, and the length may
406 	 * wind up 0.
407 	 *
408 	 * If sack_rxmit is true we are retransmitting from the scoreboard
409 	 * in which case len is already set.
410 	 */
411 	if (sack_rxmit == 0) {
412 		if (sack_bytes_rxmt == 0)
413 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
414 			    off);
415 		else {
416 			int32_t cwin;
417 
418                         /*
419 			 * We are inside of a SACK recovery episode and are
420 			 * sending new data, having retransmitted all the
421 			 * data possible in the scoreboard.
422 			 */
423 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
424 			    off);
425 			/*
426 			 * Don't remove this (len > 0) check !
427 			 * We explicitly check for len > 0 here (although it
428 			 * isn't really necessary), to work around a gcc
429 			 * optimization issue - to force gcc to compute
430 			 * len above. Without this check, the computation
431 			 * of len is bungled by the optimizer.
432 			 */
433 			if (len > 0) {
434 				cwin = tp->snd_cwnd -
435 					(tp->snd_nxt - tp->snd_recover) -
436 					sack_bytes_rxmt;
437 				if (cwin < 0)
438 					cwin = 0;
439 				len = imin(len, cwin);
440 			}
441 		}
442 	}
443 
444 	/*
445 	 * Lop off SYN bit if it has already been sent.  However, if this
446 	 * is SYN-SENT state and if segment contains data and if we don't
447 	 * know that foreign host supports TAO, suppress sending segment.
448 	 */
449 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
450 		if (tp->t_state != TCPS_SYN_RECEIVED)
451 			flags &= ~TH_SYN;
452 		/*
453 		 * When sending additional segments following a TFO SYN|ACK,
454 		 * do not include the SYN bit.
455 		 */
456 		if (IS_FASTOPEN(tp->t_flags) &&
457 		    (tp->t_state == TCPS_SYN_RECEIVED))
458 			flags &= ~TH_SYN;
459 		off--, len++;
460 	}
461 
462 	/*
463 	 * Be careful not to send data and/or FIN on SYN segments.
464 	 * This measure is needed to prevent interoperability problems
465 	 * with not fully conformant TCP implementations.
466 	 */
467 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
468 		len = 0;
469 		flags &= ~TH_FIN;
470 	}
471 
472 	/*
473 	 * On TFO sockets, ensure no data is sent in the following cases:
474 	 *
475 	 *  - When retransmitting SYN|ACK on a passively-created socket
476 	 *
477 	 *  - When retransmitting SYN on an actively created socket
478 	 *
479 	 *  - When sending a zero-length cookie (cookie request) on an
480 	 *    actively created socket
481 	 *
482 	 *  - When the socket is in the CLOSED state (RST is being sent)
483 	 */
484 	if (IS_FASTOPEN(tp->t_flags) &&
485 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
486 	     ((tp->t_state == TCPS_SYN_SENT) &&
487 	      (tp->t_tfo_client_cookie_len == 0)) ||
488 	     (flags & TH_RST)))
489 		len = 0;
490 	if (len <= 0) {
491 		/*
492 		 * If FIN has been sent but not acked,
493 		 * but we haven't been called to retransmit,
494 		 * len will be < 0.  Otherwise, window shrank
495 		 * after we sent into it.  If window shrank to 0,
496 		 * cancel pending retransmit, pull snd_nxt back
497 		 * to (closed) window, and set the persist timer
498 		 * if it isn't already going.  If the window didn't
499 		 * close completely, just wait for an ACK.
500 		 *
501 		 * We also do a general check here to ensure that
502 		 * we will set the persist timer when we have data
503 		 * to send, but a 0-byte window. This makes sure
504 		 * the persist timer is set even if the packet
505 		 * hits one of the "goto send" lines below.
506 		 */
507 		len = 0;
508 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
509 			(off < (int) sbavail(&so->so_snd))) {
510 			tcp_timer_activate(tp, TT_REXMT, 0);
511 			tp->t_rxtshift = 0;
512 			tp->snd_nxt = tp->snd_una;
513 			if (!tcp_timer_active(tp, TT_PERSIST))
514 				tcp_setpersist(tp);
515 		}
516 	}
517 
518 	/* len will be >= 0 after this point. */
519 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
520 
521 	tcp_sndbuf_autoscale(tp, so, sendwin);
522 
523 	/*
524 	 * Decide if we can use TCP Segmentation Offloading (if supported by
525 	 * hardware).
526 	 *
527 	 * TSO may only be used if we are in a pure bulk sending state.  The
528 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
529 	 * IP options prevent using TSO.  With TSO the TCP header is the same
530 	 * (except for the sequence number) for all generated packets.  This
531 	 * makes it impossible to transmit any options which vary per generated
532 	 * segment or packet.
533 	 *
534 	 * IPv4 handling has a clear separation of ip options and ip header
535 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
536 	 * the right thing below to provide length of just ip options and thus
537 	 * checking for ipoptlen is enough to decide if ip options are present.
538 	 */
539 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
540 	/*
541 	 * Pre-calculate here as we save another lookup into the darknesses
542 	 * of IPsec that way and can actually decide if TSO is ok.
543 	 */
544 #ifdef INET6
545 	if (isipv6 && IPSEC_ENABLED(ipv6))
546 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
547 #ifdef INET
548 	else
549 #endif
550 #endif /* INET6 */
551 #ifdef INET
552 	if (IPSEC_ENABLED(ipv4))
553 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
554 #endif /* INET */
555 #endif /* IPSEC */
556 #ifdef INET6
557 	if (isipv6)
558 		ipoptlen = ip6_optlen(tp->t_inpcb);
559 	else
560 #endif
561 	if (tp->t_inpcb->inp_options)
562 		ipoptlen = tp->t_inpcb->inp_options->m_len -
563 				offsetof(struct ipoption, ipopt_list);
564 	else
565 		ipoptlen = 0;
566 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
567 	ipoptlen += ipsec_optlen;
568 #endif
569 
570 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
571 	    (tp->t_port == 0) &&
572 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
573 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
574 	    ipoptlen == 0 && !(flags & TH_SYN))
575 		tso = 1;
576 
577 	if (sack_rxmit) {
578 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
579 			flags &= ~TH_FIN;
580 	} else {
581 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
582 		    sbused(&so->so_snd)))
583 			flags &= ~TH_FIN;
584 	}
585 
586 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
587 	    (long)TCP_MAXWIN << tp->rcv_scale);
588 
589 	/*
590 	 * Sender silly window avoidance.   We transmit under the following
591 	 * conditions when len is non-zero:
592 	 *
593 	 *	- We have a full segment (or more with TSO)
594 	 *	- This is the last buffer in a write()/send() and we are
595 	 *	  either idle or running NODELAY
596 	 *	- we've timed out (e.g. persist timer)
597 	 *	- we have more then 1/2 the maximum send window's worth of
598 	 *	  data (receiver may be limited the window size)
599 	 *	- we need to retransmit
600 	 */
601 	if (len) {
602 		if (len >= tp->t_maxseg)
603 			goto send;
604 		/*
605 		 * As the TCP header options are now
606 		 * considered when setting up the initial
607 		 * window, we would not send the last segment
608 		 * if we skip considering the option length here.
609 		 * Note: this may not work when tcp headers change
610 		 * very dynamically in the future.
611 		 */
612 		if ((((tp->t_flags & TF_SIGNATURE) ?
613 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
614 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
615 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
616 		    len) >= tp->t_maxseg)
617 			goto send;
618 		/*
619 		 * NOTE! on localhost connections an 'ack' from the remote
620 		 * end may occur synchronously with the output and cause
621 		 * us to flush a buffer queued with moretocome.  XXX
622 		 *
623 		 * note: the len + off check is almost certainly unnecessary.
