xref: /freebsd/sys/netinet/tcp_output.c (revision cd675bb60e39395c92b8d08704bc01cdbdc03fac)
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/vnet.h>
68 
69 #include <netinet/in.h>
70 #include <netinet/in_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
76 #ifdef INET6
77 #include <netinet6/in6_pcb.h>
78 #include <netinet/ip6.h>
79 #include <netinet6/ip6_var.h>
80 #endif
81 #include <netinet/tcp.h>
82 #define	TCPOUTFLAGS
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_log_buf.h>
85 #include <netinet/tcp_seq.h>
86 #include <netinet/tcp_timer.h>
87 #include <netinet/tcp_var.h>
88 #include <netinet/tcpip.h>
89 #include <netinet/cc/cc.h>
90 #include <netinet/tcp_fastopen.h>
91 #ifdef TCPPCAP
92 #include <netinet/tcp_pcap.h>
93 #endif
94 #ifdef TCPDEBUG
95 #include <netinet/tcp_debug.h>
96 #endif
97 #ifdef TCP_OFFLOAD
98 #include <netinet/tcp_offload.h>
99 #endif
100 
101 #include <netipsec/ipsec_support.h>
102 
103 #include <machine/in_cksum.h>
104 
105 #include <security/mac/mac_framework.h>
106 
107 VNET_DEFINE(int, path_mtu_discovery) = 1;
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
109 	&VNET_NAME(path_mtu_discovery), 1,
110 	"Enable Path MTU Discovery");
111 
112 VNET_DEFINE(int, tcp_do_tso) = 1;
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
114 	&VNET_NAME(tcp_do_tso), 0,
115 	"Enable TCP Segmentation Offload");
116 
117 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
118 #define	V_tcp_sendspace	VNET(tcp_sendspace)
119 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
120 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
121 
122 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
124 	&VNET_NAME(tcp_do_autosndbuf), 0,
125 	"Enable automatic send buffer sizing");
126 
127 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
129 	&VNET_NAME(tcp_autosndbuf_inc), 0,
130 	"Incrementor step size of automatic send buffer");
131 
132 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
134 	&VNET_NAME(tcp_autosndbuf_max), 0,
135 	"Max size of automatic send buffer");
136 
137 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
138 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
140 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
141 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
142 
143 /*
144  * Make sure that either retransmit or persist timer is set for SYN, FIN and
145  * non-ACK.
146  */
147 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
148 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
149 	    tcp_timer_active((tp), TT_REXMT) ||				\
150 	    tcp_timer_active((tp), TT_PERSIST),				\
151 	    ("neither rexmt nor persist timer is set"))
152 
153 static void inline	cc_after_idle(struct tcpcb *tp);
154 
155 #ifdef TCP_HHOOK
156 /*
157  * Wrapper for the TCP established output helper hook.
158  */
159 void
160 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
161     struct tcpopt *to, uint32_t len, int tso)
162 {
163 	struct tcp_hhook_data hhook_data;
164 
165 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
166 		hhook_data.tp = tp;
167 		hhook_data.th = th;
168 		hhook_data.to = to;
169 		hhook_data.len = len;
170 		hhook_data.tso = tso;
171 
172 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
173 		    tp->osd);
174 	}
175 }
176 #endif
177 
178 /*
179  * CC wrapper hook functions
180  */
181 static void inline
182 cc_after_idle(struct tcpcb *tp)
183 {
184 	INP_WLOCK_ASSERT(tp->t_inpcb);
185 
186 	if (CC_ALGO(tp)->after_idle != NULL)
187 		CC_ALGO(tp)->after_idle(tp->ccv);
188 }
189 
190 /*
191  * Tcp output routine: figure out what should be sent and send it.
192  */
193 int
194 tcp_output(struct tcpcb *tp)
195 {
196 	struct socket *so = tp->t_inpcb->inp_socket;
197 	int32_t len;
198 	uint32_t recwin, sendwin;
199 	int off, flags, error = 0;	/* Keep compiler happy */
200 	u_int if_hw_tsomaxsegcount = 0;
201 	u_int if_hw_tsomaxsegsize = 0;
202 	struct mbuf *m;
203 	struct ip *ip = NULL;
204 #ifdef TCPDEBUG
205 	struct ipovly *ipov = NULL;
206 #endif
207 	struct tcphdr *th;
208 	u_char opt[TCP_MAXOLEN];
209 	unsigned ipoptlen, optlen, hdrlen;
210 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
211 	unsigned ipsec_optlen = 0;
212 #endif
213 	int idle, sendalot, curticks;
214 	int sack_rxmit, sack_bytes_rxmt;
215 	struct sackhole *p;
216 	int tso, mtu;
217 	struct tcpopt to;
218 	unsigned int wanted_cookie = 0;
219 	unsigned int dont_sendalot = 0;
220 #if 0
221 	int maxburst = TCP_MAXBURST;
222 #endif
223 #ifdef INET6
224 	struct ip6_hdr *ip6 = NULL;
225 	int isipv6;
226 
227 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
228 #endif
229 #ifdef KERN_TLS
230 	const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
231 #else
232 	const bool hw_tls = false;
233 #endif
234 
235 	NET_EPOCH_ASSERT();
236 	INP_WLOCK_ASSERT(tp->t_inpcb);
237 
238 #ifdef TCP_OFFLOAD
239 	if (tp->t_flags & TF_TOE)
240 		return (tcp_offload_output(tp));
241 #endif
242 
243 	/*
244 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
245 	 * only allow the initial SYN or SYN|ACK and those sent
246 	 * by the retransmit timer.
247 	 */
248 	if (IS_FASTOPEN(tp->t_flags) &&
249 	    ((tp->t_state == TCPS_SYN_SENT) ||
250 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
251 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
252 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
253 		return (0);
254 
255 	/*
256 	 * Determine length of data that should be transmitted,
257 	 * and flags that will be used.
258 	 * If there is some data or critical controls (SYN, RST)
259 	 * to send, then transmit; otherwise, investigate further.
260 	 */
261 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
262 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
263 		cc_after_idle(tp);
264 	tp->t_flags &= ~TF_LASTIDLE;
265 	if (idle) {
266 		if (tp->t_flags & TF_MORETOCOME) {
267 			tp->t_flags |= TF_LASTIDLE;
268 			idle = 0;
269 		}
270 	}
271 again:
272 	/*
273 	 * If we've recently taken a timeout, snd_max will be greater than
274 	 * snd_nxt.  There may be SACK information that allows us to avoid
275 	 * resending already delivered data.  Adjust snd_nxt accordingly.
276 	 */
277 	if ((tp->t_flags & TF_SACK_PERMIT) &&
278 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
279 		tcp_sack_adjust(tp);
280 	sendalot = 0;
281 	tso = 0;
282 	mtu = 0;
283 	off = tp->snd_nxt - tp->snd_una;
284 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
285 
286 	flags = tcp_outflags[tp->t_state];
287 	/*
288 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
289 	 * to send out new data (when sendalot is 1), bypass this function.
290 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
291 	 * we're replacing a (future) new transmission with a retransmission
292 	 * now, and we previously incremented snd_cwnd in tcp_input().
293 	 */
294 	/*
295 	 * Still in sack recovery , reset rxmit flag to zero.
296 	 */
297 	sack_rxmit = 0;
298 	sack_bytes_rxmt = 0;
299 	len = 0;
300 	p = NULL;
301 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
302 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
303 		uint32_t cwin;
304 
305 		cwin =
306 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
307 		/* Do not retransmit SACK segments beyond snd_recover */
308 		if (SEQ_GT(p->end, tp->snd_recover)) {
309 			/*
310 			 * (At least) part of sack hole extends beyond
311 			 * snd_recover. Check to see if we can rexmit data
312 			 * for this hole.
313 			 */
314 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
315 				/*
316 				 * Can't rexmit any more data for this hole.
317 				 * That data will be rexmitted in the next
318 				 * sack recovery episode, when snd_recover
319 				 * moves past p->rxmit.
320 				 */
321 				p = NULL;
322 				goto after_sack_rexmit;
323 			} else
324 				/* Can rexmit part of the current hole */
325 				len = ((int32_t)ulmin(cwin,
326 						   tp->snd_recover - p->rxmit));
327 		} else
328 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
329 		off = p->rxmit - tp->snd_una;
330 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
331 		    __func__, off));
332 		if (len > 0) {
333 			sack_rxmit = 1;
334 			sendalot = 1;
335 			TCPSTAT_INC(tcps_sack_rexmits);
336 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
337 			    min(len, tp->t_maxseg));
338 		}
339 	}
340 after_sack_rexmit:
341 	/*
342 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
343 	 * state flags.
