xref: /freebsd/sys/netinet/tcp_output.c (revision 752c1d14e398faab8e6f19f99ee29df87ecb05e1)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_kern_tls.h"
41 #include "opt_tcpdebug.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/arb.h>
46 #include <sys/domain.h>
47 #ifdef TCP_HHOOK
48 #include <sys/hhook.h>
49 #endif
50 #include <sys/kernel.h>
51 #ifdef KERN_TLS
52 #include <sys/ktls.h>
53 #endif
54 #include <sys/lock.h>
55 #include <sys/mbuf.h>
56 #include <sys/mutex.h>
57 #include <sys/protosw.h>
58 #include <sys/qmath.h>
59 #include <sys/sdt.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/sysctl.h>
63 #include <sys/stats.h>
64 
65 #include <net/if.h>
66 #include <net/route.h>
67 #include <net/route/nhop.h>
68 #include <net/vnet.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
77 #ifdef INET6
78 #include <netinet6/in6_pcb.h>
79 #include <netinet/ip6.h>
80 #include <netinet6/ip6_var.h>
81 #endif
82 #include <netinet/tcp.h>
83 #define	TCPOUTFLAGS
84 #include <netinet/tcp_fsm.h>
85 #include <netinet/tcp_log_buf.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcpip.h>
90 #include <netinet/cc/cc.h>
91 #include <netinet/tcp_fastopen.h>
92 #ifdef TCPPCAP
93 #include <netinet/tcp_pcap.h>
94 #endif
95 #ifdef TCPDEBUG
96 #include <netinet/tcp_debug.h>
97 #endif
98 #ifdef TCP_OFFLOAD
99 #include <netinet/tcp_offload.h>
100 #endif
101 
102 #include <netipsec/ipsec_support.h>
103 
104 #include <machine/in_cksum.h>
105 
106 #include <security/mac/mac_framework.h>
107 
108 VNET_DEFINE(int, path_mtu_discovery) = 1;
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
110 	&VNET_NAME(path_mtu_discovery), 1,
111 	"Enable Path MTU Discovery");
112 
113 VNET_DEFINE(int, tcp_do_tso) = 1;
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
115 	&VNET_NAME(tcp_do_tso), 0,
116 	"Enable TCP Segmentation Offload");
117 
118 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
119 #define	V_tcp_sendspace	VNET(tcp_sendspace)
120 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
121 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
122 
123 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
125 	&VNET_NAME(tcp_do_autosndbuf), 0,
126 	"Enable automatic send buffer sizing");
127 
128 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
130 	&VNET_NAME(tcp_autosndbuf_inc), 0,
131 	"Incrementor step size of automatic send buffer");
132 
133 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
134 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
135 	&VNET_NAME(tcp_autosndbuf_max), 0,
136 	"Max size of automatic send buffer");
137 
138 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
139 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
141 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
142 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
143 
144 /*
145  * Make sure that either retransmit or persist timer is set for SYN, FIN and
146  * non-ACK.
147  */
148 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
149 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
150 	    tcp_timer_active((tp), TT_REXMT) ||				\
151 	    tcp_timer_active((tp), TT_PERSIST),				\
152 	    ("neither rexmt nor persist timer is set"))
153 
154 static void inline	cc_after_idle(struct tcpcb *tp);
155 
156 #ifdef TCP_HHOOK
157 /*
158  * Wrapper for the TCP established output helper hook.
159  */
160 void
161 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
162     struct tcpopt *to, uint32_t len, int tso)
163 {
164 	struct tcp_hhook_data hhook_data;
165 
166 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
167 		hhook_data.tp = tp;
168 		hhook_data.th = th;
169 		hhook_data.to = to;
170 		hhook_data.len = len;
171 		hhook_data.tso = tso;
172 
173 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
174 		    tp->osd);
175 	}
176 }
177 #endif
178 
179 /*
180  * CC wrapper hook functions
181  */
182 static void inline
183 cc_after_idle(struct tcpcb *tp)
184 {
185 	INP_WLOCK_ASSERT(tp->t_inpcb);
186 
187 	if (CC_ALGO(tp)->after_idle != NULL)
188 		CC_ALGO(tp)->after_idle(tp->ccv);
189 }
190 
191 /*
192  * Tcp output routine: figure out what should be sent and send it.
