xref: /freebsd/sys/netinet/tcp_output.c (revision 87bf66d4a7488c496af110d4d05cc0273d49f82e)
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 	unsigned ipsec_optlen = 0;
205 	int idle, sendalot, curticks;
206 	int sack_rxmit, sack_bytes_rxmt;
207 	struct sackhole *p;
208 	int tso, mtu;
209 	struct tcpopt to;
210 	struct udphdr *udp = NULL;
211 	struct tcp_log_buffer *lgb;
212 	unsigned int wanted_cookie = 0;
213 	unsigned int dont_sendalot = 0;
214 #ifdef INET6
215 	struct ip6_hdr *ip6 = NULL;
216 	const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
217 #endif
218 #ifdef KERN_TLS
219 	const bool hw_tls = tp->t_nic_ktls_xmit != 0;
220 #else
221 	const bool hw_tls = false;
222 #endif
223 
224 	NET_EPOCH_ASSERT();
225 	INP_WLOCK_ASSERT(inp);
226 
227 #ifdef TCP_OFFLOAD
228 	if (tp->t_flags & TF_TOE)
229 		return (tcp_offload_output(tp));
230 #endif
231 
232 	/*
233 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
234 	 * only allow the initial SYN or SYN|ACK and those sent
235 	 * by the retransmit timer.
236 	 */
237 	if ((tp->t_flags & TF_FASTOPEN) &&
238 	    ((tp->t_state == TCPS_SYN_SENT) ||
239 	    (tp->t_state == TCPS_SYN_RECEIVED)) &&
240 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* SYN or SYN|ACK sent */
241 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
242 		return (0);
243 
244 	/*
245 	 * Determine length of data that should be transmitted,
246 	 * and flags that will be used.
247 	 * If there is some data or critical controls (SYN, RST)
248 	 * to send, then transmit; otherwise, investigate further.
249 	 */
250 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
251 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
252 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
253 		cc_after_idle(tp);
254 	tp->t_flags &= ~TF_LASTIDLE;
255 	if (idle) {
256 		if (tp->t_flags & TF_MORETOCOME) {
257 			tp->t_flags |= TF_LASTIDLE;
258 			idle = 0;
259 		}
260 	}
261 again:
262 	sendwin = 0;
263 	/*
264 	 * If we've recently taken a timeout, snd_max will be greater than
265 	 * snd_nxt.  There may be SACK information that allows us to avoid
266 	 * resending already delivered data.  Adjust snd_nxt accordingly.
267 	 */
268 	if ((tp->t_flags & TF_SACK_PERMIT) &&
269 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
270 		sendwin = tcp_sack_adjust(tp);
271 	sendalot = 0;
272 	tso = 0;
273 	mtu = 0;
274 	off = tp->snd_nxt - tp->snd_una;
275 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + sendwin);
276 
277 	flags = tcp_outflags[tp->t_state];
278 	/*
279 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
280 	 * to send out new data (when sendalot is 1), bypass this function.
281 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
282 	 * we're replacing a (future) new transmission with a retransmission
283 	 * now, and we previously incremented snd_cwnd in tcp_input().
284 	 */
285 	/*
286 	 * Still in sack recovery , reset rxmit flag to zero.
287 	 */
288 	sack_rxmit = 0;
289 	sack_bytes_rxmt = 0;
290 	len = 0;
291 	p = NULL;
292 	if ((tp->t_flags & TF_SACK_PERMIT) &&
293 	    (IN_FASTRECOVERY(tp->t_flags) || SEQ_LT(tp->snd_nxt, tp->snd_max)) &&
294 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
295 		uint32_t cwin;
296 
297 		cwin =
298 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
299 		/* Do not retransmit SACK segments beyond snd_recover */
300 		if (SEQ_GT(p->end, tp->snd_recover)) {
301 			/*
302 			 * (At least) part of sack hole extends beyond
303 			 * snd_recover. Check to see if we can rexmit data
304 			 * for this hole.
305 			 */
306 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
307 				/*
308 				 * Can't rexmit any more data for this hole.
309 				 * That data will be rexmitted in the next
310 				 * sack recovery episode, when snd_recover
311 				 * moves past p->rxmit.
312 				 */
313 				p = NULL;
314 				goto after_sack_rexmit;
315 			} else {
316 				/* Can rexmit part of the current hole */
317 				len = ((int32_t)ulmin(cwin,
318 				    SEQ_SUB(tp->snd_recover, p->rxmit)));
319 			}
320 		} else {
321 			len = ((int32_t)ulmin(cwin,
322 			    SEQ_SUB(p->end, p->rxmit)));
323 		}
324 		if (len > 0) {
325 			off = SEQ_SUB(p->rxmit, tp->snd_una);
326 			KASSERT(off >= 0,("%s: sack block to the left of una : %d",
327 			    __func__, off));
328 			sack_rxmit = 1;
329 			sendalot = 1;
330 		}
331 	}
332 after_sack_rexmit:
333 	/*
334 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
335 	 * state flags.
336 	 */
337 	if (tp->t_flags & TF_NEEDFIN)
338 		flags |= TH_FIN;
339 	if (tp->t_flags & TF_NEEDSYN)
340 		flags |= TH_SYN;
341 
342 	SOCKBUF_LOCK(&so->so_snd);
343 	/*
344 	 * If in persist timeout with window of 0, send 1 byte.
345 	 * Otherwise, if window is small but nonzero
346 	 * and timer expired, we will send what we can
347 	 * and go to transmit state.
348 	 */
349 	if (tp->t_flags & TF_FORCEDATA) {
350 		if (sendwin == 0) {
351 			/*
352 			 * If we still have some data to send, then
353 			 * clear the FIN bit.  Usually this would
354 			 * happen below when it realizes that we
355 			 * aren't sending all the data.  However,
356 			 * if we have exactly 1 byte of unsent data,
357 			 * then it won't clear the FIN bit below,
358 			 * and if we are in persist state, we wind
359 			 * up sending the packet without recording
360 			 * that we sent the FIN bit.
361 			 *
362 			 * We can't just blindly clear the FIN bit,
363 			 * because if we don't have any more data
364 			 * to send then the probe will be the FIN
365 			 * itself.
366 			 */
367 			if (off < sbused(&so->so_snd))
368 				flags &= ~TH_FIN;
369 			sendwin = 1;
370 		} else {
371 			tcp_timer_activate(tp, TT_PERSIST, 0);
372 			tp->t_rxtshift = 0;
373 		}
374 	}
375 
376 	/*
377 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
378 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
379 	 * a negative length.  This can also occur when TCP opens up
380 	 * its congestion window while receiving additional duplicate
381 	 * acks after fast-retransmit because TCP will reset snd_nxt
382 	 * to snd_max after the fast-retransmit.
383 	 *
384 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
385 	 * be set to snd_una, the offset will be 0, and the length may
386 	 * wind up 0.
387 	 *
388 	 * If sack_rxmit is true we are retransmitting from the scoreboard
389 	 * in which case len is already set.
390 	 */
391 	if (sack_rxmit == 0) {
392 		if ((sack_bytes_rxmt == 0) || SEQ_LT(tp->snd_nxt, tp->snd_max)) {
393 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
394 			    off);
395 		} else {
396 			int32_t cwin;
397 
398 			/*
399 			 * We are inside of a SACK recovery episode and are
400 			 * sending new data, having retransmitted all the
401 			 * data possible in the scoreboard.
402 			 */
403 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
404 			    off);
405 			/*
406 			 * Don't remove this (len > 0) check !
407 			 * We explicitly check for len > 0 here (although it
408 			 * isn't really necessary), to work around a gcc
409 			 * optimization issue - to force gcc to compute
410 			 * len above. Without this check, the computation
411 			 * of len is bungled by the optimizer.
412 			 */
413 			if (len > 0) {
414 				cwin = tp->snd_cwnd - imax(0, (int32_t)
415 					(tp->snd_nxt - tp->snd_recover)) -
416 					sack_bytes_rxmt;
417 				if (cwin < 0)
418 					cwin = 0;
419 				len = imin(len, cwin);
420 			}
421 		}
422 	}
423 
424 	/*
425 	 * Lop off SYN bit if it has already been sent.  However, if this
426 	 * is SYN-SENT state and if segment contains data and if we don't
427 	 * know that foreign host supports TAO, suppress sending segment.
