xref: /freebsd/sys/netinet/tcp_output.c (revision 045c8f526484cb3b97f5fd693987f4376fa43c5f)
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 	unsigned int wanted_cookie = 0;
226 	unsigned int dont_sendalot = 0;
227 #if 0
228 	int maxburst = TCP_MAXBURST;
229 #endif
230 #ifdef INET6
231 	struct ip6_hdr *ip6 = NULL;
232 	int isipv6;
233 
234 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
235 #endif
236 #ifdef KERN_TLS
237 	const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
238 #else
239 	const bool hw_tls = false;
240 #endif
241 
242 	NET_EPOCH_ASSERT();
243 	INP_WLOCK_ASSERT(tp->t_inpcb);
244 
245 #ifdef TCP_OFFLOAD
246 	if (tp->t_flags & TF_TOE)
247 		return (tcp_offload_output(tp));
248 #endif
249 
250 	/*
251 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
252 	 * only allow the initial SYN or SYN|ACK and those sent
253 	 * by the retransmit timer.
254 	 */
255 	if (IS_FASTOPEN(tp->t_flags) &&
256 	    ((tp->t_state == TCPS_SYN_SENT) ||
257 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
258 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
259 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
260 		return (0);
261 
262 	/*
263 	 * Determine length of data that should be transmitted,
264 	 * and flags that will be used.
265 	 * If there is some data or critical controls (SYN, RST)
266 	 * to send, then transmit; otherwise, investigate further.
267 	 */
268 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
269 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
270 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
271 		cc_after_idle(tp);
272 	tp->t_flags &= ~TF_LASTIDLE;
273 	if (idle) {
274 		if (tp->t_flags & TF_MORETOCOME) {
275 			tp->t_flags |= TF_LASTIDLE;
276 			idle = 0;
277 		}
278 	}
279 again:
280 	/*
281 	 * If we've recently taken a timeout, snd_max will be greater than
282 	 * snd_nxt.  There may be SACK information that allows us to avoid
283 	 * resending already delivered data.  Adjust snd_nxt accordingly.
284 	 */
285 	if ((tp->t_flags & TF_SACK_PERMIT) &&
286 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
287 		tcp_sack_adjust(tp);
288 	sendalot = 0;
289 	tso = 0;
290 	mtu = 0;
291 	off = tp->snd_nxt - tp->snd_una;
292 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
293 
294 	flags = tcp_outflags[tp->t_state];
295 	/*
296 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
297 	 * to send out new data (when sendalot is 1), bypass this function.
298 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
299 	 * we're replacing a (future) new transmission with a retransmission
300 	 * now, and we previously incremented snd_cwnd in tcp_input().
301 	 */
302 	/*
303 	 * Still in sack recovery , reset rxmit flag to zero.
304 	 */
305 	sack_rxmit = 0;
306 	sack_bytes_rxmt = 0;
307 	len = 0;
308 	p = NULL;
309 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
310 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
311 		uint32_t cwin;
312 
313 		cwin =
314 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
315 		/* Do not retransmit SACK segments beyond snd_recover */
316 		if (SEQ_GT(p->end, tp->snd_recover)) {
317 			/*
318 			 * (At least) part of sack hole extends beyond
319 			 * snd_recover. Check to see if we can rexmit data
320 			 * for this hole.
321 			 */
322 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
323 				/*
324 				 * Can't rexmit any more data for this hole.
325 				 * That data will be rexmitted in the next
326 				 * sack recovery episode, when snd_recover
327 				 * moves past p->rxmit.
328 				 */
329 				p = NULL;
330 				goto after_sack_rexmit;
331 			} else
332 				/* Can rexmit part of the current hole */
333 				len = ((int32_t)ulmin(cwin,
334 						   tp->snd_recover - p->rxmit));
335 		} else
336 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
337 		off = p->rxmit - tp->snd_una;
338 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
339 		    __func__, off));
340 		if (len > 0) {
341 			sack_rxmit = 1;
342 			sendalot = 1;
343 			TCPSTAT_INC(tcps_sack_rexmits);
344 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
345 			    min(len, tcp_maxseg(tp)));
346 		}
347 	}
348 after_sack_rexmit:
349 	/*
350 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
351 	 * state flags.
352 	 */
353 	if (tp->t_flags & TF_NEEDFIN)
354 		flags |= TH_FIN;
355 	if (tp->t_flags & TF_NEEDSYN)
356 		flags |= TH_SYN;
357 
358 	SOCKBUF_LOCK(&so->so_snd);
359 	/*
360 	 * If in persist timeout with window of 0, send 1 byte.
361 	 * Otherwise, if window is small but nonzero
362 	 * and timer expired, we will send what we can
363 	 * and go to transmit state.
364 	 */
365 	if (tp->t_flags & TF_FORCEDATA) {
366 		if (sendwin == 0) {
367 			/*
368 			 * If we still have some data to send, then
369 			 * clear the FIN bit.  Usually this would
370 			 * happen below when it realizes that we
371 			 * aren't sending all the data.  However,
372 			 * if we have exactly 1 byte of unsent data,
373 			 * then it won't clear the FIN bit below,
374 			 * and if we are in persist state, we wind
375 			 * up sending the packet without recording
376 			 * that we sent the FIN bit.
377 			 *
378 			 * We can't just blindly clear the FIN bit,
379 			 * because if we don't have any more data
380 			 * to send then the probe will be the FIN
381 			 * itself.
382 			 */
383 			if (off < sbused(&so->so_snd))
384 				flags &= ~TH_FIN;
385 			sendwin = 1;
386 		} else {
387 			tcp_timer_activate(tp, TT_PERSIST, 0);
388 			tp->t_rxtshift = 0;
389 		}
390 	}
391 
392 	/*
393 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
394 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
395 	 * a negative length.  This can also occur when TCP opens up
396 	 * its congestion window while receiving additional duplicate
397 	 * acks after fast-retransmit because TCP will reset snd_nxt
398 	 * to snd_max after the fast-retransmit.
399 	 *
400 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
401 	 * be set to snd_una, the offset will be 0, and the length may
402 	 * wind up 0.
403 	 *
404 	 * If sack_rxmit is true we are retransmitting from the scoreboard
405 	 * in which case len is already set.
406 	 */
407 	if (sack_rxmit == 0) {
408 		if (sack_bytes_rxmt == 0)
409 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
410 			    off);
411 		else {
412 			int32_t cwin;
413 
414                         /*
415 			 * We are inside of a SACK recovery episode and are
416 			 * sending new data, having retransmitted all the
417 			 * data possible in the scoreboard.
418 			 */
419 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
420 			    off);
421 			/*
422 			 * Don't remove this (len > 0) check !
423 			 * We explicitly check for len > 0 here (although it
424 			 * isn't really necessary), to work around a gcc
425 			 * optimization issue - to force gcc to compute
426 			 * len above. Without this check, the computation
427 			 * of len is bungled by the optimizer.
428 			 */
429 			if (len > 0) {
430 				cwin = tp->snd_cwnd -
431 					(tp->snd_nxt - tp->snd_recover) -
432 					sack_bytes_rxmt;
433 				if (cwin < 0)
434 					cwin = 0;
435 				len = imin(len, cwin);
436 			}
437 		}
438 	}
439 
440 	/*
441 	 * Lop off SYN bit if it has already been sent.  However, if this
442 	 * is SYN-SENT state and if segment contains data and if we don't
443 	 * know that foreign host supports TAO, suppress sending segment.
444 	 */
445 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
446 		if (tp->t_state != TCPS_SYN_RECEIVED)
447 			flags &= ~TH_SYN;
448 		/*
449 		 * When sending additional segments following a TFO SYN|ACK,
450 		 * do not include the SYN bit.
451 		 */
452 		if (IS_FASTOPEN(tp->t_flags) &&
453 		    (tp->t_state == TCPS_SYN_RECEIVED))
454 			flags &= ~TH_SYN;
455 		off--, len++;
456 	}
457 
458 	/*
459 	 * Be careful not to send data and/or FIN on SYN segments.
460 	 * This measure is needed to prevent interoperability problems
461 	 * with not fully conformant TCP implementations.