193  */
194 int
195 tcp_output(struct tcpcb *tp)
196 {
197 	struct socket *so = tp->t_inpcb->inp_socket;
198 	int32_t len;
199 	uint32_t recwin, sendwin;
200 	int off, flags, error = 0;	/* Keep compiler happy */
201 	u_int if_hw_tsomaxsegcount = 0;
202 	u_int if_hw_tsomaxsegsize = 0;
203 	struct mbuf *m;
204 	struct ip *ip = NULL;
205 #ifdef TCPDEBUG
206 	struct ipovly *ipov = NULL;
207 #endif
208 	struct tcphdr *th;
209 	u_char opt[TCP_MAXOLEN];
210 	unsigned ipoptlen, optlen, hdrlen;
211 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
212 	unsigned ipsec_optlen = 0;
213 #endif
214 	int idle, sendalot, curticks;
215 	int sack_rxmit, sack_bytes_rxmt;
216 	struct sackhole *p;
217 	int tso, mtu;
218 	struct tcpopt to;
219 	unsigned int wanted_cookie = 0;
220 	unsigned int dont_sendalot = 0;
221 #if 0
222 	int maxburst = TCP_MAXBURST;
223 #endif
224 #ifdef INET6
225 	struct ip6_hdr *ip6 = NULL;
226 	int isipv6;
227 
228 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
229 #endif
230 #ifdef KERN_TLS
231 	const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
232 #else
233 	const bool hw_tls = false;
234 #endif
235 
236 	NET_EPOCH_ASSERT();
237 	INP_WLOCK_ASSERT(tp->t_inpcb);
238 
239 #ifdef TCP_OFFLOAD
240 	if (tp->t_flags & TF_TOE)
241 		return (tcp_offload_output(tp));
242 #endif
243 
244 	/*
245 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
246 	 * only allow the initial SYN or SYN|ACK and those sent
247 	 * by the retransmit timer.
248 	 */
249 	if (IS_FASTOPEN(tp->t_flags) &&
250 	    ((tp->t_state == TCPS_SYN_SENT) ||
251 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
252 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
253 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
254 		return (0);
255 
256 	/*
257 	 * Determine length of data that should be transmitted,
258 	 * and flags that will be used.
259 	 * If there is some data or critical controls (SYN, RST)
260 	 * to send, then transmit; otherwise, investigate further.
261 	 */
262 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
263 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
264 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
265 		cc_after_idle(tp);
266 	tp->t_flags &= ~TF_LASTIDLE;
267 	if (idle) {
268 		if (tp->t_flags & TF_MORETOCOME) {
269 			tp->t_flags |= TF_LASTIDLE;
270 			idle = 0;
271 		}
272 	}
273 again:
274 	/*
275 	 * If we've recently taken a timeout, snd_max will be greater than
276 	 * snd_nxt.  There may be SACK information that allows us to avoid
277 	 * resending already delivered data.  Adjust snd_nxt accordingly.
278 	 */
279 	if ((tp->t_flags & TF_SACK_PERMIT) &&
280 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
281 		tcp_sack_adjust(tp);
282 	sendalot = 0;
283 	tso = 0;
284 	mtu = 0;
285 	off = tp->snd_nxt - tp->snd_una;
286 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
287 
288 	flags = tcp_outflags[tp->t_state];
289 	/*
290 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
291 	 * to send out new data (when sendalot is 1), bypass this function.
292 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
293 	 * we're replacing a (future) new transmission with a retransmission
294 	 * now, and we previously incremented snd_cwnd in tcp_input().
295 	 */
296 	/*
297 	 * Still in sack recovery , reset rxmit flag to zero.
298 	 */
299 	sack_rxmit = 0;
300 	sack_bytes_rxmt = 0;
301 	len = 0;
302 	p = NULL;
303 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
304 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
305 		uint32_t cwin;
306 
307 		cwin =
308 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
309 		/* Do not retransmit SACK segments beyond snd_recover */
310 		if (SEQ_GT(p->end, tp->snd_recover)) {
311 			/*
312 			 * (At least) part of sack hole extends beyond
313 			 * snd_recover. Check to see if we can rexmit data
314 			 * for this hole.
315 			 */
316 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
317 				/*
318 				 * Can't rexmit any more data for this hole.
319 				 * That data will be rexmitted in the next
320 				 * sack recovery episode, when snd_recover
321 				 * moves past p->rxmit.