344 	 */
345 	if (tp->t_flags & TF_NEEDFIN)
346 		flags |= TH_FIN;
347 	if (tp->t_flags & TF_NEEDSYN)
348 		flags |= TH_SYN;
349 
350 	SOCKBUF_LOCK(&so->so_snd);
351 	/*
352 	 * If in persist timeout with window of 0, send 1 byte.
353 	 * Otherwise, if window is small but nonzero
354 	 * and timer expired, we will send what we can
355 	 * and go to transmit state.
356 	 */
357 	if (tp->t_flags & TF_FORCEDATA) {
358 		if (sendwin == 0) {
359 			/*
360 			 * If we still have some data to send, then
361 			 * clear the FIN bit.  Usually this would
362 			 * happen below when it realizes that we
363 			 * aren't sending all the data.  However,
364 			 * if we have exactly 1 byte of unsent data,
365 			 * then it won't clear the FIN bit below,
366 			 * and if we are in persist state, we wind
367 			 * up sending the packet without recording
368 			 * that we sent the FIN bit.
369 			 *
370 			 * We can't just blindly clear the FIN bit,
371 			 * because if we don't have any more data
372 			 * to send then the probe will be the FIN
373 			 * itself.
374 			 */
375 			if (off < sbused(&so->so_snd))
376 				flags &= ~TH_FIN;
377 			sendwin = 1;
378 		} else {
379 			tcp_timer_activate(tp, TT_PERSIST, 0);
380 			tp->t_rxtshift = 0;
381 		}
382 	}
383 
384 	/*
385 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
386 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
387 	 * a negative length.  This can also occur when TCP opens up
388 	 * its congestion window while receiving additional duplicate
389 	 * acks after fast-retransmit because TCP will reset snd_nxt
390 	 * to snd_max after the fast-retransmit.
391 	 *
392 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
393 	 * be set to snd_una, the offset will be 0, and the length may
394 	 * wind up 0.
395 	 *
396 	 * If sack_rxmit is true we are retransmitting from the scoreboard
397 	 * in which case len is already set.
398 	 */
399 	if (sack_rxmit == 0) {
400 		if (sack_bytes_rxmt == 0)
401 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
402 			    off);
403 		else {
404 			int32_t cwin;
405 
406                         /*
407 			 * We are inside of a SACK recovery episode and are
408 			 * sending new data, having retransmitted all the
409 			 * data possible in the scoreboard.
410 			 */
411 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
412 			    off);
413 			/*
414 			 * Don't remove this (len > 0) check !
415 			 * We explicitly check for len > 0 here (although it
416 			 * isn't really necessary), to work around a gcc
417 			 * optimization issue - to force gcc to compute
418 			 * len above. Without this check, the computation
419 			 * of len is bungled by the optimizer.
420 			 */
421 			if (len > 0) {
422 				cwin = tp->snd_cwnd -
423 					(tp->snd_nxt - tp->snd_recover) -
424 					sack_bytes_rxmt;
425 				if (cwin < 0)
426 					cwin = 0;
427 				len = imin(len, cwin);
428 			}
429 		}
430 	}
431 
432 	/*
433 	 * Lop off SYN bit if it has already been sent.  However, if this
434 	 * is SYN-SENT state and if segment contains data and if we don't
435 	 * know that foreign host supports TAO, suppress sending segment.
436 	 */
437 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
438 		if (tp->t_state != TCPS_SYN_RECEIVED)
439 			flags &= ~TH_SYN;
440 		/*
441 		 * When sending additional segments following a TFO SYN|ACK,
442 		 * do not include the SYN bit.
443 		 */
444 		if (IS_FASTOPEN(tp->t_flags) &&
445 		    (tp->t_state == TCPS_SYN_RECEIVED))
446 			flags &= ~TH_SYN;
447 		off--, len++;
448 	}
449 
450 	/*
451 	 * Be careful not to send data and/or FIN on SYN segments.
452 	 * This measure is needed to prevent interoperability problems
453 	 * with not fully conformant TCP implementations.
454 	 */
455 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
456 		len = 0;
457 		flags &= ~TH_FIN;
458 	}
459 
460 	/*
461 	 * On TFO sockets, ensure no data is sent in the following cases:
462 	 *
463 	 *  - When retransmitting SYN|ACK on a passively-created socket
464 	 *
465 	 *  - When retransmitting SYN on an actively created socket
466 	 *
467 	 *  - When sending a zero-length cookie (cookie request) on an
468 	 *    actively created socket
469 	 *
470 	 *  - When the socket is in the CLOSED state (RST is being sent)
471 	 */
472 	if (IS_FASTOPEN(tp->t_flags) &&
473 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
474 	     ((tp->t_state == TCPS_SYN_SENT) &&
475 	      (tp->t_tfo_client_cookie_len == 0)) ||
476 	     (flags & TH_RST)))
477 		len = 0;
478 	if (len <= 0) {
479 		/*
480 		 * If FIN has been sent but not acked,
481 		 * but we haven't been called to retransmit,
482 		 * len will be < 0.  Otherwise, window shrank
483 		 * after we sent into it.  If window shrank to 0,
484 		 * cancel pending retransmit, pull snd_nxt back
485 		 * to (closed) window, and set the persist timer
486 		 * if it isn't already going.  If the window didn't
487 		 * close completely, just wait for an ACK.
488 		 *
489 		 * We also do a general check here to ensure that
490 		 * we will set the persist timer when we have data
491 		 * to send, but a 0-byte window. This makes sure
492 		 * the persist timer is set even if the packet
493 		 * hits one of the "goto send" lines below.
494 		 */
495 		len = 0;
496 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
497 			(off < (int) sbavail(&so->so_snd))) {
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, tp->t_inpcb);
535 #ifdef INET
536 	else
537 #endif
538 #endif /* INET6 */
539 #ifdef INET
540 	if (IPSEC_ENABLED(ipv4))
541 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
542 #endif /* INET */
543 #endif /* IPSEC */
544 #ifdef INET6
545 	if (isipv6)
546 		ipoptlen = ip6_optlen(tp->t_inpcb);
547 	else
548 #endif
549 	if (tp->t_inpcb->inp_options)
550 		ipoptlen = tp->t_inpcb->inp_options->m_len -
551 				offsetof(struct ipoption, ipopt_list);
552 	else
553 		ipoptlen = 0;
554 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
555 	ipoptlen += ipsec_optlen;
556 #endif
557 
558 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
559 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
560 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
561 	    ipoptlen == 0 && !(flags & TH_SYN))
562 		tso = 1;
563 
564 	if (sack_rxmit) {
565 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
566 			flags &= ~TH_FIN;
567 	} else {
568 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
569 		    sbused(&so->so_snd)))
570 			flags &= ~TH_FIN;
571 	}
572 
573 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
574 	    (long)TCP_MAXWIN << tp->rcv_scale);
575 
576 	/*
577 	 * Sender silly window avoidance.   We transmit under the following
578 	 * conditions when len is non-zero:
579 	 *
580 	 *	- We have a full segment (or more with TSO)
581 	 *	- This is the last buffer in a write()/send() and we are
582 	 *	  either idle or running NODELAY
583 	 *	- we've timed out (e.g. persist timer)
584 	 *	- we have more then 1/2 the maximum send window's worth of
585 	 *	  data (receiver may be limited the window size)
586 	 *	- we need to retransmit
587 	 */
588 	if (len) {
589 		if (len >= tp->t_maxseg)
590 			goto send;
591 		/*
592 		 * NOTE! on localhost connections an 'ack' from the remote
593 		 * end may occur synchronously with the output and cause
594 		 * us to flush a buffer queued with moretocome.  XXX
595 		 *
596 		 * note: the len + off check is almost certainly unnecessary.
597 		 */
598 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
599 		    (idle || (tp->t_flags & TF_NODELAY)) &&
600 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
601 		    (tp->t_flags & TF_NOPUSH) == 0) {
602 			goto send;
603 		}
604 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
605 			goto send;
606 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
607 			goto send;
608 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
609 			goto send;
610 		if (sack_rxmit)
611 			goto send;
612 	}
613 
614 	/*
615 	 * Sending of standalone window updates.
616 	 *
617 	 * Window updates are important when we close our window due to a
618 	 * full socket buffer and are opening it again after the application
619 	 * reads data from it.  Once the window has opened again and the
620 	 * remote end starts to send again the ACK clock takes over and
621 	 * provides the most current window information.