462 	 */
463 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
464 		len = 0;
465 		flags &= ~TH_FIN;
466 	}
467 
468 	/*
469 	 * On TFO sockets, ensure no data is sent in the following cases:
470 	 *
471 	 *  - When retransmitting SYN|ACK on a passively-created socket
472 	 *
473 	 *  - When retransmitting SYN on an actively created socket
474 	 *
475 	 *  - When sending a zero-length cookie (cookie request) on an
476 	 *    actively created socket
477 	 *
478 	 *  - When the socket is in the CLOSED state (RST is being sent)
479 	 */
480 	if (IS_FASTOPEN(tp->t_flags) &&
481 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
482 	     ((tp->t_state == TCPS_SYN_SENT) &&
483 	      (tp->t_tfo_client_cookie_len == 0)) ||
484 	     (flags & TH_RST)))
485 		len = 0;
486 	if (len <= 0) {
487 		/*
488 		 * If FIN has been sent but not acked,
489 		 * but we haven't been called to retransmit,
490 		 * len will be < 0.  Otherwise, window shrank
491 		 * after we sent into it.  If window shrank to 0,
492 		 * cancel pending retransmit, pull snd_nxt back
493 		 * to (closed) window, and set the persist timer
494 		 * if it isn't already going.  If the window didn't
495 		 * close completely, just wait for an ACK.
496 		 *
497 		 * We also do a general check here to ensure that
498 		 * we will set the persist timer when we have data
499 		 * to send, but a 0-byte window. This makes sure
500 		 * the persist timer is set even if the packet
501 		 * hits one of the "goto send" lines below.
502 		 */
503 		len = 0;
504 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
505 			(off < (int) sbavail(&so->so_snd))) {
506 			tcp_timer_activate(tp, TT_REXMT, 0);
507 			tp->t_rxtshift = 0;
508 			tp->snd_nxt = tp->snd_una;
509 			if (!tcp_timer_active(tp, TT_PERSIST))
510 				tcp_setpersist(tp);
511 		}
512 	}
513 
514 	/* len will be >= 0 after this point. */
515 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
516 
517 	tcp_sndbuf_autoscale(tp, so, sendwin);
518 
519 	/*
520 	 * Decide if we can use TCP Segmentation Offloading (if supported by
521 	 * hardware).
522 	 *
523 	 * TSO may only be used if we are in a pure bulk sending state.  The
524 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
525 	 * IP options prevent using TSO.  With TSO the TCP header is the same
526 	 * (except for the sequence number) for all generated packets.  This
527 	 * makes it impossible to transmit any options which vary per generated
528 	 * segment or packet.
529 	 *
530 	 * IPv4 handling has a clear separation of ip options and ip header
531 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
532 	 * the right thing below to provide length of just ip options and thus
533 	 * checking for ipoptlen is enough to decide if ip options are present.
534 	 */
535 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
536 	/*
537 	 * Pre-calculate here as we save another lookup into the darknesses
538 	 * of IPsec that way and can actually decide if TSO is ok.
539 	 */
540 #ifdef INET6
541 	if (isipv6 && IPSEC_ENABLED(ipv6))
542 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
543 #ifdef INET
544 	else
545 #endif
546 #endif /* INET6 */
547 #ifdef INET
548 	if (IPSEC_ENABLED(ipv4))
549 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
550 #endif /* INET */
551 #endif /* IPSEC */
552 #ifdef INET6
553 	if (isipv6)
554 		ipoptlen = ip6_optlen(tp->t_inpcb);
555 	else
556 #endif
557 	if (tp->t_inpcb->inp_options)
558 		ipoptlen = tp->t_inpcb->inp_options->m_len -
559 				offsetof(struct ipoption, ipopt_list);
560 	else
561 		ipoptlen = 0;
562 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
563 	ipoptlen += ipsec_optlen;
564 #endif
565 
566 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
567 	    (tp->t_port == 0) &&
568 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
569 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
570 	    ipoptlen == 0 && !(flags & TH_SYN))
571 		tso = 1;
572 
573 	if (sack_rxmit) {
574 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
575 			flags &= ~TH_FIN;
576 	} else {
577 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
578 		    sbused(&so->so_snd)))
579 			flags &= ~TH_FIN;
580 	}
581 
582 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
583 	    (long)TCP_MAXWIN << tp->rcv_scale);
584 
585 	/*
586 	 * Sender silly window avoidance.   We transmit under the following
587 	 * conditions when len is non-zero:
588 	 *
589 	 *	- We have a full segment (or more with TSO)
590 	 *	- This is the last buffer in a write()/send() and we are
591 	 *	  either idle or running NODELAY
592 	 *	- we've timed out (e.g. persist timer)
593 	 *	- we have more then 1/2 the maximum send window's worth of
594 	 *	  data (receiver may be limited the window size)
595 	 *	- we need to retransmit
596 	 */
597 	if (len) {
598 		if (len >= tp->t_maxseg)
599 			goto send;
600 		/*
601 		 * As the TCP header options are now
602 		 * considered when setting up the initial
603 		 * window, we would not send the last segment
604 		 * if we skip considering the option length here.
605 		 * Note: this may not work when tcp headers change
606 		 * very dynamically in the future.
607 		 */
608 		if ((((tp->t_flags & TF_SIGNATURE) ?
609 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
610 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
611 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
612 		    len) >= tp->t_maxseg)
613 			goto send;
614 		/*
615 		 * NOTE! on localhost connections an 'ack' from the remote
616 		 * end may occur synchronously with the output and cause
617 		 * us to flush a buffer queued with moretocome.  XXX
618 		 *
619 		 * note: the len + off check is almost certainly unnecessary.
620 		 */
621 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
622 		    (idle || (tp->t_flags & TF_NODELAY)) &&
623 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
624 		    (tp->t_flags & TF_NOPUSH) == 0) {
625 			goto send;
626 		}
627 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
628 			goto send;
629 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
630 			goto send;
631 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
632 			goto send;
633 		if (sack_rxmit)
634 			goto send;
635 	}
636 
637 	/*
638 	 * Sending of standalone window updates.
639 	 *
640 	 * Window updates are important when we close our window due to a
641 	 * full socket buffer and are opening it again after the application
642 	 * reads data from it.  Once the window has opened again and the
643 	 * remote end starts to send again the ACK clock takes over and
644 	 * provides the most current window information.
645 	 *
646 	 * We must avoid the silly window syndrome whereas every read
647 	 * from the receive buffer, no matter how small, causes a window
648 	 * update to be sent.  We also should avoid sending a flurry of
649 	 * window updates when the socket buffer had queued a lot of data
650 	 * and the application is doing small reads.
651 	 *
652 	 * Prevent a flurry of pointless window updates by only sending
653 	 * an update when we can increase the advertized window by more
654 	 * than 1/4th of the socket buffer capacity.  When the buffer is
655 	 * getting full or is very small be more aggressive and send an
656 	 * update whenever we can increase by two mss sized segments.
657 	 * In all other situations the ACK's to new incoming data will
658 	 * carry further window increases.
659 	 *
660 	 * Don't send an independent window update if a delayed
661 	 * ACK is pending (it will get piggy-backed on it) or the
662 	 * remote side already has done a half-close and won't send
663 	 * more data.  Skip this if the connection is in T/TCP
664 	 * half-open state.
665 	 */
666 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
667 	    !(tp->t_flags & TF_DELACK) &&
668 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
669 		/*
670 		 * "adv" is the amount we could increase the window,
671 		 * taking into account that we are limited by
672 		 * TCP_MAXWIN << tp->rcv_scale.
673 		 */
674 		int32_t adv;
675 		int oldwin;
676 
677 		adv = recwin;
678 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
679 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
680 			if (adv > oldwin)
681 				adv -= oldwin;
682 			else
683 				adv = 0;
684 		} else
685 			oldwin = 0;
686 
687 		/*
688 		 * If the new window size ends up being the same as or less
689 		 * than the old size when it is scaled, then don't force
690 		 * a window update.
691 		 */
692 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
693 			goto dontupdate;
694 
695 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
696 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
697 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
698 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
699 		     adv >= TCP_MAXWIN << tp->rcv_scale))
700 			goto send;
701 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
702 			goto send;
703 	}
704 dontupdate:
705 
706 	/*
707 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
708 	 * is also a catch-all for the retransmit timer timeout case.
709 	 */
710 	if (tp->t_flags & TF_ACKNOW)
711 		goto send;
712 	if ((flags & TH_RST) ||
713 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
714 		goto send;
715 	if (SEQ_GT(tp->snd_up, tp->snd_una))
716 		goto send;
717 	/*
718 	 * If our state indicates that FIN should be sent
719 	 * and we have not yet done so, then we need to send.