624 		 */
625 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
626 		    (idle || (tp->t_flags & TF_NODELAY)) &&
627 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
628 		    (tp->t_flags & TF_NOPUSH) == 0) {
629 			goto send;
630 		}
631 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
632 			goto send;
633 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
634 			goto send;
635 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
636 			goto send;
637 		if (sack_rxmit)
638 			goto send;
639 	}
640 
641 	/*
642 	 * Sending of standalone window updates.
643 	 *
644 	 * Window updates are important when we close our window due to a
645 	 * full socket buffer and are opening it again after the application
646 	 * reads data from it.  Once the window has opened again and the
647 	 * remote end starts to send again the ACK clock takes over and
648 	 * provides the most current window information.
649 	 *
650 	 * We must avoid the silly window syndrome whereas every read
651 	 * from the receive buffer, no matter how small, causes a window
652 	 * update to be sent.  We also should avoid sending a flurry of
653 	 * window updates when the socket buffer had queued a lot of data
654 	 * and the application is doing small reads.
655 	 *
656 	 * Prevent a flurry of pointless window updates by only sending
657 	 * an update when we can increase the advertized window by more
658 	 * than 1/4th of the socket buffer capacity.  When the buffer is
659 	 * getting full or is very small be more aggressive and send an
660 	 * update whenever we can increase by two mss sized segments.
661 	 * In all other situations the ACK's to new incoming data will
662 	 * carry further window increases.
663 	 *
664 	 * Don't send an independent window update if a delayed
665 	 * ACK is pending (it will get piggy-backed on it) or the
666 	 * remote side already has done a half-close and won't send
667 	 * more data.  Skip this if the connection is in T/TCP
668 	 * half-open state.
669 	 */
670 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
671 	    !(tp->t_flags & TF_DELACK) &&
672 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
673 		/*
674 		 * "adv" is the amount we could increase the window,
675 		 * taking into account that we are limited by
676 		 * TCP_MAXWIN << tp->rcv_scale.
677 		 */
678 		int32_t adv;
679 		int oldwin;
680 
681 		adv = recwin;
682 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
683 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
684 			if (adv > oldwin)
685 				adv -= oldwin;
686 			else
687 				adv = 0;
688 		} else
689 			oldwin = 0;
690 
691 		/*
692 		 * If the new window size ends up being the same as or less
693 		 * than the old size when it is scaled, then don't force
694 		 * a window update.
695 		 */
696 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
697 			goto dontupdate;
698 
699 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
700 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
701 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
702 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
703 		     adv >= TCP_MAXWIN << tp->rcv_scale))
704 			goto send;
705 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
706 			goto send;
707 	}
708 dontupdate:
709 
710 	/*
711 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
712 	 * is also a catch-all for the retransmit timer timeout case.
713 	 */
714 	if (tp->t_flags & TF_ACKNOW)
715 		goto send;
716 	if ((flags & TH_RST) ||
717 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
718 		goto send;
719 	if (SEQ_GT(tp->snd_up, tp->snd_una))
720 		goto send;
721 	/*
722 	 * If our state indicates that FIN should be sent
723 	 * and we have not yet done so, then we need to send.
724 	 */
725 	if (flags & TH_FIN &&
726 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
727 		goto send;
728 	/*
729 	 * In SACK, it is possible for tcp_output to fail to send a segment
730 	 * after the retransmission timer has been turned off.  Make sure
731 	 * that the retransmission timer is set.
732 	 */
733 	if ((tp->t_flags & TF_SACK_PERMIT) &&
734 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
735 	    !tcp_timer_active(tp, TT_REXMT) &&
736 	    !tcp_timer_active(tp, TT_PERSIST)) {
737 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
738 		goto just_return;
739 	}
740 	/*
741 	 * TCP window updates are not reliable, rather a polling protocol
742 	 * using ``persist'' packets is used to insure receipt of window
743 	 * updates.  The three ``states'' for the output side are:
744 	 *	idle			not doing retransmits or persists
745 	 *	persisting		to move a small or zero window
746 	 *	(re)transmitting	and thereby not persisting
747 	 *
748 	 * tcp_timer_active(tp, TT_PERSIST)
749 	 *	is true when we are in persist state.
750 	 * (tp->t_flags & TF_FORCEDATA)
751 	 *	is set when we are called to send a persist packet.
752 	 * tcp_timer_active(tp, TT_REXMT)
753 	 *	is set when we are retransmitting
754 	 * The output side is idle when both timers are zero.
755 	 *
756 	 * If send window is too small, there is data to transmit, and no
757 	 * retransmit or persist is pending, then go to persist state.
758 	 * If nothing happens soon, send when timer expires:
759 	 * if window is nonzero, transmit what we can,
760 	 * otherwise force out a byte.
761 	 */
762 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
763 	    !tcp_timer_active(tp, TT_PERSIST)) {
764 		tp->t_rxtshift = 0;
765 		tcp_setpersist(tp);
766 	}
767 
768 	/*
769 	 * No reason to send a segment, just return.
770 	 */
771 just_return:
772 	SOCKBUF_UNLOCK(&so->so_snd);
773 	return (0);
774 
775 send:
776 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
777 	if (len > 0) {
778 		if (len >= tp->t_maxseg)
779 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
780 		else
781 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
782 	}
783 	/*
784 	 * Before ESTABLISHED, force sending of initial options
785 	 * unless TCP set not to do any options.
786 	 * NOTE: we assume that the IP/TCP header plus TCP options
787 	 * always fit in a single mbuf, leaving room for a maximum
788 	 * link header, i.e.
789 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
790 	 */
791 	optlen = 0;
792 #ifdef INET6
793 	if (isipv6)
794 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
795 	else
796 #endif
797 		hdrlen = sizeof (struct tcpiphdr);
798 
799 	if (flags & TH_SYN) {
800 		tp->snd_nxt = tp->iss;
801 	}
802 
803 	/*
804 	 * Compute options for segment.
805 	 * We only have to care about SYN and established connection
806 	 * segments.  Options for SYN-ACK segments are handled in TCP
807 	 * syncache.
808 	 */
809 	to.to_flags = 0;
810 	if ((tp->t_flags & TF_NOOPT) == 0) {
811 		/* Maximum segment size. */
812 		if (flags & TH_SYN) {
813 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
814 			if (tp->t_port)
815 				to.to_mss -= V_tcp_udp_tunneling_overhead;
816 			to.to_flags |= TOF_MSS;
817 
818 			/*
819 			 * On SYN or SYN|ACK transmits on TFO connections,
820 			 * only include the TFO option if it is not a
821 			 * retransmit, as the presence of the TFO option may
822 			 * have caused the original SYN or SYN|ACK to have
823 			 * been dropped by a middlebox.
824 			 */
825 			if (IS_FASTOPEN(tp->t_flags) &&
826 			    (tp->t_rxtshift == 0)) {
827 				if (tp->t_state == TCPS_SYN_RECEIVED) {
828 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
829 					to.to_tfo_cookie =
830 					    (u_int8_t *)&tp->t_tfo_cookie.server;
831 					to.to_flags |= TOF_FASTOPEN;
832 					wanted_cookie = 1;
833 				} else if (tp->t_state == TCPS_SYN_SENT) {
834 					to.to_tfo_len =
835 					    tp->t_tfo_client_cookie_len;
836 					to.to_tfo_cookie =
837 					    tp->t_tfo_cookie.client;
838 					to.to_flags |= TOF_FASTOPEN;
839 					wanted_cookie = 1;
840 					/*
841 					 * If we wind up having more data to
842 					 * send with the SYN than can fit in
843 					 * one segment, don't send any more
844 					 * until the SYN|ACK comes back from
845 					 * the other end.