322 				 */
323 				p = NULL;
324 				goto after_sack_rexmit;
325 			} else
326 				/* Can rexmit part of the current hole */
327 				len = ((int32_t)ulmin(cwin,
328 						   tp->snd_recover - p->rxmit));
329 		} else
330 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
331 		off = p->rxmit - tp->snd_una;
332 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
333 		    __func__, off));
334 		if (len > 0) {
335 			sack_rxmit = 1;
336 			sendalot = 1;
337 			TCPSTAT_INC(tcps_sack_rexmits);
338 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
339 			    min(len, tp->t_maxseg));
340 		}
341 	}
342 after_sack_rexmit:
343 	/*
344 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
345 	 * state flags.
346 	 */
347 	if (tp->t_flags & TF_NEEDFIN)
348 		flags |= TH_FIN;
349 	if (tp->t_flags & TF_NEEDSYN)
350 		flags |= TH_SYN;
351 
352 	SOCKBUF_LOCK(&so->so_snd);
353 	/*
354 	 * If in persist timeout with window of 0, send 1 byte.
355 	 * Otherwise, if window is small but nonzero
356 	 * and timer expired, we will send what we can
357 	 * and go to transmit state.
358 	 */
359 	if (tp->t_flags & TF_FORCEDATA) {
360 		if (sendwin == 0) {
361 			/*
362 			 * If we still have some data to send, then
363 			 * clear the FIN bit.  Usually this would
364 			 * happen below when it realizes that we
365 			 * aren't sending all the data.  However,
366 			 * if we have exactly 1 byte of unsent data,
367 			 * then it won't clear the FIN bit below,
368 			 * and if we are in persist state, we wind
369 			 * up sending the packet without recording
370 			 * that we sent the FIN bit.
371 			 *
372 			 * We can't just blindly clear the FIN bit,
373 			 * because if we don't have any more data
374 			 * to send then the probe will be the FIN
375 			 * itself.
376 			 */
377 			if (off < sbused(&so->so_snd))
378 				flags &= ~TH_FIN;
379 			sendwin = 1;
380 		} else {
381 			tcp_timer_activate(tp, TT_PERSIST, 0);
382 			tp->t_rxtshift = 0;
383 		}
384 	}
385 
386 	/*
387 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
388 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
389 	 * a negative length.  This can also occur when TCP opens up
390 	 * its congestion window while receiving additional duplicate
391 	 * acks after fast-retransmit because TCP will reset snd_nxt
392 	 * to snd_max after the fast-retransmit.
393 	 *
394 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
395 	 * be set to snd_una, the offset will be 0, and the length may
396 	 * wind up 0.
397 	 *
398 	 * If sack_rxmit is true we are retransmitting from the scoreboard
399 	 * in which case len is already set.
400 	 */
401 	if (sack_rxmit == 0) {
402 		if (sack_bytes_rxmt == 0)
403 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
404 			    off);
405 		else {
406 			int32_t cwin;
407 
408                         /*
409 			 * We are inside of a SACK recovery episode and are
410 			 * sending new data, having retransmitted all the
411 			 * data possible in the scoreboard.
412 			 */
413 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
414 			    off);
415 			/*
416 			 * Don't remove this (len > 0) check !
417 			 * We explicitly check for len > 0 here (although it
418 			 * isn't really necessary), to work around a gcc
419 			 * optimization issue - to force gcc to compute
420 			 * len above. Without this check, the computation
421 			 * of len is bungled by the optimizer.
422 			 */
423 			if (len > 0) {
424 				cwin = tp->snd_cwnd -
425 					(tp->snd_nxt - tp->snd_recover) -
426 					sack_bytes_rxmt;
427 				if (cwin < 0)
428 					cwin = 0;
429 				len = imin(len, cwin);
430 			}
431 		}
432 	}
433 
434 	/*
435 	 * Lop off SYN bit if it has already been sent.  However, if this
436 	 * is SYN-SENT state and if segment contains data and if we don't
437 	 * know that foreign host supports TAO, suppress sending segment.
438 	 */
439 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
440 		if (tp->t_state != TCPS_SYN_RECEIVED)
441 			flags &= ~TH_SYN;
442 		/*
443 		 * When sending additional segments following a TFO SYN|ACK,
444 		 * do not include the SYN bit.
445 		 */
446 		if (IS_FASTOPEN(tp->t_flags) &&
447 		    (tp->t_state == TCPS_SYN_RECEIVED))
448 			flags &= ~TH_SYN;
449 		off--, len++;
450 	}
451 
452 	/*
453 	 * Be careful not to send data and/or FIN on SYN segments.
454 	 * This measure is needed to prevent interoperability problems
455 	 * with not fully conformant TCP implementations.
