xref: /freebsd/sys/netinet/tcp_output.c (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
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 	    (tp->sackhint.nexthole != NULL) &&
270 	    !IN_FASTRECOVERY(tp->t_flags)) {
271 		sendwin = tcp_sack_adjust(tp);
272 	}
273 	sendalot = 0;
274 	tso = 0;
275 	mtu = 0;
276 	off = tp->snd_nxt - tp->snd_una;
277 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + sendwin);
278 
279 	flags = tcp_outflags[tp->t_state];
280 	/*
281 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
282 	 * to send out new data (when sendalot is 1), bypass this function.
283 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
284 	 * we're replacing a (future) new transmission with a retransmission
285 	 * now, and we previously incremented snd_cwnd in tcp_input().
286 	 */
287 	/*
288 	 * Still in sack recovery , reset rxmit flag to zero.
289 	 */
290 	sack_rxmit = 0;
291 	sack_bytes_rxmt = 0;
292 	len = 0;
293 	p = NULL;
294 	if ((tp->t_flags & TF_SACK_PERMIT) &&
295 	    (IN_FASTRECOVERY(tp->t_flags) || SEQ_LT(tp->snd_nxt, tp->snd_max)) &&
296 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
297 		int32_t cwin;
298 
299 		if (IN_FASTRECOVERY(tp->t_flags)) {
300 			cwin = imax(sendwin - tcp_compute_pipe(tp), 0);
301 		} else {
302 			cwin = imax(sendwin - off, 0);
303 		}
304 		/* Do not retransmit SACK segments beyond snd_recover */
305 		if (SEQ_GT(p->end, tp->snd_recover)) {
306 			/*
307 			 * (At least) part of sack hole extends beyond
308 			 * snd_recover. Check to see if we can rexmit data
309 			 * for this hole.
310 			 */
311 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
312 				/*
313 				 * Can't rexmit any more data for this hole.
314 				 * That data will be rexmitted in the next
315 				 * sack recovery episode, when snd_recover
316 				 * moves past p->rxmit.
317 				 */
318 				p = NULL;
319 				goto after_sack_rexmit;
320 			} else {
321 				/* Can rexmit part of the current hole */
322 				len = SEQ_SUB(tp->snd_recover, p->rxmit);
323 				if (cwin <= len) {
324 					len = cwin;
325 				} else {
326 					sendalot = 1;
327 				}
328 			}
329 		} else {
330 			len = SEQ_SUB(p->end, p->rxmit);
331 			if (cwin <= len) {
332 				len = cwin;
333 			} else {
334 				sendalot = 1;
335 			}
336 		}
337 		/* we could have transmitted from the scoreboard,
338 		 * but sendwin (expected flightsize) - pipe didn't
339 		 * allow any transmission.
340 		 * Bypass recalculating the possible transmission
341 		 * length further down by setting sack_rxmit.
342 		 * Wouldn't be here if there would have been
343 		 * nothing in the scoreboard to transmit.
344 		 */
345 		sack_rxmit = 1;
346 		if (len > 0) {
347 			off = SEQ_SUB(p->rxmit, tp->snd_una);
348 			KASSERT(off >= 0,("%s: sack block to the left of una : %d",
349 			    __func__, off));
350 		}
351 	}
352 after_sack_rexmit:
353 	/*
354 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
355 	 * state flags.
356 	 */
357 	if (tp->t_flags & TF_NEEDFIN)
358 		flags |= TH_FIN;
359 	if (tp->t_flags & TF_NEEDSYN)
360 		flags |= TH_SYN;
361 
362 	SOCKBUF_LOCK(&so->so_snd);
363 	/*
364 	 * If in persist timeout with window of 0, send 1 byte.
365 	 * Otherwise, if window is small but nonzero
366 	 * and timer expired, we will send what we can
367 	 * and go to transmit state.
368 	 */
369 	if (tp->t_flags & TF_FORCEDATA) {
370 		if (sendwin == 0) {
371 			/*
372 			 * If we still have some data to send, then
373 			 * clear the FIN bit.  Usually this would
374 			 * happen below when it realizes that we
375 			 * aren't sending all the data.  However,
376 			 * if we have exactly 1 byte of unsent data,
377 			 * then it won't clear the FIN bit below,
378 			 * and if we are in persist state, we wind
379 			 * up sending the packet without recording
380 			 * that we sent the FIN bit.
381 			 *
382 			 * We can't just blindly clear the FIN bit,
383 			 * because if we don't have any more data
384 			 * to send then the probe will be the FIN
385 			 * itself.
386 			 */
387 			if (off < sbused(&so->so_snd))
388 				flags &= ~TH_FIN;
389 			sendwin = 1;
390 		} else {
391 			tcp_timer_activate(tp, TT_PERSIST, 0);
392 			tp->t_rxtshift = 0;
393 		}
394 	}
395 
396 	/*
397 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
398 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
399 	 * a negative length.  This can also occur when TCP opens up
400 	 * its congestion window while receiving additional duplicate
401 	 * acks after fast-retransmit because TCP will reset snd_nxt
402 	 * to snd_max after the fast-retransmit.
403 	 *
404 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
405 	 * be set to snd_una, the offset will be 0, and the length may
406 	 * wind up 0.
407 	 *
408 	 * If sack_rxmit is true we are retransmitting from the scoreboard
409 	 * in which case len is already set.
410 	 */
411 	if (sack_rxmit == 0) {
412 		if ((sack_bytes_rxmt == 0) || SEQ_LT(tp->snd_nxt, tp->snd_max)) {
413 			len = imin(sbavail(&so->so_snd), sendwin) - off;
414 		} else {
415 			/*
416 			 * We are inside of a SACK recovery episode and are
417 			 * sending new data, having retransmitted all the
418 			 * data possible in the scoreboard.
419 			 */
420 			len = imin(sbavail(&so->so_snd) - off,
421 				sendwin - tcp_compute_pipe(tp));
422 		}
423 	}
424 
425 	/*
426 	 * Lop off SYN bit if it has already been sent.  However, if this
427 	 * is SYN-SENT state and if segment contains data and if we don't
428 	 * know that foreign host supports TAO, suppress sending segment.
429 	 */
430 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
431 		if (tp->t_state != TCPS_SYN_RECEIVED)
432 			flags &= ~TH_SYN;
433 		/*
434 		 * When sending additional segments following a TFO SYN|ACK,
435 		 * do not include the SYN bit.
436 		 */
437 		if ((tp->t_flags & TF_FASTOPEN) &&
438 		    (tp->t_state == TCPS_SYN_RECEIVED))
439 			flags &= ~TH_SYN;
440 		off--, len++;
441 	}
442 
443 	/*
444 	 * Be careful not to send data and/or FIN on SYN segments.
445 	 * This measure is needed to prevent interoperability problems
446 	 * with not fully conformant TCP implementations.
447 	 */
448 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
449 		len = 0;
450 		flags &= ~TH_FIN;
451 	}
452 
453 	/*
454 	 * On TFO sockets, ensure no data is sent in the following cases:
455 	 *
456 	 *  - When retransmitting SYN|ACK on a passively-created socket
457 	 *
458 	 *  - When retransmitting SYN on an actively created socket
459 	 *
460 	 *  - When sending a zero-length cookie (cookie request) on an
461 	 *    actively created socket
462 	 *
463 	 *  - When the socket is in the CLOSED state (RST is being sent)
464 	 */
465 	if ((tp->t_flags & TF_FASTOPEN) &&
466 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
467 	     ((tp->t_state == TCPS_SYN_SENT) &&
468 	      (tp->t_tfo_client_cookie_len == 0)) ||
469 	     (flags & TH_RST)))
470 		len = 0;
471 
472 	/* Without fast-open there should never be data sent on a SYN. */
473 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
474 		len = 0;
475 	}
476 
477 	if (len <= 0) {
478 		/*
479 		 * If FIN has been sent but not acked,
480 		 * but we haven't been called to retransmit,
481 		 * len will be < 0.  Otherwise, window shrank
482 		 * after we sent into it.  If window shrank to 0,
483 		 * cancel pending retransmit, pull snd_nxt back
484 		 * to (closed) window, and set the persist timer
485 		 * if it isn't already going.  If the window didn't
486 		 * close completely, just wait for an ACK.
487 		 *
488 		 * We also do a general check here to ensure that
489 		 * we will set the persist timer when we have data
490 		 * to send, but a 0-byte window. This makes sure
491 		 * the persist timer is set even if the packet
492 		 * hits one of the "goto send" lines below.