720 	 */
721 	if (flags & TH_FIN &&
722 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
723 		goto send;
724 	/*
725 	 * In SACK, it is possible for tcp_output to fail to send a segment
726 	 * after the retransmission timer has been turned off.  Make sure
727 	 * that the retransmission timer is set.
728 	 */
729 	if ((tp->t_flags & TF_SACK_PERMIT) &&
730 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
731 	    !tcp_timer_active(tp, TT_REXMT) &&
732 	    !tcp_timer_active(tp, TT_PERSIST)) {
733 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
734 		goto just_return;
735 	}
736 	/*
737 	 * TCP window updates are not reliable, rather a polling protocol
738 	 * using ``persist'' packets is used to insure receipt of window
739 	 * updates.  The three ``states'' for the output side are:
740 	 *	idle			not doing retransmits or persists
741 	 *	persisting		to move a small or zero window
742 	 *	(re)transmitting	and thereby not persisting
743 	 *
744 	 * tcp_timer_active(tp, TT_PERSIST)
745 	 *	is true when we are in persist state.
746 	 * (tp->t_flags & TF_FORCEDATA)
747 	 *	is set when we are called to send a persist packet.
748 	 * tcp_timer_active(tp, TT_REXMT)
749 	 *	is set when we are retransmitting
750 	 * The output side is idle when both timers are zero.
751 	 *
752 	 * If send window is too small, there is data to transmit, and no
753 	 * retransmit or persist is pending, then go to persist state.
754 	 * If nothing happens soon, send when timer expires:
755 	 * if window is nonzero, transmit what we can,
756 	 * otherwise force out a byte.
757 	 */
758 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
759 	    !tcp_timer_active(tp, TT_PERSIST)) {
760 		tp->t_rxtshift = 0;
761 		tcp_setpersist(tp);
762 	}
763 
764 	/*
765 	 * No reason to send a segment, just return.
766 	 */
767 just_return:
768 	SOCKBUF_UNLOCK(&so->so_snd);
769 	return (0);
770 
771 send:
772 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
773 	if (len > 0) {
774 		if (len >= tp->t_maxseg)
775 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
776 		else
777 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
778 	}
779 	/*
780 	 * Before ESTABLISHED, force sending of initial options
781 	 * unless TCP set not to do any options.
782 	 * NOTE: we assume that the IP/TCP header plus TCP options
783 	 * always fit in a single mbuf, leaving room for a maximum
784 	 * link header, i.e.
785 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
786 	 */
787 	optlen = 0;
788 #ifdef INET6
789 	if (isipv6)
790 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
791 	else
792 #endif
793 		hdrlen = sizeof (struct tcpiphdr);
794 
795 	if (flags & TH_SYN) {
796 		tp->snd_nxt = tp->iss;
797 	}
798 
799 	/*
800 	 * Compute options for segment.
801 	 * We only have to care about SYN and established connection
802 	 * segments.  Options for SYN-ACK segments are handled in TCP
803 	 * syncache.
804 	 */
805 	to.to_flags = 0;
806 	if ((tp->t_flags & TF_NOOPT) == 0) {
807 		/* Maximum segment size. */
808 		if (flags & TH_SYN) {
809 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
810 			if (tp->t_port)
811 				to.to_mss -= V_tcp_udp_tunneling_overhead;
812 			to.to_flags |= TOF_MSS;
813 
814 			/*
815 			 * On SYN or SYN|ACK transmits on TFO connections,
816 			 * only include the TFO option if it is not a
817 			 * retransmit, as the presence of the TFO option may
818 			 * have caused the original SYN or SYN|ACK to have
819 			 * been dropped by a middlebox.
820 			 */
821 			if (IS_FASTOPEN(tp->t_flags) &&
822 			    (tp->t_rxtshift == 0)) {
823 				if (tp->t_state == TCPS_SYN_RECEIVED) {
824 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
825 					to.to_tfo_cookie =
826 					    (u_int8_t *)&tp->t_tfo_cookie.server;
827 					to.to_flags |= TOF_FASTOPEN;
828 					wanted_cookie = 1;
829 				} else if (tp->t_state == TCPS_SYN_SENT) {
830 					to.to_tfo_len =
831 					    tp->t_tfo_client_cookie_len;
832 					to.to_tfo_cookie =
833 					    tp->t_tfo_cookie.client;
834 					to.to_flags |= TOF_FASTOPEN;
835 					wanted_cookie = 1;
836 					/*
837 					 * If we wind up having more data to
838 					 * send with the SYN than can fit in
839 					 * one segment, don't send any more
840 					 * until the SYN|ACK comes back from
841 					 * the other end.
842 					 */
843 					dont_sendalot = 1;
844 				}
845 			}
846 		}
847 		/* Window scaling. */
848 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
849 			to.to_wscale = tp->request_r_scale;
850 			to.to_flags |= TOF_SCALE;
851 		}
852 		/* Timestamps. */
853 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
854 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
855 			curticks = tcp_ts_getticks();
856 			to.to_tsval = curticks + tp->ts_offset;
857 			to.to_tsecr = tp->ts_recent;
858 			to.to_flags |= TOF_TS;
859 			if (tp->t_rxtshift == 1)
860 				tp->t_badrxtwin = curticks;
861 		}
862 
863 		/* Set receive buffer autosizing timestamp. */
864 		if (tp->rfbuf_ts == 0 &&
865 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
866 			tp->rfbuf_ts = tcp_ts_getticks();
867 
868 		/* Selective ACK's. */
869 		if (tp->t_flags & TF_SACK_PERMIT) {
870 			if (flags & TH_SYN)
871 				to.to_flags |= TOF_SACKPERM;
872 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
873 			    tp->rcv_numsacks > 0) {
874 				to.to_flags |= TOF_SACK;
875 				to.to_nsacks = tp->rcv_numsacks;
876 				to.to_sacks = (u_char *)tp->sackblks;
877 			}
878 		}
879 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
880 		/* TCP-MD5 (RFC2385). */
881 		/*
882 		 * Check that TCP_MD5SIG is enabled in tcpcb to
883 		 * account the size needed to set this TCP option.
884 		 */
885 		if (tp->t_flags & TF_SIGNATURE)
886 			to.to_flags |= TOF_SIGNATURE;
887 #endif /* TCP_SIGNATURE */
888 
889 		/* Processing the options. */
890 		hdrlen += optlen = tcp_addoptions(&to, opt);
891 		/*
892 		 * If we wanted a TFO option to be added, but it was unable
893 		 * to fit, ensure no data is sent.
894 		 */
895 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
896 		    !(to.to_flags & TOF_FASTOPEN))
897 			len = 0;
898 	}
899 	if (tp->t_port) {
900 		if (V_tcp_udp_tunneling_port == 0) {
901 			/* The port was removed?? */
902 			SOCKBUF_UNLOCK(&so->so_snd);
903 			return (EHOSTUNREACH);
904 		}
905 		hdrlen += sizeof(struct udphdr);
906 	}
907 	/*
908 	 * Adjust data length if insertion of options will
909 	 * bump the packet length beyond the t_maxseg length.
910 	 * Clear the FIN bit because we cut off the tail of
911 	 * the segment.
912 	 */
913 	if (len + optlen + ipoptlen > tp->t_maxseg) {
914 		flags &= ~TH_FIN;
915 
916 		if (tso) {
917 			u_int if_hw_tsomax;
918 			u_int moff;
919 			int max_len;
920 
921 			/* extract TSO information */
922 			if_hw_tsomax = tp->t_tsomax;
923 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
924 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
925 
926 			/*
927 			 * Limit a TSO burst to prevent it from
928 			 * overflowing or exceeding the maximum length
929 			 * allowed by the network interface:
930 			 */
931 			KASSERT(ipoptlen == 0,
932 			    ("%s: TSO can't do IP options", __func__));
933 
934 			/*
935 			 * Check if we should limit by maximum payload
936 			 * length:
937 			 */
938 			if (if_hw_tsomax != 0) {
939 				/* compute maximum TSO length */
940 				max_len = (if_hw_tsomax - hdrlen -
941 				    max_linkhdr);
942 				if (max_len <= 0) {
943 					len = 0;
944 				} else if (len > max_len) {
945 					sendalot = 1;
946 					len = max_len;
947 				}
948 			}
949 
950 			/*
951 			 * Prevent the last segment from being
952 			 * fractional unless the send sockbuf can be
953 			 * emptied:
954 			 */
955 			max_len = (tp->t_maxseg - optlen);
956 			if (((uint32_t)off + (uint32_t)len) <
957 			    sbavail(&so->so_snd)) {
958 				moff = len % max_len;
959 				if (moff != 0) {
960 					len -= moff;
961 					sendalot = 1;
962 				}
963 			}
964 
965 			/*
966 			 * In case there are too many small fragments
967 			 * don't use TSO:
968 			 */
969 			if (len <= max_len) {
970 				len = max_len;
971 				sendalot = 1;
972 				tso = 0;
973 			}
974 
975 			/*
976 			 * Send the FIN in a separate segment
977 			 * after the bulk sending is done.