456 	 */
457 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
458 		len = 0;
459 		flags &= ~TH_FIN;
460 	}
461 
462 	/*
463 	 * On TFO sockets, ensure no data is sent in the following cases:
464 	 *
465 	 *  - When retransmitting SYN|ACK on a passively-created socket
466 	 *
467 	 *  - When retransmitting SYN on an actively created socket
468 	 *
469 	 *  - When sending a zero-length cookie (cookie request) on an
470 	 *    actively created socket
471 	 *
472 	 *  - When the socket is in the CLOSED state (RST is being sent)
473 	 */
474 	if (IS_FASTOPEN(tp->t_flags) &&
475 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
476 	     ((tp->t_state == TCPS_SYN_SENT) &&
477 	      (tp->t_tfo_client_cookie_len == 0)) ||
478 	     (flags & TH_RST)))
479 		len = 0;
480 	if (len <= 0) {
481 		/*
482 		 * If FIN has been sent but not acked,
483 		 * but we haven't been called to retransmit,
484 		 * len will be < 0.  Otherwise, window shrank
485 		 * after we sent into it.  If window shrank to 0,
486 		 * cancel pending retransmit, pull snd_nxt back
487 		 * to (closed) window, and set the persist timer
488 		 * if it isn't already going.  If the window didn't
489 		 * close completely, just wait for an ACK.
490 		 *
491 		 * We also do a general check here to ensure that
492 		 * we will set the persist timer when we have data
493 		 * to send, but a 0-byte window. This makes sure
494 		 * the persist timer is set even if the packet
495 		 * hits one of the "goto send" lines below.
496 		 */
497 		len = 0;
498 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
499 			(off < (int) sbavail(&so->so_snd))) {
500 			tcp_timer_activate(tp, TT_REXMT, 0);
501 			tp->t_rxtshift = 0;
502 			tp->snd_nxt = tp->snd_una;
503 			if (!tcp_timer_active(tp, TT_PERSIST))
504 				tcp_setpersist(tp);
505 		}
506 	}
507 
508 	/* len will be >= 0 after this point. */
509 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
510 
511 	tcp_sndbuf_autoscale(tp, so, sendwin);
512 
513 	/*
514 	 * Decide if we can use TCP Segmentation Offloading (if supported by
515 	 * hardware).
516 	 *
517 	 * TSO may only be used if we are in a pure bulk sending state.  The
518 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
519 	 * IP options prevent using TSO.  With TSO the TCP header is the same
520 	 * (except for the sequence number) for all generated packets.  This
521 	 * makes it impossible to transmit any options which vary per generated
522 	 * segment or packet.
523 	 *
524 	 * IPv4 handling has a clear separation of ip options and ip header
525 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
526 	 * the right thing below to provide length of just ip options and thus
527 	 * checking for ipoptlen is enough to decide if ip options are present.
528 	 */
529 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
530 	/*
531 	 * Pre-calculate here as we save another lookup into the darknesses
532 	 * of IPsec that way and can actually decide if TSO is ok.
533 	 */
534 #ifdef INET6
535 	if (isipv6 && IPSEC_ENABLED(ipv6))
536 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
537 #ifdef INET
538 	else
539 #endif
540 #endif /* INET6 */
541 #ifdef INET
542 	if (IPSEC_ENABLED(ipv4))
543 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
544 #endif /* INET */
545 #endif /* IPSEC */
546 #ifdef INET6
547 	if (isipv6)
548 		ipoptlen = ip6_optlen(tp->t_inpcb);
549 	else
550 #endif
551 	if (tp->t_inpcb->inp_options)
552 		ipoptlen = tp->t_inpcb->inp_options->m_len -
553 				offsetof(struct ipoption, ipopt_list);
554 	else
555 		ipoptlen = 0;
556 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
557 	ipoptlen += ipsec_optlen;
558 #endif
559 
560 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
561 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
562 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
563 	    ipoptlen == 0 && !(flags & TH_SYN))
564 		tso = 1;
565 
566 	if (sack_rxmit) {
567 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
568 			flags &= ~TH_FIN;
569 	} else {
570 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
571 		    sbused(&so->so_snd)))
572 			flags &= ~TH_FIN;
573 	}
574 
575 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
576 	    (long)TCP_MAXWIN << tp->rcv_scale);
577 
578 	/*
579 	 * Sender silly window avoidance.   We transmit under the following
580 	 * conditions when len is non-zero:
581 	 *
582 	 *	- We have a full segment (or more with TSO)
583 	 *	- This is the last buffer in a write()/send() and we are
584 	 *	  either idle or running NODELAY
585 	 *	- we've timed out (e.g. persist timer)
586 	 *	- we have more then 1/2 the maximum send window's worth of
587 	 *	  data (receiver may be limited the window size)
588 	 *	- we need to retransmit
589 	 */
590 	if (len) {
591 		if (len >= tp->t_maxseg)
592 			goto send;
593 		/*
594 		 * As the TCP header options are now
595 		 * considered when setting up the initial
596 		 * window, we would not send the last segment
597 		 * if we skip considering the option length here.