493 		 */
494 		len = 0;
495 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
496 		    (off < (int) sbavail(&so->so_snd)) &&
497 		    !tcp_timer_active(tp, TT_PERSIST)) {
498 			tcp_timer_activate(tp, TT_REXMT, 0);
499 			tp->t_rxtshift = 0;
500 			tp->snd_nxt = tp->snd_una;
501 			if (!tcp_timer_active(tp, TT_PERSIST))
502 				tcp_setpersist(tp);
503 		}
504 	}
505 
506 	/* len will be >= 0 after this point. */
507 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
508 
509 	tcp_sndbuf_autoscale(tp, so, sendwin);
510 
511 	/*
512 	 * Decide if we can use TCP Segmentation Offloading (if supported by
513 	 * hardware).
514 	 *
515 	 * TSO may only be used if we are in a pure bulk sending state.  The
516 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
517 	 * IP options prevent using TSO.  With TSO the TCP header is the same
518 	 * (except for the sequence number) for all generated packets.  This
519 	 * makes it impossible to transmit any options which vary per generated
520 	 * segment or packet.
521 	 *
522 	 * IPv4 handling has a clear separation of ip options and ip header
523 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
524 	 * the right thing below to provide length of just ip options and thus
525 	 * checking for ipoptlen is enough to decide if ip options are present.
526 	 */
527 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
528 	/*
529 	 * Pre-calculate here as we save another lookup into the darknesses
530 	 * of IPsec that way and can actually decide if TSO is ok.
531 	 */
532 #ifdef INET6
533 	if (isipv6 && IPSEC_ENABLED(ipv6))
534 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
535 #ifdef INET
536 	else
537 #endif
538 #endif /* INET6 */
539 #ifdef INET
540 	if (IPSEC_ENABLED(ipv4))
541 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
542 #endif /* INET */
543 #endif /* IPSEC */
544 #ifdef INET6
545 	if (isipv6)
546 		ipoptlen = ip6_optlen(inp);
547 	else
548 #endif
549 	if (inp->inp_options)
550 		ipoptlen = inp->inp_options->m_len -
551 				offsetof(struct ipoption, ipopt_list);
552 	else
553 		ipoptlen = 0;
554 	ipoptlen += ipsec_optlen;
555 
556 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
557 	    (tp->t_port == 0) &&
558 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
559 	    tp->rcv_numsacks == 0 && ((sack_rxmit == 0) || V_tcp_sack_tso) &&
560 	    (ipoptlen == 0 || (ipoptlen == ipsec_optlen &&
561 	    (tp->t_flags2 & TF2_IPSEC_TSO) != 0)) &&
562 	    !(flags & TH_SYN))
563 		tso = 1;
564 
565 	if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
566 		    tp->snd_una + sbused(&so->so_snd))) {
567 			flags &= ~TH_FIN;
568 	}
569 
570 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
571 	    (long)TCP_MAXWIN << tp->rcv_scale);
572 
573 	/*
574 	 * Sender silly window avoidance.   We transmit under the following
575 	 * conditions when len is non-zero:
576 	 *
577 	 *	- We have a full segment (or more with TSO)
578 	 *	- This is the last buffer in a write()/send() and we are
579 	 *	  either idle or running NODELAY
580 	 *	- we've timed out (e.g. persist timer)
581 	 *	- we have more then 1/2 the maximum send window's worth of
582 	 *	  data (receiver may be limited the window size)
583 	 *	- we need to retransmit
584 	 */
585 	if (len) {
586 		if (len >= tp->t_maxseg)
587 			goto send;
588 		/*
589 		 * As the TCP header options are now
590 		 * considered when setting up the initial
591 		 * window, we would not send the last segment
592 		 * if we skip considering the option length here.
593 		 * Note: this may not work when tcp headers change
594 		 * very dynamically in the future.
595 		 */
596 		if ((((tp->t_flags & TF_SIGNATURE) ?
597 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
598 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
599 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
600 		    len) >= tp->t_maxseg)
601 			goto send;
602 		/*
603 		 * NOTE! on localhost connections an 'ack' from the remote
604 		 * end may occur synchronously with the output and cause
605 		 * us to flush a buffer queued with moretocome.  XXX
606 		 *
607 		 * note: the len + off check is almost certainly unnecessary.
608 		 */
609 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
610 		    (idle || (tp->t_flags & TF_NODELAY)) &&
611 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
612 		    (tp->t_flags & TF_NOPUSH) == 0) {
613 			goto send;
614 		}
615 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
616 			goto send;
617 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
618 			goto send;
619 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
620 			goto send;
621 		if (sack_rxmit)
622 			goto send;
623 	}
624 
625 	/*
626 	 * Sending of standalone window updates.
627 	 *
628 	 * Window updates are important when we close our window due to a
629 	 * full socket buffer and are opening it again after the application
630 	 * reads data from it.  Once the window has opened again and the
631 	 * remote end starts to send again the ACK clock takes over and
632 	 * provides the most current window information.
633 	 *
634 	 * We must avoid the silly window syndrome whereas every read
635 	 * from the receive buffer, no matter how small, causes a window
636 	 * update to be sent.  We also should avoid sending a flurry of
637 	 * window updates when the socket buffer had queued a lot of data
638 	 * and the application is doing small reads.
639 	 *
640 	 * Prevent a flurry of pointless window updates by only sending
641 	 * an update when we can increase the advertized window by more
642 	 * than 1/4th of the socket buffer capacity.  When the buffer is
643 	 * getting full or is very small be more aggressive and send an
644 	 * update whenever we can increase by two mss sized segments.
645 	 * In all other situations the ACK's to new incoming data will
646 	 * carry further window increases.
647 	 *
648 	 * Don't send an independent window update if a delayed
649 	 * ACK is pending (it will get piggy-backed on it) or the
650 	 * remote side already has done a half-close and won't send
651 	 * more data.
652 	 */
653 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
654 	    !(tp->t_flags & TF_DELACK) &&
655 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
656 		/*
657 		 * "adv" is the amount we could increase the window,
658 		 * taking into account that we are limited by
659 		 * TCP_MAXWIN << tp->rcv_scale.
660 		 */
661 		int32_t adv;
662 		int oldwin;
663 
664 		adv = recwin;
665 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
666 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
667 			if (adv > oldwin)
668 				adv -= oldwin;
669 			else
670 				adv = 0;
671 		} else
672 			oldwin = 0;
673 
674 		/*
675 		 * If the new window size ends up being the same as or less
676 		 * than the old size when it is scaled, then don't force
677 		 * a window update.
678 		 */
679 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
680 			goto dontupdate;
681 
682 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
683 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
684 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
685 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
686 		     adv >= TCP_MAXWIN << tp->rcv_scale))
687 			goto send;
688 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
689 			goto send;
690 	}
691 dontupdate:
692 
693 	/*
694 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
695 	 * is also a catch-all for the retransmit timer timeout case.
696 	 */
697 	if (tp->t_flags & TF_ACKNOW)
698 		goto send;
699 	if ((flags & TH_RST) ||
700 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
701 		goto send;
702 	if (SEQ_GT(tp->snd_up, tp->snd_una))
703 		goto send;
704 	/*
705 	 * If our state indicates that FIN should be sent
706 	 * and we have not yet done so, then we need to send.
707 	 */
708 	if (flags & TH_FIN &&
709 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
710 		goto send;
711 	/*
712 	 * In SACK, it is possible for tcp_output to fail to send a segment
713 	 * after the retransmission timer has been turned off.  Make sure
714 	 * that the retransmission timer is set.
715 	 */
716 	if ((tp->t_flags & TF_SACK_PERMIT) &&
717 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
718 	    !tcp_timer_active(tp, TT_REXMT) &&
719 	    !tcp_timer_active(tp, TT_PERSIST)) {
720 		tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
721 		goto just_return;
722 	}
723 	/*
724 	 * TCP window updates are not reliable, rather a polling protocol
725 	 * using ``persist'' packets is used to insure receipt of window
726 	 * updates.  The three ``states'' for the output side are:
727 	 *	idle			not doing retransmits or persists
728 	 *	persisting		to move a small or zero window
729 	 *	(re)transmitting	and thereby not persisting
730 	 *
731 	 * tcp_timer_active(tp, TT_PERSIST)
732 	 *	is true when we are in persist state.