428 	 */
429 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
430 		if (tp->t_state != TCPS_SYN_RECEIVED)
431 			flags &= ~TH_SYN;
432 		/*
433 		 * When sending additional segments following a TFO SYN|ACK,
434 		 * do not include the SYN bit.
435 		 */
436 		if ((tp->t_flags & TF_FASTOPEN) &&
437 		    (tp->t_state == TCPS_SYN_RECEIVED))
438 			flags &= ~TH_SYN;
439 		off--, len++;
440 	}
441 
442 	/*
443 	 * Be careful not to send data and/or FIN on SYN segments.
444 	 * This measure is needed to prevent interoperability problems
445 	 * with not fully conformant TCP implementations.
446 	 */
447 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
448 		len = 0;
449 		flags &= ~TH_FIN;
450 	}
451 
452 	/*
453 	 * On TFO sockets, ensure no data is sent in the following cases:
454 	 *
455 	 *  - When retransmitting SYN|ACK on a passively-created socket
456 	 *
457 	 *  - When retransmitting SYN on an actively created socket
458 	 *
459 	 *  - When sending a zero-length cookie (cookie request) on an
460 	 *    actively created socket
461 	 *
462 	 *  - When the socket is in the CLOSED state (RST is being sent)
463 	 */
464 	if ((tp->t_flags & TF_FASTOPEN) &&
465 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
466 	     ((tp->t_state == TCPS_SYN_SENT) &&
467 	      (tp->t_tfo_client_cookie_len == 0)) ||
468 	     (flags & TH_RST)))
469 		len = 0;
470 
471 	/* Without fast-open there should never be data sent on a SYN. */
472 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
473 		len = 0;
474 	}
475 
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 	ipoptlen += ipsec_optlen;
554 
555 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
556 	    (tp->t_port == 0) &&
557 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
558 	    tp->rcv_numsacks == 0 && ((sack_rxmit == 0) || V_tcp_sack_tso) &&
559 	    (ipoptlen == 0 || (ipoptlen == ipsec_optlen &&
560 	    (tp->t_flags2 & TF2_IPSEC_TSO) != 0)) &&
561 	    !(flags & TH_SYN))
562 		tso = 1;
563 
564 	if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
565 		    tp->snd_una + sbused(&so->so_snd))) {
566 			flags &= ~TH_FIN;
567 	}
568 
569 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
570 	    (long)TCP_MAXWIN << tp->rcv_scale);
571 
572 	/*
573 	 * Sender silly window avoidance.   We transmit under the following
574 	 * conditions when len is non-zero:
575 	 *
576 	 *	- We have a full segment (or more with TSO)
577 	 *	- This is the last buffer in a write()/send() and we are
578 	 *	  either idle or running NODELAY
579 	 *	- we've timed out (e.g. persist timer)
580 	 *	- we have more then 1/2 the maximum send window's worth of
581 	 *	  data (receiver may be limited the window size)
582 	 *	- we need to retransmit
583 	 */
584 	if (len) {
585 		if (len >= tp->t_maxseg)
586 			goto send;
587 		/*
588 		 * As the TCP header options are now
589 		 * considered when setting up the initial
590 		 * window, we would not send the last segment
591 		 * if we skip considering the option length here.
592 		 * Note: this may not work when tcp headers change
593 		 * very dynamically in the future.
594 		 */
595 		if ((((tp->t_flags & TF_SIGNATURE) ?
596 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
597 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
598 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
599 		    len) >= tp->t_maxseg)
600 			goto send;
601 		/*
602 		 * NOTE! on localhost connections an 'ack' from the remote
603 		 * end may occur synchronously with the output and cause
604 		 * us to flush a buffer queued with moretocome.  XXX
605 		 *
606 		 * note: the len + off check is almost certainly unnecessary.
607 		 */
608 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
609 		    (idle || (tp->t_flags & TF_NODELAY)) &&
610 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
611 		    (tp->t_flags & TF_NOPUSH) == 0) {
612 			goto send;
613 		}
614 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
615 			goto send;
616 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
617 			goto send;
618 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
619 			goto send;
620 		if (sack_rxmit)
621 			goto send;
622 	}
623 
624 	/*
625 	 * Sending of standalone window updates.
626 	 *
627 	 * Window updates are important when we close our window due to a
628 	 * full socket buffer and are opening it again after the application
629 	 * reads data from it.  Once the window has opened again and the
630 	 * remote end starts to send again the ACK clock takes over and
631 	 * provides the most current window information.
632 	 *
633 	 * We must avoid the silly window syndrome whereas every read
634 	 * from the receive buffer, no matter how small, causes a window
635 	 * update to be sent.  We also should avoid sending a flurry of
636 	 * window updates when the socket buffer had queued a lot of data
637 	 * and the application is doing small reads.
638 	 *
639 	 * Prevent a flurry of pointless window updates by only sending
640 	 * an update when we can increase the advertized window by more
641 	 * than 1/4th of the socket buffer capacity.  When the buffer is
642 	 * getting full or is very small be more aggressive and send an
643 	 * update whenever we can increase by two mss sized segments.
644 	 * In all other situations the ACK's to new incoming data will
645 	 * carry further window increases.
646 	 *
647 	 * Don't send an independent window update if a delayed
648 	 * ACK is pending (it will get piggy-backed on it) or the
649 	 * remote side already has done a half-close and won't send
650 	 * more data.
651 	 */
652 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
653 	    !(tp->t_flags & TF_DELACK) &&
654 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
655 		/*
656 		 * "adv" is the amount we could increase the window,
657 		 * taking into account that we are limited by
658 		 * TCP_MAXWIN << tp->rcv_scale.
659 		 */
660 		int32_t adv;
661 		int oldwin;
662 
663 		adv = recwin;
664 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
665 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
666 			if (adv > oldwin)
667 				adv -= oldwin;
668 			else
669 				adv = 0;
670 		} else
671 			oldwin = 0;
672 
673 		/*
674 		 * If the new window size ends up being the same as or less
675 		 * than the old size when it is scaled, then don't force
676 		 * a window update.
677 		 */
678 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
679 			goto dontupdate;
680 
681 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
682 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
683 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
684 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
685 		     adv >= TCP_MAXWIN << tp->rcv_scale))
686 			goto send;
687 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
688 			goto send;
689 	}
690 dontupdate:
691 
692 	/*
693 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
694 	 * is also a catch-all for the retransmit timer timeout case.
695 	 */
696 	if (tp->t_flags & TF_ACKNOW)
697 		goto send;
698 	if ((flags & TH_RST) ||
699 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
700 		goto send;
701 	if (SEQ_GT(tp->snd_up, tp->snd_una))
702 		goto send;
703 	/*
704 	 * If our state indicates that FIN should be sent
705 	 * and we have not yet done so, then we need to send.
706 	 */
707 	if (flags & TH_FIN &&
708 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
709 		goto send;
710 	/*
711 	 * In SACK, it is possible for tcp_output to fail to send a segment
712 	 * after the retransmission timer has been turned off.  Make sure
713 	 * that the retransmission timer is set.
714 	 */
715 	if ((tp->t_flags & TF_SACK_PERMIT) &&
716 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
717 	    !tcp_timer_active(tp, TT_REXMT) &&
718 	    !tcp_timer_active(tp, TT_PERSIST)) {
719 		tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
720 		goto just_return;
721 	}
722 	/*
723 	 * TCP window updates are not reliable, rather a polling protocol
724 	 * using ``persist'' packets is used to insure receipt of window
725 	 * updates.  The three ``states'' for the output side are:
726 	 *	idle			not doing retransmits or persists
727 	 *	persisting		to move a small or zero window
728 	 *	(re)transmitting	and thereby not persisting
729 	 *
730 	 * tcp_timer_active(tp, TT_PERSIST)
731 	 *	is true when we are in persist state.
732 	 * (tp->t_flags & TF_FORCEDATA)
733 	 *	is set when we are called to send a persist packet.
734 	 * tcp_timer_active(tp, TT_REXMT)
735 	 *	is set when we are retransmitting
736 	 * The output side is idle when both timers are zero.
737 	 *
738 	 * If send window is too small, there is data to transmit, and no
739 	 * retransmit or persist is pending, then go to persist state.
740 	 * If nothing happens soon, send when timer expires:
741 	 * if window is nonzero, transmit what we can,
742 	 * otherwise force out a byte.