598 		 * Note: this may not work when tcp headers change
599 		 * very dynamically in the future.
600 		 */
601 		if ((((tp->t_flags & TF_SIGNATURE) ?
602 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
603 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
604 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
605 		    len) >= tp->t_maxseg)
606 			goto send;
607 		/*
608 		 * NOTE! on localhost connections an 'ack' from the remote
609 		 * end may occur synchronously with the output and cause
610 		 * us to flush a buffer queued with moretocome.  XXX
611 		 *
612 		 * note: the len + off check is almost certainly unnecessary.
613 		 */
614 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
615 		    (idle || (tp->t_flags & TF_NODELAY)) &&
616 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
617 		    (tp->t_flags & TF_NOPUSH) == 0) {
618 			goto send;
619 		}
620 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
621 			goto send;
622 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
623 			goto send;
624 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
625 			goto send;
626 		if (sack_rxmit)
627 			goto send;
628 	}
629 
630 	/*
631 	 * Sending of standalone window updates.
632 	 *
633 	 * Window updates are important when we close our window due to a
634 	 * full socket buffer and are opening it again after the application
635 	 * reads data from it.  Once the window has opened again and the
636 	 * remote end starts to send again the ACK clock takes over and
637 	 * provides the most current window information.
638 	 *
639 	 * We must avoid the silly window syndrome whereas every read
640 	 * from the receive buffer, no matter how small, causes a window
641 	 * update to be sent.  We also should avoid sending a flurry of
642 	 * window updates when the socket buffer had queued a lot of data
643 	 * and the application is doing small reads.
644 	 *
645 	 * Prevent a flurry of pointless window updates by only sending
646 	 * an update when we can increase the advertized window by more
647 	 * than 1/4th of the socket buffer capacity.  When the buffer is
648 	 * getting full or is very small be more aggressive and send an
649 	 * update whenever we can increase by two mss sized segments.
650 	 * In all other situations the ACK's to new incoming data will
651 	 * carry further window increases.
652 	 *
653 	 * Don't send an independent window update if a delayed
654 	 * ACK is pending (it will get piggy-backed on it) or the
655 	 * remote side already has done a half-close and won't send
656 	 * more data.  Skip this if the connection is in T/TCP
657 	 * half-open state.
658 	 */
659 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
660 	    !(tp->t_flags & TF_DELACK) &&
661 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
662 		/*
663 		 * "adv" is the amount we could increase the window,
664 		 * taking into account that we are limited by
665 		 * TCP_MAXWIN << tp->rcv_scale.
666 		 */
667 		int32_t adv;
668 		int oldwin;
669 
670 		adv = recwin;
671 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
672 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
673 			if (adv > oldwin)
674 				adv -= oldwin;
675 			else
676 				adv = 0;
677 		} else
678 			oldwin = 0;
679 
680 		/*
681 		 * If the new window size ends up being the same as or less
682 		 * than the old size when it is scaled, then don't force
683 		 * a window update.
684 		 */
685 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
686 			goto dontupdate;
687 
688 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
689 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
690 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
691 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
692 		     adv >= TCP_MAXWIN << tp->rcv_scale))
693 			goto send;
694 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
695 			goto send;
696 	}
697 dontupdate:
698 
699 	/*
700 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
701 	 * is also a catch-all for the retransmit timer timeout case.
702 	 */
703 	if (tp->t_flags & TF_ACKNOW)
704 		goto send;
705 	if ((flags & TH_RST) ||
706 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
707 		goto send;
708 	if (SEQ_GT(tp->snd_up, tp->snd_una))
709 		goto send;
710 	/*
711 	 * If our state indicates that FIN should be sent
712 	 * and we have not yet done so, then we need to send.
713 	 */
714 	if (flags & TH_FIN &&
715 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
716 		goto send;
717 	/*
718 	 * In SACK, it is possible for tcp_output to fail to send a segment
719 	 * after the retransmission timer has been turned off.  Make sure
720 	 * that the retransmission timer is set.