978 			 * We don't trust the TSO implementations
979 			 * to clear the FIN flag on all but the
980 			 * last segment.
981 			 */
982 			if (tp->t_flags & TF_NEEDFIN)
983 				sendalot = 1;
984 		} else {
985 			if (optlen + ipoptlen >= tp->t_maxseg) {
986 				/*
987 				 * Since we don't have enough space to put
988 				 * the IP header chain and the TCP header in
989 				 * one packet as required by RFC 7112, don't
990 				 * send it. Also ensure that at least one
991 				 * byte of the payload can be put into the
992 				 * TCP segment.
993 				 */
994 				SOCKBUF_UNLOCK(&so->so_snd);
995 				error = EMSGSIZE;
996 				sack_rxmit = 0;
997 				goto out;
998 			}
999 			len = tp->t_maxseg - optlen - ipoptlen;
1000 			sendalot = 1;
1001 			if (dont_sendalot)
1002 				sendalot = 0;
1003 		}
1004 	} else
1005 		tso = 0;
1006 
1007 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1008 	    ("%s: len > IP_MAXPACKET", __func__));
1009 
1010 /*#ifdef DIAGNOSTIC*/
1011 #ifdef INET6
1012 	if (max_linkhdr + hdrlen > MCLBYTES)
1013 #else
1014 	if (max_linkhdr + hdrlen > MHLEN)
1015 #endif
1016 		panic("tcphdr too big");
1017 /*#endif*/
1018 
1019 	/*
1020 	 * This KASSERT is here to catch edge cases at a well defined place.
1021 	 * Before, those had triggered (random) panic conditions further down.
1022 	 */
1023 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1024 
1025 	/*
1026 	 * Grab a header mbuf, attaching a copy of data to
1027 	 * be transmitted, and initialize the header from
1028 	 * the template for sends on this connection.
1029 	 */
1030 	if (len) {
1031 		struct mbuf *mb;
1032 		struct sockbuf *msb;
1033 		u_int moff;
1034 
1035 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1036 			TCPSTAT_INC(tcps_sndprobe);
1037 #ifdef STATS
1038 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1039 				stats_voi_update_abs_u32(tp->t_stats,
1040 				VOI_TCP_RETXPB, len);
1041 			else
1042 				stats_voi_update_abs_u64(tp->t_stats,
1043 				    VOI_TCP_TXPB, len);
1044 #endif /* STATS */
1045 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1046 			tp->t_sndrexmitpack++;
1047 			TCPSTAT_INC(tcps_sndrexmitpack);
1048 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1049 #ifdef STATS
1050 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1051 			    len);
1052 #endif /* STATS */
1053 		} else {
1054 			TCPSTAT_INC(tcps_sndpack);
1055 			TCPSTAT_ADD(tcps_sndbyte, len);
1056 #ifdef STATS
1057 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1058 			    len);
1059 #endif /* STATS */
1060 		}
1061 #ifdef INET6
1062 		if (MHLEN < hdrlen + max_linkhdr)
1063 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1064 		else
1065 #endif
1066 			m = m_gethdr(M_NOWAIT, MT_DATA);
1067 
1068 		if (m == NULL) {
1069 			SOCKBUF_UNLOCK(&so->so_snd);
1070 			error = ENOBUFS;
1071 			sack_rxmit = 0;
1072 			goto out;
1073 		}
1074 
1075 		m->m_data += max_linkhdr;
1076 		m->m_len = hdrlen;
1077 
1078 		/*
1079 		 * Start the m_copy functions from the closest mbuf
1080 		 * to the offset in the socket buffer chain.
1081 		 */
1082 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1083 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1084 			m_copydata(mb, moff, len,
1085 			    mtod(m, caddr_t) + hdrlen);
1086 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1087 				sbsndptr_adv(&so->so_snd, mb, len);
1088 			m->m_len += len;
1089 		} else {
1090 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1091 				msb = NULL;
1092 			else
1093 				msb = &so->so_snd;
1094 			m->m_next = tcp_m_copym(mb, moff,
1095 			    &len, if_hw_tsomaxsegcount,
1096 			    if_hw_tsomaxsegsize, msb, hw_tls);
1097 			if (len <= (tp->t_maxseg - optlen)) {
1098 				/*
1099 				 * Must have ran out of mbufs for the copy
1100 				 * shorten it to no longer need tso. Lets
1101 				 * not put on sendalot since we are low on
1102 				 * mbufs.
1103 				 */
1104 				tso = 0;
1105 			}
1106 			if (m->m_next == NULL) {
1107 				SOCKBUF_UNLOCK(&so->so_snd);
1108 				(void) m_free(m);
1109 				error = ENOBUFS;
1110 				sack_rxmit = 0;
1111 				goto out;
1112 			}
1113 		}
1114 
1115 		/*
1116 		 * If we're sending everything we've got, set PUSH.
1117 		 * (This will keep happy those implementations which only
1118 		 * give data to the user when a buffer fills or
1119 		 * a PUSH comes in.)
1120 		 */
1121 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1122 		    !(flags & TH_SYN))
1123 			flags |= TH_PUSH;
1124 		SOCKBUF_UNLOCK(&so->so_snd);
1125 	} else {
1126 		SOCKBUF_UNLOCK(&so->so_snd);
1127 		if (tp->t_flags & TF_ACKNOW)
1128 			TCPSTAT_INC(tcps_sndacks);
1129 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1130 			TCPSTAT_INC(tcps_sndctrl);
1131 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1132 			TCPSTAT_INC(tcps_sndurg);
1133 		else
1134 			TCPSTAT_INC(tcps_sndwinup);
1135 
1136 		m = m_gethdr(M_NOWAIT, MT_DATA);
1137 		if (m == NULL) {
1138 			error = ENOBUFS;
1139 			sack_rxmit = 0;
1140 			goto out;
1141 		}
1142 #ifdef INET6
1143 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1144 		    MHLEN >= hdrlen) {
1145 			M_ALIGN(m, hdrlen);
1146 		} else
1147 #endif
1148 		m->m_data += max_linkhdr;
1149 		m->m_len = hdrlen;
1150 	}
1151 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1152 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1153 #ifdef MAC
1154 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1155 #endif
1156 #ifdef INET6
1157 	if (isipv6) {
1158 		ip6 = mtod(m, struct ip6_hdr *);
1159 		if (tp->t_port) {
1160 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1161 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1162 			udp->uh_dport = tp->t_port;
1163 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1164 			udp->uh_ulen = htons(ulen);
1165 			th = (struct tcphdr *)(udp + 1);
1166 		} else {
1167 			th = (struct tcphdr *)(ip6 + 1);
1168 		}
1169 		tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th);
1170 	} else
1171 #endif /* INET6 */
1172 	{
1173 		ip = mtod(m, struct ip *);
1174 #ifdef TCPDEBUG
1175 		ipov = (struct ipovly *)ip;
1176 #endif
1177 		if (tp->t_port) {
1178 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1179 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1180 			udp->uh_dport = tp->t_port;
1181 			ulen = hdrlen + len - sizeof(struct ip);
1182 			udp->uh_ulen = htons(ulen);
1183 			th = (struct tcphdr *)(udp + 1);
1184 		} else
1185 			th = (struct tcphdr *)(ip + 1);
1186 		tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th);
1187 	}
1188 
1189 	/*
1190 	 * Fill in fields, remembering maximum advertised
1191 	 * window for use in delaying messages about window sizes.
1192 	 * If resending a FIN, be sure not to use a new sequence number.