743 	 */
744 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
745 	    !tcp_timer_active(tp, TT_PERSIST)) {
746 		tp->t_rxtshift = 0;
747 		tcp_setpersist(tp);
748 	}
749 
750 	/*
751 	 * No reason to send a segment, just return.
752 	 */
753 just_return:
754 	SOCKBUF_UNLOCK(&so->so_snd);
755 	return (0);
756 
757 send:
758 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
759 	if (len > 0) {
760 		if (len >= tp->t_maxseg)
761 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
762 		else
763 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
764 	}
765 	/*
766 	 * Before ESTABLISHED, force sending of initial options
767 	 * unless TCP set not to do any options.
768 	 * NOTE: we assume that the IP/TCP header plus TCP options
769 	 * always fit in a single mbuf, leaving room for a maximum
770 	 * link header, i.e.
771 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
772 	 */
773 	optlen = 0;
774 #ifdef INET6
775 	if (isipv6)
776 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
777 	else
778 #endif
779 		hdrlen = sizeof (struct tcpiphdr);
780 
781 	if (flags & TH_SYN) {
782 		tp->snd_nxt = tp->iss;
783 	}
784 
785 	/*
786 	 * Compute options for segment.
787 	 * We only have to care about SYN and established connection
788 	 * segments.  Options for SYN-ACK segments are handled in TCP
789 	 * syncache.
790 	 */
791 	to.to_flags = 0;
792 	if ((tp->t_flags & TF_NOOPT) == 0) {
793 		/* Maximum segment size. */
794 		if (flags & TH_SYN) {
795 			to.to_mss = tcp_mssopt(&inp->inp_inc);
796 			if (tp->t_port)
797 				to.to_mss -= V_tcp_udp_tunneling_overhead;
798 			to.to_flags |= TOF_MSS;
799 
800 			/*
801 			 * On SYN or SYN|ACK transmits on TFO connections,
802 			 * only include the TFO option if it is not a
803 			 * retransmit, as the presence of the TFO option may
804 			 * have caused the original SYN or SYN|ACK to have
805 			 * been dropped by a middlebox.
806 			 */
807 			if ((tp->t_flags & TF_FASTOPEN) &&
808 			    (tp->t_rxtshift == 0)) {
809 				if (tp->t_state == TCPS_SYN_RECEIVED) {
810 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
811 					to.to_tfo_cookie =
812 					    (u_int8_t *)&tp->t_tfo_cookie.server;
813 					to.to_flags |= TOF_FASTOPEN;
814 					wanted_cookie = 1;
815 				} else if (tp->t_state == TCPS_SYN_SENT) {
816 					to.to_tfo_len =
817 					    tp->t_tfo_client_cookie_len;
818 					to.to_tfo_cookie =
819 					    tp->t_tfo_cookie.client;
820 					to.to_flags |= TOF_FASTOPEN;
821 					wanted_cookie = 1;
822 					/*
823 					 * If we wind up having more data to
824 					 * send with the SYN than can fit in
825 					 * one segment, don't send any more
826 					 * until the SYN|ACK comes back from
827 					 * the other end.
828 					 */
829 					dont_sendalot = 1;
830 				}
831 			}
832 		}
833 		/* Window scaling. */
834 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
835 			to.to_wscale = tp->request_r_scale;
836 			to.to_flags |= TOF_SCALE;
837 		}
838 		/* Timestamps. */
839 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
840 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
841 			curticks = tcp_ts_getticks();
842 			to.to_tsval = curticks + tp->ts_offset;
843 			to.to_tsecr = tp->ts_recent;
844 			to.to_flags |= TOF_TS;
845 			if (tp->t_rxtshift == 1)
846 				tp->t_badrxtwin = curticks;
847 		}
848 
849 		/* Set receive buffer autosizing timestamp. */
850 		if (tp->rfbuf_ts == 0 &&
851 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
852 			tp->rfbuf_ts = tcp_ts_getticks();
853 
854 		/* Selective ACK's. */
855 		if (tp->t_flags & TF_SACK_PERMIT) {
856 			if (flags & TH_SYN)
857 				to.to_flags |= TOF_SACKPERM;
858 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
859 			    tp->rcv_numsacks > 0) {
860 				to.to_flags |= TOF_SACK;
861 				to.to_nsacks = tp->rcv_numsacks;
862 				to.to_sacks = (u_char *)tp->sackblks;
863 			}
864 		}
865 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
866 		/* TCP-MD5 (RFC2385). */
867 		/*
868 		 * Check that TCP_MD5SIG is enabled in tcpcb to
869 		 * account the size needed to set this TCP option.
870 		 */
871 		if (tp->t_flags & TF_SIGNATURE)
872 			to.to_flags |= TOF_SIGNATURE;
873 #endif /* TCP_SIGNATURE */
874 
875 		/* Processing the options. */
876 		hdrlen += optlen = tcp_addoptions(&to, opt);
877 		/*
878 		 * If we wanted a TFO option to be added, but it was unable
879 		 * to fit, ensure no data is sent.
880 		 */
881 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
882 		    !(to.to_flags & TOF_FASTOPEN))
883 			len = 0;
884 	}
885 	if (tp->t_port) {
886 		if (V_tcp_udp_tunneling_port == 0) {
887 			/* The port was removed?? */
888 			SOCKBUF_UNLOCK(&so->so_snd);
889 			return (EHOSTUNREACH);
890 		}
891 		hdrlen += sizeof(struct udphdr);
892 	}
893 	/*
894 	 * Adjust data length if insertion of options will
895 	 * bump the packet length beyond the t_maxseg length.
896 	 * Clear the FIN bit because we cut off the tail of
897 	 * the segment.
898 	 */
899 	if (len + optlen + ipoptlen > tp->t_maxseg) {
900 		flags &= ~TH_FIN;
901 
902 		if (tso) {
903 			u_int if_hw_tsomax;
904 			u_int moff;
905 			int max_len;
906 
907 			/* extract TSO information */
908 			if_hw_tsomax = tp->t_tsomax;
909 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
910 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
911 
912 			/*
913 			 * Limit a TSO burst to prevent it from
914 			 * overflowing or exceeding the maximum length
915 			 * allowed by the network interface:
916 			 */
917 			KASSERT(ipoptlen ==  ipsec_optlen,
918 			    ("%s: TSO can't do IP options", __func__));
919 
920 			/*
921 			 * Check if we should limit by maximum payload
922 			 * length:
923 			 */
924 			if (if_hw_tsomax != 0) {
925 				/* compute maximum TSO length */
926 				max_len = if_hw_tsomax - hdrlen -
927 				    ipsec_optlen - max_linkhdr;
928 				if (max_len <= 0) {
929 					len = 0;
930 				} else if (len > max_len) {
931 					sendalot = 1;
932 					len = max_len;
933 				}
934 			}
935 
936 			/*
937 			 * Prevent the last segment from being
938 			 * fractional unless the send sockbuf can be
939 			 * emptied:
940 			 */
941 			max_len = tp->t_maxseg - optlen - ipsec_optlen;
942 			if (((uint32_t)off + (uint32_t)len) <
943 			    sbavail(&so->so_snd)) {
944 				moff = len % max_len;
945 				if (moff != 0) {
946 					len -= moff;
947 					sendalot = 1;
948 				}
949 			}
950 
951 			/*
952 			 * In case there are too many small fragments
953 			 * don't use TSO:
954 			 */
955 			if (len <= max_len) {
956 				len = max_len;
957 				sendalot = 1;
958 				tso = 0;
959 			}
960 
961 			/*
962 			 * Send the FIN in a separate segment
963 			 * after the bulk sending is done.
964 			 * We don't trust the TSO implementations
965 			 * to clear the FIN flag on all but the
966 			 * last segment.
967 			 */
968 			if (tp->t_flags & TF_NEEDFIN)
969 				sendalot = 1;
970 		} else {
971 			if (optlen + ipoptlen >= tp->t_maxseg) {
972 				/*
973 				 * Since we don't have enough space to put
974 				 * the IP header chain and the TCP header in
975 				 * one packet as required by RFC 7112, don't
976 				 * send it. Also ensure that at least one
977 				 * byte of the payload can be put into the
978 				 * TCP segment.