721 	 */
722 	if ((tp->t_flags & TF_SACK_PERMIT) &&
723 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
724 	    !tcp_timer_active(tp, TT_REXMT) &&
725 	    !tcp_timer_active(tp, TT_PERSIST)) {
726 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
727 		goto just_return;
728 	}
729 	/*
730 	 * TCP window updates are not reliable, rather a polling protocol
731 	 * using ``persist'' packets is used to insure receipt of window
732 	 * updates.  The three ``states'' for the output side are:
733 	 *	idle			not doing retransmits or persists
734 	 *	persisting		to move a small or zero window
735 	 *	(re)transmitting	and thereby not persisting
736 	 *
737 	 * tcp_timer_active(tp, TT_PERSIST)
738 	 *	is true when we are in persist state.
739 	 * (tp->t_flags & TF_FORCEDATA)
740 	 *	is set when we are called to send a persist packet.
741 	 * tcp_timer_active(tp, TT_REXMT)
742 	 *	is set when we are retransmitting
743 	 * The output side is idle when both timers are zero.
744 	 *
745 	 * If send window is too small, there is data to transmit, and no
746 	 * retransmit or persist is pending, then go to persist state.
747 	 * If nothing happens soon, send when timer expires:
748 	 * if window is nonzero, transmit what we can,
749 	 * otherwise force out a byte.
750 	 */
751 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
752 	    !tcp_timer_active(tp, TT_PERSIST)) {
753 		tp->t_rxtshift = 0;
754 		tcp_setpersist(tp);
755 	}
756 
757 	/*
758 	 * No reason to send a segment, just return.
759 	 */
760 just_return:
761 	SOCKBUF_UNLOCK(&so->so_snd);
762 	return (0);
763 
764 send:
765 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
766 	if (len > 0) {
767 		if (len >= tp->t_maxseg)
768 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
769 		else
770 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
771 	}
772 	/*
773 	 * Before ESTABLISHED, force sending of initial options
774 	 * unless TCP set not to do any options.
775 	 * NOTE: we assume that the IP/TCP header plus TCP options
776 	 * always fit in a single mbuf, leaving room for a maximum
777 	 * link header, i.e.
778 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
779 	 */
780 	optlen = 0;
781 #ifdef INET6
782 	if (isipv6)
783 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
784 	else
785 #endif
786 		hdrlen = sizeof (struct tcpiphdr);
787 
788 	/*
789 	 * Compute options for segment.
790 	 * We only have to care about SYN and established connection
791 	 * segments.  Options for SYN-ACK segments are handled in TCP
792 	 * syncache.
793 	 */
794 	to.to_flags = 0;
795 	if ((tp->t_flags & TF_NOOPT) == 0) {
796 		/* Maximum segment size. */
797 		if (flags & TH_SYN) {
798 			tp->snd_nxt = tp->iss;
799 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
800 			to.to_flags |= TOF_MSS;
801 
802 			/*
803 			 * On SYN or SYN|ACK transmits on TFO connections,
804 			 * only include the TFO option if it is not a
805 			 * retransmit, as the presence of the TFO option may
806 			 * have caused the original SYN or SYN|ACK to have
807 			 * been dropped by a middlebox.
808 			 */
809 			if (IS_FASTOPEN(tp->t_flags) &&
810 			    (tp->t_rxtshift == 0)) {
811 				if (tp->t_state == TCPS_SYN_RECEIVED) {
812 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
813 					to.to_tfo_cookie =
814 					    (u_int8_t *)&tp->t_tfo_cookie.server;
815 					to.to_flags |= TOF_FASTOPEN;
816 					wanted_cookie = 1;
817 				} else if (tp->t_state == TCPS_SYN_SENT) {
818 					to.to_tfo_len =
819 					    tp->t_tfo_client_cookie_len;
820 					to.to_tfo_cookie =
821 					    tp->t_tfo_cookie.client;
822 					to.to_flags |= TOF_FASTOPEN;
823 					wanted_cookie = 1;
824 					/*
825 					 * If we wind up having more data to
826 					 * send with the SYN than can fit in
827 					 * one segment, don't send any more
828 					 * until the SYN|ACK comes back from
829 					 * the other end.