733 	 * (tp->t_flags & TF_FORCEDATA)
734 	 *	is set when we are called to send a persist packet.
735 	 * tcp_timer_active(tp, TT_REXMT)
736 	 *	is set when we are retransmitting
737 	 * The output side is idle when both timers are zero.
738 	 *
739 	 * If send window is too small, there is data to transmit, and no
740 	 * retransmit or persist is pending, then go to persist state.
741 	 * If nothing happens soon, send when timer expires:
742 	 * if window is nonzero, transmit what we can,
743 	 * otherwise force out a byte.
744 	 */
745 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
746 	    !tcp_timer_active(tp, TT_PERSIST)) {
747 		tp->t_rxtshift = 0;
748 		tcp_setpersist(tp);
749 	}
750 
751 	/*
752 	 * No reason to send a segment, just return.
753 	 */
754 just_return:
755 	SOCKBUF_UNLOCK(&so->so_snd);
756 	return (0);
757 
758 send:
759 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
760 	if (len > 0) {
761 		if (len >= tp->t_maxseg)
762 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
763 		else
764 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
765 	}
766 	/*
767 	 * Before ESTABLISHED, force sending of initial options
768 	 * unless TCP set not to do any options.
769 	 * NOTE: we assume that the IP/TCP header plus TCP options
770 	 * always fit in a single mbuf, leaving room for a maximum
771 	 * link header, i.e.
772 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
773 	 */
774 	optlen = 0;
775 #ifdef INET6
776 	if (isipv6)
777 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
778 	else
779 #endif
780 		hdrlen = sizeof (struct tcpiphdr);
781 
782 	if (flags & TH_SYN) {
783 		tp->snd_nxt = tp->iss;
784 	}
785 
786 	/*
787 	 * Compute options for segment.
788 	 * We only have to care about SYN and established connection
789 	 * segments.  Options for SYN-ACK segments are handled in TCP
790 	 * syncache.
791 	 */
792 	to.to_flags = 0;
793 	if ((tp->t_flags & TF_NOOPT) == 0) {
794 		/* Maximum segment size. */
795 		if (flags & TH_SYN) {
796 			to.to_mss = tcp_mssopt(&inp->inp_inc);
797 			if (tp->t_port)
798 				to.to_mss -= V_tcp_udp_tunneling_overhead;
799 			to.to_flags |= TOF_MSS;
800 
801 			/*
802 			 * On SYN or SYN|ACK transmits on TFO connections,
803 			 * only include the TFO option if it is not a
804 			 * retransmit, as the presence of the TFO option may
805 			 * have caused the original SYN or SYN|ACK to have
806 			 * been dropped by a middlebox.
807 			 */
808 			if ((tp->t_flags & TF_FASTOPEN) &&
809 			    (tp->t_rxtshift == 0)) {
810 				if (tp->t_state == TCPS_SYN_RECEIVED) {
811 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
812 					to.to_tfo_cookie =
813 					    (u_int8_t *)&tp->t_tfo_cookie.server;
814 					to.to_flags |= TOF_FASTOPEN;
815 					wanted_cookie = 1;
816 				} else if (tp->t_state == TCPS_SYN_SENT) {
817 					to.to_tfo_len =
818 					    tp->t_tfo_client_cookie_len;
819 					to.to_tfo_cookie =
820 					    tp->t_tfo_cookie.client;
821 					to.to_flags |= TOF_FASTOPEN;
822 					wanted_cookie = 1;
823 					/*
824 					 * If we wind up having more data to
825 					 * send with the SYN than can fit in
826 					 * one segment, don't send any more
827 					 * until the SYN|ACK comes back from
828 					 * the other end.
829 					 */
830 					dont_sendalot = 1;
831 				}
832 			}
833 		}
834 		/* Window scaling. */
835 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
836 			to.to_wscale = tp->request_r_scale;
837 			to.to_flags |= TOF_SCALE;
838 		}
839 		/* Timestamps. */
840 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
841 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
842 			curticks = tcp_ts_getticks();
843 			to.to_tsval = curticks + tp->ts_offset;
844 			to.to_tsecr = tp->ts_recent;
845 			to.to_flags |= TOF_TS;
846 			if (tp->t_rxtshift == 1)
847 				tp->t_badrxtwin = curticks;
848 		}
849 
850 		/* Set receive buffer autosizing timestamp. */
851 		if (tp->rfbuf_ts == 0 &&
852 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
853 			tp->rfbuf_ts = tcp_ts_getticks();
854 
855 		/* Selective ACK's. */
856 		if (tp->t_flags & TF_SACK_PERMIT) {
857 			if (flags & TH_SYN)
858 				to.to_flags |= TOF_SACKPERM;
859 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
860 			    tp->rcv_numsacks > 0) {
861 				to.to_flags |= TOF_SACK;
862 				to.to_nsacks = tp->rcv_numsacks;
863 				to.to_sacks = (u_char *)tp->sackblks;
864 			}
865 		}
866 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
867 		/* TCP-MD5 (RFC2385). */
868 		/*
869 		 * Check that TCP_MD5SIG is enabled in tcpcb to
870 		 * account the size needed to set this TCP option.
871 		 */
872 		if (tp->t_flags & TF_SIGNATURE)
873 			to.to_flags |= TOF_SIGNATURE;
874 #endif /* TCP_SIGNATURE */
875 
876 		/* Processing the options. */
877 		hdrlen += optlen = tcp_addoptions(&to, opt);
878 		/*
879 		 * If we wanted a TFO option to be added, but it was unable
880 		 * to fit, ensure no data is sent.
881 		 */
882 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
883 		    !(to.to_flags & TOF_FASTOPEN))
884 			len = 0;
885 	}
886 	if (tp->t_port) {
887 		if (V_tcp_udp_tunneling_port == 0) {
888 			/* The port was removed?? */
889 			SOCKBUF_UNLOCK(&so->so_snd);
890 			return (EHOSTUNREACH);
891 		}
892 		hdrlen += sizeof(struct udphdr);
893 	}
894 	/*
895 	 * Adjust data length if insertion of options will
896 	 * bump the packet length beyond the t_maxseg length.
897 	 * Clear the FIN bit because we cut off the tail of
898 	 * the segment.
899 	 */
900 	if (len + optlen + ipoptlen > tp->t_maxseg) {
901 		flags &= ~TH_FIN;
902 
903 		if (tso) {
904 			u_int if_hw_tsomax;
905 			u_int moff;
906 			int max_len;
907 
908 			/* extract TSO information */
909 			if_hw_tsomax = tp->t_tsomax;
910 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
911 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
912 
913 			/*
914 			 * Limit a TSO burst to prevent it from
915 			 * overflowing or exceeding the maximum length
916 			 * allowed by the network interface:
917 			 */
918 			KASSERT(ipoptlen ==  ipsec_optlen,
919 			    ("%s: TSO can't do IP options", __func__));
920 
921 			/*
922 			 * Check if we should limit by maximum payload
923 			 * length:
924 			 */
925 			if (if_hw_tsomax != 0) {
926 				/* compute maximum TSO length */
927 				max_len = if_hw_tsomax - hdrlen -
928 				    ipsec_optlen - max_linkhdr;
929 				if (max_len <= 0) {
930 					len = 0;
931 				} else if (len > max_len) {
932 					sendalot = 1;
933 					len = max_len;
934 				}
935 			}
936 
937 			/*
938 			 * Prevent the last segment from being
939 			 * fractional unless the send sockbuf can be
940 			 * emptied:
941 			 */
942 			max_len = tp->t_maxseg - optlen - ipsec_optlen;
943 			if (((uint32_t)off + (uint32_t)len) <
944 			    sbavail(&so->so_snd)) {
945 				moff = len % max_len;
946 				if (moff != 0) {
947 					len -= moff;
948 					sendalot = 1;
949 				}
950 			}
951 
952 			/*
953 			 * In case there are too many small fragments
954 			 * don't use TSO:
955 			 */
956 			if (len <= max_len) {
957 				len = max_len;
958 				sendalot = 1;
959 				tso = 0;
960 			}
961 
962 			/*
963 			 * Send the FIN in a separate segment
964 			 * after the bulk sending is done.
965 			 * We don't trust the TSO implementations
966 			 * to clear the FIN flag on all but the
967 			 * last segment.