846 					 */
847 					dont_sendalot = 1;
848 				}
849 			}
850 		}
851 		/* Window scaling. */
852 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
853 			to.to_wscale = tp->request_r_scale;
854 			to.to_flags |= TOF_SCALE;
855 		}
856 		/* Timestamps. */
857 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
858 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
859 			curticks = tcp_ts_getticks();
860 			to.to_tsval = curticks + tp->ts_offset;
861 			to.to_tsecr = tp->ts_recent;
862 			to.to_flags |= TOF_TS;
863 			if (tp->t_rxtshift == 1)
864 				tp->t_badrxtwin = curticks;
865 		}
866 
867 		/* Set receive buffer autosizing timestamp. */
868 		if (tp->rfbuf_ts == 0 &&
869 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
870 			tp->rfbuf_ts = tcp_ts_getticks();
871 
872 		/* Selective ACK's. */
873 		if (tp->t_flags & TF_SACK_PERMIT) {
874 			if (flags & TH_SYN)
875 				to.to_flags |= TOF_SACKPERM;
876 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
877 			    tp->rcv_numsacks > 0) {
878 				to.to_flags |= TOF_SACK;
879 				to.to_nsacks = tp->rcv_numsacks;
880 				to.to_sacks = (u_char *)tp->sackblks;
881 			}
882 		}
883 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
884 		/* TCP-MD5 (RFC2385). */
885 		/*
886 		 * Check that TCP_MD5SIG is enabled in tcpcb to
887 		 * account the size needed to set this TCP option.
888 		 */
889 		if (tp->t_flags & TF_SIGNATURE)
890 			to.to_flags |= TOF_SIGNATURE;
891 #endif /* TCP_SIGNATURE */
892 
893 		/* Processing the options. */
894 		hdrlen += optlen = tcp_addoptions(&to, opt);
895 		/*
896 		 * If we wanted a TFO option to be added, but it was unable
897 		 * to fit, ensure no data is sent.
898 		 */
899 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
900 		    !(to.to_flags & TOF_FASTOPEN))
901 			len = 0;
902 	}
903 	if (tp->t_port) {
904 		if (V_tcp_udp_tunneling_port == 0) {
905 			/* The port was removed?? */
906 			SOCKBUF_UNLOCK(&so->so_snd);
907 			return (EHOSTUNREACH);
908 		}
909 		hdrlen += sizeof(struct udphdr);
910 	}
911 	/*
912 	 * Adjust data length if insertion of options will
913 	 * bump the packet length beyond the t_maxseg length.
914 	 * Clear the FIN bit because we cut off the tail of
915 	 * the segment.
916 	 */
917 	if (len + optlen + ipoptlen > tp->t_maxseg) {
918 		flags &= ~TH_FIN;
919 
920 		if (tso) {
921 			u_int if_hw_tsomax;
922 			u_int moff;
923 			int max_len;
924 
925 			/* extract TSO information */
926 			if_hw_tsomax = tp->t_tsomax;
927 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
928 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
929 
930 			/*
931 			 * Limit a TSO burst to prevent it from
932 			 * overflowing or exceeding the maximum length
933 			 * allowed by the network interface:
934 			 */
935 			KASSERT(ipoptlen == 0,
936 			    ("%s: TSO can't do IP options", __func__));
937 
938 			/*
939 			 * Check if we should limit by maximum payload
940 			 * length:
941 			 */
942 			if (if_hw_tsomax != 0) {
943 				/* compute maximum TSO length */
944 				max_len = (if_hw_tsomax - hdrlen -
945 				    max_linkhdr);
946 				if (max_len <= 0) {
947 					len = 0;
948 				} else if (len > max_len) {
949 					sendalot = 1;
950 					len = max_len;
951 				}
952 			}
953 
954 			/*
955 			 * Prevent the last segment from being
956 			 * fractional unless the send sockbuf can be
957 			 * emptied:
958 			 */
959 			max_len = (tp->t_maxseg - optlen);
960 			if (((uint32_t)off + (uint32_t)len) <
961 			    sbavail(&so->so_snd)) {
962 				moff = len % max_len;
963 				if (moff != 0) {
964 					len -= moff;
965 					sendalot = 1;
966 				}
967 			}
968 
969 			/*
970 			 * In case there are too many small fragments
971 			 * don't use TSO:
972 			 */
973 			if (len <= max_len) {
974 				len = max_len;
975 				sendalot = 1;
976 				tso = 0;
977 			}
978 
979 			/*
980 			 * Send the FIN in a separate segment
981 			 * after the bulk sending is done.
982 			 * We don't trust the TSO implementations
983 			 * to clear the FIN flag on all but the
984 			 * last segment.
985 			 */
986 			if (tp->t_flags & TF_NEEDFIN)
987 				sendalot = 1;
988 		} else {
989 			if (optlen + ipoptlen >= tp->t_maxseg) {
990 				/*
991 				 * Since we don't have enough space to put
992 				 * the IP header chain and the TCP header in
993 				 * one packet as required by RFC 7112, don't
994 				 * send it. Also ensure that at least one
995 				 * byte of the payload can be put into the
996 				 * TCP segment.
997 				 */
998 				SOCKBUF_UNLOCK(&so->so_snd);
999 				error = EMSGSIZE;
1000 				sack_rxmit = 0;
1001 				goto out;
1002 			}
1003 			len = tp->t_maxseg - optlen - ipoptlen;
1004 			sendalot = 1;
1005 			if (dont_sendalot)
1006 				sendalot = 0;
1007 		}
1008 	} else
1009 		tso = 0;
1010 
1011 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1012 	    ("%s: len > IP_MAXPACKET", __func__));
1013 
1014 /*#ifdef DIAGNOSTIC*/
1015 #ifdef INET6
1016 	if (max_linkhdr + hdrlen > MCLBYTES)
1017 #else
1018 	if (max_linkhdr + hdrlen > MHLEN)
1019 #endif
1020 		panic("tcphdr too big");
1021 /*#endif*/
1022 
1023 	/*
1024 	 * This KASSERT is here to catch edge cases at a well defined place.
1025 	 * Before, those had triggered (random) panic conditions further down.
1026 	 */
1027 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1028 
1029 	/*
1030 	 * Grab a header mbuf, attaching a copy of data to
1031 	 * be transmitted, and initialize the header from
1032 	 * the template for sends on this connection.
1033 	 */
1034 	if (len) {
1035 		struct mbuf *mb;
1036 		struct sockbuf *msb;
1037 		u_int moff;
1038 
1039 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1040 			TCPSTAT_INC(tcps_sndprobe);
1041 #ifdef STATS
1042 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1043 				stats_voi_update_abs_u32(tp->t_stats,
1044 				VOI_TCP_RETXPB, len);
1045 			else
1046 				stats_voi_update_abs_u64(tp->t_stats,
1047 				    VOI_TCP_TXPB, len);
1048 #endif /* STATS */
1049 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1050 			tp->t_sndrexmitpack++;
1051 			TCPSTAT_INC(tcps_sndrexmitpack);
1052 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1053 #ifdef STATS
1054 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1055 			    len);
1056 #endif /* STATS */
1057 		} else {
1058 			TCPSTAT_INC(tcps_sndpack);
1059 			TCPSTAT_ADD(tcps_sndbyte, len);
1060 #ifdef STATS
1061 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1062 			    len);
1063 #endif /* STATS */
1064 		}
1065 #ifdef INET6
1066 		if (MHLEN < hdrlen + max_linkhdr)
1067 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1068 		else
1069 #endif
1070 			m = m_gethdr(M_NOWAIT, MT_DATA);
1071 
1072 		if (m == NULL) {
1073 			SOCKBUF_UNLOCK(&so->so_snd);
1074 			error = ENOBUFS;
1075 			sack_rxmit = 0;
1076 			goto out;
1077 		}
1078 
1079 		m->m_data += max_linkhdr;
1080 		m->m_len = hdrlen;
1081 
1082 		/*
1083 		 * Start the m_copy functions from the closest mbuf
1084 		 * to the offset in the socket buffer chain.