830 					 */
831 					dont_sendalot = 1;
832 				}
833 			}
834 		}
835 		/* Window scaling. */
836 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
837 			to.to_wscale = tp->request_r_scale;
838 			to.to_flags |= TOF_SCALE;
839 		}
840 		/* Timestamps. */
841 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
842 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
843 			curticks = tcp_ts_getticks();
844 			to.to_tsval = curticks + tp->ts_offset;
845 			to.to_tsecr = tp->ts_recent;
846 			to.to_flags |= TOF_TS;
847 			if (tp->t_rxtshift == 1)
848 				tp->t_badrxtwin = curticks;
849 		}
850 
851 		/* Set receive buffer autosizing timestamp. */
852 		if (tp->rfbuf_ts == 0 &&
853 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
854 			tp->rfbuf_ts = tcp_ts_getticks();
855 
856 		/* Selective ACK's. */
857 		if (tp->t_flags & TF_SACK_PERMIT) {
858 			if (flags & TH_SYN)
859 				to.to_flags |= TOF_SACKPERM;
860 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
861 			    (tp->t_flags & TF_SACK_PERMIT) &&
862 			    tp->rcv_numsacks > 0) {
863 				to.to_flags |= TOF_SACK;
864 				to.to_nsacks = tp->rcv_numsacks;
865 				to.to_sacks = (u_char *)tp->sackblks;
866 			}
867 		}
868 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
869 		/* TCP-MD5 (RFC2385). */
870 		/*
871 		 * Check that TCP_MD5SIG is enabled in tcpcb to
872 		 * account the size needed to set this TCP option.
873 		 */
874 		if (tp->t_flags & TF_SIGNATURE)
875 			to.to_flags |= TOF_SIGNATURE;
876 #endif /* TCP_SIGNATURE */
877 
878 		/* Processing the options. */
879 		hdrlen += optlen = tcp_addoptions(&to, opt);
880 		/*
881 		 * If we wanted a TFO option to be added, but it was unable
882 		 * to fit, ensure no data is sent.
883 		 */
884 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
885 		    !(to.to_flags & TOF_FASTOPEN))
886 			len = 0;
887 	}
888 
889 	/*
890 	 * Adjust data length if insertion of options will
891 	 * bump the packet length beyond the t_maxseg length.
892 	 * Clear the FIN bit because we cut off the tail of
893 	 * the segment.
894 	 */
895 	if (len + optlen + ipoptlen > tp->t_maxseg) {
896 		flags &= ~TH_FIN;
897 
898 		if (tso) {
899 			u_int if_hw_tsomax;
900 			u_int moff;
901 			int max_len;
902 
903 			/* extract TSO information */
904 			if_hw_tsomax = tp->t_tsomax;
905 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
906 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
907 
908 			/*
909 			 * Limit a TSO burst to prevent it from
910 			 * overflowing or exceeding the maximum length
911 			 * allowed by the network interface:
912 			 */
913 			KASSERT(ipoptlen == 0,
914 			    ("%s: TSO can't do IP options", __func__));
915 
916 			/*
917 			 * Check if we should limit by maximum payload
918 			 * length:
919 			 */
920 			if (if_hw_tsomax != 0) {
921 				/* compute maximum TSO length */
922 				max_len = (if_hw_tsomax - hdrlen -
923 				    max_linkhdr);
924 				if (max_len <= 0) {
925 					len = 0;
926 				} else if (len > max_len) {
927 					sendalot = 1;
928 					len = max_len;
929 				}
930 			}
931 
932 			/*
933 			 * Prevent the last segment from being
934 			 * fractional unless the send sockbuf can be
935 			 * emptied:
936 			 */
937 			max_len = (tp->t_maxseg - optlen);
938 			if (((uint32_t)off + (uint32_t)len) <
939 			    sbavail(&so->so_snd)) {
940 				moff = len % max_len;
941 				if (moff != 0) {
942 					len -= moff;
943 					sendalot = 1;
944 				}
945 			}
946 
947 			/*
948 			 * In case there are too many small fragments
949 			 * don't use TSO:
950 			 */
951 			if (len <= max_len) {
952 				len = max_len;
953 				sendalot = 1;
954 				tso = 0;
955 			}
956 
957 			/*
958 			 * Send the FIN in a separate segment
959 			 * after the bulk sending is done.
960 			 * We don't trust the TSO implementations
961 			 * to clear the FIN flag on all but the
962 			 * last segment.
963 			 */
964 			if (tp->t_flags & TF_NEEDFIN)
965 				sendalot = 1;
966 		} else {
967 			if (optlen + ipoptlen >= tp->t_maxseg) {
968 				/*
969 				 * Since we don't have enough space to put
970 				 * the IP header chain and the TCP header in
971 				 * one packet as required by RFC 7112, don't
972 				 * send it. Also ensure that at least one
973 				 * byte of the payload can be put into the
974 				 * TCP segment.