622 	 *
623 	 * We must avoid the silly window syndrome whereas every read
624 	 * from the receive buffer, no matter how small, causes a window
625 	 * update to be sent.  We also should avoid sending a flurry of
626 	 * window updates when the socket buffer had queued a lot of data
627 	 * and the application is doing small reads.
628 	 *
629 	 * Prevent a flurry of pointless window updates by only sending
630 	 * an update when we can increase the advertized window by more
631 	 * than 1/4th of the socket buffer capacity.  When the buffer is
632 	 * getting full or is very small be more aggressive and send an
633 	 * update whenever we can increase by two mss sized segments.
634 	 * In all other situations the ACK's to new incoming data will
635 	 * carry further window increases.
636 	 *
637 	 * Don't send an independent window update if a delayed
638 	 * ACK is pending (it will get piggy-backed on it) or the
639 	 * remote side already has done a half-close and won't send
640 	 * more data.  Skip this if the connection is in T/TCP
641 	 * half-open state.
642 	 */
643 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
644 	    !(tp->t_flags & TF_DELACK) &&
645 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
646 		/*
647 		 * "adv" is the amount we could increase the window,
648 		 * taking into account that we are limited by
649 		 * TCP_MAXWIN << tp->rcv_scale.
650 		 */
651 		int32_t adv;
652 		int oldwin;
653 
654 		adv = recwin;
655 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
656 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
657 			adv -= oldwin;
658 		} else
659 			oldwin = 0;
660 
661 		/*
662 		 * If the new window size ends up being the same as or less
663 		 * than the old size when it is scaled, then don't force
664 		 * a window update.
665 		 */
666 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
667 			goto dontupdate;
668 
669 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
670 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
671 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
672 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
673 		     adv >= TCP_MAXWIN << tp->rcv_scale))
674 			goto send;
675 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
676 			goto send;
677 	}
678 dontupdate:
679 
680 	/*
681 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
682 	 * is also a catch-all for the retransmit timer timeout case.
683 	 */
684 	if (tp->t_flags & TF_ACKNOW)
685 		goto send;
686 	if ((flags & TH_RST) ||
687 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
688 		goto send;
689 	if (SEQ_GT(tp->snd_up, tp->snd_una))
690 		goto send;
691 	/*
692 	 * If our state indicates that FIN should be sent
693 	 * and we have not yet done so, then we need to send.
694 	 */
695 	if (flags & TH_FIN &&
696 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
697 		goto send;
698 	/*
699 	 * In SACK, it is possible for tcp_output to fail to send a segment
700 	 * after the retransmission timer has been turned off.  Make sure
701 	 * that the retransmission timer is set.
702 	 */
703 	if ((tp->t_flags & TF_SACK_PERMIT) &&
704 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
705 	    !tcp_timer_active(tp, TT_REXMT) &&
706 	    !tcp_timer_active(tp, TT_PERSIST)) {
707 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
708 		goto just_return;
709 	}
710 	/*
711 	 * TCP window updates are not reliable, rather a polling protocol
712 	 * using ``persist'' packets is used to insure receipt of window
713 	 * updates.  The three ``states'' for the output side are:
714 	 *	idle			not doing retransmits or persists
715 	 *	persisting		to move a small or zero window
716 	 *	(re)transmitting	and thereby not persisting
717 	 *
718 	 * tcp_timer_active(tp, TT_PERSIST)
719 	 *	is true when we are in persist state.
720 	 * (tp->t_flags & TF_FORCEDATA)
721 	 *	is set when we are called to send a persist packet.
722 	 * tcp_timer_active(tp, TT_REXMT)
723 	 *	is set when we are retransmitting
724 	 * The output side is idle when both timers are zero.
725 	 *
726 	 * If send window is too small, there is data to transmit, and no
727 	 * retransmit or persist is pending, then go to persist state.
728 	 * If nothing happens soon, send when timer expires:
729 	 * if window is nonzero, transmit what we can,
730 	 * otherwise force out a byte.
731 	 */
732 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
733 	    !tcp_timer_active(tp, TT_PERSIST)) {
734 		tp->t_rxtshift = 0;
735 		tcp_setpersist(tp);
736 	}
737 
738 	/*
739 	 * No reason to send a segment, just return.
740 	 */
741 just_return:
742 	SOCKBUF_UNLOCK(&so->so_snd);
743 	return (0);
744 
745 send:
746 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
747 	if (len > 0) {
748 		if (len >= tp->t_maxseg)
749 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
750 		else
751 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
752 	}
753 	/*
754 	 * Before ESTABLISHED, force sending of initial options
755 	 * unless TCP set not to do any options.
756 	 * NOTE: we assume that the IP/TCP header plus TCP options
757 	 * always fit in a single mbuf, leaving room for a maximum
758 	 * link header, i.e.
759 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
760 	 */
761 	optlen = 0;
762 #ifdef INET6
763 	if (isipv6)
764 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
765 	else
766 #endif
767 		hdrlen = sizeof (struct tcpiphdr);
768 
769 	/*
770 	 * Compute options for segment.
771 	 * We only have to care about SYN and established connection
772 	 * segments.  Options for SYN-ACK segments are handled in TCP
773 	 * syncache.
774 	 */
775 	to.to_flags = 0;
776 	if ((tp->t_flags & TF_NOOPT) == 0) {
777 		/* Maximum segment size. */
778 		if (flags & TH_SYN) {
779 			tp->snd_nxt = tp->iss;
780 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
781 			to.to_flags |= TOF_MSS;
782 
783 			/*
784 			 * On SYN or SYN|ACK transmits on TFO connections,
785 			 * only include the TFO option if it is not a
786 			 * retransmit, as the presence of the TFO option may
787 			 * have caused the original SYN or SYN|ACK to have
788 			 * been dropped by a middlebox.
789 			 */
790 			if (IS_FASTOPEN(tp->t_flags) &&
791 			    (tp->t_rxtshift == 0)) {
792 				if (tp->t_state == TCPS_SYN_RECEIVED) {
793 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
794 					to.to_tfo_cookie =
795 					    (u_int8_t *)&tp->t_tfo_cookie.server;
796 					to.to_flags |= TOF_FASTOPEN;
797 					wanted_cookie = 1;
798 				} else if (tp->t_state == TCPS_SYN_SENT) {
799 					to.to_tfo_len =
800 					    tp->t_tfo_client_cookie_len;
801 					to.to_tfo_cookie =
802 					    tp->t_tfo_cookie.client;
803 					to.to_flags |= TOF_FASTOPEN;
804 					wanted_cookie = 1;
805 					/*
806 					 * If we wind up having more data to
807 					 * send with the SYN than can fit in
808 					 * one segment, don't send any more
809 					 * until the SYN|ACK comes back from
810 					 * the other end.
811 					 */
812 					dont_sendalot = 1;
813 				}
814 			}
815 		}
816 		/* Window scaling. */
817 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
818 			to.to_wscale = tp->request_r_scale;
819 			to.to_flags |= TOF_SCALE;
820 		}
821 		/* Timestamps. */
822 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
823 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
824 			curticks = tcp_ts_getticks();
825 			to.to_tsval = curticks + tp->ts_offset;
826 			to.to_tsecr = tp->ts_recent;
827 			to.to_flags |= TOF_TS;
828 			if (tp->t_rxtshift == 1)
829 				tp->t_badrxtwin = curticks;
830 		}
831 
832 		/* Set receive buffer autosizing timestamp. */
833 		if (tp->rfbuf_ts == 0 &&
834 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
835 			tp->rfbuf_ts = tcp_ts_getticks();
836 
837 		/* Selective ACK's. */
838 		if (tp->t_flags & TF_SACK_PERMIT) {
839 			if (flags & TH_SYN)
840 				to.to_flags |= TOF_SACKPERM;
841 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
842 			    (tp->t_flags & TF_SACK_PERMIT) &&
843 			    tp->rcv_numsacks > 0) {
844 				to.to_flags |= TOF_SACK;
845 				to.to_nsacks = tp->rcv_numsacks;
846 				to.to_sacks = (u_char *)tp->sackblks;
847 			}
848 		}
849 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
850 		/* TCP-MD5 (RFC2385). */
851 		/*
852 		 * Check that TCP_MD5SIG is enabled in tcpcb to
853 		 * account the size needed to set this TCP option.