979 				 */
980 				SOCKBUF_UNLOCK(&so->so_snd);
981 				error = EMSGSIZE;
982 				sack_rxmit = 0;
983 				goto out;
984 			}
985 			len = tp->t_maxseg - optlen - ipoptlen;
986 			sendalot = 1;
987 			if (dont_sendalot)
988 				sendalot = 0;
989 		}
990 	} else
991 		tso = 0;
992 
993 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
994 	    ("%s: len > IP_MAXPACKET", __func__));
995 
996 /*#ifdef DIAGNOSTIC*/
997 #ifdef INET6
998 	if (max_linkhdr + hdrlen > MCLBYTES)
999 #else
1000 	if (max_linkhdr + hdrlen > MHLEN)
1001 #endif
1002 		panic("tcphdr too big");
1003 /*#endif*/
1004 
1005 	/*
1006 	 * This KASSERT is here to catch edge cases at a well defined place.
1007 	 * Before, those had triggered (random) panic conditions further down.
1008 	 */
1009 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1010 
1011 	/*
1012 	 * Grab a header mbuf, attaching a copy of data to
1013 	 * be transmitted, and initialize the header from
1014 	 * the template for sends on this connection.
1015 	 */
1016 	if (len) {
1017 		struct mbuf *mb;
1018 		struct sockbuf *msb;
1019 		u_int moff;
1020 
1021 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1022 			TCPSTAT_INC(tcps_sndprobe);
1023 #ifdef STATS
1024 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1025 				stats_voi_update_abs_u32(tp->t_stats,
1026 				VOI_TCP_RETXPB, len);
1027 			else
1028 				stats_voi_update_abs_u64(tp->t_stats,
1029 				    VOI_TCP_TXPB, len);
1030 #endif /* STATS */
1031 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1032 			tp->t_sndrexmitpack++;
1033 			TCPSTAT_INC(tcps_sndrexmitpack);
1034 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1035 			if (sack_rxmit) {
1036 				TCPSTAT_INC(tcps_sack_rexmits);
1037 				if (tso) {
1038 					TCPSTAT_INC(tcps_sack_rexmits_tso);
1039 				}
1040 				TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1041 			}
1042 #ifdef STATS
1043 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1044 			    len);
1045 #endif /* STATS */
1046 		} else {
1047 			TCPSTAT_INC(tcps_sndpack);
1048 			TCPSTAT_ADD(tcps_sndbyte, len);
1049 #ifdef STATS
1050 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1051 			    len);
1052 #endif /* STATS */
1053 		}
1054 #ifdef INET6
1055 		if (MHLEN < hdrlen + max_linkhdr)
1056 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1057 		else
1058 #endif
1059 			m = m_gethdr(M_NOWAIT, MT_DATA);
1060 
1061 		if (m == NULL) {
1062 			SOCKBUF_UNLOCK(&so->so_snd);
1063 			error = ENOBUFS;
1064 			sack_rxmit = 0;
1065 			goto out;
1066 		}
1067 
1068 		m->m_data += max_linkhdr;
1069 		m->m_len = hdrlen;
1070 
1071 		/*
1072 		 * Start the m_copy functions from the closest mbuf
1073 		 * to the offset in the socket buffer chain.
1074 		 */
1075 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1076 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1077 			m_copydata(mb, moff, len,
1078 			    mtod(m, caddr_t) + hdrlen);
1079 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1080 				sbsndptr_adv(&so->so_snd, mb, len);
1081 			m->m_len += len;
1082 		} else {
1083 			int32_t old_len;
1084 
1085 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1086 				msb = NULL;
1087 			else
1088 				msb = &so->so_snd;
1089 			old_len = len;
1090 			m->m_next = tcp_m_copym(mb, moff,
1091 			    &len, if_hw_tsomaxsegcount,
1092 			    if_hw_tsomaxsegsize, msb, hw_tls);
1093 			if (old_len != len)
1094 				flags &= ~TH_FIN;
1095 			if (len <= (tp->t_maxseg - optlen)) {
1096 				/*
1097 				 * Must have ran out of mbufs for the copy
1098 				 * shorten it to no longer need tso. Lets
1099 				 * not put on sendalot since we are low on
1100 				 * mbufs.
1101 				 */
1102 				tso = 0;
1103 			}
1104 			if (m->m_next == NULL) {
1105 				SOCKBUF_UNLOCK(&so->so_snd);
1106 				(void) m_free(m);
1107 				error = ENOBUFS;
1108 				sack_rxmit = 0;
1109 				goto out;
1110 			}
1111 		}
1112 
1113 		/*
1114 		 * If we're sending everything we've got, set PUSH.
1115 		 * (This will keep happy those implementations which only
1116 		 * give data to the user when a buffer fills or
1117 		 * a PUSH comes in.)
1118 		 */
1119 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1120 		    !(flags & TH_SYN))
1121 			flags |= TH_PUSH;
1122 		SOCKBUF_UNLOCK(&so->so_snd);
1123 	} else {
1124 		SOCKBUF_UNLOCK(&so->so_snd);
1125 		if (tp->t_flags & TF_ACKNOW)
1126 			TCPSTAT_INC(tcps_sndacks);
1127 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1128 			TCPSTAT_INC(tcps_sndctrl);
1129 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1130 			TCPSTAT_INC(tcps_sndurg);
1131 		else
1132 			TCPSTAT_INC(tcps_sndwinup);
1133 
1134 		m = m_gethdr(M_NOWAIT, MT_DATA);
1135 		if (m == NULL) {
1136 			error = ENOBUFS;
1137 			sack_rxmit = 0;
1138 			goto out;
1139 		}
1140 #ifdef INET6
1141 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1142 		    MHLEN >= hdrlen) {
1143 			M_ALIGN(m, hdrlen);
1144 		} else
1145 #endif
1146 		m->m_data += max_linkhdr;
1147 		m->m_len = hdrlen;
1148 	}
1149 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1150 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1151 #ifdef MAC
1152 	mac_inpcb_create_mbuf(inp, m);
1153 #endif
1154 #ifdef INET6
1155 	if (isipv6) {
1156 		ip6 = mtod(m, struct ip6_hdr *);
1157 		if (tp->t_port) {
1158 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1159 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1160 			udp->uh_dport = tp->t_port;
1161 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1162 			udp->uh_ulen = htons(ulen);
1163 			th = (struct tcphdr *)(udp + 1);
1164 		} else {
1165 			th = (struct tcphdr *)(ip6 + 1);
1166 		}
1167 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
1168 	} else
1169 #endif /* INET6 */
1170 	{
1171 		ip = mtod(m, struct ip *);
1172 		if (tp->t_port) {
1173 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1174 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1175 			udp->uh_dport = tp->t_port;
1176 			ulen = hdrlen + len - sizeof(struct ip);
1177 			udp->uh_ulen = htons(ulen);
1178 			th = (struct tcphdr *)(udp + 1);
1179 		} else
1180 			th = (struct tcphdr *)(ip + 1);
1181 		tcpip_fillheaders(inp, tp->t_port, ip, th);
1182 	}
1183 
1184 	/*
1185 	 * Fill in fields, remembering maximum advertised
1186 	 * window for use in delaying messages about window sizes.
1187 	 * If resending a FIN, be sure not to use a new sequence number.
1188 	 */
1189 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1190 	    tp->snd_nxt == tp->snd_max)
1191 		tp->snd_nxt--;
1192 	/*
1193 	 * If we are starting a connection, send ECN setup
1194 	 * SYN packet. If we are on a retransmit, we may
1195 	 * resend those bits a number of times as per
1196 	 * RFC 3168.
1197 	 */
1198 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1199 		flags |= tcp_ecn_output_syn_sent(tp);
1200 	}
1201 	/* Also handle parallel SYN for ECN */
1202 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1203 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1204 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1205 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1206 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1207 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1208 #ifdef INET6
1209 		if (isipv6) {
1210 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1211 			ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1212 		}
1213 		else
1214 #endif
1215 		{
1216 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1217 			ip->ip_tos |= ect;
1218 		}
1219 	}
1220 
1221 	/*
1222 	 * If we are doing retransmissions, then snd_nxt will
1223 	 * not reflect the first unsent octet.  For ACK only
1224 	 * packets, we do not want the sequence number of the
1225 	 * retransmitted packet, we want the sequence number
1226 	 * of the next unsent octet.  So, if there is no data
1227 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1228 	 * when filling in ti_seq.  But if we are in persist
1229 	 * state, snd_max might reflect one byte beyond the
1230 	 * right edge of the window, so use snd_nxt in that
1231 	 * case, since we know we aren't doing a retransmission.