1193 	 */
1194 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1195 	    tp->snd_nxt == tp->snd_max)
1196 		tp->snd_nxt--;
1197 	/*
1198 	 * If we are starting a connection, send ECN setup
1199 	 * SYN packet. If we are on a retransmit, we may
1200 	 * resend those bits a number of times as per
1201 	 * RFC 3168.
1202 	 */
1203 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1204 		flags |= tcp_ecn_output_syn_sent(tp);
1205 	}
1206 	/* Also handle parallel SYN for ECN */
1207 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1208 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
1209 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1210 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1211 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1212 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1213 #ifdef INET6
1214 		if (isipv6) {
1215 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
1216 			ip6->ip6_flow |= htonl(ect << 20);
1217 		}
1218 		else
1219 #endif
1220 		{
1221 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1222 			ip->ip_tos |= ect;
1223 		}
1224 	}
1225 
1226 	/*
1227 	 * If we are doing retransmissions, then snd_nxt will
1228 	 * not reflect the first unsent octet.  For ACK only
1229 	 * packets, we do not want the sequence number of the
1230 	 * retransmitted packet, we want the sequence number
1231 	 * of the next unsent octet.  So, if there is no data
1232 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1233 	 * when filling in ti_seq.  But if we are in persist
1234 	 * state, snd_max might reflect one byte beyond the
1235 	 * right edge of the window, so use snd_nxt in that
1236 	 * case, since we know we aren't doing a retransmission.
1237 	 * (retransmit and persist are mutually exclusive...)
1238 	 */
1239 	if (sack_rxmit == 0) {
1240 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1241 		    tcp_timer_active(tp, TT_PERSIST))
1242 			th->th_seq = htonl(tp->snd_nxt);
1243 		else
1244 			th->th_seq = htonl(tp->snd_max);
1245 	} else {
1246 		th->th_seq = htonl(p->rxmit);
1247 		p->rxmit += len;
1248 		/*
1249 		 * Lost Retransmission Detection
1250 		 * trigger resending of a (then
1251 		 * still existing) hole, when
1252 		 * fack acks recoverypoint.
1253 		 */
1254 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1255 			p->rxmit = tp->snd_recover;
1256 		tp->sackhint.sack_bytes_rexmit += len;
1257 	}
1258 	if (IN_RECOVERY(tp->t_flags)) {
1259 		/*
1260 		 * Account all bytes transmitted while
1261 		 * IN_RECOVERY, simplifying PRR and
1262 		 * Lost Retransmit Detection
1263 		 */
1264 		tp->sackhint.prr_out += len;
1265 	}
1266 	th->th_ack = htonl(tp->rcv_nxt);
1267 	if (optlen) {
1268 		bcopy(opt, th + 1, optlen);
1269 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1270 	}
1271 	tcp_set_flags(th, flags);
1272 	/*
1273 	 * Calculate receive window.  Don't shrink window,
1274 	 * but avoid silly window syndrome.
1275 	 * If a RST segment is sent, advertise a window of zero.
1276 	 */
1277 	if (flags & TH_RST) {
1278 		recwin = 0;
1279 	} else {
1280 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1281 		    recwin < tp->t_maxseg)
1282 			recwin = 0;
1283 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1284 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1285 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1286 	}
1287 	/*
1288 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1289 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1290 	 * case is handled in syncache.
1291 	 */
1292 	if (flags & TH_SYN)
1293 		th->th_win = htons((u_short)
1294 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1295 	else {
1296 		/* Avoid shrinking window with window scaling. */
1297 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1298 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1299 	}
1300 
1301 	/*
1302 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1303 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1304 	 * flag tells soreceive() to disable delayed acknowledgements when
1305 	 * draining the buffer.  This can occur if the receiver is attempting
1306 	 * to read more data than can be buffered prior to transmitting on
1307 	 * the connection.
1308 	 */
1309 	if (th->th_win == 0) {
1310 		tp->t_sndzerowin++;
1311 		tp->t_flags |= TF_RXWIN0SENT;
1312 	} else
1313 		tp->t_flags &= ~TF_RXWIN0SENT;
1314 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1315 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1316 		th->th_flags |= TH_URG;
1317 	} else
1318 		/*
1319 		 * If no urgent pointer to send, then we pull
1320 		 * the urgent pointer to the left edge of the send window
1321 		 * so that it doesn't drift into the send window on sequence
1322 		 * number wraparound.
1323 		 */
1324 		tp->snd_up = tp->snd_una;		/* drag it along */
1325 
1326 	/*
1327 	 * Put TCP length in extended header, and then
1328 	 * checksum extended header and data.
1329 	 */
1330 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1331 
1332 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1333 	if (to.to_flags & TOF_SIGNATURE) {
1334 		/*
1335 		 * Calculate MD5 signature and put it into the place
1336 		 * determined before.
1337 		 * NOTE: since TCP options buffer doesn't point into
1338 		 * mbuf's data, calculate offset and use it.
1339 		 */
1340 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1341 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1342 			/*
1343 			 * Do not send segment if the calculation of MD5
1344 			 * digest has failed.
1345 			 */
1346 			m_freem(m);
1347 			goto out;
1348 		}
1349 	}
1350 #endif
1351 #ifdef INET6
1352 	if (isipv6) {
1353 		/*
1354 		 * There is no need to fill in ip6_plen right now.
1355 		 * It will be filled later by ip6_output.
1356 		 */
1357 		if (tp->t_port) {
1358 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1359 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1360 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1361 			th->th_sum = htons(0);
1362 			UDPSTAT_INC(udps_opackets);
1363 		} else {
1364 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1365 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1366 			th->th_sum = in6_cksum_pseudo(ip6,
1367 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1368 			    0);
1369 		}
1370 	}
1371 #endif
1372 #if defined(INET6) && defined(INET)
1373 	else
1374 #endif
1375 #ifdef INET
1376 	{
1377 		if (tp->t_port) {
1378 			m->m_pkthdr.csum_flags = CSUM_UDP;
1379 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1380 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1381 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1382 			th->th_sum = htons(0);
1383 			UDPSTAT_INC(udps_opackets);
1384 		} else {
1385 			m->m_pkthdr.csum_flags = CSUM_TCP;
1386 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1387 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1388 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1389 			    IPPROTO_TCP + len + optlen));
1390 		}
1391 
1392 		/* IP version must be set here for ipv4/ipv6 checking later */
1393 		KASSERT(ip->ip_v == IPVERSION,
1394 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1395 	}
1396 #endif
1397 
1398 	/*
1399 	 * Enable TSO and specify the size of the segments.
1400 	 * The TCP pseudo header checksum is always provided.
1401 	 */
1402 	if (tso) {
1403 		KASSERT(len > tp->t_maxseg - optlen,
1404 		    ("%s: len <= tso_segsz", __func__));
1405 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1406 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1407 	}
1408 
1409 	KASSERT(len + hdrlen == m_length(m, NULL),
1410 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1411 	    __func__, len, hdrlen, m_length(m, NULL)));
1412 
1413 #ifdef TCP_HHOOK
1414 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1415 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1416 #endif
1417 
1418 #ifdef TCPDEBUG
1419 	/*
1420 	 * Trace.
1421 	 */
1422 	if (so->so_options & SO_DEBUG) {
1423 		u_short save = 0;
1424 #ifdef INET6
1425 		if (!isipv6)
1426 #endif
1427 		{
1428 			save = ipov->ih_len;
1429 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1430 		}
1431 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1432 #ifdef INET6
1433 		if (!isipv6)
1434 #endif
1435 		ipov->ih_len = save;
1436 	}
1437 #endif /* TCPDEBUG */
1438 	TCP_PROBE3(debug__output, tp, th, m);
1439 
1440 	/* We're getting ready to send; log now. */
1441 	TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1442 	    len, NULL, false);
1443 
1444 	/*
1445 	 * Fill in IP length and desired time to live and
1446 	 * send to IP level.  There should be a better way
1447 	 * to handle ttl and tos; we could keep them in
1448 	 * the template, but need a way to checksum without them.
1449 	 */
1450 	/*
1451 	 * m->m_pkthdr.len should have been set before checksum calculation,
1452 	 * because in6_cksum() need it.
1453 	 */
1454 #ifdef INET6
1455 	if (isipv6) {
1456 		/*
1457 		 * we separately set hoplimit for every segment, since the
1458 		 * user might want to change the value via setsockopt.