1085 		 */
1086 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1087 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1088 			m_copydata(mb, moff, len,
1089 			    mtod(m, caddr_t) + hdrlen);
1090 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1091 				sbsndptr_adv(&so->so_snd, mb, len);
1092 			m->m_len += len;
1093 		} else {
1094 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1095 				msb = NULL;
1096 			else
1097 				msb = &so->so_snd;
1098 			m->m_next = tcp_m_copym(mb, moff,
1099 			    &len, if_hw_tsomaxsegcount,
1100 			    if_hw_tsomaxsegsize, msb, hw_tls);
1101 			if (len <= (tp->t_maxseg - optlen)) {
1102 				/*
1103 				 * Must have ran out of mbufs for the copy
1104 				 * shorten it to no longer need tso. Lets
1105 				 * not put on sendalot since we are low on
1106 				 * mbufs.
1107 				 */
1108 				tso = 0;
1109 			}
1110 			if (m->m_next == NULL) {
1111 				SOCKBUF_UNLOCK(&so->so_snd);
1112 				(void) m_free(m);
1113 				error = ENOBUFS;
1114 				sack_rxmit = 0;
1115 				goto out;
1116 			}
1117 		}
1118 
1119 		/*
1120 		 * If we're sending everything we've got, set PUSH.
1121 		 * (This will keep happy those implementations which only
1122 		 * give data to the user when a buffer fills or
1123 		 * a PUSH comes in.)
1124 		 */
1125 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1126 		    !(flags & TH_SYN))
1127 			flags |= TH_PUSH;
1128 		SOCKBUF_UNLOCK(&so->so_snd);
1129 	} else {
1130 		SOCKBUF_UNLOCK(&so->so_snd);
1131 		if (tp->t_flags & TF_ACKNOW)
1132 			TCPSTAT_INC(tcps_sndacks);
1133 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1134 			TCPSTAT_INC(tcps_sndctrl);
1135 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1136 			TCPSTAT_INC(tcps_sndurg);
1137 		else
1138 			TCPSTAT_INC(tcps_sndwinup);
1139 
1140 		m = m_gethdr(M_NOWAIT, MT_DATA);
1141 		if (m == NULL) {
1142 			error = ENOBUFS;
1143 			sack_rxmit = 0;
1144 			goto out;
1145 		}
1146 #ifdef INET6
1147 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1148 		    MHLEN >= hdrlen) {
1149 			M_ALIGN(m, hdrlen);
1150 		} else
1151 #endif
1152 		m->m_data += max_linkhdr;
1153 		m->m_len = hdrlen;
1154 	}
1155 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1156 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1157 #ifdef MAC
1158 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1159 #endif
1160 #ifdef INET6
1161 	if (isipv6) {
1162 		ip6 = mtod(m, struct ip6_hdr *);
1163 		if (tp->t_port) {
1164 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1165 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1166 			udp->uh_dport = tp->t_port;
1167 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1168 			udp->uh_ulen = htons(ulen);
1169 			th = (struct tcphdr *)(udp + 1);
1170 		} else {
1171 			th = (struct tcphdr *)(ip6 + 1);
1172 		}
1173 		tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th);
1174 	} else
1175 #endif /* INET6 */
1176 	{
1177 		ip = mtod(m, struct ip *);
1178 #ifdef TCPDEBUG
1179 		ipov = (struct ipovly *)ip;
1180 #endif
1181 		if (tp->t_port) {
1182 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1183 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1184 			udp->uh_dport = tp->t_port;
1185 			ulen = hdrlen + len - sizeof(struct ip);
1186 			udp->uh_ulen = htons(ulen);
1187 			th = (struct tcphdr *)(udp + 1);
1188 		} else
1189 			th = (struct tcphdr *)(ip + 1);
1190 		tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th);
1191 	}
1192 
1193 	/*
1194 	 * Fill in fields, remembering maximum advertised
1195 	 * window for use in delaying messages about window sizes.
1196 	 * If resending a FIN, be sure not to use a new sequence number.
1197 	 */
1198 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1199 	    tp->snd_nxt == tp->snd_max)
1200 		tp->snd_nxt--;
1201 	/*
1202 	 * If we are starting a connection, send ECN setup
1203 	 * SYN packet. If we are on a retransmit, we may
1204 	 * resend those bits a number of times as per
1205 	 * RFC 3168.
1206 	 */
1207 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1208 		flags |= tcp_ecn_output_syn_sent(tp);
1209 	}
1210 	/* Also handle parallel SYN for ECN */
1211 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1212 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1213 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1214 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1215 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1216 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1217 #ifdef INET6
1218 		if (isipv6) {
1219 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
1220 			ip6->ip6_flow |= htonl(ect << 20);
1221 		}
1222 		else
1223 #endif
1224 		{
1225 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1226 			ip->ip_tos |= ect;
1227 		}
1228 	}
1229 
1230 	/*
1231 	 * If we are doing retransmissions, then snd_nxt will
1232 	 * not reflect the first unsent octet.  For ACK only
1233 	 * packets, we do not want the sequence number of the
1234 	 * retransmitted packet, we want the sequence number
1235 	 * of the next unsent octet.  So, if there is no data
1236 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1237 	 * when filling in ti_seq.  But if we are in persist
1238 	 * state, snd_max might reflect one byte beyond the
1239 	 * right edge of the window, so use snd_nxt in that
1240 	 * case, since we know we aren't doing a retransmission.
1241 	 * (retransmit and persist are mutually exclusive...)
1242 	 */
1243 	if (sack_rxmit == 0) {
1244 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1245 		    tcp_timer_active(tp, TT_PERSIST))
1246 			th->th_seq = htonl(tp->snd_nxt);
1247 		else
1248 			th->th_seq = htonl(tp->snd_max);
1249 	} else {
1250 		th->th_seq = htonl(p->rxmit);
1251 		p->rxmit += len;
1252 		/*
1253 		 * Lost Retransmission Detection
1254 		 * trigger resending of a (then
1255 		 * still existing) hole, when
1256 		 * fack acks recoverypoint.
1257 		 */
1258 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1259 			p->rxmit = tp->snd_recover;
1260 		tp->sackhint.sack_bytes_rexmit += len;
1261 	}
1262 	if (IN_RECOVERY(tp->t_flags)) {
1263 		/*
1264 		 * Account all bytes transmitted while
1265 		 * IN_RECOVERY, simplifying PRR and
1266 		 * Lost Retransmit Detection
1267 		 */
1268 		tp->sackhint.prr_out += len;
1269 	}
1270 	th->th_ack = htonl(tp->rcv_nxt);
1271 	if (optlen) {
1272 		bcopy(opt, th + 1, optlen);
1273 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1274 	}
1275 	tcp_set_flags(th, flags);
1276 	/*
1277 	 * Calculate receive window.  Don't shrink window,
1278 	 * but avoid silly window syndrome.
1279 	 * If a RST segment is sent, advertise a window of zero.
1280 	 */
1281 	if (flags & TH_RST) {
1282 		recwin = 0;
1283 	} else {
1284 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1285 		    recwin < tp->t_maxseg)
1286 			recwin = 0;
1287 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1288 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1289 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1290 	}
1291 	/*
1292 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1293 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1294 	 * case is handled in syncache.