1232 	 * (retransmit and persist are mutually exclusive...)
1233 	 */
1234 	if (sack_rxmit == 0) {
1235 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1236 		    tcp_timer_active(tp, TT_PERSIST))
1237 			th->th_seq = htonl(tp->snd_nxt);
1238 		else
1239 			th->th_seq = htonl(tp->snd_max);
1240 	} else {
1241 		th->th_seq = htonl(p->rxmit);
1242 		p->rxmit += len;
1243 		/*
1244 		 * Lost Retransmission Detection
1245 		 * trigger resending of a (then
1246 		 * still existing) hole, when
1247 		 * fack acks recoverypoint.
1248 		 */
1249 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1250 			p->rxmit = tp->snd_recover;
1251 		tp->sackhint.sack_bytes_rexmit += len;
1252 	}
1253 	if (IN_RECOVERY(tp->t_flags)) {
1254 		/*
1255 		 * Account all bytes transmitted while
1256 		 * IN_RECOVERY, simplifying PRR and
1257 		 * Lost Retransmit Detection
1258 		 */
1259 		tp->sackhint.prr_out += len;
1260 	}
1261 	th->th_ack = htonl(tp->rcv_nxt);
1262 	if (optlen) {
1263 		bcopy(opt, th + 1, optlen);
1264 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1265 	}
1266 	tcp_set_flags(th, flags);
1267 	/*
1268 	 * Calculate receive window.  Don't shrink window,
1269 	 * but avoid silly window syndrome.
1270 	 * If a RST segment is sent, advertise a window of zero.
1271 	 */
1272 	if (flags & TH_RST) {
1273 		recwin = 0;
1274 	} else {
1275 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1276 		    recwin < tp->t_maxseg)
1277 			recwin = 0;
1278 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1279 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1280 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1281 	}
1282 	/*
1283 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1284 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1285 	 * case is handled in syncache.
1286 	 */
1287 	if (flags & TH_SYN)
1288 		th->th_win = htons((u_short)
1289 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1290 	else {
1291 		/* Avoid shrinking window with window scaling. */
1292 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1293 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1294 	}
1295 
1296 	/*
1297 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1298 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1299 	 * flag tells soreceive() to disable delayed acknowledgements when
1300 	 * draining the buffer.  This can occur if the receiver is attempting
1301 	 * to read more data than can be buffered prior to transmitting on
1302 	 * the connection.
1303 	 */
1304 	if (th->th_win == 0) {
1305 		tp->t_sndzerowin++;
1306 		tp->t_flags |= TF_RXWIN0SENT;
1307 	} else
1308 		tp->t_flags &= ~TF_RXWIN0SENT;
1309 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1310 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1311 		th->th_flags |= TH_URG;
1312 	} else
1313 		/*
1314 		 * If no urgent pointer to send, then we pull
1315 		 * the urgent pointer to the left edge of the send window
1316 		 * so that it doesn't drift into the send window on sequence
1317 		 * number wraparound.
1318 		 */
1319 		tp->snd_up = tp->snd_una;		/* drag it along */
1320 
1321 	/*
1322 	 * Put TCP length in extended header, and then
1323 	 * checksum extended header and data.
1324 	 */
1325 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1326 
1327 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1328 	if (to.to_flags & TOF_SIGNATURE) {
1329 		/*
1330 		 * Calculate MD5 signature and put it into the place
1331 		 * determined before.
1332 		 * NOTE: since TCP options buffer doesn't point into
1333 		 * mbuf's data, calculate offset and use it.
1334 		 */
1335 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1336 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1337 			/*
1338 			 * Do not send segment if the calculation of MD5
1339 			 * digest has failed.
1340 			 */
1341 			m_freem(m);
1342 			goto out;
1343 		}
1344 	}
1345 #endif
1346 #ifdef INET6
1347 	if (isipv6) {
1348 		/*
1349 		 * There is no need to fill in ip6_plen right now.
1350 		 * It will be filled later by ip6_output.
1351 		 */
1352 		if (tp->t_port) {
1353 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1354 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1355 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1356 			th->th_sum = htons(0);
1357 			UDPSTAT_INC(udps_opackets);
1358 		} else {
1359 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1360 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1361 			th->th_sum = in6_cksum_pseudo(ip6,
1362 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1363 			    0);
1364 		}
1365 	}
1366 #endif
1367 #if defined(INET6) && defined(INET)
1368 	else
1369 #endif
1370 #ifdef INET
1371 	{
1372 		if (tp->t_port) {
1373 			m->m_pkthdr.csum_flags = CSUM_UDP;
1374 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1375 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1376 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1377 			th->th_sum = htons(0);
1378 			UDPSTAT_INC(udps_opackets);
1379 		} else {
1380 			m->m_pkthdr.csum_flags = CSUM_TCP;
1381 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1382 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1383 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1384 			    IPPROTO_TCP + len + optlen));
1385 		}
1386 
1387 		/* IP version must be set here for ipv4/ipv6 checking later */
1388 		KASSERT(ip->ip_v == IPVERSION,
1389 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1390 	}
1391 #endif
1392 
1393 	/*
1394 	 * Enable TSO and specify the size of the segments.
1395 	 * The TCP pseudo header checksum is always provided.
1396 	 */
1397 	if (tso) {
1398 		KASSERT(len > tp->t_maxseg - optlen - ipsec_optlen,
1399 		    ("%s: len <= tso_segsz", __func__));
1400 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1401 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen - ipsec_optlen;
1402 	}
1403 
1404 	KASSERT(len + hdrlen == m_length(m, NULL),
1405 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1406 	    __func__, len, hdrlen, m_length(m, NULL)));
1407 
1408 #ifdef TCP_HHOOK
1409 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1410 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1411 #endif
1412 
1413 	TCP_PROBE3(debug__output, tp, th, m);
1414 
1415 	/* We're getting ready to send; log now. */
1416 	/* XXXMT: We are not honoring verbose logging. */
1417 
1418 	if (tcp_bblogging_on(tp))
1419 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1420 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1421 		    NULL);
1422 	else
1423 		lgb = NULL;
1424 
1425 	/*
1426 	 * Fill in IP length and desired time to live and
1427 	 * send to IP level.  There should be a better way
1428 	 * to handle ttl and tos; we could keep them in
1429 	 * the template, but need a way to checksum without them.
1430 	 */
1431 	/*
1432 	 * m->m_pkthdr.len should have been set before checksum calculation,
1433 	 * because in6_cksum() need it.
1434 	 */
1435 #ifdef INET6
1436 	if (isipv6) {
1437 		/*
1438 		 * we separately set hoplimit for every segment, since the
1439 		 * user might want to change the value via setsockopt.
1440 		 * Also, desired default hop limit might be changed via
1441 		 * Neighbor Discovery.
1442 		 */
1443 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1444 
1445 		/*
1446 		 * Set the packet size here for the benefit of DTrace probes.
1447 		 * ip6_output() will set it properly; it's supposed to include
1448 		 * the option header lengths as well.
1449 		 */
1450 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1451 
1452 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1453 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1454 		else
1455 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1456 
1457 		if (tp->t_state == TCPS_SYN_SENT)
1458 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1459 
1460 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1461 
1462 #ifdef TCPPCAP
1463 		/* Save packet, if requested. */
1464 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1465 #endif
1466 
1467 		/* TODO: IPv6 IP6TOS_ECT bit on */
1468 		error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1469 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1470 		    NULL, NULL, inp);
1471 
1472 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1473 			mtu = inp->inp_route6.ro_nh->nh_mtu;
1474 	}
1475 #endif /* INET6 */
1476 #if defined(INET) && defined(INET6)
1477 	else
1478 #endif
1479 #ifdef INET
1480     {
1481 	ip->ip_len = htons(m->m_pkthdr.len);
1482 #ifdef INET6
1483 	if (inp->inp_vflag & INP_IPV6PROTO)
1484 		ip->ip_ttl = in6_selecthlim(inp, NULL);
1485 #endif /* INET6 */
1486 	/*
1487 	 * If we do path MTU discovery, then we set DF on every packet.