1459 		 * Also, desired default hop limit might be changed via
1460 		 * Neighbor Discovery.
1461 		 */
1462 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1463 
1464 		/*
1465 		 * Set the packet size here for the benefit of DTrace probes.
1466 		 * ip6_output() will set it properly; it's supposed to include
1467 		 * the option header lengths as well.
1468 		 */
1469 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1470 
1471 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1472 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1473 		else
1474 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1475 
1476 		if (tp->t_state == TCPS_SYN_SENT)
1477 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1478 
1479 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1480 
1481 #ifdef TCPPCAP
1482 		/* Save packet, if requested. */
1483 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1484 #endif
1485 
1486 		/* TODO: IPv6 IP6TOS_ECT bit on */
1487 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1488 		    &tp->t_inpcb->inp_route6,
1489 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1490 		    NULL, NULL, tp->t_inpcb);
1491 
1492 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1493 			mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1494 	}
1495 #endif /* INET6 */
1496 #if defined(INET) && defined(INET6)
1497 	else
1498 #endif
1499 #ifdef INET
1500     {
1501 	ip->ip_len = htons(m->m_pkthdr.len);
1502 #ifdef INET6
1503 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1504 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1505 #endif /* INET6 */
1506 	/*
1507 	 * If we do path MTU discovery, then we set DF on every packet.
1508 	 * This might not be the best thing to do according to RFC3390
1509 	 * Section 2. However the tcp hostcache migitates the problem
1510 	 * so it affects only the first tcp connection with a host.
1511 	 *
1512 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1513 	 */
1514 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1515 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1516 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1517 			ip->ip_off |= htons(IP_DF);
1518 		}
1519 	} else {
1520 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1521 	}
1522 
1523 	if (tp->t_state == TCPS_SYN_SENT)
1524 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1525 
1526 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1527 
1528 #ifdef TCPPCAP
1529 	/* Save packet, if requested. */
1530 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1531 #endif
1532 
1533 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1534 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1535 	    tp->t_inpcb);
1536 
1537 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1538 		mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1539     }
1540 #endif /* INET */
1541 
1542 out:
1543 	if (error == 0)
1544 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1545 	/*
1546 	 * In transmit state, time the transmission and arrange for
1547 	 * the retransmit.  In persist state, just set snd_max.
1548 	 */
1549 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1550 	    !tcp_timer_active(tp, TT_PERSIST)) {
1551 		tcp_seq startseq = tp->snd_nxt;
1552 
1553 		/*
1554 		 * Advance snd_nxt over sequence space of this segment.
1555 		 */
1556 		if (flags & (TH_SYN|TH_FIN)) {
1557 			if (flags & TH_SYN)
1558 				tp->snd_nxt++;
1559 			if (flags & TH_FIN) {
1560 				tp->snd_nxt++;
1561 				tp->t_flags |= TF_SENTFIN;
1562 			}
1563 		}
1564 		if (sack_rxmit)
1565 			goto timer;
1566 		tp->snd_nxt += len;
1567 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1568 			tp->snd_max = tp->snd_nxt;
1569 			/*
1570 			 * Time this transmission if not a retransmission and
1571 			 * not currently timing anything.
1572 			 */
1573 			tp->t_sndtime = ticks;
1574 			if (tp->t_rtttime == 0) {
1575 				tp->t_rtttime = ticks;
1576 				tp->t_rtseq = startseq;
1577 				TCPSTAT_INC(tcps_segstimed);
1578 			}
1579 #ifdef STATS
1580 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1581 				tp->t_flags |= TF_GPUTINPROG;
1582 				tp->gput_seq = startseq;
1583 				tp->gput_ack = startseq +
1584 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1585 				tp->gput_ts = tcp_ts_getticks();
1586 			}
1587 #endif /* STATS */
1588 		}
1589 
1590 		/*
1591 		 * Set retransmit timer if not currently set,
1592 		 * and not doing a pure ack or a keep-alive probe.
1593 		 * Initial value for retransmit timer is smoothed
1594 		 * round-trip time + 2 * round-trip time variance.
1595 		 * Initialize shift counter which is used for backoff
1596 		 * of retransmit time.
1597 		 */
1598 timer:
1599 		if (!tcp_timer_active(tp, TT_REXMT) &&
1600 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1601 		     (tp->snd_nxt != tp->snd_una))) {
1602 			if (tcp_timer_active(tp, TT_PERSIST)) {
1603 				tcp_timer_activate(tp, TT_PERSIST, 0);
1604 				tp->t_rxtshift = 0;
1605 			}
1606 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1607 		} else if (len == 0 && sbavail(&so->so_snd) &&
1608 		    !tcp_timer_active(tp, TT_REXMT) &&
1609 		    !tcp_timer_active(tp, TT_PERSIST)) {
1610 			/*
1611 			 * Avoid a situation where we do not set persist timer
1612 			 * after a zero window condition. For example:
1613 			 * 1) A -> B: packet with enough data to fill the window
1614 			 * 2) B -> A: ACK for #1 + new data (0 window
1615 			 *    advertisement)
1616 			 * 3) A -> B: ACK for #2, 0 len packet
1617 			 *
1618 			 * In this case, A will not activate the persist timer,
1619 			 * because it chose to send a packet. Unless tcp_output
1620 			 * is called for some other reason (delayed ack timer,
1621 			 * another input packet from B, socket syscall), A will
1622 			 * not send zero window probes.
1623 			 *
1624 			 * So, if you send a 0-length packet, but there is data
1625 			 * in the socket buffer, and neither the rexmt or
1626 			 * persist timer is already set, then activate the
1627 			 * persist timer.
1628 			 */
1629 			tp->t_rxtshift = 0;
1630 			tcp_setpersist(tp);
1631 		}
1632 	} else {
1633 		/*
1634 		 * Persist case, update snd_max but since we are in
1635 		 * persist mode (no window) we do not update snd_nxt.
1636 		 */
1637 		int xlen = len;
1638 		if (flags & TH_SYN)
1639 			++xlen;
1640 		if (flags & TH_FIN) {
1641 			++xlen;
1642 			tp->t_flags |= TF_SENTFIN;
1643 		}
1644 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1645 			tp->snd_max = tp->snd_nxt + xlen;
1646 	}
1647 	if ((error == 0) &&
1648 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1649 	     (tp->t_flags & TF_SACK_PERMIT) &&
1650 	     tp->rcv_numsacks > 0)) {
1651 		    /* Clean up any DSACK's sent */
1652 		    tcp_clean_dsack_blocks(tp);
1653 	}
1654 	if (error) {
1655 		/* Record the error. */
1656 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1657 		    error, 0, NULL, false);
1658 
1659 		/*
1660 		 * We know that the packet was lost, so back out the
1661 		 * sequence number advance, if any.
1662 		 *
1663 		 * If the error is EPERM the packet got blocked by the
1664 		 * local firewall.  Normally we should terminate the
1665 		 * connection but the blocking may have been spurious
1666 		 * due to a firewall reconfiguration cycle.  So we treat
1667 		 * it like a packet loss and let the retransmit timer and
1668 		 * timeouts do their work over time.
1669 		 * XXX: It is a POLA question whether calling tcp_drop right
1670 		 * away would be the really correct behavior instead.
1671 		 */
1672 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1673 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1674 		    ((flags & TH_SYN) == 0) &&
1675 		    (error != EPERM)) {
1676 			if (sack_rxmit) {
1677 				p->rxmit -= len;
1678 				tp->sackhint.sack_bytes_rexmit -= len;
1679 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1680 				    ("sackhint bytes rtx >= 0"));
1681 			} else
1682 				tp->snd_nxt -= len;
1683 		}
1684 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1685 		switch (error) {
1686 		case EACCES:
1687 		case EPERM:
1688 			tp->t_softerror = error;
1689 			return (error);
1690 		case ENOBUFS:
1691 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1692 			tp->snd_cwnd = tp->t_maxseg;
1693 			return (0);
1694 		case EMSGSIZE:
1695 			/*
1696 			 * For some reason the interface we used initially
1697 			 * to send segments changed to another or lowered
1698 			 * its MTU.
1699 			 * If TSO was active we either got an interface
1700 			 * without TSO capabilits or TSO was turned off.
1701 			 * If we obtained mtu from ip_output() then update
1702 			 * it and try again.