968 			 */
969 			if (tp->t_flags & TF_NEEDFIN)
970 				sendalot = 1;
971 		} else {
972 			if (optlen + ipoptlen >= tp->t_maxseg) {
973 				/*
974 				 * Since we don't have enough space to put
975 				 * the IP header chain and the TCP header in
976 				 * one packet as required by RFC 7112, don't
977 				 * send it. Also ensure that at least one
978 				 * byte of the payload can be put into the
979 				 * TCP segment.
980 				 */
981 				SOCKBUF_UNLOCK(&so->so_snd);
982 				error = EMSGSIZE;
983 				sack_rxmit = 0;
984 				goto out;
985 			}
986 			len = tp->t_maxseg - optlen - ipoptlen;
987 			sendalot = 1;
988 			if (dont_sendalot)
989 				sendalot = 0;
990 		}
991 	} else
992 		tso = 0;
993 
994 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
995 	    ("%s: len > IP_MAXPACKET", __func__));
996 
997 /*#ifdef DIAGNOSTIC*/
998 #ifdef INET6
999 	if (max_linkhdr + hdrlen > MCLBYTES)
1000 #else
1001 	if (max_linkhdr + hdrlen > MHLEN)
1002 #endif
1003 		panic("tcphdr too big");
1004 /*#endif*/
1005 
1006 	/*
1007 	 * This KASSERT is here to catch edge cases at a well defined place.
1008 	 * Before, those had triggered (random) panic conditions further down.
1009 	 */
1010 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1011 
1012 	/*
1013 	 * Grab a header mbuf, attaching a copy of data to
1014 	 * be transmitted, and initialize the header from
1015 	 * the template for sends on this connection.
1016 	 */
1017 	if (len) {
1018 		struct mbuf *mb;
1019 		struct sockbuf *msb;
1020 		u_int moff;
1021 
1022 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1023 			TCPSTAT_INC(tcps_sndprobe);
1024 #ifdef STATS
1025 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1026 				stats_voi_update_abs_u32(tp->t_stats,
1027 				VOI_TCP_RETXPB, len);
1028 			else
1029 				stats_voi_update_abs_u64(tp->t_stats,
1030 				    VOI_TCP_TXPB, len);
1031 #endif /* STATS */
1032 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1033 			tp->t_sndrexmitpack++;
1034 			TCPSTAT_INC(tcps_sndrexmitpack);
1035 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1036 			if (sack_rxmit) {
1037 				TCPSTAT_INC(tcps_sack_rexmits);
1038 				if (tso) {
1039 					TCPSTAT_INC(tcps_sack_rexmits_tso);
1040 				}
1041 				TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1042 			}
1043 #ifdef STATS
1044 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1045 			    len);
1046 #endif /* STATS */
1047 		} else {
1048 			TCPSTAT_INC(tcps_sndpack);
1049 			TCPSTAT_ADD(tcps_sndbyte, len);
1050 #ifdef STATS
1051 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1052 			    len);
1053 #endif /* STATS */
1054 		}
1055 #ifdef INET6
1056 		if (MHLEN < hdrlen + max_linkhdr)
1057 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1058 		else
1059 #endif
1060 			m = m_gethdr(M_NOWAIT, MT_DATA);
1061 
1062 		if (m == NULL) {
1063 			SOCKBUF_UNLOCK(&so->so_snd);
1064 			error = ENOBUFS;
1065 			sack_rxmit = 0;
1066 			goto out;
1067 		}
1068 
1069 		m->m_data += max_linkhdr;
1070 		m->m_len = hdrlen;
1071 
1072 		/*
1073 		 * Start the m_copy functions from the closest mbuf
1074 		 * to the offset in the socket buffer chain.
1075 		 */
1076 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1077 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1078 			m_copydata(mb, moff, len,
1079 			    mtod(m, caddr_t) + hdrlen);
1080 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1081 				sbsndptr_adv(&so->so_snd, mb, len);
1082 			m->m_len += len;
1083 		} else {
1084 			int32_t old_len;
1085 
1086 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1087 				msb = NULL;
1088 			else
1089 				msb = &so->so_snd;
1090 			old_len = len;
1091 			m->m_next = tcp_m_copym(mb, moff,
1092 			    &len, if_hw_tsomaxsegcount,
1093 			    if_hw_tsomaxsegsize, msb, hw_tls);
1094 			if (old_len != len)
1095 				flags &= ~TH_FIN;
1096 			if (len <= (tp->t_maxseg - optlen)) {
1097 				/*
1098 				 * Must have ran out of mbufs for the copy
1099 				 * shorten it to no longer need tso. Lets
1100 				 * not put on sendalot since we are low on
1101 				 * mbufs.
1102 				 */
1103 				tso = 0;
1104 			}
1105 			if (m->m_next == NULL) {
1106 				SOCKBUF_UNLOCK(&so->so_snd);
1107 				(void) m_free(m);
1108 				error = ENOBUFS;
1109 				sack_rxmit = 0;
1110 				goto out;
1111 			}
1112 		}
1113 
1114 		/*
1115 		 * If we're sending everything we've got, set PUSH.
1116 		 * (This will keep happy those implementations which only
1117 		 * give data to the user when a buffer fills or
1118 		 * a PUSH comes in.)
1119 		 */
1120 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1121 		    !(flags & TH_SYN))
1122 			flags |= TH_PUSH;
1123 		SOCKBUF_UNLOCK(&so->so_snd);
1124 	} else {
1125 		SOCKBUF_UNLOCK(&so->so_snd);
1126 		if (tp->t_flags & TF_ACKNOW)
1127 			TCPSTAT_INC(tcps_sndacks);
1128 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1129 			TCPSTAT_INC(tcps_sndctrl);
1130 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1131 			TCPSTAT_INC(tcps_sndurg);
1132 		else
1133 			TCPSTAT_INC(tcps_sndwinup);
1134 
1135 		m = m_gethdr(M_NOWAIT, MT_DATA);
1136 		if (m == NULL) {
1137 			error = ENOBUFS;
1138 			sack_rxmit = 0;
1139 			goto out;
1140 		}
1141 #ifdef INET6
1142 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1143 		    MHLEN >= hdrlen) {
1144 			M_ALIGN(m, hdrlen);
1145 		} else
1146 #endif
1147 		m->m_data += max_linkhdr;
1148 		m->m_len = hdrlen;
1149 	}
1150 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1151 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1152 #ifdef MAC
1153 	mac_inpcb_create_mbuf(inp, m);
1154 #endif
1155 #ifdef INET6
1156 	if (isipv6) {
1157 		ip6 = mtod(m, struct ip6_hdr *);
1158 		if (tp->t_port) {
1159 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1160 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1161 			udp->uh_dport = tp->t_port;
1162 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1163 			udp->uh_ulen = htons(ulen);
1164 			th = (struct tcphdr *)(udp + 1);
1165 		} else {
1166 			th = (struct tcphdr *)(ip6 + 1);
1167 		}
1168 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
1169 	} else
1170 #endif /* INET6 */
1171 	{
1172 		ip = mtod(m, struct ip *);
1173 		if (tp->t_port) {
1174 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1175 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1176 			udp->uh_dport = tp->t_port;
1177 			ulen = hdrlen + len - sizeof(struct ip);
1178 			udp->uh_ulen = htons(ulen);
1179 			th = (struct tcphdr *)(udp + 1);
1180 		} else
1181 			th = (struct tcphdr *)(ip + 1);
1182 		tcpip_fillheaders(inp, tp->t_port, ip, th);
1183 	}
1184 
1185 	/*
1186 	 * Fill in fields, remembering maximum advertised
1187 	 * window for use in delaying messages about window sizes.
1188 	 * If resending a FIN, be sure not to use a new sequence number.
1189 	 */
1190 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1191 	    tp->snd_nxt == tp->snd_max)
1192 		tp->snd_nxt--;
1193 	/*
1194 	 * If we are starting a connection, send ECN setup
1195 	 * SYN packet. If we are on a retransmit, we may
1196 	 * resend those bits a number of times as per
1197 	 * RFC 3168.
1198 	 */
1199 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1200 		flags |= tcp_ecn_output_syn_sent(tp);
1201 	}
1202 	/* Also handle parallel SYN for ECN */
1203 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1204 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1205 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1206 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1207 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1208 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1209 #ifdef INET6
1210 		if (isipv6) {
1211 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1212 			ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1213 		}
1214 		else
1215 #endif
1216 		{
1217 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1218 			ip->ip_tos |= ect;
1219 		}
1220 	}
1221 
1222 	/*
1223 	 * If we are doing retransmissions, then snd_nxt will
1224 	 * not reflect the first unsent octet.  For ACK only
1225 	 * packets, we do not want the sequence number of the
1226 	 * retransmitted packet, we want the sequence number
1227 	 * of the next unsent octet.  So, if there is no data
1228 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1229 	 * when filling in ti_seq.  But if we are in persist
1230 	 * state, snd_max might reflect one byte beyond the
1231 	 * right edge of the window, so use snd_nxt in that
1232 	 * case, since we know we aren't doing a retransmission.