854 		 */
855 		if (tp->t_flags & TF_SIGNATURE)
856 			to.to_flags |= TOF_SIGNATURE;
857 #endif /* TCP_SIGNATURE */
858 
859 		/* Processing the options. */
860 		hdrlen += optlen = tcp_addoptions(&to, opt);
861 		/*
862 		 * If we wanted a TFO option to be added, but it was unable
863 		 * to fit, ensure no data is sent.
864 		 */
865 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
866 		    !(to.to_flags & TOF_FASTOPEN))
867 			len = 0;
868 	}
869 
870 	/*
871 	 * Adjust data length if insertion of options will
872 	 * bump the packet length beyond the t_maxseg length.
873 	 * Clear the FIN bit because we cut off the tail of
874 	 * the segment.
875 	 */
876 	if (len + optlen + ipoptlen > tp->t_maxseg) {
877 		flags &= ~TH_FIN;
878 
879 		if (tso) {
880 			u_int if_hw_tsomax;
881 			u_int moff;
882 			int max_len;
883 
884 			/* extract TSO information */
885 			if_hw_tsomax = tp->t_tsomax;
886 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
887 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
888 
889 			/*
890 			 * Limit a TSO burst to prevent it from
891 			 * overflowing or exceeding the maximum length
892 			 * allowed by the network interface:
893 			 */
894 			KASSERT(ipoptlen == 0,
895 			    ("%s: TSO can't do IP options", __func__));
896 
897 			/*
898 			 * Check if we should limit by maximum payload
899 			 * length:
900 			 */
901 			if (if_hw_tsomax != 0) {
902 				/* compute maximum TSO length */
903 				max_len = (if_hw_tsomax - hdrlen -
904 				    max_linkhdr);
905 				if (max_len <= 0) {
906 					len = 0;
907 				} else if (len > max_len) {
908 					sendalot = 1;
909 					len = max_len;
910 				}
911 			}
912 
913 			/*
914 			 * Prevent the last segment from being
915 			 * fractional unless the send sockbuf can be
916 			 * emptied:
917 			 */
918 			max_len = (tp->t_maxseg - optlen);
919 			if (((uint32_t)off + (uint32_t)len) <
920 			    sbavail(&so->so_snd)) {
921 				moff = len % max_len;
922 				if (moff != 0) {
923 					len -= moff;
924 					sendalot = 1;
925 				}
926 			}
927 
928 			/*
929 			 * In case there are too many small fragments
930 			 * don't use TSO:
931 			 */
932 			if (len <= max_len) {
933 				len = max_len;
934 				sendalot = 1;
935 				tso = 0;
936 			}
937 
938 			/*
939 			 * Send the FIN in a separate segment
940 			 * after the bulk sending is done.
941 			 * We don't trust the TSO implementations
942 			 * to clear the FIN flag on all but the
943 			 * last segment.
944 			 */
945 			if (tp->t_flags & TF_NEEDFIN)
946 				sendalot = 1;
947 		} else {
948 			if (optlen + ipoptlen >= tp->t_maxseg) {
949 				/*
950 				 * Since we don't have enough space to put
951 				 * the IP header chain and the TCP header in
952 				 * one packet as required by RFC 7112, don't
953 				 * send it. Also ensure that at least one
954 				 * byte of the payload can be put into the
955 				 * TCP segment.
956 				 */
957 				SOCKBUF_UNLOCK(&so->so_snd);
958 				error = EMSGSIZE;
959 				sack_rxmit = 0;
960 				goto out;
961 			}
962 			len = tp->t_maxseg - optlen - ipoptlen;
963 			sendalot = 1;
964 			if (dont_sendalot)
965 				sendalot = 0;
966 		}
967 	} else
968 		tso = 0;
969 
970 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
971 	    ("%s: len > IP_MAXPACKET", __func__));
972 
973 /*#ifdef DIAGNOSTIC*/
974 #ifdef INET6
975 	if (max_linkhdr + hdrlen > MCLBYTES)
976 #else
977 	if (max_linkhdr + hdrlen > MHLEN)
978 #endif
979 		panic("tcphdr too big");
980 /*#endif*/
981 
982 	/*
983 	 * This KASSERT is here to catch edge cases at a well defined place.
984 	 * Before, those had triggered (random) panic conditions further down.
985 	 */
986 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
987 
988 	/*
989 	 * Grab a header mbuf, attaching a copy of data to
990 	 * be transmitted, and initialize the header from
991 	 * the template for sends on this connection.
992 	 */
993 	if (len) {
994 		struct mbuf *mb;
995 		struct sockbuf *msb;
996 		u_int moff;
997 
998 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
999 			TCPSTAT_INC(tcps_sndprobe);
1000 #ifdef STATS
1001 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1002 				stats_voi_update_abs_u32(tp->t_stats,
1003 				VOI_TCP_RETXPB, len);
1004 			else
1005 				stats_voi_update_abs_u64(tp->t_stats,
1006 				    VOI_TCP_TXPB, len);
1007 #endif /* STATS */
1008 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1009 			tp->t_sndrexmitpack++;
1010 			TCPSTAT_INC(tcps_sndrexmitpack);
1011 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1012 #ifdef STATS
1013 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1014 			    len);
1015 #endif /* STATS */
1016 		} else {
1017 			TCPSTAT_INC(tcps_sndpack);
1018 			TCPSTAT_ADD(tcps_sndbyte, len);
1019 #ifdef STATS
1020 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1021 			    len);
1022 #endif /* STATS */
1023 		}
1024 #ifdef INET6
1025 		if (MHLEN < hdrlen + max_linkhdr)
1026 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1027 		else
1028 #endif
1029 			m = m_gethdr(M_NOWAIT, MT_DATA);
1030 
1031 		if (m == NULL) {
1032 			SOCKBUF_UNLOCK(&so->so_snd);
1033 			error = ENOBUFS;
1034 			sack_rxmit = 0;
1035 			goto out;
1036 		}
1037 
1038 		m->m_data += max_linkhdr;
1039 		m->m_len = hdrlen;
1040 
1041 		/*
1042 		 * Start the m_copy functions from the closest mbuf
1043 		 * to the offset in the socket buffer chain.
1044 		 */
1045 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1046 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1047 			m_copydata(mb, moff, len,
1048 			    mtod(m, caddr_t) + hdrlen);
1049 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1050 				sbsndptr_adv(&so->so_snd, mb, len);
1051 			m->m_len += len;
1052 		} else {
1053 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1054 				msb = NULL;
1055 			else
1056 				msb = &so->so_snd;
1057 			m->m_next = tcp_m_copym(mb, moff,
1058 			    &len, if_hw_tsomaxsegcount,
1059 			    if_hw_tsomaxsegsize, msb, hw_tls);
1060 			if (len <= (tp->t_maxseg - optlen)) {
1061 				/*
1062 				 * Must have ran out of mbufs for the copy
1063 				 * shorten it to no longer need tso. Lets
1064 				 * not put on sendalot since we are low on
1065 				 * mbufs.
1066 				 */
1067 				tso = 0;
1068 			}
1069 			if (m->m_next == NULL) {
1070 				SOCKBUF_UNLOCK(&so->so_snd);
1071 				(void) m_free(m);
1072 				error = ENOBUFS;
1073 				sack_rxmit = 0;
1074 				goto out;
1075 			}
1076 		}
1077 
1078 		/*
1079 		 * If we're sending everything we've got, set PUSH.
1080 		 * (This will keep happy those implementations which only
1081 		 * give data to the user when a buffer fills or
1082 		 * a PUSH comes in.)
1083 		 */
1084 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1085 		    !(flags & TH_SYN))
1086 			flags |= TH_PUSH;
1087 		SOCKBUF_UNLOCK(&so->so_snd);
1088 	} else {
1089 		SOCKBUF_UNLOCK(&so->so_snd);
1090 		if (tp->t_flags & TF_ACKNOW)
1091 			TCPSTAT_INC(tcps_sndacks);
1092 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1093 			TCPSTAT_INC(tcps_sndctrl);
1094 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1095 			TCPSTAT_INC(tcps_sndurg);
1096 		else
1097 			TCPSTAT_INC(tcps_sndwinup);
1098 
1099 		m = m_gethdr(M_NOWAIT, MT_DATA);
1100 		if (m == NULL) {
1101 			error = ENOBUFS;
1102 			sack_rxmit = 0;
1103 			goto out;
1104 		}
1105 #ifdef INET6
1106 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1107 		    MHLEN >= hdrlen) {
1108 			M_ALIGN(m, hdrlen);
1109 		} else
1110 #endif
1111 		m->m_data += max_linkhdr;
1112 		m->m_len = hdrlen;
1113 	}
1114 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1115 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1116 #ifdef MAC
1117 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1118 #endif
1119 #ifdef INET6
1120 	if (isipv6) {
1121 		ip6 = mtod(m, struct ip6_hdr *);
1122 		th = (struct tcphdr *)(ip6 + 1);
1123 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1124 	} else
1125 #endif /* INET6 */
1126 	{
1127 		ip = mtod(m, struct ip *);
1128 #ifdef TCPDEBUG
1129 		ipov = (struct ipovly *)ip;
1130 #endif
1131 		th = (struct tcphdr *)(ip + 1);
1132 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1133 	}
1134 
1135 	/*
1136 	 * Fill in fields, remembering maximum advertised
1137 	 * window for use in delaying messages about window sizes.