1703 			 */
1704 			if (tso)
1705 				tp->t_flags &= ~TF_TSO;
1706 			if (mtu != 0) {
1707 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1708 				goto again;
1709 			}
1710 			return (error);
1711 		case EHOSTDOWN:
1712 		case EHOSTUNREACH:
1713 		case ENETDOWN:
1714 		case ENETUNREACH:
1715 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1716 				tp->t_softerror = error;
1717 				return (0);
1718 			}
1719 			/* FALLTHROUGH */
1720 		default:
1721 			return (error);
1722 		}
1723 	}
1724 	TCPSTAT_INC(tcps_sndtotal);
1725 
1726 	/*
1727 	 * Data sent (as far as we can tell).
1728 	 * If this advertises a larger window than any other segment,
1729 	 * then remember the size of the advertised window.
1730 	 * Any pending ACK has now been sent.
1731 	 */
1732 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1733 		tp->rcv_adv = tp->rcv_nxt + recwin;
1734 	tp->last_ack_sent = tp->rcv_nxt;
1735 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1736 	if (tcp_timer_active(tp, TT_DELACK))
1737 		tcp_timer_activate(tp, TT_DELACK, 0);
1738 #if 0
1739 	/*
1740 	 * This completely breaks TCP if newreno is turned on.  What happens
1741 	 * is that if delayed-acks are turned on on the receiver, this code
1742 	 * on the transmitter effectively destroys the TCP window, forcing
1743 	 * it to four packets (1.5Kx4 = 6K window).
1744 	 */
1745 	if (sendalot && --maxburst)
1746 		goto again;
1747 #endif
1748 	if (sendalot)
1749 		goto again;
1750 	return (0);
1751 }
1752 
1753 void
1754 tcp_setpersist(struct tcpcb *tp)
1755 {
1756 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1757 	int tt;
1758 
1759 	tp->t_flags &= ~TF_PREVVALID;
1760 	if (tcp_timer_active(tp, TT_REXMT))
1761 		panic("tcp_setpersist: retransmit pending");
1762 	/*
1763 	 * Start/restart persistence timer.
1764 	 */
1765 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1766 		      tcp_persmin, tcp_persmax);
1767 	tcp_timer_activate(tp, TT_PERSIST, tt);
1768 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1769 		tp->t_rxtshift++;
1770 }
1771 
1772 /*
1773  * Insert TCP options according to the supplied parameters to the place
1774  * optp in a consistent way.  Can handle unaligned destinations.
1775  *
1776  * The order of the option processing is crucial for optimal packing and
1777  * alignment for the scarce option space.
1778  *
1779  * The optimal order for a SYN/SYN-ACK segment is:
1780  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1781  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1782  *
1783  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1784  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1785  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1786  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1787  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1788  */
1789 int
1790 tcp_addoptions(struct tcpopt *to, u_char *optp)
1791 {
1792 	u_int32_t mask, optlen = 0;
1793 
1794 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1795 		if ((to->to_flags & mask) != mask)
1796 			continue;
1797 		if (optlen == TCP_MAXOLEN)
1798 			break;
1799 		switch (to->to_flags & mask) {
1800 		case TOF_MSS:
1801 			while (optlen % 4) {
1802 				optlen += TCPOLEN_NOP;
1803 				*optp++ = TCPOPT_NOP;
1804 			}
1805 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1806 				continue;
1807 			optlen += TCPOLEN_MAXSEG;
1808 			*optp++ = TCPOPT_MAXSEG;
1809 			*optp++ = TCPOLEN_MAXSEG;
1810 			to->to_mss = htons(to->to_mss);
1811 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1812 			optp += sizeof(to->to_mss);
1813 			break;
1814 		case TOF_SCALE:
1815 			while (!optlen || optlen % 2 != 1) {
1816 				optlen += TCPOLEN_NOP;
1817 				*optp++ = TCPOPT_NOP;
1818 			}
1819 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1820 				continue;
1821 			optlen += TCPOLEN_WINDOW;
1822 			*optp++ = TCPOPT_WINDOW;
1823 			*optp++ = TCPOLEN_WINDOW;
1824 			*optp++ = to->to_wscale;
1825 			break;
1826 		case TOF_SACKPERM:
1827 			while (optlen % 2) {
1828 				optlen += TCPOLEN_NOP;
1829 				*optp++ = TCPOPT_NOP;
1830 			}
1831 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1832 				continue;
1833 			optlen += TCPOLEN_SACK_PERMITTED;
1834 			*optp++ = TCPOPT_SACK_PERMITTED;
1835 			*optp++ = TCPOLEN_SACK_PERMITTED;
1836 			break;
1837 		case TOF_TS:
1838 			while (!optlen || optlen % 4 != 2) {
1839 				optlen += TCPOLEN_NOP;
1840 				*optp++ = TCPOPT_NOP;
1841 			}
1842 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1843 				continue;
1844 			optlen += TCPOLEN_TIMESTAMP;
1845 			*optp++ = TCPOPT_TIMESTAMP;
1846 			*optp++ = TCPOLEN_TIMESTAMP;
1847 			to->to_tsval = htonl(to->to_tsval);
1848 			to->to_tsecr = htonl(to->to_tsecr);
1849 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1850 			optp += sizeof(to->to_tsval);
1851 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1852 			optp += sizeof(to->to_tsecr);
1853 			break;
1854 		case TOF_SIGNATURE:
1855 			{
1856 			int siglen = TCPOLEN_SIGNATURE - 2;
1857 
1858 			while (!optlen || optlen % 4 != 2) {
1859 				optlen += TCPOLEN_NOP;
1860 				*optp++ = TCPOPT_NOP;
1861 			}
1862 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1863 				to->to_flags &= ~TOF_SIGNATURE;
1864 				continue;
1865 			}
1866 			optlen += TCPOLEN_SIGNATURE;
1867 			*optp++ = TCPOPT_SIGNATURE;
1868 			*optp++ = TCPOLEN_SIGNATURE;
1869 			to->to_signature = optp;
1870 			while (siglen--)
1871 				 *optp++ = 0;
1872 			break;
1873 			}
1874 		case TOF_SACK:
1875 			{
1876 			int sackblks = 0;
1877 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1878 			tcp_seq sack_seq;
1879 
1880 			while (!optlen || optlen % 4 != 2) {
1881 				optlen += TCPOLEN_NOP;
1882 				*optp++ = TCPOPT_NOP;
1883 			}
1884 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1885 				continue;
1886 			optlen += TCPOLEN_SACKHDR;
1887 			*optp++ = TCPOPT_SACK;
1888 			sackblks = min(to->to_nsacks,
1889 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1890 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1891 			while (sackblks--) {
1892 				sack_seq = htonl(sack->start);
1893 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1894 				optp += sizeof(sack_seq);
1895 				sack_seq = htonl(sack->end);
1896 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1897 				optp += sizeof(sack_seq);
1898 				optlen += TCPOLEN_SACK;
1899 				sack++;
1900 			}
1901 			TCPSTAT_INC(tcps_sack_send_blocks);
1902 			break;
1903 			}
1904 		case TOF_FASTOPEN:
1905 			{
1906 			int total_len;
1907 
1908 			/* XXX is there any point to aligning this option? */
1909 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1910 			if (TCP_MAXOLEN - optlen < total_len) {
1911 				to->to_flags &= ~TOF_FASTOPEN;
1912 				continue;
1913 			}
1914 			*optp++ = TCPOPT_FAST_OPEN;
1915 			*optp++ = total_len;
1916 			if (to->to_tfo_len > 0) {
1917 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1918 				optp += to->to_tfo_len;
1919 			}
1920 			optlen += total_len;
1921 			break;
1922 			}
1923 		default:
1924 			panic("%s: unknown TCP option type", __func__);
1925 			break;
1926 		}
1927 	}
1928 
1929 	/* Terminate and pad TCP options to a 4 byte boundary. */
1930 	if (optlen % 4) {
1931 		optlen += TCPOLEN_EOL;
1932 		*optp++ = TCPOPT_EOL;
1933 	}
1934 	/*
1935 	 * According to RFC 793 (STD0007):
1936 	 *   "The content of the header beyond the End-of-Option option
1937 	 *    must be header padding (i.e., zero)."
1938 	 *   and later: "The padding is composed of zeros."