1233 	 * (retransmit and persist are mutually exclusive...)
1234 	 */
1235 	if (sack_rxmit == 0) {
1236 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1237 		    tcp_timer_active(tp, TT_PERSIST))
1238 			th->th_seq = htonl(tp->snd_nxt);
1239 		else
1240 			th->th_seq = htonl(tp->snd_max);
1241 	} else {
1242 		th->th_seq = htonl(p->rxmit);
1243 		p->rxmit += len;
1244 		/*
1245 		 * Lost Retransmission Detection
1246 		 * trigger resending of a (then
1247 		 * still existing) hole, when
1248 		 * fack acks recoverypoint.
1249 		 */
1250 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1251 			p->rxmit = tp->snd_recover;
1252 		tp->sackhint.sack_bytes_rexmit += len;
1253 	}
1254 	if (IN_RECOVERY(tp->t_flags)) {
1255 		/*
1256 		 * Account all bytes transmitted while
1257 		 * IN_RECOVERY, simplifying PRR and
1258 		 * Lost Retransmit Detection
1259 		 */
1260 		tp->sackhint.prr_out += len;
1261 	}
1262 	th->th_ack = htonl(tp->rcv_nxt);
1263 	if (optlen) {
1264 		bcopy(opt, th + 1, optlen);
1265 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1266 	}
1267 	tcp_set_flags(th, flags);
1268 	/*
1269 	 * Calculate receive window.  Don't shrink window,
1270 	 * but avoid silly window syndrome.
1271 	 * If a RST segment is sent, advertise a window of zero.
1272 	 */
1273 	if (flags & TH_RST) {
1274 		recwin = 0;
1275 	} else {
1276 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1277 		    recwin < tp->t_maxseg)
1278 			recwin = 0;
1279 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1280 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1281 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1282 	}
1283 	/*
1284 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1285 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1286 	 * case is handled in syncache.
1287 	 */
1288 	if (flags & TH_SYN)
1289 		th->th_win = htons((u_short)
1290 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1291 	else {
1292 		/* Avoid shrinking window with window scaling. */
1293 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1294 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1295 	}
1296 
1297 	/*
1298 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1299 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1300 	 * flag tells soreceive() to disable delayed acknowledgements when
1301 	 * draining the buffer.  This can occur if the receiver is attempting
1302 	 * to read more data than can be buffered prior to transmitting on
1303 	 * the connection.
1304 	 */
1305 	if (th->th_win == 0) {
1306 		tp->t_sndzerowin++;
1307 		tp->t_flags |= TF_RXWIN0SENT;
1308 	} else
1309 		tp->t_flags &= ~TF_RXWIN0SENT;
1310 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1311 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1312 		th->th_flags |= TH_URG;
1313 	} else
1314 		/*
1315 		 * If no urgent pointer to send, then we pull
1316 		 * the urgent pointer to the left edge of the send window
1317 		 * so that it doesn't drift into the send window on sequence
1318 		 * number wraparound.
1319 		 */
1320 		tp->snd_up = tp->snd_una;		/* drag it along */
1321 
1322 	/*
1323 	 * Put TCP length in extended header, and then
1324 	 * checksum extended header and data.
1325 	 */
1326 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1327 
1328 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1329 	if (to.to_flags & TOF_SIGNATURE) {
1330 		/*
1331 		 * Calculate MD5 signature and put it into the place
1332 		 * determined before.
1333 		 * NOTE: since TCP options buffer doesn't point into
1334 		 * mbuf's data, calculate offset and use it.
1335 		 */
1336 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1337 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1338 			/*
1339 			 * Do not send segment if the calculation of MD5
1340 			 * digest has failed.
1341 			 */
1342 			m_freem(m);
1343 			goto out;
1344 		}
1345 	}
1346 #endif
1347 #ifdef INET6
1348 	if (isipv6) {
1349 		/*
1350 		 * There is no need to fill in ip6_plen right now.
1351 		 * It will be filled later by ip6_output.
1352 		 */
1353 		if (tp->t_port) {
1354 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1355 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1356 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1357 			th->th_sum = htons(0);
1358 			UDPSTAT_INC(udps_opackets);
1359 		} else {
1360 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1361 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1362 			th->th_sum = in6_cksum_pseudo(ip6,
1363 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1364 			    0);
1365 		}
1366 	}
1367 #endif
1368 #if defined(INET6) && defined(INET)
1369 	else
1370 #endif
1371 #ifdef INET
1372 	{
1373 		if (tp->t_port) {
1374 			m->m_pkthdr.csum_flags = CSUM_UDP;
1375 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1376 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1377 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1378 			th->th_sum = htons(0);
1379 			UDPSTAT_INC(udps_opackets);
1380 		} else {
1381 			m->m_pkthdr.csum_flags = CSUM_TCP;
1382 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1383 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1384 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1385 			    IPPROTO_TCP + len + optlen));
1386 		}
1387 
1388 		/* IP version must be set here for ipv4/ipv6 checking later */
1389 		KASSERT(ip->ip_v == IPVERSION,
1390 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1391 	}
1392 #endif
1393 
1394 	/*
1395 	 * Enable TSO and specify the size of the segments.
1396 	 * The TCP pseudo header checksum is always provided.
1397 	 */
1398 	if (tso) {
1399 		KASSERT(len > tp->t_maxseg - optlen - ipsec_optlen,
1400 		    ("%s: len <= tso_segsz", __func__));
1401 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1402 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen - ipsec_optlen;
1403 	}
1404 
1405 	KASSERT(len + hdrlen == m_length(m, NULL),
1406 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1407 	    __func__, len, hdrlen, m_length(m, NULL)));
1408 
1409 #ifdef TCP_HHOOK
1410 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1411 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1412 #endif
1413 
1414 	TCP_PROBE3(debug__output, tp, th, m);
1415 
1416 	/* We're getting ready to send; log now. */
1417 	/* XXXMT: We are not honoring verbose logging. */
1418 
1419 	if (tcp_bblogging_on(tp))
1420 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1421 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1422 		    NULL);
1423 	else
1424 		lgb = NULL;
1425 
1426 	/*
1427 	 * Fill in IP length and desired time to live and
1428 	 * send to IP level.  There should be a better way
1429 	 * to handle ttl and tos; we could keep them in
1430 	 * the template, but need a way to checksum without them.
1431 	 */
1432 	/*
1433 	 * m->m_pkthdr.len should have been set before checksum calculation,
1434 	 * because in6_cksum() need it.
1435 	 */
1436 #ifdef INET6
1437 	if (isipv6) {
1438 		/*
1439 		 * we separately set hoplimit for every segment, since the
1440 		 * user might want to change the value via setsockopt.
1441 		 * Also, desired default hop limit might be changed via
1442 		 * Neighbor Discovery.
1443 		 */
1444 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1445 
1446 		/*
1447 		 * Set the packet size here for the benefit of DTrace probes.
1448 		 * ip6_output() will set it properly; it's supposed to include
1449 		 * the option header lengths as well.
1450 		 */
1451 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1452 
1453 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1454 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1455 		else
1456 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1457 
1458 		if (tp->t_state == TCPS_SYN_SENT)
1459 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1460 
1461 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1462 
1463 #ifdef TCPPCAP
1464 		/* Save packet, if requested. */
1465 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1466 #endif
1467 
1468 		/* TODO: IPv6 IP6TOS_ECT bit on */
1469 		error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1470 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1471 		    NULL, NULL, inp);
1472 
1473 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1474 			mtu = inp->inp_route6.ro_nh->nh_mtu;
1475 	}
1476 #endif /* INET6 */
1477 #if defined(INET) && defined(INET6)
1478 	else
1479 #endif
1480 #ifdef INET
1481     {
1482 	ip->ip_len = htons(m->m_pkthdr.len);
1483 #ifdef INET6
1484 	if (inp->inp_vflag & INP_IPV6PROTO)
1485 		ip->ip_ttl = in6_selecthlim(inp, NULL);
1486 #endif /* INET6 */
1487 	/*
1488 	 * If we do path MTU discovery, then we set DF on every packet.