1138 	 * If resending a FIN, be sure not to use a new sequence number.
1139 	 */
1140 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1141 	    tp->snd_nxt == tp->snd_max)
1142 		tp->snd_nxt--;
1143 	/*
1144 	 * If we are starting a connection, send ECN setup
1145 	 * SYN packet. If we are on a retransmit, we may
1146 	 * resend those bits a number of times as per
1147 	 * RFC 3168.
1148 	 */
1149 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1150 		if (tp->t_rxtshift >= 1) {
1151 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1152 				flags |= TH_ECE|TH_CWR;
1153 		} else
1154 			flags |= TH_ECE|TH_CWR;
1155 	}
1156 
1157 	if (tp->t_state == TCPS_ESTABLISHED &&
1158 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
1159 		/*
1160 		 * If the peer has ECN, mark data packets with
1161 		 * ECN capable transmission (ECT).
1162 		 * Ignore pure ack packets, retransmissions and window probes.
1163 		 */
1164 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1165 		    (sack_rxmit == 0) &&
1166 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1167 #ifdef INET6
1168 			if (isipv6)
1169 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1170 			else
1171 #endif
1172 				ip->ip_tos |= IPTOS_ECN_ECT0;
1173 			TCPSTAT_INC(tcps_ecn_ect0);
1174 		}
1175 
1176 		/*
1177 		 * Reply with proper ECN notifications.
1178 		 */
1179 		if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1180 			flags |= TH_CWR;
1181 			tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1182 		}
1183 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
1184 			flags |= TH_ECE;
1185 	}
1186 
1187 	/*
1188 	 * If we are doing retransmissions, then snd_nxt will
1189 	 * not reflect the first unsent octet.  For ACK only
1190 	 * packets, we do not want the sequence number of the
1191 	 * retransmitted packet, we want the sequence number
1192 	 * of the next unsent octet.  So, if there is no data
1193 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1194 	 * when filling in ti_seq.  But if we are in persist
1195 	 * state, snd_max might reflect one byte beyond the
1196 	 * right edge of the window, so use snd_nxt in that
1197 	 * case, since we know we aren't doing a retransmission.
1198 	 * (retransmit and persist are mutually exclusive...)
1199 	 */
1200 	if (sack_rxmit == 0) {
1201 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1202 		    tcp_timer_active(tp, TT_PERSIST))
1203 			th->th_seq = htonl(tp->snd_nxt);
1204 		else
1205 			th->th_seq = htonl(tp->snd_max);
1206 	} else {
1207 		th->th_seq = htonl(p->rxmit);
1208 		p->rxmit += len;
1209 		tp->sackhint.sack_bytes_rexmit += len;
1210 	}
1211 	th->th_ack = htonl(tp->rcv_nxt);
1212 	if (optlen) {
1213 		bcopy(opt, th + 1, optlen);
1214 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1215 	}
1216 	th->th_flags = flags;
1217 	/*
1218 	 * Calculate receive window.  Don't shrink window,
1219 	 * but avoid silly window syndrome.
1220 	 * If a RST segment is sent, advertise a window of zero.
1221 	 */
1222 	if (flags & TH_RST) {
1223 		recwin = 0;
1224 	} else {
1225 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1226 		    recwin < tp->t_maxseg)
1227 			recwin = 0;
1228 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1229 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1230 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1231 	}
1232 	/*
1233 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1234 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1235 	 * case is handled in syncache.
1236 	 */
1237 	if (flags & TH_SYN)
1238 		th->th_win = htons((u_short)
1239 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1240 	else
1241 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1242 
1243 	/*
1244 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1245 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1246 	 * flag tells soreceive() to disable delayed acknowledgements when
1247 	 * draining the buffer.  This can occur if the receiver is attempting
1248 	 * to read more data than can be buffered prior to transmitting on
1249 	 * the connection.
1250 	 */
1251 	if (th->th_win == 0) {
1252 		tp->t_sndzerowin++;
1253 		tp->t_flags |= TF_RXWIN0SENT;
1254 	} else
1255 		tp->t_flags &= ~TF_RXWIN0SENT;
1256 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1257 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1258 		th->th_flags |= TH_URG;
1259 	} else
1260 		/*
1261 		 * If no urgent pointer to send, then we pull
1262 		 * the urgent pointer to the left edge of the send window
1263 		 * so that it doesn't drift into the send window on sequence
1264 		 * number wraparound.
1265 		 */
1266 		tp->snd_up = tp->snd_una;		/* drag it along */
1267 
1268 	/*
1269 	 * Put TCP length in extended header, and then
1270 	 * checksum extended header and data.
1271 	 */
1272 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1273 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1274 
1275 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1276 	if (to.to_flags & TOF_SIGNATURE) {
1277 		/*
1278 		 * Calculate MD5 signature and put it into the place
1279 		 * determined before.
1280 		 * NOTE: since TCP options buffer doesn't point into
1281 		 * mbuf's data, calculate offset and use it.
1282 		 */
1283 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1284 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1285 			/*
1286 			 * Do not send segment if the calculation of MD5
1287 			 * digest has failed.
1288 			 */
1289 			m_freem(m);
1290 			goto out;
1291 		}
1292 	}
1293 #endif
1294 #ifdef INET6
1295 	if (isipv6) {
1296 		/*
1297 		 * There is no need to fill in ip6_plen right now.
1298 		 * It will be filled later by ip6_output.
1299 		 */
1300 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1301 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1302 		    optlen + len, IPPROTO_TCP, 0);
1303 	}
1304 #endif
1305 #if defined(INET6) && defined(INET)
1306 	else
1307 #endif
1308 #ifdef INET
1309 	{
1310 		m->m_pkthdr.csum_flags = CSUM_TCP;
1311 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1312 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1313 
1314 		/* IP version must be set here for ipv4/ipv6 checking later */
1315 		KASSERT(ip->ip_v == IPVERSION,
1316 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1317 	}
1318 #endif
1319 
1320 	/*
1321 	 * Enable TSO and specify the size of the segments.
1322 	 * The TCP pseudo header checksum is always provided.
1323 	 */
1324 	if (tso) {
1325 		KASSERT(len > tp->t_maxseg - optlen,
1326 		    ("%s: len <= tso_segsz", __func__));
1327 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1328 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1329 	}
1330 
1331 	KASSERT(len + hdrlen == m_length(m, NULL),
1332 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1333 	    __func__, len, hdrlen, m_length(m, NULL)));
1334 
1335 #ifdef TCP_HHOOK
1336 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1337 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1338 #endif
1339 
1340 #ifdef TCPDEBUG
1341 	/*
1342 	 * Trace.
1343 	 */
1344 	if (so->so_options & SO_DEBUG) {
1345 		u_short save = 0;
1346 #ifdef INET6
1347 		if (!isipv6)
1348 #endif
1349 		{
1350 			save = ipov->ih_len;
1351 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1352 		}
1353 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1354 #ifdef INET6
1355 		if (!isipv6)
1356 #endif
1357 		ipov->ih_len = save;
1358 	}
1359 #endif /* TCPDEBUG */
1360 	TCP_PROBE3(debug__output, tp, th, m);
1361 
1362 	/* We're getting ready to send; log now. */
1363 	TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1364 	    len, NULL, false);
1365 
1366 	/*
1367 	 * Fill in IP length and desired time to live and
1368 	 * send to IP level.  There should be a better way
1369 	 * to handle ttl and tos; we could keep them in
1370 	 * the template, but need a way to checksum without them.
1371 	 */
1372 	/*
1373 	 * m->m_pkthdr.len should have been set before checksum calculation,
1374 	 * because in6_cksum() need it.