1939 	 */
1940 	while (optlen % 4) {
1941 		optlen += TCPOLEN_PAD;
1942 		*optp++ = TCPOPT_PAD;
1943 	}
1944 
1945 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1946 	return (optlen);
1947 }
1948 
1949 /*
1950  * This is a copy of m_copym(), taking the TSO segment size/limit
1951  * constraints into account, and advancing the sndptr as it goes.
1952  */
1953 struct mbuf *
1954 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1955     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1956 {
1957 #ifdef KERN_TLS
1958 	struct ktls_session *tls, *ntls;
1959 	struct mbuf *start;
1960 #endif
1961 	struct mbuf *n, **np;
1962 	struct mbuf *top;
1963 	int32_t off = off0;
1964 	int32_t len = *plen;
1965 	int32_t fragsize;
1966 	int32_t len_cp = 0;
1967 	int32_t *pkthdrlen;
1968 	uint32_t mlen, frags;
1969 	bool copyhdr;
1970 
1971 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1972 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1973 	if (off == 0 && m->m_flags & M_PKTHDR)
1974 		copyhdr = true;
1975 	else
1976 		copyhdr = false;
1977 	while (off > 0) {
1978 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1979 		if (off < m->m_len)
1980 			break;
1981 		off -= m->m_len;
1982 		if ((sb) && (m == sb->sb_sndptr)) {
1983 			sb->sb_sndptroff += m->m_len;
1984 			sb->sb_sndptr = m->m_next;
1985 		}
1986 		m = m->m_next;
1987 	}
1988 	np = &top;
1989 	top = NULL;
1990 	pkthdrlen = NULL;
1991 #ifdef KERN_TLS
1992 	if (hw_tls && (m->m_flags & M_EXTPG))
1993 		tls = m->m_epg_tls;
1994 	else
1995 		tls = NULL;
1996 	start = m;
1997 #endif
1998 	while (len > 0) {
1999 		if (m == NULL) {
2000 			KASSERT(len == M_COPYALL,
2001 			    ("tcp_m_copym, length > size of mbuf chain"));
2002 			*plen = len_cp;
2003 			if (pkthdrlen != NULL)
2004 				*pkthdrlen = len_cp;
2005 			break;
2006 		}
2007 #ifdef KERN_TLS
2008 		if (hw_tls) {
2009 			if (m->m_flags & M_EXTPG)
2010 				ntls = m->m_epg_tls;
2011 			else
2012 				ntls = NULL;
2013 
2014 			/*
2015 			 * Avoid mixing TLS records with handshake
2016 			 * data or TLS records from different
2017 			 * sessions.
2018 			 */
2019 			if (tls != ntls) {
2020 				MPASS(m != start);
2021 				*plen = len_cp;
2022 				if (pkthdrlen != NULL)
2023 					*pkthdrlen = len_cp;
2024 				break;
2025 			}
2026 		}
2027 #endif
2028 		mlen = min(len, m->m_len - off);
2029 		if (seglimit) {
2030 			/*
2031 			 * For M_EXTPG mbufs, add 3 segments
2032 			 * + 1 in case we are crossing page boundaries
2033 			 * + 2 in case the TLS hdr/trailer are used
2034 			 * It is cheaper to just add the segments
2035 			 * than it is to take the cache miss to look
2036 			 * at the mbuf ext_pgs state in detail.
2037 			 */
2038 			if (m->m_flags & M_EXTPG) {
2039 				fragsize = min(segsize, PAGE_SIZE);
2040 				frags = 3;
2041 			} else {
2042 				fragsize = segsize;
2043 				frags = 0;
2044 			}
2045 
2046 			/* Break if we really can't fit anymore. */
2047 			if ((frags + 1) >= seglimit) {
2048 				*plen =	len_cp;
2049 				if (pkthdrlen != NULL)
2050 					*pkthdrlen = len_cp;
2051 				break;
2052 			}
2053 
2054 			/*
2055 			 * Reduce size if you can't copy the whole
2056 			 * mbuf. If we can't copy the whole mbuf, also
2057 			 * adjust len so the loop will end after this
2058 			 * mbuf.
2059 			 */
2060 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2061 				mlen = (seglimit - frags - 1) * fragsize;
2062 				len = mlen;
2063 				*plen = len_cp + len;
2064 				if (pkthdrlen != NULL)
2065 					*pkthdrlen = *plen;
2066 			}
2067 			frags += howmany(mlen, fragsize);
2068 			if (frags == 0)
2069 				frags++;
2070 			seglimit -= frags;
2071 			KASSERT(seglimit > 0,
2072 			    ("%s: seglimit went too low", __func__));
2073 		}
2074 		if (copyhdr)
2075 			n = m_gethdr(M_NOWAIT, m->m_type);
2076 		else
2077 			n = m_get(M_NOWAIT, m->m_type);
2078 		*np = n;
2079 		if (n == NULL)
2080 			goto nospace;
2081 		if (copyhdr) {
2082 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2083 				goto nospace;
2084 			if (len == M_COPYALL)
2085 				n->m_pkthdr.len -= off0;
2086 			else
2087 				n->m_pkthdr.len = len;
2088 			pkthdrlen = &n->m_pkthdr.len;
2089 			copyhdr = false;
2090 		}
2091 		n->m_len = mlen;
2092 		len_cp += n->m_len;
2093 		if (m->m_flags & (M_EXT|M_EXTPG)) {
2094 			n->m_data = m->m_data + off;
2095 			mb_dupcl(n, m);
2096 		} else
2097 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2098 			    (u_int)n->m_len);
2099 
2100 		if (sb && (sb->sb_sndptr == m) &&
2101 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2102 			sb->sb_sndptroff += m->m_len;
2103 			sb->sb_sndptr = m->m_next;
2104 		}
2105 		off = 0;
2106 		if (len != M_COPYALL) {
2107 			len -= n->m_len;
2108 		}
2109 		m = m->m_next;
2110 		np = &n->m_next;
2111 	}
2112 	return (top);
2113 nospace:
2114 	m_freem(top);
2115 	return (NULL);
2116 }
2117 
2118 void
2119 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2120 {
2121 
2122 	/*
2123 	 * Automatic sizing of send socket buffer.  Often the send buffer
2124 	 * size is not optimally adjusted to the actual network conditions
2125 	 * at hand (delay bandwidth product).  Setting the buffer size too
2126 	 * small limits throughput on links with high bandwidth and high
2127 	 * delay (eg. trans-continental/oceanic links).  Setting the
2128 	 * buffer size too big consumes too much real kernel memory,
2129 	 * especially with many connections on busy servers.
2130 	 *
2131 	 * The criteria to step up the send buffer one notch are:
2132 	 *  1. receive window of remote host is larger than send buffer
2133 	 *     (with a fudge factor of 5/4th);
2134 	 *  2. send buffer is filled to 7/8th with data (so we actually
2135 	 *     have data to make use of it);
2136 	 *  3. send buffer fill has not hit maximal automatic size;
2137 	 *  4. our send window (slow start and cogestion controlled) is
2138 	 *     larger than sent but unacknowledged data in send buffer.
2139 	 *
2140 	 * The remote host receive window scaling factor may limit the
2141 	 * growing of the send buffer before it reaches its allowed
2142 	 * maximum.
2143 	 *
2144 	 * It scales directly with slow start or congestion window
2145 	 * and does at most one step per received ACK.  This fast
2146 	 * scaling has the drawback of growing the send buffer beyond
2147 	 * what is strictly necessary to make full use of a given
2148 	 * delay*bandwidth product.  However testing has shown this not
2149 	 * to be much of an problem.  At worst we are trading wasting
2150 	 * of available bandwidth (the non-use of it) for wasting some
2151 	 * socket buffer memory.
2152 	 *
2153 	 * TODO: Shrink send buffer during idle periods together
2154 	 * with congestion window.  Requires another timer.  Has to
2155 	 * wait for upcoming tcp timer rewrite.
2156 	 *
2157 	 * XXXGL: should there be used sbused() or sbavail()?
2158 	 */
2159 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2160 		int lowat;
2161 
2162 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2163 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2164 		    sbused(&so->so_snd) >=
2165 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2166 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2167 		    sendwin >= (sbused(&so->so_snd) -
2168 		    (tp->snd_nxt - tp->snd_una))) {
2169 			if (!sbreserve_locked(&so->so_snd,
2170 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2171 			     V_tcp_autosndbuf_max), so, curthread))
2172 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2173 		}
2174 	}
2175 }
2176