1489 	 * This might not be the best thing to do according to RFC3390
1490 	 * Section 2. However the tcp hostcache migitates the problem
1491 	 * so it affects only the first tcp connection with a host.
1492 	 *
1493 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1494 	 */
1495 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1496 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1497 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1498 			ip->ip_off |= htons(IP_DF);
1499 		}
1500 	} else {
1501 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1502 	}
1503 
1504 	if (tp->t_state == TCPS_SYN_SENT)
1505 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1506 
1507 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1508 
1509 #ifdef TCPPCAP
1510 	/* Save packet, if requested. */
1511 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1512 #endif
1513 
1514 	error = ip_output(m, inp->inp_options, &inp->inp_route,
1515 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1516 
1517 	if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1518 		mtu = inp->inp_route.ro_nh->nh_mtu;
1519     }
1520 #endif /* INET */
1521 
1522 	if (lgb != NULL) {
1523 		lgb->tlb_errno = error;
1524 		lgb = NULL;
1525 	}
1526 out:
1527 	if (error == 0)
1528 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1529 	/*
1530 	 * In transmit state, time the transmission and arrange for
1531 	 * the retransmit.  In persist state, just set snd_max.  In a closed
1532 	 * state just return.
1533 	 */
1534 	if (flags & TH_RST) {
1535 		TCPSTAT_INC(tcps_sndtotal);
1536 		return (0);
1537 	} else if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1538 	    !tcp_timer_active(tp, TT_PERSIST)) {
1539 		tcp_seq startseq = tp->snd_nxt;
1540 
1541 		/*
1542 		 * Advance snd_nxt over sequence space of this segment.
1543 		 */
1544 		if (flags & (TH_SYN|TH_FIN)) {
1545 			if (flags & TH_SYN)
1546 				tp->snd_nxt++;
1547 			if (flags & TH_FIN) {
1548 				tp->snd_nxt++;
1549 				tp->t_flags |= TF_SENTFIN;
1550 			}
1551 		}
1552 		if (sack_rxmit)
1553 			goto timer;
1554 		tp->snd_nxt += len;
1555 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1556 			/*
1557 			 * Update "made progress" indication if we just
1558 			 * added new data to an empty socket buffer.
1559 			 */
1560 			if (tp->snd_una == tp->snd_max)
1561 				tp->t_acktime = ticks;
1562 			tp->snd_max = tp->snd_nxt;
1563 			/*
1564 			 * Time this transmission if not a retransmission and
1565 			 * not currently timing anything.
1566 			 */
1567 			tp->t_sndtime = ticks;
1568 			if (tp->t_rtttime == 0) {
1569 				tp->t_rtttime = ticks;
1570 				tp->t_rtseq = startseq;
1571 				TCPSTAT_INC(tcps_segstimed);
1572 			}
1573 #ifdef STATS
1574 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1575 				tp->t_flags |= TF_GPUTINPROG;
1576 				tp->gput_seq = startseq;
1577 				tp->gput_ack = startseq +
1578 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1579 				tp->gput_ts = tcp_ts_getticks();
1580 			}
1581 #endif /* STATS */
1582 		}
1583 
1584 		/*
1585 		 * Set retransmit timer if not currently set,
1586 		 * and not doing a pure ack or a keep-alive probe.
1587 		 * Initial value for retransmit timer is smoothed
1588 		 * round-trip time + 2 * round-trip time variance.
1589 		 * Initialize shift counter which is used for backoff
1590 		 * of retransmit time.
1591 		 */
1592 timer:
1593 		if (!tcp_timer_active(tp, TT_REXMT) &&
1594 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1595 		     (tp->snd_nxt != tp->snd_una))) {
1596 			if (tcp_timer_active(tp, TT_PERSIST)) {
1597 				tcp_timer_activate(tp, TT_PERSIST, 0);
1598 				tp->t_rxtshift = 0;
1599 			}
1600 			tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1601 		} else if (len == 0 && sbavail(&so->so_snd) &&
1602 		    !tcp_timer_active(tp, TT_REXMT) &&
1603 		    !tcp_timer_active(tp, TT_PERSIST)) {
1604 			/*
1605 			 * Avoid a situation where we do not set persist timer
1606 			 * after a zero window condition. For example:
1607 			 * 1) A -> B: packet with enough data to fill the window
1608 			 * 2) B -> A: ACK for #1 + new data (0 window
1609 			 *    advertisement)
1610 			 * 3) A -> B: ACK for #2, 0 len packet
1611 			 *
1612 			 * In this case, A will not activate the persist timer,
1613 			 * because it chose to send a packet. Unless tcp_output
1614 			 * is called for some other reason (delayed ack timer,
1615 			 * another input packet from B, socket syscall), A will
1616 			 * not send zero window probes.
1617 			 *
1618 			 * So, if you send a 0-length packet, but there is data
1619 			 * in the socket buffer, and neither the rexmt or
1620 			 * persist timer is already set, then activate the
1621 			 * persist timer.
1622 			 */
1623 			tp->t_rxtshift = 0;
1624 			tcp_setpersist(tp);
1625 		}
1626 	} else {
1627 		/*
1628 		 * Persist case, update snd_max but since we are in
1629 		 * persist mode (no window) we do not update snd_nxt.
1630 		 */
1631 		int xlen = len;
1632 		if (flags & TH_SYN)
1633 			++xlen;
1634 		if (flags & TH_FIN) {
1635 			++xlen;
1636 			tp->t_flags |= TF_SENTFIN;
1637 		}
1638 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1639 			tp->snd_max = tp->snd_nxt + xlen;
1640 	}
1641 	if ((error == 0) &&
1642 	    (tp->rcv_numsacks > 0) &&
1643 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
1644 	    (tp->t_flags & TF_SACK_PERMIT)) {
1645 		/* Clean up any DSACK's sent */
1646 		tcp_clean_dsack_blocks(tp);
1647 	}
1648 	if ((error == 0) &&
1649 	    sack_rxmit &&
1650 	    SEQ_LT(tp->snd_nxt, SEQ_MIN(p->rxmit, p->end))) {
1651 		/*
1652 		 * When transmitting from SACK scoreboard
1653 		 * after an RTO, pull snd_nxt along.
1654 		 */
1655 		tp->snd_nxt = SEQ_MIN(p->rxmit, p->end);
1656 	}
1657 	if (error) {
1658 		/*
1659 		 * We know that the packet was lost, so back out the
1660 		 * sequence number advance, if any.
1661 		 *
1662 		 * If the error is EPERM the packet got blocked by the
1663 		 * local firewall.  Normally we should terminate the
1664 		 * connection but the blocking may have been spurious
1665 		 * due to a firewall reconfiguration cycle.  So we treat
1666 		 * it like a packet loss and let the retransmit timer and
1667 		 * timeouts do their work over time.
1668 		 * XXX: It is a POLA question whether calling tcp_drop right
1669 		 * away would be the really correct behavior instead.
1670 		 */
1671 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1672 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1673 		    ((flags & TH_SYN) == 0) &&
1674 		    (error != EPERM)) {
1675 			if (sack_rxmit) {
1676 				p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1677 				tp->sackhint.sack_bytes_rexmit -= len;
1678 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1679 				    ("sackhint bytes rtx >= 0"));
1680 				KASSERT((flags & TH_FIN) == 0,
1681 				    ("error while FIN with SACK rxmit"));
1682 			} else {
1683 				tp->snd_nxt -= len;
1684 				if (flags & TH_FIN)
1685 					tp->snd_nxt--;
1686 			}
1687 			if (IN_RECOVERY(tp->t_flags))
1688 				tp->sackhint.prr_out -= len;
1689 		}
1690 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1691 		switch (error) {
1692 		case EACCES:
1693 		case EPERM:
1694 			tp->t_softerror = error;
1695 			return (error);
1696 		case ENOBUFS:
1697 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1698 			tp->snd_cwnd = tp->t_maxseg;
1699 			return (0);
1700 		case EMSGSIZE:
1701 			/*
1702 			 * For some reason the interface we used initially
1703 			 * to send segments changed to another or lowered
1704 			 * its MTU.
1705 			 * If TSO was active we either got an interface
1706 			 * without TSO capabilits or TSO was turned off.
1707 			 * If we obtained mtu from ip_output() then update
1708 			 * it and try again.