1375 	 */
1376 #ifdef INET6
1377 	if (isipv6) {
1378 		/*
1379 		 * we separately set hoplimit for every segment, since the
1380 		 * user might want to change the value via setsockopt.
1381 		 * Also, desired default hop limit might be changed via
1382 		 * Neighbor Discovery.
1383 		 */
1384 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1385 
1386 		/*
1387 		 * Set the packet size here for the benefit of DTrace probes.
1388 		 * ip6_output() will set it properly; it's supposed to include
1389 		 * the option header lengths as well.
1390 		 */
1391 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1392 
1393 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1394 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1395 		else
1396 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1397 
1398 		if (tp->t_state == TCPS_SYN_SENT)
1399 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1400 
1401 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1402 
1403 #ifdef TCPPCAP
1404 		/* Save packet, if requested. */
1405 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1406 #endif
1407 
1408 		/* TODO: IPv6 IP6TOS_ECT bit on */
1409 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1410 		    &tp->t_inpcb->inp_route6,
1411 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1412 		    NULL, NULL, tp->t_inpcb);
1413 
1414 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1415 			mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1416 	}
1417 #endif /* INET6 */
1418 #if defined(INET) && defined(INET6)
1419 	else
1420 #endif
1421 #ifdef INET
1422     {
1423 	ip->ip_len = htons(m->m_pkthdr.len);
1424 #ifdef INET6
1425 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1426 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1427 #endif /* INET6 */
1428 	/*
1429 	 * If we do path MTU discovery, then we set DF on every packet.
1430 	 * This might not be the best thing to do according to RFC3390
1431 	 * Section 2. However the tcp hostcache migitates the problem
1432 	 * so it affects only the first tcp connection with a host.
1433 	 *
1434 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1435 	 */
1436 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1437 		ip->ip_off |= htons(IP_DF);
1438 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1439 	} else {
1440 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1441 	}
1442 
1443 	if (tp->t_state == TCPS_SYN_SENT)
1444 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1445 
1446 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1447 
1448 #ifdef TCPPCAP
1449 	/* Save packet, if requested. */
1450 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1451 #endif
1452 
1453 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1454 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1455 	    tp->t_inpcb);
1456 
1457 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1458 		mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1459     }
1460 #endif /* INET */
1461 
1462 out:
1463 	/*
1464 	 * In transmit state, time the transmission and arrange for
1465 	 * the retransmit.  In persist state, just set snd_max.
1466 	 */
1467 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1468 	    !tcp_timer_active(tp, TT_PERSIST)) {
1469 		tcp_seq startseq = tp->snd_nxt;
1470 
1471 		/*
1472 		 * Advance snd_nxt over sequence space of this segment.
1473 		 */
1474 		if (flags & (TH_SYN|TH_FIN)) {
1475 			if (flags & TH_SYN)
1476 				tp->snd_nxt++;
1477 			if (flags & TH_FIN) {
1478 				tp->snd_nxt++;
1479 				tp->t_flags |= TF_SENTFIN;
1480 			}
1481 		}
1482 		if (sack_rxmit)
1483 			goto timer;
1484 		tp->snd_nxt += len;
1485 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1486 			tp->snd_max = tp->snd_nxt;
1487 			/*
1488 			 * Time this transmission if not a retransmission and
1489 			 * not currently timing anything.
1490 			 */
1491 			if (tp->t_rtttime == 0) {
1492 				tp->t_rtttime = ticks;
1493 				tp->t_rtseq = startseq;
1494 				TCPSTAT_INC(tcps_segstimed);
1495 			}
1496 #ifdef STATS
1497 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1498 				tp->t_flags |= TF_GPUTINPROG;
1499 				tp->gput_seq = startseq;
1500 				tp->gput_ack = startseq +
1501 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1502 				tp->gput_ts = tcp_ts_getticks();
1503 			}
1504 #endif /* STATS */
1505 		}
1506 
1507 		/*
1508 		 * Set retransmit timer if not currently set,
1509 		 * and not doing a pure ack or a keep-alive probe.
1510 		 * Initial value for retransmit timer is smoothed
1511 		 * round-trip time + 2 * round-trip time variance.
1512 		 * Initialize shift counter which is used for backoff
1513 		 * of retransmit time.
1514 		 */
1515 timer:
1516 		if (!tcp_timer_active(tp, TT_REXMT) &&
1517 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1518 		     (tp->snd_nxt != tp->snd_una))) {
1519 			if (tcp_timer_active(tp, TT_PERSIST)) {
1520 				tcp_timer_activate(tp, TT_PERSIST, 0);
1521 				tp->t_rxtshift = 0;
1522 			}
1523 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1524 		} else if (len == 0 && sbavail(&so->so_snd) &&
1525 		    !tcp_timer_active(tp, TT_REXMT) &&
1526 		    !tcp_timer_active(tp, TT_PERSIST)) {
1527 			/*
1528 			 * Avoid a situation where we do not set persist timer
1529 			 * after a zero window condition. For example:
1530 			 * 1) A -> B: packet with enough data to fill the window
1531 			 * 2) B -> A: ACK for #1 + new data (0 window
1532 			 *    advertisement)
1533 			 * 3) A -> B: ACK for #2, 0 len packet
1534 			 *
1535 			 * In this case, A will not activate the persist timer,
1536 			 * because it chose to send a packet. Unless tcp_output
1537 			 * is called for some other reason (delayed ack timer,
1538 			 * another input packet from B, socket syscall), A will
1539 			 * not send zero window probes.
1540 			 *
1541 			 * So, if you send a 0-length packet, but there is data
1542 			 * in the socket buffer, and neither the rexmt or
1543 			 * persist timer is already set, then activate the
1544 			 * persist timer.
1545 			 */
1546 			tp->t_rxtshift = 0;
1547 			tcp_setpersist(tp);
1548 		}
1549 	} else {
1550 		/*
1551 		 * Persist case, update snd_max but since we are in
1552 		 * persist mode (no window) we do not update snd_nxt.
1553 		 */
1554 		int xlen = len;
1555 		if (flags & TH_SYN)
1556 			++xlen;
1557 		if (flags & TH_FIN) {
1558 			++xlen;
1559 			tp->t_flags |= TF_SENTFIN;
1560 		}
1561 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1562 			tp->snd_max = tp->snd_nxt + xlen;
1563 	}
1564 	if ((error == 0) &&
1565 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1566 	     (tp->t_flags & TF_SACK_PERMIT) &&
1567 	     tp->rcv_numsacks > 0)) {
1568 		    /* Clean up any DSACK's sent */
1569 		    tcp_clean_dsack_blocks(tp);
1570 	}
1571 	if (error) {
1572 		/* Record the error. */
1573 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1574 		    error, 0, NULL, false);
1575 
1576 		/*
1577 		 * We know that the packet was lost, so back out the
1578 		 * sequence number advance, if any.
1579 		 *
1580 		 * If the error is EPERM the packet got blocked by the
1581 		 * local firewall.  Normally we should terminate the
1582 		 * connection but the blocking may have been spurious
1583 		 * due to a firewall reconfiguration cycle.  So we treat
1584 		 * it like a packet loss and let the retransmit timer and
1585 		 * timeouts do their work over time.
1586 		 * XXX: It is a POLA question whether calling tcp_drop right
1587 		 * away would be the really correct behavior instead.
1588 		 */
1589 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1590 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1591 		    ((flags & TH_SYN) == 0) &&
1592 		    (error != EPERM)) {
1593 			if (sack_rxmit) {
1594 				p->rxmit -= len;
1595 				tp->sackhint.sack_bytes_rexmit -= len;
1596 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1597 				    ("sackhint bytes rtx >= 0"));
1598 			} else
1599 				tp->snd_nxt -= len;
1600 		}
1601 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1602 		switch (error) {
1603 		case EACCES:
1604 		case EPERM:
1605 			tp->t_softerror = error;
1606 			return (error);
1607 		case ENOBUFS:
1608 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1609 			tp->snd_cwnd = tp->t_maxseg;
1610 			return (0);
1611 		case EMSGSIZE:
1612 			/*
1613 			 * For some reason the interface we used initially
1614 			 * to send segments changed to another or lowered
1615 			 * its MTU.
1616 			 * If TSO was active we either got an interface
1617 			 * without TSO capabilits or TSO was turned off.
1618 			 * If we obtained mtu from ip_output() then update
1619 			 * it and try again.