1488 	 * This might not be the best thing to do according to RFC3390
1489 	 * Section 2. However the tcp hostcache migitates the problem
1490 	 * so it affects only the first tcp connection with a host.
1491 	 *
1492 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1493 	 */
1494 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1495 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1496 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1497 			ip->ip_off |= htons(IP_DF);
1498 		}
1499 	} else {
1500 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1501 	}
1502 
1503 	if (tp->t_state == TCPS_SYN_SENT)
1504 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1505 
1506 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1507 
1508 #ifdef TCPPCAP
1509 	/* Save packet, if requested. */
1510 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1511 #endif
1512 
1513 	error = ip_output(m, inp->inp_options, &inp->inp_route,
1514 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1515 
1516 	if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1517 		mtu = inp->inp_route.ro_nh->nh_mtu;
1518     }
1519 #endif /* INET */
1520 
1521 	if (lgb != NULL) {
1522 		lgb->tlb_errno = error;
1523 		lgb = NULL;
1524 	}
1525 out:
1526 	if (error == 0)
1527 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1528 	/*
1529 	 * In transmit state, time the transmission and arrange for
1530 	 * the retransmit.  In persist state, just set snd_max.  In a closed
1531 	 * state just return.
1532 	 */
1533 	if (flags & TH_RST) {
1534 		TCPSTAT_INC(tcps_sndtotal);
1535 		return (0);
1536 	} else if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1537 	    !tcp_timer_active(tp, TT_PERSIST)) {
1538 		tcp_seq startseq = tp->snd_nxt;
1539 
1540 		/*
1541 		 * Advance snd_nxt over sequence space of this segment.
1542 		 */
1543 		if (flags & (TH_SYN|TH_FIN)) {
1544 			if (flags & TH_SYN)
1545 				tp->snd_nxt++;
1546 			if (flags & TH_FIN) {
1547 				tp->snd_nxt++;
1548 				tp->t_flags |= TF_SENTFIN;
1549 			}
1550 		}
1551 		if (sack_rxmit)
1552 			goto timer;
1553 		tp->snd_nxt += len;
1554 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1555 			/*
1556 			 * Update "made progress" indication if we just
1557 			 * added new data to an empty socket buffer.
1558 			 */
1559 			if (tp->snd_una == tp->snd_max)
1560 				tp->t_acktime = ticks;
1561 			tp->snd_max = tp->snd_nxt;
1562 			/*
1563 			 * Time this transmission if not a retransmission and
1564 			 * not currently timing anything.
1565 			 */
1566 			tp->t_sndtime = ticks;
1567 			if (tp->t_rtttime == 0) {
1568 				tp->t_rtttime = ticks;
1569 				tp->t_rtseq = startseq;
1570 				TCPSTAT_INC(tcps_segstimed);
1571 			}
1572 #ifdef STATS
1573 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1574 				tp->t_flags |= TF_GPUTINPROG;
1575 				tp->gput_seq = startseq;
1576 				tp->gput_ack = startseq +
1577 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1578 				tp->gput_ts = tcp_ts_getticks();
1579 			}
1580 #endif /* STATS */
1581 		}
1582 
1583 		/*
1584 		 * Set retransmit timer if not currently set,
1585 		 * and not doing a pure ack or a keep-alive probe.
1586 		 * Initial value for retransmit timer is smoothed
1587 		 * round-trip time + 2 * round-trip time variance.
1588 		 * Initialize shift counter which is used for backoff
1589 		 * of retransmit time.
1590 		 */
1591 timer:
1592 		if (!tcp_timer_active(tp, TT_REXMT) &&
1593 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1594 		     (tp->snd_nxt != tp->snd_una))) {
1595 			if (tcp_timer_active(tp, TT_PERSIST)) {
1596 				tcp_timer_activate(tp, TT_PERSIST, 0);
1597 				tp->t_rxtshift = 0;
1598 			}
1599 			tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1600 		} else if (len == 0 && sbavail(&so->so_snd) &&
1601 		    !tcp_timer_active(tp, TT_REXMT) &&
1602 		    !tcp_timer_active(tp, TT_PERSIST)) {
1603 			/*
1604 			 * Avoid a situation where we do not set persist timer
1605 			 * after a zero window condition. For example:
1606 			 * 1) A -> B: packet with enough data to fill the window
1607 			 * 2) B -> A: ACK for #1 + new data (0 window
1608 			 *    advertisement)
1609 			 * 3) A -> B: ACK for #2, 0 len packet
1610 			 *
1611 			 * In this case, A will not activate the persist timer,
1612 			 * because it chose to send a packet. Unless tcp_output
1613 			 * is called for some other reason (delayed ack timer,
1614 			 * another input packet from B, socket syscall), A will
1615 			 * not send zero window probes.
1616 			 *
1617 			 * So, if you send a 0-length packet, but there is data
1618 			 * in the socket buffer, and neither the rexmt or
1619 			 * persist timer is already set, then activate the
1620 			 * persist timer.
1621 			 */
1622 			tp->t_rxtshift = 0;
1623 			tcp_setpersist(tp);
1624 		}
1625 	} else {
1626 		/*
1627 		 * Persist case, update snd_max but since we are in
1628 		 * persist mode (no window) we do not update snd_nxt.
1629 		 */
1630 		int xlen = len;
1631 		if (flags & TH_SYN)
1632 			++xlen;
1633 		if (flags & TH_FIN) {
1634 			++xlen;
1635 			tp->t_flags |= TF_SENTFIN;
1636 		}
1637 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1638 			tp->snd_max = tp->snd_nxt + xlen;
1639 	}
1640 	if ((error == 0) &&
1641 	    (tp->rcv_numsacks > 0) &&
1642 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
1643 	    (tp->t_flags & TF_SACK_PERMIT)) {
1644 		/* Clean up any DSACK's sent */
1645 		tcp_clean_dsack_blocks(tp);
1646 	}
1647 	if ((error == 0) &&
1648 	    sack_rxmit &&
1649 	    SEQ_LT(tp->snd_nxt, SEQ_MIN(p->rxmit, p->end))) {
1650 		tp->snd_nxt = SEQ_MIN(p->rxmit, p->end);
1651 	}
1652 	if (error) {
1653 		/*
1654 		 * We know that the packet was lost, so back out the
1655 		 * sequence number advance, if any.
1656 		 *
1657 		 * If the error is EPERM the packet got blocked by the
1658 		 * local firewall.  Normally we should terminate the
1659 		 * connection but the blocking may have been spurious
1660 		 * due to a firewall reconfiguration cycle.  So we treat
1661 		 * it like a packet loss and let the retransmit timer and
1662 		 * timeouts do their work over time.
1663 		 * XXX: It is a POLA question whether calling tcp_drop right
1664 		 * away would be the really correct behavior instead.
1665 		 */
1666 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1667 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1668 		    ((flags & TH_SYN) == 0) &&
1669 		    (error != EPERM)) {
1670 			if (sack_rxmit) {
1671 				p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1672 				tp->sackhint.sack_bytes_rexmit -= len;
1673 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1674 				    ("sackhint bytes rtx >= 0"));
1675 				KASSERT((flags & TH_FIN) == 0,
1676 				    ("error while FIN with SACK rxmit"));
1677 			} else {
1678 				tp->snd_nxt -= len;
1679 				if (flags & TH_FIN)
1680 					tp->snd_nxt--;
1681 			}
1682 			if (IN_RECOVERY(tp->t_flags))
1683 				tp->sackhint.prr_out -= len;
1684 		}
1685 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1686 		switch (error) {
1687 		case EACCES:
1688 		case EPERM:
1689 			tp->t_softerror = error;
1690 			return (error);
1691 		case ENOBUFS:
1692 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1693 			tp->snd_cwnd = tp->t_maxseg;
1694 			return (0);
1695 		case EMSGSIZE:
1696 			/*
1697 			 * For some reason the interface we used initially
1698 			 * to send segments changed to another or lowered
1699 			 * its MTU.
1700 			 * If TSO was active we either got an interface
1701 			 * without TSO capabilits or TSO was turned off.
1702 			 * If we obtained mtu from ip_output() then update
1703 			 * it and try again.