1295 	 */
1296 	if (flags & TH_SYN)
1297 		th->th_win = htons((u_short)
1298 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1299 	else {
1300 		/* Avoid shrinking window with window scaling. */
1301 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1302 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1303 	}
1304 
1305 	/*
1306 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1307 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1308 	 * flag tells soreceive() to disable delayed acknowledgements when
1309 	 * draining the buffer.  This can occur if the receiver is attempting
1310 	 * to read more data than can be buffered prior to transmitting on
1311 	 * the connection.
1312 	 */
1313 	if (th->th_win == 0) {
1314 		tp->t_sndzerowin++;
1315 		tp->t_flags |= TF_RXWIN0SENT;
1316 	} else
1317 		tp->t_flags &= ~TF_RXWIN0SENT;
1318 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1319 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1320 		th->th_flags |= TH_URG;
1321 	} else
1322 		/*
1323 		 * If no urgent pointer to send, then we pull
1324 		 * the urgent pointer to the left edge of the send window
1325 		 * so that it doesn't drift into the send window on sequence
1326 		 * number wraparound.
1327 		 */
1328 		tp->snd_up = tp->snd_una;		/* drag it along */
1329 
1330 	/*
1331 	 * Put TCP length in extended header, and then
1332 	 * checksum extended header and data.
1333 	 */
1334 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1335 
1336 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1337 	if (to.to_flags & TOF_SIGNATURE) {
1338 		/*
1339 		 * Calculate MD5 signature and put it into the place
1340 		 * determined before.
1341 		 * NOTE: since TCP options buffer doesn't point into
1342 		 * mbuf's data, calculate offset and use it.
1343 		 */
1344 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1345 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1346 			/*
1347 			 * Do not send segment if the calculation of MD5
1348 			 * digest has failed.
1349 			 */
1350 			m_freem(m);
1351 			goto out;
1352 		}
1353 	}
1354 #endif
1355 #ifdef INET6
1356 	if (isipv6) {
1357 		/*
1358 		 * There is no need to fill in ip6_plen right now.
1359 		 * It will be filled later by ip6_output.
1360 		 */
1361 		if (tp->t_port) {
1362 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1363 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1364 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1365 			th->th_sum = htons(0);
1366 			UDPSTAT_INC(udps_opackets);
1367 		} else {
1368 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1369 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1370 			th->th_sum = in6_cksum_pseudo(ip6,
1371 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1372 			    0);
1373 		}
1374 	}
1375 #endif
1376 #if defined(INET6) && defined(INET)
1377 	else
1378 #endif
1379 #ifdef INET
1380 	{
1381 		if (tp->t_port) {
1382 			m->m_pkthdr.csum_flags = CSUM_UDP;
1383 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1384 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1385 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1386 			th->th_sum = htons(0);
1387 			UDPSTAT_INC(udps_opackets);
1388 		} else {
1389 			m->m_pkthdr.csum_flags = CSUM_TCP;
1390 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1391 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1392 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1393 			    IPPROTO_TCP + len + optlen));
1394 		}
1395 
1396 		/* IP version must be set here for ipv4/ipv6 checking later */
1397 		KASSERT(ip->ip_v == IPVERSION,
1398 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1399 	}
1400 #endif
1401 
1402 	/*
1403 	 * Enable TSO and specify the size of the segments.
1404 	 * The TCP pseudo header checksum is always provided.
1405 	 */
1406 	if (tso) {
1407 		KASSERT(len > tp->t_maxseg - optlen,
1408 		    ("%s: len <= tso_segsz", __func__));
1409 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1410 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1411 	}
1412 
1413 	KASSERT(len + hdrlen == m_length(m, NULL),
1414 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1415 	    __func__, len, hdrlen, m_length(m, NULL)));
1416 
1417 #ifdef TCP_HHOOK
1418 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1419 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1420 #endif
1421 
1422 #ifdef TCPDEBUG
1423 	/*
1424 	 * Trace.
1425 	 */
1426 	if (so->so_options & SO_DEBUG) {
1427 		u_short save = 0;
1428 #ifdef INET6
1429 		if (!isipv6)
1430 #endif
1431 		{
1432 			save = ipov->ih_len;
1433 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1434 		}
1435 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1436 #ifdef INET6
1437 		if (!isipv6)
1438 #endif
1439 		ipov->ih_len = save;
1440 	}
1441 #endif /* TCPDEBUG */
1442 	TCP_PROBE3(debug__output, tp, th, m);
1443 
1444 	/* We're getting ready to send; log now. */
1445 	/* XXXMT: We are not honoring verbose logging. */
1446 	if (tp->t_logstate != TCP_LOG_STATE_OFF)
1447 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd,
1448 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1449 		    NULL);
1450 	else
1451 		lgb = NULL;
1452 
1453 	/*
1454 	 * Fill in IP length and desired time to live and
1455 	 * send to IP level.  There should be a better way
1456 	 * to handle ttl and tos; we could keep them in
1457 	 * the template, but need a way to checksum without them.
1458 	 */
1459 	/*
1460 	 * m->m_pkthdr.len should have been set before checksum calculation,
1461 	 * because in6_cksum() need it.
1462 	 */
1463 #ifdef INET6
1464 	if (isipv6) {
1465 		/*
1466 		 * we separately set hoplimit for every segment, since the
1467 		 * user might want to change the value via setsockopt.
1468 		 * Also, desired default hop limit might be changed via
1469 		 * Neighbor Discovery.
1470 		 */
1471 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1472 
1473 		/*
1474 		 * Set the packet size here for the benefit of DTrace probes.
1475 		 * ip6_output() will set it properly; it's supposed to include
1476 		 * the option header lengths as well.
1477 		 */
1478 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1479 
1480 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1481 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1482 		else
1483 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1484 
1485 		if (tp->t_state == TCPS_SYN_SENT)
1486 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1487 
1488 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1489 
1490 #ifdef TCPPCAP
1491 		/* Save packet, if requested. */
1492 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1493 #endif
1494 
1495 		/* TODO: IPv6 IP6TOS_ECT bit on */
1496 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1497 		    &tp->t_inpcb->inp_route6,
1498 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1499 		    NULL, NULL, tp->t_inpcb);
1500 
1501 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1502 			mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1503 	}
1504 #endif /* INET6 */
1505 #if defined(INET) && defined(INET6)
1506 	else
1507 #endif
1508 #ifdef INET
1509     {
1510 	ip->ip_len = htons(m->m_pkthdr.len);
1511 #ifdef INET6
1512 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1513 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1514 #endif /* INET6 */
1515 	/*
1516 	 * If we do path MTU discovery, then we set DF on every packet.
1517 	 * This might not be the best thing to do according to RFC3390
1518 	 * Section 2. However the tcp hostcache migitates the problem
1519 	 * so it affects only the first tcp connection with a host.
1520 	 *
1521 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1522 	 */
1523 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1524 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1525 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1526 			ip->ip_off |= htons(IP_DF);
1527 		}
1528 	} else {
1529 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1530 	}
1531 
1532 	if (tp->t_state == TCPS_SYN_SENT)
1533 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1534 
1535 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1536 
1537 #ifdef TCPPCAP
1538 	/* Save packet, if requested. */
1539 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1540 #endif
1541 
1542 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1543 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1544 	    tp->t_inpcb);
1545 
1546 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1547 		mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1548     }
1549 #endif /* INET */
1550 
1551 	if (lgb != NULL) {
1552 		lgb->tlb_errno = error;
1553 		lgb = NULL;
1554 	}
1555 out:
1556 	if (error == 0)
1557 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1558 	/*
1559 	 * In transmit state, time the transmission and arrange for
1560 	 * the retransmit.  In persist state, just set snd_max.