1620 			 */
1621 			if (tso)
1622 				tp->t_flags &= ~TF_TSO;
1623 			if (mtu != 0) {
1624 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1625 				goto again;
1626 			}
1627 			return (error);
1628 		case EHOSTDOWN:
1629 		case EHOSTUNREACH:
1630 		case ENETDOWN:
1631 		case ENETUNREACH:
1632 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1633 				tp->t_softerror = error;
1634 				return (0);
1635 			}
1636 			/* FALLTHROUGH */
1637 		default:
1638 			return (error);
1639 		}
1640 	}
1641 	TCPSTAT_INC(tcps_sndtotal);
1642 
1643 	/*
1644 	 * Data sent (as far as we can tell).
1645 	 * If this advertises a larger window than any other segment,
1646 	 * then remember the size of the advertised window.
1647 	 * Any pending ACK has now been sent.
1648 	 */
1649 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1650 		tp->rcv_adv = tp->rcv_nxt + recwin;
1651 	tp->last_ack_sent = tp->rcv_nxt;
1652 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1653 	if (tcp_timer_active(tp, TT_DELACK))
1654 		tcp_timer_activate(tp, TT_DELACK, 0);
1655 #if 0
1656 	/*
1657 	 * This completely breaks TCP if newreno is turned on.  What happens
1658 	 * is that if delayed-acks are turned on on the receiver, this code
1659 	 * on the transmitter effectively destroys the TCP window, forcing
1660 	 * it to four packets (1.5Kx4 = 6K window).
1661 	 */
1662 	if (sendalot && --maxburst)
1663 		goto again;
1664 #endif
1665 	if (sendalot)
1666 		goto again;
1667 	return (0);
1668 }
1669 
1670 void
1671 tcp_setpersist(struct tcpcb *tp)
1672 {
1673 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1674 	int tt;
1675 
1676 	tp->t_flags &= ~TF_PREVVALID;
1677 	if (tcp_timer_active(tp, TT_REXMT))
1678 		panic("tcp_setpersist: retransmit pending");
1679 	/*
1680 	 * Start/restart persistence timer.
1681 	 */
1682 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1683 		      tcp_persmin, tcp_persmax);
1684 	tcp_timer_activate(tp, TT_PERSIST, tt);
1685 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1686 		tp->t_rxtshift++;
1687 }
1688 
1689 /*
1690  * Insert TCP options according to the supplied parameters to the place
1691  * optp in a consistent way.  Can handle unaligned destinations.
1692  *
1693  * The order of the option processing is crucial for optimal packing and
1694  * alignment for the scarce option space.
1695  *
1696  * The optimal order for a SYN/SYN-ACK segment is:
1697  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1698  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1699  *
1700  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1701  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1702  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1703  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1704  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1705  */
1706 int
1707 tcp_addoptions(struct tcpopt *to, u_char *optp)
1708 {
1709 	u_int32_t mask, optlen = 0;
1710 
1711 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1712 		if ((to->to_flags & mask) != mask)
1713 			continue;
1714 		if (optlen == TCP_MAXOLEN)
1715 			break;
1716 		switch (to->to_flags & mask) {
1717 		case TOF_MSS:
1718 			while (optlen % 4) {
1719 				optlen += TCPOLEN_NOP;
1720 				*optp++ = TCPOPT_NOP;
1721 			}
1722 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1723 				continue;
1724 			optlen += TCPOLEN_MAXSEG;
1725 			*optp++ = TCPOPT_MAXSEG;
1726 			*optp++ = TCPOLEN_MAXSEG;
1727 			to->to_mss = htons(to->to_mss);
1728 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1729 			optp += sizeof(to->to_mss);
1730 			break;
1731 		case TOF_SCALE:
1732 			while (!optlen || optlen % 2 != 1) {
1733 				optlen += TCPOLEN_NOP;
1734 				*optp++ = TCPOPT_NOP;
1735 			}
1736 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1737 				continue;
1738 			optlen += TCPOLEN_WINDOW;
1739 			*optp++ = TCPOPT_WINDOW;
1740 			*optp++ = TCPOLEN_WINDOW;
1741 			*optp++ = to->to_wscale;
1742 			break;
1743 		case TOF_SACKPERM:
1744 			while (optlen % 2) {
1745 				optlen += TCPOLEN_NOP;
1746 				*optp++ = TCPOPT_NOP;
1747 			}
1748 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1749 				continue;
1750 			optlen += TCPOLEN_SACK_PERMITTED;
1751 			*optp++ = TCPOPT_SACK_PERMITTED;
1752 			*optp++ = TCPOLEN_SACK_PERMITTED;
1753 			break;
1754 		case TOF_TS:
1755 			while (!optlen || optlen % 4 != 2) {
1756 				optlen += TCPOLEN_NOP;
1757 				*optp++ = TCPOPT_NOP;
1758 			}
1759 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1760 				continue;
1761 			optlen += TCPOLEN_TIMESTAMP;
1762 			*optp++ = TCPOPT_TIMESTAMP;
1763 			*optp++ = TCPOLEN_TIMESTAMP;
1764 			to->to_tsval = htonl(to->to_tsval);
1765 			to->to_tsecr = htonl(to->to_tsecr);
1766 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1767 			optp += sizeof(to->to_tsval);
1768 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1769 			optp += sizeof(to->to_tsecr);
1770 			break;
1771 		case TOF_SIGNATURE:
1772 			{
1773 			int siglen = TCPOLEN_SIGNATURE - 2;
1774 
1775 			while (!optlen || optlen % 4 != 2) {
1776 				optlen += TCPOLEN_NOP;
1777 				*optp++ = TCPOPT_NOP;
1778 			}
1779 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1780 				to->to_flags &= ~TOF_SIGNATURE;
1781 				continue;
1782 			}
1783 			optlen += TCPOLEN_SIGNATURE;
1784 			*optp++ = TCPOPT_SIGNATURE;
1785 			*optp++ = TCPOLEN_SIGNATURE;
1786 			to->to_signature = optp;
1787 			while (siglen--)
1788 				 *optp++ = 0;
1789 			break;
1790 			}
1791 		case TOF_SACK:
1792 			{
1793 			int sackblks = 0;
1794 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1795 			tcp_seq sack_seq;
1796 
1797 			while (!optlen || optlen % 4 != 2) {
1798 				optlen += TCPOLEN_NOP;
1799 				*optp++ = TCPOPT_NOP;
1800 			}
1801 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1802 				continue;
1803 			optlen += TCPOLEN_SACKHDR;
1804 			*optp++ = TCPOPT_SACK;
1805 			sackblks = min(to->to_nsacks,
1806 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1807 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1808 			while (sackblks--) {
1809 				sack_seq = htonl(sack->start);
1810 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1811 				optp += sizeof(sack_seq);
1812 				sack_seq = htonl(sack->end);
1813 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1814 				optp += sizeof(sack_seq);
1815 				optlen += TCPOLEN_SACK;
1816 				sack++;
1817 			}
1818 			TCPSTAT_INC(tcps_sack_send_blocks);
1819 			break;
1820 			}
1821 		case TOF_FASTOPEN:
1822 			{
1823 			int total_len;
1824 
1825 			/* XXX is there any point to aligning this option? */
1826 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1827 			if (TCP_MAXOLEN - optlen < total_len) {
1828 				to->to_flags &= ~TOF_FASTOPEN;
1829 				continue;
1830 			}
1831 			*optp++ = TCPOPT_FAST_OPEN;
1832 			*optp++ = total_len;
1833 			if (to->to_tfo_len > 0) {
1834 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1835 				optp += to->to_tfo_len;
1836 			}
1837 			optlen += total_len;
1838 			break;
1839 			}
1840 		default:
1841 			panic("%s: unknown TCP option type", __func__);
1842 			break;
1843 		}
1844 	}
1845 
1846 	/* Terminate and pad TCP options to a 4 byte boundary. */
1847 	if (optlen % 4) {
1848 		optlen += TCPOLEN_EOL;
1849 		*optp++ = TCPOPT_EOL;
1850 	}
1851 	/*
1852 	 * According to RFC 793 (STD0007):
1853 	 *   "The content of the header beyond the End-of-Option option
1854 	 *    must be header padding (i.e., zero)."
1855 	 *   and later: "The padding is composed of zeros."
1856 	 */
1857 	while (optlen % 4) {
1858 		optlen += TCPOLEN_PAD;
1859 		*optp++ = TCPOPT_PAD;
1860 	}
1861 
1862 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1863 	return (optlen);
1864 }
1865 
1866 /*
1867  * This is a copy of m_copym(), taking the TSO segment size/limit
1868  * constraints into account, and advancing the sndptr as it goes.