1704 			 */
1705 			if (tso)
1706 				tp->t_flags &= ~TF_TSO;
1707 			if (mtu != 0) {
1708 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1709 				goto again;
1710 			}
1711 			return (error);
1712 		case EHOSTDOWN:
1713 		case EHOSTUNREACH:
1714 		case ENETDOWN:
1715 		case ENETUNREACH:
1716 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1717 				tp->t_softerror = error;
1718 				return (0);
1719 			}
1720 			/* FALLTHROUGH */
1721 		default:
1722 			return (error);
1723 		}
1724 	}
1725 	TCPSTAT_INC(tcps_sndtotal);
1726 
1727 	/*
1728 	 * Data sent (as far as we can tell).
1729 	 * If this advertises a larger window than any other segment,
1730 	 * then remember the size of the advertised window.
1731 	 * Any pending ACK has now been sent.
1732 	 */
1733 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1734 		tp->rcv_adv = tp->rcv_nxt + recwin;
1735 	tp->last_ack_sent = tp->rcv_nxt;
1736 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1737 	if (tcp_timer_active(tp, TT_DELACK))
1738 		tcp_timer_activate(tp, TT_DELACK, 0);
1739 	if (sendalot)
1740 		goto again;
1741 	return (0);
1742 }
1743 
1744 void
1745 tcp_setpersist(struct tcpcb *tp)
1746 {
1747 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1748 	int tt;
1749 	int maxunacktime;
1750 
1751 	tp->t_flags &= ~TF_PREVVALID;
1752 	if (tcp_timer_active(tp, TT_REXMT))
1753 		panic("tcp_setpersist: retransmit pending");
1754 	/*
1755 	 * If the state is already closed, don't bother.
1756 	 */
1757 	if (tp->t_state == TCPS_CLOSED)
1758 		return;
1759 
1760 	/*
1761 	 * Start/restart persistence timer.
1762 	 */
1763 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1764 		      tcp_persmin, tcp_persmax);
1765 	if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1766 		maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1767 		if (maxunacktime < 1)
1768 			maxunacktime = 1;
1769 		if (maxunacktime < tt)
1770 			tt = maxunacktime;
1771 	}
1772 	tcp_timer_activate(tp, TT_PERSIST, tt);
1773 	if (tp->t_rxtshift < V_tcp_retries)
1774 		tp->t_rxtshift++;
1775 }
1776 
1777 /*
1778  * Insert TCP options according to the supplied parameters to the place
1779  * optp in a consistent way.  Can handle unaligned destinations.
1780  *
1781  * The order of the option processing is crucial for optimal packing and
1782  * alignment for the scarce option space.
1783  *
1784  * The optimal order for a SYN/SYN-ACK segment is:
1785  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1786  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1787  *
1788  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1789  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1790  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1791  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1792  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1793  */
1794 int
1795 tcp_addoptions(struct tcpopt *to, u_char *optp)
1796 {
1797 	u_int32_t mask, optlen = 0;
1798 
1799 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1800 		if ((to->to_flags & mask) != mask)
1801 			continue;
1802 		if (optlen == TCP_MAXOLEN)
1803 			break;
1804 		switch (to->to_flags & mask) {
1805 		case TOF_MSS:
1806 			while (optlen % 4) {
1807 				optlen += TCPOLEN_NOP;
1808 				*optp++ = TCPOPT_NOP;
1809 			}
1810 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1811 				continue;
1812 			optlen += TCPOLEN_MAXSEG;
1813 			*optp++ = TCPOPT_MAXSEG;
1814 			*optp++ = TCPOLEN_MAXSEG;
1815 			to->to_mss = htons(to->to_mss);
1816 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1817 			optp += sizeof(to->to_mss);
1818 			break;
1819 		case TOF_SCALE:
1820 			while (!optlen || optlen % 2 != 1) {
1821 				optlen += TCPOLEN_NOP;
1822 				*optp++ = TCPOPT_NOP;
1823 			}
1824 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1825 				continue;
1826 			optlen += TCPOLEN_WINDOW;
1827 			*optp++ = TCPOPT_WINDOW;
1828 			*optp++ = TCPOLEN_WINDOW;
1829 			*optp++ = to->to_wscale;
1830 			break;
1831 		case TOF_SACKPERM:
1832 			while (optlen % 2) {
1833 				optlen += TCPOLEN_NOP;
1834 				*optp++ = TCPOPT_NOP;
1835 			}
1836 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1837 				continue;
1838 			optlen += TCPOLEN_SACK_PERMITTED;
1839 			*optp++ = TCPOPT_SACK_PERMITTED;
1840 			*optp++ = TCPOLEN_SACK_PERMITTED;
1841 			break;
1842 		case TOF_TS:
1843 			while (!optlen || optlen % 4 != 2) {
1844 				optlen += TCPOLEN_NOP;
1845 				*optp++ = TCPOPT_NOP;
1846 			}
1847 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1848 				continue;
1849 			optlen += TCPOLEN_TIMESTAMP;
1850 			*optp++ = TCPOPT_TIMESTAMP;
1851 			*optp++ = TCPOLEN_TIMESTAMP;
1852 			to->to_tsval = htonl(to->to_tsval);
1853 			to->to_tsecr = htonl(to->to_tsecr);
1854 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1855 			optp += sizeof(to->to_tsval);
1856 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1857 			optp += sizeof(to->to_tsecr);
1858 			break;
1859 		case TOF_SIGNATURE:
1860 			{
1861 			int siglen = TCPOLEN_SIGNATURE - 2;
1862 
1863 			while (!optlen || optlen % 4 != 2) {
1864 				optlen += TCPOLEN_NOP;
1865 				*optp++ = TCPOPT_NOP;
1866 			}
1867 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1868 				to->to_flags &= ~TOF_SIGNATURE;
1869 				continue;
1870 			}
1871 			optlen += TCPOLEN_SIGNATURE;
1872 			*optp++ = TCPOPT_SIGNATURE;
1873 			*optp++ = TCPOLEN_SIGNATURE;
1874 			to->to_signature = optp;
1875 			while (siglen--)
1876 				 *optp++ = 0;
1877 			break;
1878 			}
1879 		case TOF_SACK:
1880 			{
1881 			int sackblks = 0;
1882 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1883 			tcp_seq sack_seq;
1884 
1885 			while (!optlen || optlen % 4 != 2) {
1886 				optlen += TCPOLEN_NOP;
1887 				*optp++ = TCPOPT_NOP;
1888 			}
1889 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1890 				continue;
1891 			optlen += TCPOLEN_SACKHDR;
1892 			*optp++ = TCPOPT_SACK;
1893 			sackblks = min(to->to_nsacks,
1894 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1895 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1896 			while (sackblks--) {
1897 				sack_seq = htonl(sack->start);
1898 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1899 				optp += sizeof(sack_seq);
1900 				sack_seq = htonl(sack->end);
1901 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1902 				optp += sizeof(sack_seq);
1903 				optlen += TCPOLEN_SACK;
1904 				sack++;
1905 			}
1906 			TCPSTAT_INC(tcps_sack_send_blocks);
1907 			break;
1908 			}
1909 		case TOF_FASTOPEN:
1910 			{
1911 			int total_len;
1912 
1913 			/* XXX is there any point to aligning this option? */
1914 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1915 			if (TCP_MAXOLEN - optlen < total_len) {
1916 				to->to_flags &= ~TOF_FASTOPEN;
1917 				continue;
1918 			}
1919 			*optp++ = TCPOPT_FAST_OPEN;
1920 			*optp++ = total_len;
1921 			if (to->to_tfo_len > 0) {
1922 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1923 				optp += to->to_tfo_len;
1924 			}
1925 			optlen += total_len;
1926 			break;
1927 			}
1928 		default:
1929 			panic("%s: unknown TCP option type", __func__);
1930 			break;
1931 		}
1932 	}
1933 
1934 	/* Terminate and pad TCP options to a 4 byte boundary. */
1935 	if (optlen % 4) {
1936 		optlen += TCPOLEN_EOL;
1937 		*optp++ = TCPOPT_EOL;
1938 	}
1939 	/*
1940 	 * According to RFC 793 (STD0007):
1941 	 *   "The content of the header beyond the End-of-Option option
1942 	 *    must be header padding (i.e., zero)."
1943 	 *   and later: "The padding is composed of zeros."