1561 	 */
1562 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1563 	    !tcp_timer_active(tp, TT_PERSIST)) {
1564 		tcp_seq startseq = tp->snd_nxt;
1565 
1566 		/*
1567 		 * Advance snd_nxt over sequence space of this segment.
1568 		 */
1569 		if (flags & (TH_SYN|TH_FIN)) {
1570 			if (flags & TH_SYN)
1571 				tp->snd_nxt++;
1572 			if (flags & TH_FIN) {
1573 				tp->snd_nxt++;
1574 				tp->t_flags |= TF_SENTFIN;
1575 			}
1576 		}
1577 		if (sack_rxmit)
1578 			goto timer;
1579 		tp->snd_nxt += len;
1580 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1581 			tp->snd_max = tp->snd_nxt;
1582 			/*
1583 			 * Time this transmission if not a retransmission and
1584 			 * not currently timing anything.
1585 			 */
1586 			tp->t_sndtime = ticks;
1587 			if (tp->t_rtttime == 0) {
1588 				tp->t_rtttime = ticks;
1589 				tp->t_rtseq = startseq;
1590 				TCPSTAT_INC(tcps_segstimed);
1591 			}
1592 #ifdef STATS
1593 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1594 				tp->t_flags |= TF_GPUTINPROG;
1595 				tp->gput_seq = startseq;
1596 				tp->gput_ack = startseq +
1597 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1598 				tp->gput_ts = tcp_ts_getticks();
1599 			}
1600 #endif /* STATS */
1601 		}
1602 
1603 		/*
1604 		 * Set retransmit timer if not currently set,
1605 		 * and not doing a pure ack or a keep-alive probe.
1606 		 * Initial value for retransmit timer is smoothed
1607 		 * round-trip time + 2 * round-trip time variance.
1608 		 * Initialize shift counter which is used for backoff
1609 		 * of retransmit time.
1610 		 */
1611 timer:
1612 		if (!tcp_timer_active(tp, TT_REXMT) &&
1613 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1614 		     (tp->snd_nxt != tp->snd_una))) {
1615 			if (tcp_timer_active(tp, TT_PERSIST)) {
1616 				tcp_timer_activate(tp, TT_PERSIST, 0);
1617 				tp->t_rxtshift = 0;
1618 			}
1619 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1620 		} else if (len == 0 && sbavail(&so->so_snd) &&
1621 		    !tcp_timer_active(tp, TT_REXMT) &&
1622 		    !tcp_timer_active(tp, TT_PERSIST)) {
1623 			/*
1624 			 * Avoid a situation where we do not set persist timer
1625 			 * after a zero window condition. For example:
1626 			 * 1) A -> B: packet with enough data to fill the window
1627 			 * 2) B -> A: ACK for #1 + new data (0 window
1628 			 *    advertisement)
1629 			 * 3) A -> B: ACK for #2, 0 len packet
1630 			 *
1631 			 * In this case, A will not activate the persist timer,
1632 			 * because it chose to send a packet. Unless tcp_output
1633 			 * is called for some other reason (delayed ack timer,
1634 			 * another input packet from B, socket syscall), A will
1635 			 * not send zero window probes.
1636 			 *
1637 			 * So, if you send a 0-length packet, but there is data
1638 			 * in the socket buffer, and neither the rexmt or
1639 			 * persist timer is already set, then activate the
1640 			 * persist timer.
1641 			 */
1642 			tp->t_rxtshift = 0;
1643 			tcp_setpersist(tp);
1644 		}
1645 	} else {
1646 		/*
1647 		 * Persist case, update snd_max but since we are in
1648 		 * persist mode (no window) we do not update snd_nxt.
1649 		 */
1650 		int xlen = len;
1651 		if (flags & TH_SYN)
1652 			++xlen;
1653 		if (flags & TH_FIN) {
1654 			++xlen;
1655 			tp->t_flags |= TF_SENTFIN;
1656 		}
1657 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1658 			tp->snd_max = tp->snd_nxt + xlen;
1659 	}
1660 	if ((error == 0) &&
1661 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1662 	     (tp->t_flags & TF_SACK_PERMIT) &&
1663 	     tp->rcv_numsacks > 0)) {
1664 		    /* Clean up any DSACK's sent */
1665 		    tcp_clean_dsack_blocks(tp);
1666 	}
1667 	if (error) {
1668 		/*
1669 		 * We know that the packet was lost, so back out the
1670 		 * sequence number advance, if any.
1671 		 *
1672 		 * If the error is EPERM the packet got blocked by the
1673 		 * local firewall.  Normally we should terminate the
1674 		 * connection but the blocking may have been spurious
1675 		 * due to a firewall reconfiguration cycle.  So we treat
1676 		 * it like a packet loss and let the retransmit timer and
1677 		 * timeouts do their work over time.
1678 		 * XXX: It is a POLA question whether calling tcp_drop right
1679 		 * away would be the really correct behavior instead.
1680 		 */
1681 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1682 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1683 		    ((flags & TH_SYN) == 0) &&
1684 		    (error != EPERM)) {
1685 			if (sack_rxmit) {
1686 				p->rxmit -= len;
1687 				tp->sackhint.sack_bytes_rexmit -= len;
1688 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1689 				    ("sackhint bytes rtx >= 0"));
1690 				KASSERT((flags & TH_FIN) == 0,
1691 				    ("error while FIN with SACK rxmit"));
1692 			} else {
1693 				tp->snd_nxt -= len;
1694 				if (flags & TH_FIN)
1695 					tp->snd_nxt--;
1696 			}
1697 		}
1698 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1699 		switch (error) {
1700 		case EACCES:
1701 		case EPERM:
1702 			tp->t_softerror = error;
1703 			return (error);
1704 		case ENOBUFS:
1705 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1706 			tp->snd_cwnd = tp->t_maxseg;
1707 			return (0);
1708 		case EMSGSIZE:
1709 			/*
1710 			 * For some reason the interface we used initially
1711 			 * to send segments changed to another or lowered
1712 			 * its MTU.
1713 			 * If TSO was active we either got an interface
1714 			 * without TSO capabilits or TSO was turned off.
1715 			 * If we obtained mtu from ip_output() then update
1716 			 * it and try again.
1717 			 */
1718 			if (tso)
1719 				tp->t_flags &= ~TF_TSO;
1720 			if (mtu != 0) {
1721 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1722 				goto again;
1723 			}
1724 			return (error);
1725 		case EHOSTDOWN:
1726 		case EHOSTUNREACH:
1727 		case ENETDOWN:
1728 		case ENETUNREACH:
1729 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1730 				tp->t_softerror = error;
1731 				return (0);
1732 			}
1733 			/* FALLTHROUGH */
1734 		default:
1735 			return (error);
1736 		}
1737 	}
1738 	TCPSTAT_INC(tcps_sndtotal);
1739 
1740 	/*
1741 	 * Data sent (as far as we can tell).
1742 	 * If this advertises a larger window than any other segment,
1743 	 * then remember the size of the advertised window.
1744 	 * Any pending ACK has now been sent.
1745 	 */
1746 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1747 		tp->rcv_adv = tp->rcv_nxt + recwin;
1748 	tp->last_ack_sent = tp->rcv_nxt;
1749 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1750 	if (tcp_timer_active(tp, TT_DELACK))
1751 		tcp_timer_activate(tp, TT_DELACK, 0);
1752 #if 0
1753 	/*
1754 	 * This completely breaks TCP if newreno is turned on.  What happens
1755 	 * is that if delayed-acks are turned on on the receiver, this code
1756 	 * on the transmitter effectively destroys the TCP window, forcing
1757 	 * it to four packets (1.5Kx4 = 6K window).