1869  */
1870 struct mbuf *
1871 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1872     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1873 {
1874 #ifdef KERN_TLS
1875 	struct ktls_session *tls, *ntls;
1876 	struct mbuf *start;
1877 #endif
1878 	struct mbuf *n, **np;
1879 	struct mbuf *top;
1880 	int32_t off = off0;
1881 	int32_t len = *plen;
1882 	int32_t fragsize;
1883 	int32_t len_cp = 0;
1884 	int32_t *pkthdrlen;
1885 	uint32_t mlen, frags;
1886 	bool copyhdr;
1887 
1888 
1889 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1890 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1891 	if (off == 0 && m->m_flags & M_PKTHDR)
1892 		copyhdr = true;
1893 	else
1894 		copyhdr = false;
1895 	while (off > 0) {
1896 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1897 		if (off < m->m_len)
1898 			break;
1899 		off -= m->m_len;
1900 		if ((sb) && (m == sb->sb_sndptr)) {
1901 			sb->sb_sndptroff += m->m_len;
1902 			sb->sb_sndptr = m->m_next;
1903 		}
1904 		m = m->m_next;
1905 	}
1906 	np = &top;
1907 	top = NULL;
1908 	pkthdrlen = NULL;
1909 #ifdef KERN_TLS
1910 	if (hw_tls && (m->m_flags & M_NOMAP))
1911 		tls = m->m_ext.ext_pgs->tls;
1912 	else
1913 		tls = NULL;
1914 	start = m;
1915 #endif
1916 	while (len > 0) {
1917 		if (m == NULL) {
1918 			KASSERT(len == M_COPYALL,
1919 			    ("tcp_m_copym, length > size of mbuf chain"));
1920 			*plen = len_cp;
1921 			if (pkthdrlen != NULL)
1922 				*pkthdrlen = len_cp;
1923 			break;
1924 		}
1925 #ifdef KERN_TLS
1926 		if (hw_tls) {
1927 			if (m->m_flags & M_NOMAP)
1928 				ntls = m->m_ext.ext_pgs->tls;
1929 			else
1930 				ntls = NULL;
1931 
1932 			/*
1933 			 * Avoid mixing TLS records with handshake
1934 			 * data or TLS records from different
1935 			 * sessions.
1936 			 */
1937 			if (tls != ntls) {
1938 				MPASS(m != start);
1939 				*plen = len_cp;
1940 				if (pkthdrlen != NULL)
1941 					*pkthdrlen = len_cp;
1942 				break;
1943 			}
1944 
1945 			/*
1946 			 * Don't end a send in the middle of a TLS
1947 			 * record if it spans multiple TLS records.
1948 			 */
1949 			if (tls != NULL && (m != start) && len < m->m_len) {
1950 				*plen = len_cp;
1951 				if (pkthdrlen != NULL)
1952 					*pkthdrlen = len_cp;
1953 				break;
1954 			}
1955 		}
1956 #endif
1957 		mlen = min(len, m->m_len - off);
1958 		if (seglimit) {
1959 			/*
1960 			 * For M_NOMAP mbufs, add 3 segments
1961 			 * + 1 in case we are crossing page boundaries
1962 			 * + 2 in case the TLS hdr/trailer are used
1963 			 * It is cheaper to just add the segments
1964 			 * than it is to take the cache miss to look
1965 			 * at the mbuf ext_pgs state in detail.
1966 			 */
1967 			if (m->m_flags & M_NOMAP) {
1968 				fragsize = min(segsize, PAGE_SIZE);
1969 				frags = 3;
1970 			} else {
1971 				fragsize = segsize;
1972 				frags = 0;
1973 			}
1974 
1975 			/* Break if we really can't fit anymore. */
1976 			if ((frags + 1) >= seglimit) {
1977 				*plen =	len_cp;
1978 				if (pkthdrlen != NULL)
1979 					*pkthdrlen = len_cp;
1980 				break;
1981 			}
1982 
1983 			/*
1984 			 * Reduce size if you can't copy the whole
1985 			 * mbuf. If we can't copy the whole mbuf, also
1986 			 * adjust len so the loop will end after this
1987 			 * mbuf.
1988 			 */
1989 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1990 				mlen = (seglimit - frags - 1) * fragsize;
1991 				len = mlen;
1992 				*plen = len_cp + len;
1993 				if (pkthdrlen != NULL)
1994 					*pkthdrlen = *plen;
1995 			}
1996 			frags += howmany(mlen, fragsize);
1997 			if (frags == 0)
1998 				frags++;
1999 			seglimit -= frags;
2000 			KASSERT(seglimit > 0,
2001 			    ("%s: seglimit went too low", __func__));
2002 		}
2003 		if (copyhdr)
2004 			n = m_gethdr(M_NOWAIT, m->m_type);
2005 		else
2006 			n = m_get(M_NOWAIT, m->m_type);
2007 		*np = n;
2008 		if (n == NULL)
2009 			goto nospace;
2010 		if (copyhdr) {
2011 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2012 				goto nospace;
2013 			if (len == M_COPYALL)
2014 				n->m_pkthdr.len -= off0;
2015 			else
2016 				n->m_pkthdr.len = len;
2017 			pkthdrlen = &n->m_pkthdr.len;
2018 			copyhdr = false;
2019 		}
2020 		n->m_len = mlen;
2021 		len_cp += n->m_len;
2022 		if (m->m_flags & M_EXT) {
2023 			n->m_data = m->m_data + off;
2024 			mb_dupcl(n, m);
2025 		} else
2026 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2027 			    (u_int)n->m_len);
2028 
2029 		if (sb && (sb->sb_sndptr == m) &&
2030 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2031 			sb->sb_sndptroff += m->m_len;
2032 			sb->sb_sndptr = m->m_next;
2033 		}
2034 		off = 0;
2035 		if (len != M_COPYALL) {
2036 			len -= n->m_len;
2037 		}
2038 		m = m->m_next;
2039 		np = &n->m_next;
2040 	}
2041 	return (top);
2042 nospace:
2043 	m_freem(top);
2044 	return (NULL);
2045 }
2046 
2047 void
2048 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2049 {
2050 
2051 	/*
2052 	 * Automatic sizing of send socket buffer.  Often the send buffer
2053 	 * size is not optimally adjusted to the actual network conditions
2054 	 * at hand (delay bandwidth product).  Setting the buffer size too
2055 	 * small limits throughput on links with high bandwidth and high
2056 	 * delay (eg. trans-continental/oceanic links).  Setting the
2057 	 * buffer size too big consumes too much real kernel memory,
2058 	 * especially with many connections on busy servers.
2059 	 *
2060 	 * The criteria to step up the send buffer one notch are:
2061 	 *  1. receive window of remote host is larger than send buffer
2062 	 *     (with a fudge factor of 5/4th);
2063 	 *  2. send buffer is filled to 7/8th with data (so we actually
2064 	 *     have data to make use of it);
2065 	 *  3. send buffer fill has not hit maximal automatic size;
2066 	 *  4. our send window (slow start and cogestion controlled) is
2067 	 *     larger than sent but unacknowledged data in send buffer.
2068 	 *
2069 	 * The remote host receive window scaling factor may limit the
2070 	 * growing of the send buffer before it reaches its allowed
2071 	 * maximum.
2072 	 *
2073 	 * It scales directly with slow start or congestion window
2074 	 * and does at most one step per received ACK.  This fast
2075 	 * scaling has the drawback of growing the send buffer beyond
2076 	 * what is strictly necessary to make full use of a given
2077 	 * delay*bandwidth product.  However testing has shown this not
2078 	 * to be much of an problem.  At worst we are trading wasting
2079 	 * of available bandwidth (the non-use of it) for wasting some
2080 	 * socket buffer memory.
2081 	 *
2082 	 * TODO: Shrink send buffer during idle periods together
2083 	 * with congestion window.  Requires another timer.  Has to
2084 	 * wait for upcoming tcp timer rewrite.
2085 	 *
2086 	 * XXXGL: should there be used sbused() or sbavail()?
2087 	 */
2088 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2089 		int lowat;
2090 
2091 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2092 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2093 		    sbused(&so->so_snd) >=
2094 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2095 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2096 		    sendwin >= (sbused(&so->so_snd) -
2097 		    (tp->snd_nxt - tp->snd_una))) {
2098 			if (!sbreserve_locked(&so->so_snd,
2099 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2100 			     V_tcp_autosndbuf_max), so, curthread))
2101 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2102 		}
2103 	}
2104 }
2105