1944 	 */
1945 	while (optlen % 4) {
1946 		optlen += TCPOLEN_PAD;
1947 		*optp++ = TCPOPT_PAD;
1948 	}
1949 
1950 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1951 	return (optlen);
1952 }
1953 
1954 /*
1955  * This is a copy of m_copym(), taking the TSO segment size/limit
1956  * constraints into account, and advancing the sndptr as it goes.
1957  */
1958 struct mbuf *
1959 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1960     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1961 {
1962 #ifdef KERN_TLS
1963 	struct ktls_session *tls, *ntls;
1964 	struct mbuf *start __diagused;
1965 #endif
1966 	struct mbuf *n, **np;
1967 	struct mbuf *top;
1968 	int32_t off = off0;
1969 	int32_t len = *plen;
1970 	int32_t fragsize;
1971 	int32_t len_cp = 0;
1972 	int32_t *pkthdrlen;
1973 	uint32_t mlen, frags;
1974 	bool copyhdr;
1975 
1976 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1977 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1978 	if (off == 0 && m->m_flags & M_PKTHDR)
1979 		copyhdr = true;
1980 	else
1981 		copyhdr = false;
1982 	while (off > 0) {
1983 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1984 		if (off < m->m_len)
1985 			break;
1986 		off -= m->m_len;
1987 		if ((sb) && (m == sb->sb_sndptr)) {
1988 			sb->sb_sndptroff += m->m_len;
1989 			sb->sb_sndptr = m->m_next;
1990 		}
1991 		m = m->m_next;
1992 	}
1993 	np = &top;
1994 	top = NULL;
1995 	pkthdrlen = NULL;
1996 #ifdef KERN_TLS
1997 	if (hw_tls && (m->m_flags & M_EXTPG))
1998 		tls = m->m_epg_tls;
1999 	else
2000 		tls = NULL;
2001 	start = m;
2002 #endif
2003 	while (len > 0) {
2004 		if (m == NULL) {
2005 			KASSERT(len == M_COPYALL,
2006 			    ("tcp_m_copym, length > size of mbuf chain"));
2007 			*plen = len_cp;
2008 			if (pkthdrlen != NULL)
2009 				*pkthdrlen = len_cp;
2010 			break;
2011 		}
2012 #ifdef KERN_TLS
2013 		if (hw_tls) {
2014 			if (m->m_flags & M_EXTPG)
2015 				ntls = m->m_epg_tls;
2016 			else
2017 				ntls = NULL;
2018 
2019 			/*
2020 			 * Avoid mixing TLS records with handshake
2021 			 * data or TLS records from different
2022 			 * sessions.
2023 			 */
2024 			if (tls != ntls) {
2025 				MPASS(m != start);
2026 				*plen = len_cp;
2027 				if (pkthdrlen != NULL)
2028 					*pkthdrlen = len_cp;
2029 				break;
2030 			}
2031 		}
2032 #endif
2033 		mlen = min(len, m->m_len - off);
2034 		if (seglimit) {
2035 			/*
2036 			 * For M_EXTPG mbufs, add 3 segments
2037 			 * + 1 in case we are crossing page boundaries
2038 			 * + 2 in case the TLS hdr/trailer are used
2039 			 * It is cheaper to just add the segments
2040 			 * than it is to take the cache miss to look
2041 			 * at the mbuf ext_pgs state in detail.
2042 			 */
2043 			if (m->m_flags & M_EXTPG) {
2044 				fragsize = min(segsize, PAGE_SIZE);
2045 				frags = 3;
2046 			} else {
2047 				fragsize = segsize;
2048 				frags = 0;
2049 			}
2050 
2051 			/* Break if we really can't fit anymore. */
2052 			if ((frags + 1) >= seglimit) {
2053 				*plen =	len_cp;
2054 				if (pkthdrlen != NULL)
2055 					*pkthdrlen = len_cp;
2056 				break;
2057 			}
2058 
2059 			/*
2060 			 * Reduce size if you can't copy the whole
2061 			 * mbuf. If we can't copy the whole mbuf, also
2062 			 * adjust len so the loop will end after this
2063 			 * mbuf.
2064 			 */
2065 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2066 				mlen = (seglimit - frags - 1) * fragsize;
2067 				len = mlen;
2068 				*plen = len_cp + len;
2069 				if (pkthdrlen != NULL)
2070 					*pkthdrlen = *plen;
2071 			}
2072 			frags += howmany(mlen, fragsize);
2073 			if (frags == 0)
2074 				frags++;
2075 			seglimit -= frags;
2076 			KASSERT(seglimit > 0,
2077 			    ("%s: seglimit went too low", __func__));
2078 		}
2079 		if (copyhdr)
2080 			n = m_gethdr(M_NOWAIT, m->m_type);
2081 		else
2082 			n = m_get(M_NOWAIT, m->m_type);
2083 		*np = n;
2084 		if (n == NULL)
2085 			goto nospace;
2086 		if (copyhdr) {
2087 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2088 				goto nospace;
2089 			if (len == M_COPYALL)
2090 				n->m_pkthdr.len -= off0;
2091 			else
2092 				n->m_pkthdr.len = len;
2093 			pkthdrlen = &n->m_pkthdr.len;
2094 			copyhdr = false;
2095 		}
2096 		n->m_len = mlen;
2097 		len_cp += n->m_len;
2098 		if (m->m_flags & (M_EXT | M_EXTPG)) {
2099 			n->m_data = m->m_data + off;
2100 			mb_dupcl(n, m);
2101 		} else
2102 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2103 			    (u_int)n->m_len);
2104 
2105 		if (sb && (sb->sb_sndptr == m) &&
2106 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2107 			sb->sb_sndptroff += m->m_len;
2108 			sb->sb_sndptr = m->m_next;
2109 		}
2110 		off = 0;
2111 		if (len != M_COPYALL) {
2112 			len -= n->m_len;
2113 		}
2114 		m = m->m_next;
2115 		np = &n->m_next;
2116 	}
2117 	return (top);
2118 nospace:
2119 	m_freem(top);
2120 	return (NULL);
2121 }
2122 
2123 void
2124 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2125 {
2126 
2127 	/*
2128 	 * Automatic sizing of send socket buffer.  Often the send buffer
2129 	 * size is not optimally adjusted to the actual network conditions
2130 	 * at hand (delay bandwidth product).  Setting the buffer size too
2131 	 * small limits throughput on links with high bandwidth and high
2132 	 * delay (eg. trans-continental/oceanic links).  Setting the
2133 	 * buffer size too big consumes too much real kernel memory,
2134 	 * especially with many connections on busy servers.
2135 	 *
2136 	 * The criteria to step up the send buffer one notch are:
2137 	 *  1. receive window of remote host is larger than send buffer
2138 	 *     (with a fudge factor of 5/4th);
2139 	 *  2. send buffer is filled to 7/8th with data (so we actually
2140 	 *     have data to make use of it);
2141 	 *  3. send buffer fill has not hit maximal automatic size;
2142 	 *  4. our send window (slow start and cogestion controlled) is
2143 	 *     larger than sent but unacknowledged data in send buffer.
2144 	 *
2145 	 * The remote host receive window scaling factor may limit the
2146 	 * growing of the send buffer before it reaches its allowed
2147 	 * maximum.
2148 	 *
2149 	 * It scales directly with slow start or congestion window
2150 	 * and does at most one step per received ACK.  This fast
2151 	 * scaling has the drawback of growing the send buffer beyond
2152 	 * what is strictly necessary to make full use of a given
2153 	 * delay*bandwidth product.  However testing has shown this not
2154 	 * to be much of an problem.  At worst we are trading wasting
2155 	 * of available bandwidth (the non-use of it) for wasting some
2156 	 * socket buffer memory.
2157 	 *
2158 	 * TODO: Shrink send buffer during idle periods together
2159 	 * with congestion window.  Requires another timer.  Has to
2160 	 * wait for upcoming tcp timer rewrite.
2161 	 *
2162 	 * XXXGL: should there be used sbused() or sbavail()?
2163 	 */
2164 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2165 		int lowat;
2166 
2167 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2168 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2169 		    sbused(&so->so_snd) >=
2170 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2171 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2172 		    sendwin >= (sbused(&so->so_snd) -
2173 		    (tp->snd_nxt - tp->snd_una))) {
2174 			if (!sbreserve_locked(so, SO_SND,
2175 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2176 			     V_tcp_autosndbuf_max), curthread))
2177 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2178 		}
2179 	}
2180 }
2181