1709 			 */
1710 			if (tso)
1711 				tp->t_flags &= ~TF_TSO;
1712 			if (mtu != 0) {
1713 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1714 				goto again;
1715 			}
1716 			return (error);
1717 		case EHOSTDOWN:
1718 		case EHOSTUNREACH:
1719 		case ENETDOWN:
1720 		case ENETUNREACH:
1721 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1722 				tp->t_softerror = error;
1723 				return (0);
1724 			}
1725 			/* FALLTHROUGH */
1726 		default:
1727 			return (error);
1728 		}
1729 	}
1730 	TCPSTAT_INC(tcps_sndtotal);
1731 
1732 	/*
1733 	 * Data sent (as far as we can tell).
1734 	 * If this advertises a larger window than any other segment,
1735 	 * then remember the size of the advertised window.
1736 	 * Any pending ACK has now been sent.
1737 	 */
1738 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1739 		tp->rcv_adv = tp->rcv_nxt + recwin;
1740 	tp->last_ack_sent = tp->rcv_nxt;
1741 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1742 	if (tcp_timer_active(tp, TT_DELACK))
1743 		tcp_timer_activate(tp, TT_DELACK, 0);
1744 	if (sendalot)
1745 		goto again;
1746 	return (0);
1747 }
1748 
1749 void
1750 tcp_setpersist(struct tcpcb *tp)
1751 {
1752 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1753 	int tt;
1754 	int maxunacktime;
1755 
1756 	tp->t_flags &= ~TF_PREVVALID;
1757 	if (tcp_timer_active(tp, TT_REXMT))
1758 		panic("tcp_setpersist: retransmit pending");
1759 	/*
1760 	 * If the state is already closed, don't bother.
1761 	 */
1762 	if (tp->t_state == TCPS_CLOSED)
1763 		return;
1764 
1765 	/*
1766 	 * Start/restart persistence timer.
1767 	 */
1768 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1769 		      tcp_persmin, tcp_persmax);
1770 	if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1771 		maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1772 		if (maxunacktime < 1)
1773 			maxunacktime = 1;
1774 		if (maxunacktime < tt)
1775 			tt = maxunacktime;
1776 	}
1777 	tcp_timer_activate(tp, TT_PERSIST, tt);
1778 	if (tp->t_rxtshift < V_tcp_retries)
1779 		tp->t_rxtshift++;
1780 }
1781 
1782 /*
1783  * Insert TCP options according to the supplied parameters to the place
1784  * optp in a consistent way.  Can handle unaligned destinations.
1785  *
1786  * The order of the option processing is crucial for optimal packing and
1787  * alignment for the scarce option space.
1788  *
1789  * The optimal order for a SYN/SYN-ACK segment is:
1790  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1791  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1792  *
1793  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1794  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1795  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1796  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1797  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1798  */
1799 int
1800 tcp_addoptions(struct tcpopt *to, u_char *optp)
1801 {
1802 	u_int32_t mask, optlen = 0;
1803 
1804 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1805 		if ((to->to_flags & mask) != mask)
1806 			continue;
1807 		if (optlen == TCP_MAXOLEN)
1808 			break;
1809 		switch (to->to_flags & mask) {
1810 		case TOF_MSS:
1811 			while (optlen % 4) {
1812 				optlen += TCPOLEN_NOP;
1813 				*optp++ = TCPOPT_NOP;
1814 			}
1815 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1816 				continue;
1817 			optlen += TCPOLEN_MAXSEG;
1818 			*optp++ = TCPOPT_MAXSEG;
1819 			*optp++ = TCPOLEN_MAXSEG;
1820 			to->to_mss = htons(to->to_mss);
1821 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1822 			optp += sizeof(to->to_mss);
1823 			break;
1824 		case TOF_SCALE:
1825 			while (!optlen || optlen % 2 != 1) {
1826 				optlen += TCPOLEN_NOP;
1827 				*optp++ = TCPOPT_NOP;
1828 			}
1829 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1830 				continue;
1831 			optlen += TCPOLEN_WINDOW;
1832 			*optp++ = TCPOPT_WINDOW;
1833 			*optp++ = TCPOLEN_WINDOW;
1834 			*optp++ = to->to_wscale;
1835 			break;
1836 		case TOF_SACKPERM:
1837 			while (optlen % 2) {
1838 				optlen += TCPOLEN_NOP;
1839 				*optp++ = TCPOPT_NOP;
1840 			}
1841 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1842 				continue;
1843 			optlen += TCPOLEN_SACK_PERMITTED;
1844 			*optp++ = TCPOPT_SACK_PERMITTED;
1845 			*optp++ = TCPOLEN_SACK_PERMITTED;
1846 			break;
1847 		case TOF_TS:
1848 			while (!optlen || optlen % 4 != 2) {
1849 				optlen += TCPOLEN_NOP;
1850 				*optp++ = TCPOPT_NOP;
1851 			}
1852 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1853 				continue;
1854 			optlen += TCPOLEN_TIMESTAMP;
1855 			*optp++ = TCPOPT_TIMESTAMP;
1856 			*optp++ = TCPOLEN_TIMESTAMP;
1857 			to->to_tsval = htonl(to->to_tsval);
1858 			to->to_tsecr = htonl(to->to_tsecr);
1859 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1860 			optp += sizeof(to->to_tsval);
1861 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1862 			optp += sizeof(to->to_tsecr);
1863 			break;
1864 		case TOF_SIGNATURE:
1865 			{
1866 			int siglen = TCPOLEN_SIGNATURE - 2;
1867 
1868 			while (!optlen || optlen % 4 != 2) {
1869 				optlen += TCPOLEN_NOP;
1870 				*optp++ = TCPOPT_NOP;
1871 			}
1872 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1873 				to->to_flags &= ~TOF_SIGNATURE;
1874 				continue;
1875 			}
1876 			optlen += TCPOLEN_SIGNATURE;
1877 			*optp++ = TCPOPT_SIGNATURE;
1878 			*optp++ = TCPOLEN_SIGNATURE;
1879 			to->to_signature = optp;
1880 			while (siglen--)
1881 				 *optp++ = 0;
1882 			break;
1883 			}
1884 		case TOF_SACK:
1885 			{
1886 			int sackblks = 0;
1887 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1888 			tcp_seq sack_seq;
1889 
1890 			while (!optlen || optlen % 4 != 2) {
1891 				optlen += TCPOLEN_NOP;
1892 				*optp++ = TCPOPT_NOP;
1893 			}
1894 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1895 				continue;
1896 			optlen += TCPOLEN_SACKHDR;
1897 			*optp++ = TCPOPT_SACK;
1898 			sackblks = min(to->to_nsacks,
1899 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1900 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1901 			while (sackblks--) {
1902 				sack_seq = htonl(sack->start);
1903 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1904 				optp += sizeof(sack_seq);
1905 				sack_seq = htonl(sack->end);
1906 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1907 				optp += sizeof(sack_seq);
1908 				optlen += TCPOLEN_SACK;
1909 				sack++;
1910 			}
1911 			TCPSTAT_INC(tcps_sack_send_blocks);
1912 			break;
1913 			}
1914 		case TOF_FASTOPEN:
1915 			{
1916 			int total_len;
1917 
1918 			/* XXX is there any point to aligning this option? */
1919 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1920 			if (TCP_MAXOLEN - optlen < total_len) {
1921 				to->to_flags &= ~TOF_FASTOPEN;
1922 				continue;
1923 			}
1924 			*optp++ = TCPOPT_FAST_OPEN;
1925 			*optp++ = total_len;
1926 			if (to->to_tfo_len > 0) {
1927 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1928 				optp += to->to_tfo_len;
1929 			}
1930 			optlen += total_len;
1931 			break;
1932 			}
1933 		default:
1934 			panic("%s: unknown TCP option type", __func__);
1935 			break;
1936 		}
1937 	}
1938 
1939 	/* Terminate and pad TCP options to a 4 byte boundary. */
1940 	if (optlen % 4) {
1941 		optlen += TCPOLEN_EOL;
1942 		*optp++ = TCPOPT_EOL;
1943 	}
1944 	/*
1945 	 * According to RFC 793 (STD0007):
1946 	 *   "The content of the header beyond the End-of-Option option
1947 	 *    must be header padding (i.e., zero)."
1948 	 *   and later: "The padding is composed of zeros."