1758 	 */
1759 	if (sendalot && --maxburst)
1760 		goto again;
1761 #endif
1762 	if (sendalot)
1763 		goto again;
1764 	return (0);
1765 }
1766 
1767 void
1768 tcp_setpersist(struct tcpcb *tp)
1769 {
1770 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1771 	int tt;
1772 
1773 	tp->t_flags &= ~TF_PREVVALID;
1774 	if (tcp_timer_active(tp, TT_REXMT))
1775 		panic("tcp_setpersist: retransmit pending");
1776 	/*
1777 	 * Start/restart persistence timer.
1778 	 */
1779 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1780 		      tcp_persmin, tcp_persmax);
1781 	tcp_timer_activate(tp, TT_PERSIST, tt);
1782 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1783 		tp->t_rxtshift++;
1784 }
1785 
1786 /*
1787  * Insert TCP options according to the supplied parameters to the place
1788  * optp in a consistent way.  Can handle unaligned destinations.
1789  *
1790  * The order of the option processing is crucial for optimal packing and
1791  * alignment for the scarce option space.
1792  *
1793  * The optimal order for a SYN/SYN-ACK segment is:
1794  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1795  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1796  *
1797  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1798  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1799  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1800  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1801  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1802  */
1803 int
1804 tcp_addoptions(struct tcpopt *to, u_char *optp)
1805 {
1806 	u_int32_t mask, optlen = 0;
1807 
1808 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1809 		if ((to->to_flags & mask) != mask)
1810 			continue;
1811 		if (optlen == TCP_MAXOLEN)
1812 			break;
1813 		switch (to->to_flags & mask) {
1814 		case TOF_MSS:
1815 			while (optlen % 4) {
1816 				optlen += TCPOLEN_NOP;
1817 				*optp++ = TCPOPT_NOP;
1818 			}
1819 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1820 				continue;
1821 			optlen += TCPOLEN_MAXSEG;
1822 			*optp++ = TCPOPT_MAXSEG;
1823 			*optp++ = TCPOLEN_MAXSEG;
1824 			to->to_mss = htons(to->to_mss);
1825 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1826 			optp += sizeof(to->to_mss);
1827 			break;
1828 		case TOF_SCALE:
1829 			while (!optlen || optlen % 2 != 1) {
1830 				optlen += TCPOLEN_NOP;
1831 				*optp++ = TCPOPT_NOP;
1832 			}
1833 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1834 				continue;
1835 			optlen += TCPOLEN_WINDOW;
1836 			*optp++ = TCPOPT_WINDOW;
1837 			*optp++ = TCPOLEN_WINDOW;
1838 			*optp++ = to->to_wscale;
1839 			break;
1840 		case TOF_SACKPERM:
1841 			while (optlen % 2) {
1842 				optlen += TCPOLEN_NOP;
1843 				*optp++ = TCPOPT_NOP;
1844 			}
1845 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1846 				continue;
1847 			optlen += TCPOLEN_SACK_PERMITTED;
1848 			*optp++ = TCPOPT_SACK_PERMITTED;
1849 			*optp++ = TCPOLEN_SACK_PERMITTED;
1850 			break;
1851 		case TOF_TS:
1852 			while (!optlen || optlen % 4 != 2) {
1853 				optlen += TCPOLEN_NOP;
1854 				*optp++ = TCPOPT_NOP;
1855 			}
1856 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1857 				continue;
1858 			optlen += TCPOLEN_TIMESTAMP;
1859 			*optp++ = TCPOPT_TIMESTAMP;
1860 			*optp++ = TCPOLEN_TIMESTAMP;
1861 			to->to_tsval = htonl(to->to_tsval);
1862 			to->to_tsecr = htonl(to->to_tsecr);
1863 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1864 			optp += sizeof(to->to_tsval);
1865 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1866 			optp += sizeof(to->to_tsecr);
1867 			break;
1868 		case TOF_SIGNATURE:
1869 			{
1870 			int siglen = TCPOLEN_SIGNATURE - 2;
1871 
1872 			while (!optlen || optlen % 4 != 2) {
1873 				optlen += TCPOLEN_NOP;
1874 				*optp++ = TCPOPT_NOP;
1875 			}
1876 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1877 				to->to_flags &= ~TOF_SIGNATURE;
1878 				continue;
1879 			}
1880 			optlen += TCPOLEN_SIGNATURE;
1881 			*optp++ = TCPOPT_SIGNATURE;
1882 			*optp++ = TCPOLEN_SIGNATURE;
1883 			to->to_signature = optp;
1884 			while (siglen--)
1885 				 *optp++ = 0;
1886 			break;
1887 			}
1888 		case TOF_SACK:
1889 			{
1890 			int sackblks = 0;
1891 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1892 			tcp_seq sack_seq;
1893 
1894 			while (!optlen || optlen % 4 != 2) {
1895 				optlen += TCPOLEN_NOP;
1896 				*optp++ = TCPOPT_NOP;
1897 			}
1898 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1899 				continue;
1900 			optlen += TCPOLEN_SACKHDR;
1901 			*optp++ = TCPOPT_SACK;
1902 			sackblks = min(to->to_nsacks,
1903 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1904 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1905 			while (sackblks--) {
1906 				sack_seq = htonl(sack->start);
1907 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1908 				optp += sizeof(sack_seq);
1909 				sack_seq = htonl(sack->end);
1910 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1911 				optp += sizeof(sack_seq);
1912 				optlen += TCPOLEN_SACK;
1913 				sack++;
1914 			}
1915 			TCPSTAT_INC(tcps_sack_send_blocks);
1916 			break;
1917 			}
1918 		case TOF_FASTOPEN:
1919 			{
1920 			int total_len;
1921 
1922 			/* XXX is there any point to aligning this option? */
1923 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1924 			if (TCP_MAXOLEN - optlen < total_len) {
1925 				to->to_flags &= ~TOF_FASTOPEN;
1926 				continue;
1927 			}
1928 			*optp++ = TCPOPT_FAST_OPEN;
1929 			*optp++ = total_len;
1930 			if (to->to_tfo_len > 0) {
1931 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1932 				optp += to->to_tfo_len;
1933 			}
1934 			optlen += total_len;
1935 			break;
1936 			}
1937 		default:
1938 			panic("%s: unknown TCP option type", __func__);
1939 			break;
1940 		}
1941 	}
1942 
1943 	/* Terminate and pad TCP options to a 4 byte boundary. */
1944 	if (optlen % 4) {
1945 		optlen += TCPOLEN_EOL;
1946 		*optp++ = TCPOPT_EOL;
1947 	}
1948 	/*
1949 	 * According to RFC 793 (STD0007):
1950 	 *   "The content of the header beyond the End-of-Option option
1951 	 *    must be header padding (i.e., zero)."
1952 	 *   and later: "The padding is composed of zeros."
1953 	 */
1954 	while (optlen % 4) {
1955 		optlen += TCPOLEN_PAD;
1956 		*optp++ = TCPOPT_PAD;
1957 	}
1958 
1959 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1960 	return (optlen);
1961 }
1962 
1963 /*
1964  * This is a copy of m_copym(), taking the TSO segment size/limit
1965  * constraints into account, and advancing the sndptr as it goes.