1949 	 */
1950 	while (optlen % 4) {
1951 		optlen += TCPOLEN_PAD;
1952 		*optp++ = TCPOPT_PAD;
1953 	}
1954 
1955 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1956 	return (optlen);
1957 }
1958 
1959 /*
1960  * This is a copy of m_copym(), taking the TSO segment size/limit
1961  * constraints into account, and advancing the sndptr as it goes.
1962  */
1963 struct mbuf *
1964 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1965     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1966 {
1967 #ifdef KERN_TLS
1968 	struct ktls_session *tls, *ntls;
1969 	struct mbuf *start __diagused;
1970 #endif
1971 	struct mbuf *n, **np;
1972 	struct mbuf *top;
1973 	int32_t off = off0;
1974 	int32_t len = *plen;
1975 	int32_t fragsize;
1976 	int32_t len_cp = 0;
1977 	int32_t *pkthdrlen;
1978 	uint32_t mlen, frags;
1979 	bool copyhdr;
1980 
1981 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1982 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1983 	if (off == 0 && m->m_flags & M_PKTHDR)
1984 		copyhdr = true;
1985 	else
1986 		copyhdr = false;
1987 	while (off > 0) {
1988 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1989 		if (off < m->m_len)
1990 			break;
1991 		off -= m->m_len;
1992 		if ((sb) && (m == sb->sb_sndptr)) {
1993 			sb->sb_sndptroff += m->m_len;
1994 			sb->sb_sndptr = m->m_next;
1995 		}
1996 		m = m->m_next;
1997 	}
1998 	np = &top;
1999 	top = NULL;
2000 	pkthdrlen = NULL;
2001 #ifdef KERN_TLS
2002 	if (hw_tls && (m->m_flags & M_EXTPG))
2003 		tls = m->m_epg_tls;
2004 	else
2005 		tls = NULL;
2006 	start = m;
2007 #endif
2008 	while (len > 0) {
2009 		if (m == NULL) {
2010 			KASSERT(len == M_COPYALL,
2011 			    ("tcp_m_copym, length > size of mbuf chain"));
2012 			*plen = len_cp;
2013 			if (pkthdrlen != NULL)
2014 				*pkthdrlen = len_cp;
2015 			break;
2016 		}
2017 #ifdef KERN_TLS
2018 		if (hw_tls) {
2019 			if (m->m_flags & M_EXTPG)
2020 				ntls = m->m_epg_tls;
2021 			else
2022 				ntls = NULL;
2023 
2024 			/*
2025 			 * Avoid mixing TLS records with handshake
2026 			 * data or TLS records from different
2027 			 * sessions.
2028 			 */
2029 			if (tls != ntls) {
2030 				MPASS(m != start);
2031 				*plen = len_cp;
2032 				if (pkthdrlen != NULL)
2033 					*pkthdrlen = len_cp;
2034 				break;
2035 			}
2036 		}
2037 #endif
2038 		mlen = min(len, m->m_len - off);
2039 		if (seglimit) {
2040 			/*
2041 			 * For M_EXTPG mbufs, add 3 segments
2042 			 * + 1 in case we are crossing page boundaries
2043 			 * + 2 in case the TLS hdr/trailer are used
2044 			 * It is cheaper to just add the segments
2045 			 * than it is to take the cache miss to look
2046 			 * at the mbuf ext_pgs state in detail.
2047 			 */
2048 			if (m->m_flags & M_EXTPG) {
2049 				fragsize = min(segsize, PAGE_SIZE);
2050 				frags = 3;
2051 			} else {
2052 				fragsize = segsize;
2053 				frags = 0;
2054 			}
2055 
2056 			/* Break if we really can't fit anymore. */
2057 			if ((frags + 1) >= seglimit) {
2058 				*plen =	len_cp;
2059 				if (pkthdrlen != NULL)
2060 					*pkthdrlen = len_cp;
2061 				break;
2062 			}
2063 
2064 			/*
2065 			 * Reduce size if you can't copy the whole
2066 			 * mbuf. If we can't copy the whole mbuf, also
2067 			 * adjust len so the loop will end after this
2068 			 * mbuf.
2069 			 */
2070 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2071 				mlen = (seglimit - frags - 1) * fragsize;
2072 				len = mlen;
2073 				*plen = len_cp + len;
2074 				if (pkthdrlen != NULL)
2075 					*pkthdrlen = *plen;
2076 			}
2077 			frags += howmany(mlen, fragsize);
2078 			if (frags == 0)
2079 				frags++;
2080 			seglimit -= frags;
2081 			KASSERT(seglimit > 0,
2082 			    ("%s: seglimit went too low", __func__));
2083 		}
2084 		if (copyhdr)
2085 			n = m_gethdr(M_NOWAIT, m->m_type);
2086 		else
2087 			n = m_get(M_NOWAIT, m->m_type);
2088 		*np = n;
2089 		if (n == NULL)
2090 			goto nospace;
2091 		if (copyhdr) {
2092 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2093 				goto nospace;
2094 			if (len == M_COPYALL)
2095 				n->m_pkthdr.len -= off0;
2096 			else
2097 				n->m_pkthdr.len = len;
2098 			pkthdrlen = &n->m_pkthdr.len;
2099 			copyhdr = false;
2100 		}
2101 		n->m_len = mlen;
2102 		len_cp += n->m_len;
2103 		if (m->m_flags & (M_EXT | M_EXTPG)) {
2104 			n->m_data = m->m_data + off;
2105 			mb_dupcl(n, m);
2106 		} else
2107 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2108 			    (u_int)n->m_len);
2109 
2110 		if (sb && (sb->sb_sndptr == m) &&
2111 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2112 			sb->sb_sndptroff += m->m_len;
2113 			sb->sb_sndptr = m->m_next;
2114 		}
2115 		off = 0;
2116 		if (len != M_COPYALL) {
2117 			len -= n->m_len;
2118 		}
2119 		m = m->m_next;
2120 		np = &n->m_next;
2121 	}
2122 	return (top);
2123 nospace:
2124 	m_freem(top);
2125 	return (NULL);
2126 }
2127 
2128 void
2129 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2130 {
2131 
2132 	/*
2133 	 * Automatic sizing of send socket buffer.  Often the send buffer
2134 	 * size is not optimally adjusted to the actual network conditions
2135 	 * at hand (delay bandwidth product).  Setting the buffer size too
2136 	 * small limits throughput on links with high bandwidth and high
2137 	 * delay (eg. trans-continental/oceanic links).  Setting the
2138 	 * buffer size too big consumes too much real kernel memory,
2139 	 * especially with many connections on busy servers.
2140 	 *
2141 	 * The criteria to step up the send buffer one notch are:
2142 	 *  1. receive window of remote host is larger than send buffer
2143 	 *     (with a fudge factor of 5/4th);
2144 	 *  2. send buffer is filled to 7/8th with data (so we actually
2145 	 *     have data to make use of it);
2146 	 *  3. send buffer fill has not hit maximal automatic size;
2147 	 *  4. our send window (slow start and cogestion controlled) is
2148 	 *     larger than sent but unacknowledged data in send buffer.
2149 	 *
2150 	 * The remote host receive window scaling factor may limit the
2151 	 * growing of the send buffer before it reaches its allowed
2152 	 * maximum.
2153 	 *
2154 	 * It scales directly with slow start or congestion window
2155 	 * and does at most one step per received ACK.  This fast
2156 	 * scaling has the drawback of growing the send buffer beyond
2157 	 * what is strictly necessary to make full use of a given
2158 	 * delay*bandwidth product.  However testing has shown this not
2159 	 * to be much of an problem.  At worst we are trading wasting
2160 	 * of available bandwidth (the non-use of it) for wasting some
2161 	 * socket buffer memory.
2162 	 *
2163 	 * TODO: Shrink send buffer during idle periods together
2164 	 * with congestion window.  Requires another timer.  Has to
2165 	 * wait for upcoming tcp timer rewrite.
2166 	 *
2167 	 * XXXGL: should there be used sbused() or sbavail()?
2168 	 */
2169 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2170 		int lowat;
2171 
2172 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2173 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2174 		    sbused(&so->so_snd) >=
2175 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2176 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2177 		    sendwin >= (sbused(&so->so_snd) -
2178 		    (tp->snd_nxt - tp->snd_una))) {
2179 			if (!sbreserve_locked(so, SO_SND,
2180 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2181 			     V_tcp_autosndbuf_max), curthread))
2182 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2183 		}
2184 	}
2185 }
2186