1966  */
1967 struct mbuf *
1968 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1969     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1970 {
1971 #ifdef KERN_TLS
1972 	struct ktls_session *tls, *ntls;
1973 	struct mbuf *start __diagused;
1974 #endif
1975 	struct mbuf *n, **np;
1976 	struct mbuf *top;
1977 	int32_t off = off0;
1978 	int32_t len = *plen;
1979 	int32_t fragsize;
1980 	int32_t len_cp = 0;
1981 	int32_t *pkthdrlen;
1982 	uint32_t mlen, frags;
1983 	bool copyhdr;
1984 
1985 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1986 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1987 	if (off == 0 && m->m_flags & M_PKTHDR)
1988 		copyhdr = true;
1989 	else
1990 		copyhdr = false;
1991 	while (off > 0) {
1992 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1993 		if (off < m->m_len)
1994 			break;
1995 		off -= m->m_len;
1996 		if ((sb) && (m == sb->sb_sndptr)) {
1997 			sb->sb_sndptroff += m->m_len;
1998 			sb->sb_sndptr = m->m_next;
1999 		}
2000 		m = m->m_next;
2001 	}
2002 	np = &top;
2003 	top = NULL;
2004 	pkthdrlen = NULL;
2005 #ifdef KERN_TLS
2006 	if (hw_tls && (m->m_flags & M_EXTPG))
2007 		tls = m->m_epg_tls;
2008 	else
2009 		tls = NULL;
2010 	start = m;
2011 #endif
2012 	while (len > 0) {
2013 		if (m == NULL) {
2014 			KASSERT(len == M_COPYALL,
2015 			    ("tcp_m_copym, length > size of mbuf chain"));
2016 			*plen = len_cp;
2017 			if (pkthdrlen != NULL)
2018 				*pkthdrlen = len_cp;
2019 			break;
2020 		}
2021 #ifdef KERN_TLS
2022 		if (hw_tls) {
2023 			if (m->m_flags & M_EXTPG)
2024 				ntls = m->m_epg_tls;
2025 			else
2026 				ntls = NULL;
2027 
2028 			/*
2029 			 * Avoid mixing TLS records with handshake
2030 			 * data or TLS records from different
2031 			 * sessions.
2032 			 */
2033 			if (tls != ntls) {
2034 				MPASS(m != start);
2035 				*plen = len_cp;
2036 				if (pkthdrlen != NULL)
2037 					*pkthdrlen = len_cp;
2038 				break;
2039 			}
2040 		}
2041 #endif
2042 		mlen = min(len, m->m_len - off);
2043 		if (seglimit) {
2044 			/*
2045 			 * For M_EXTPG mbufs, add 3 segments
2046 			 * + 1 in case we are crossing page boundaries
2047 			 * + 2 in case the TLS hdr/trailer are used
2048 			 * It is cheaper to just add the segments
2049 			 * than it is to take the cache miss to look
2050 			 * at the mbuf ext_pgs state in detail.
2051 			 */
2052 			if (m->m_flags & M_EXTPG) {
2053 				fragsize = min(segsize, PAGE_SIZE);
2054 				frags = 3;
2055 			} else {
2056 				fragsize = segsize;
2057 				frags = 0;
2058 			}
2059 
2060 			/* Break if we really can't fit anymore. */
2061 			if ((frags + 1) >= seglimit) {
2062 				*plen =	len_cp;
2063 				if (pkthdrlen != NULL)
2064 					*pkthdrlen = len_cp;
2065 				break;
2066 			}
2067 
2068 			/*
2069 			 * Reduce size if you can't copy the whole
2070 			 * mbuf. If we can't copy the whole mbuf, also
2071 			 * adjust len so the loop will end after this
2072 			 * mbuf.
2073 			 */
2074 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2075 				mlen = (seglimit - frags - 1) * fragsize;
2076 				len = mlen;
2077 				*plen = len_cp + len;
2078 				if (pkthdrlen != NULL)
2079 					*pkthdrlen = *plen;
2080 			}
2081 			frags += howmany(mlen, fragsize);
2082 			if (frags == 0)
2083 				frags++;
2084 			seglimit -= frags;
2085 			KASSERT(seglimit > 0,
2086 			    ("%s: seglimit went too low", __func__));
2087 		}
2088 		if (copyhdr)
2089 			n = m_gethdr(M_NOWAIT, m->m_type);
2090 		else
2091 			n = m_get(M_NOWAIT, m->m_type);
2092 		*np = n;
2093 		if (n == NULL)
2094 			goto nospace;
2095 		if (copyhdr) {
2096 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2097 				goto nospace;
2098 			if (len == M_COPYALL)
2099 				n->m_pkthdr.len -= off0;
2100 			else
2101 				n->m_pkthdr.len = len;
2102 			pkthdrlen = &n->m_pkthdr.len;
2103 			copyhdr = false;
2104 		}
2105 		n->m_len = mlen;
2106 		len_cp += n->m_len;
2107 		if (m->m_flags & (M_EXT|M_EXTPG)) {
2108 			n->m_data = m->m_data + off;
2109 			mb_dupcl(n, m);
2110 		} else
2111 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2112 			    (u_int)n->m_len);
2113 
2114 		if (sb && (sb->sb_sndptr == m) &&
2115 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2116 			sb->sb_sndptroff += m->m_len;
2117 			sb->sb_sndptr = m->m_next;
2118 		}
2119 		off = 0;
2120 		if (len != M_COPYALL) {
2121 			len -= n->m_len;
2122 		}
2123 		m = m->m_next;
2124 		np = &n->m_next;
2125 	}
2126 	return (top);
2127 nospace:
2128 	m_freem(top);
2129 	return (NULL);
2130 }
2131 
2132 void
2133 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2134 {
2135 
2136 	/*
2137 	 * Automatic sizing of send socket buffer.  Often the send buffer
2138 	 * size is not optimally adjusted to the actual network conditions
2139 	 * at hand (delay bandwidth product).  Setting the buffer size too
2140 	 * small limits throughput on links with high bandwidth and high
2141 	 * delay (eg. trans-continental/oceanic links).  Setting the
2142 	 * buffer size too big consumes too much real kernel memory,
2143 	 * especially with many connections on busy servers.
2144 	 *
2145 	 * The criteria to step up the send buffer one notch are:
2146 	 *  1. receive window of remote host is larger than send buffer
2147 	 *     (with a fudge factor of 5/4th);
2148 	 *  2. send buffer is filled to 7/8th with data (so we actually
2149 	 *     have data to make use of it);
2150 	 *  3. send buffer fill has not hit maximal automatic size;
2151 	 *  4. our send window (slow start and cogestion controlled) is
2152 	 *     larger than sent but unacknowledged data in send buffer.
2153 	 *
2154 	 * The remote host receive window scaling factor may limit the
2155 	 * growing of the send buffer before it reaches its allowed
2156 	 * maximum.
2157 	 *
2158 	 * It scales directly with slow start or congestion window
2159 	 * and does at most one step per received ACK.  This fast
2160 	 * scaling has the drawback of growing the send buffer beyond
2161 	 * what is strictly necessary to make full use of a given
2162 	 * delay*bandwidth product.  However testing has shown this not
2163 	 * to be much of an problem.  At worst we are trading wasting
2164 	 * of available bandwidth (the non-use of it) for wasting some
2165 	 * socket buffer memory.
2166 	 *
2167 	 * TODO: Shrink send buffer during idle periods together
2168 	 * with congestion window.  Requires another timer.  Has to
2169 	 * wait for upcoming tcp timer rewrite.
2170 	 *
2171 	 * XXXGL: should there be used sbused() or sbavail()?
2172 	 */
2173 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2174 		int lowat;
2175 
2176 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2177 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2178 		    sbused(&so->so_snd) >=
2179 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2180 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2181 		    sendwin >= (sbused(&so->so_snd) -
2182 		    (tp->snd_nxt - tp->snd_una))) {
2183 			if (!sbreserve_locked(so, SO_SND,
2184 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2185 			     V_tcp_autosndbuf_max), curthread))
2186 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2187 		}
2188 	}
2189 }
2190