975 				 */
976 				SOCKBUF_UNLOCK(&so->so_snd);
977 				error = EMSGSIZE;
978 				sack_rxmit = 0;
979 				goto out;
980 			}
981 			len = tp->t_maxseg - optlen - ipoptlen;
982 			sendalot = 1;
983 			if (dont_sendalot)
984 				sendalot = 0;
985 		}
986 	} else
987 		tso = 0;
988 
989 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
990 	    ("%s: len > IP_MAXPACKET", __func__));
991 
992 /*#ifdef DIAGNOSTIC*/
993 #ifdef INET6
994 	if (max_linkhdr + hdrlen > MCLBYTES)
995 #else
996 	if (max_linkhdr + hdrlen > MHLEN)
997 #endif
998 		panic("tcphdr too big");
999 /*#endif*/
1000 
1001 	/*
1002 	 * This KASSERT is here to catch edge cases at a well defined place.
1003 	 * Before, those had triggered (random) panic conditions further down.
1004 	 */
1005 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1006 
1007 	/*
1008 	 * Grab a header mbuf, attaching a copy of data to
1009 	 * be transmitted, and initialize the header from
1010 	 * the template for sends on this connection.
1011 	 */
1012 	if (len) {
1013 		struct mbuf *mb;
1014 		struct sockbuf *msb;
1015 		u_int moff;
1016 
1017 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1018 			TCPSTAT_INC(tcps_sndprobe);
1019 #ifdef STATS
1020 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1021 				stats_voi_update_abs_u32(tp->t_stats,
1022 				VOI_TCP_RETXPB, len);
1023 			else
1024 				stats_voi_update_abs_u64(tp->t_stats,
1025 				    VOI_TCP_TXPB, len);
1026 #endif /* STATS */
1027 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1028 			tp->t_sndrexmitpack++;
1029 			TCPSTAT_INC(tcps_sndrexmitpack);
1030 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1031 #ifdef STATS
1032 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1033 			    len);
1034 #endif /* STATS */
1035 		} else {
1036 			TCPSTAT_INC(tcps_sndpack);
1037 			TCPSTAT_ADD(tcps_sndbyte, len);
1038 #ifdef STATS
1039 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1040 			    len);
1041 #endif /* STATS */
1042 		}
1043 #ifdef INET6
1044 		if (MHLEN < hdrlen + max_linkhdr)
1045 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1046 		else
1047 #endif
1048 			m = m_gethdr(M_NOWAIT, MT_DATA);
1049 
1050 		if (m == NULL) {
1051 			SOCKBUF_UNLOCK(&so->so_snd);
1052 			error = ENOBUFS;
1053 			sack_rxmit = 0;
1054 			goto out;
1055 		}
1056 
1057 		m->m_data += max_linkhdr;
1058 		m->m_len = hdrlen;
1059 
1060 		/*
1061 		 * Start the m_copy functions from the closest mbuf
1062 		 * to the offset in the socket buffer chain.
1063 		 */
1064 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1065 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1066 			m_copydata(mb, moff, len,
1067 			    mtod(m, caddr_t) + hdrlen);
1068 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1069 				sbsndptr_adv(&so->so_snd, mb, len);
1070 			m->m_len += len;
1071 		} else {
1072 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1073 				msb = NULL;
1074 			else
1075 				msb = &so->so_snd;
1076 			m->m_next = tcp_m_copym(mb, moff,
1077 			    &len, if_hw_tsomaxsegcount,
1078 			    if_hw_tsomaxsegsize, msb, hw_tls);
1079 			if (len <= (tp->t_maxseg - optlen)) {
1080 				/*
1081 				 * Must have ran out of mbufs for the copy
1082 				 * shorten it to no longer need tso. Lets
1083 				 * not put on sendalot since we are low on
1084 				 * mbufs.
1085 				 */
1086 				tso = 0;
1087 			}
1088 			if (m->m_next == NULL) {
1089 				SOCKBUF_UNLOCK(&so->so_snd);
1090 				(void) m_free(m);
1091 				error = ENOBUFS;
1092 				sack_rxmit = 0;
1093 				goto out;
1094 			}
1095 		}
1096 
1097 		/*
1098 		 * If we're sending everything we've got, set PUSH.
1099 		 * (This will keep happy those implementations which only
1100 		 * give data to the user when a buffer fills or
1101 		 * a PUSH comes in.)
1102 		 */
1103 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1104 		    !(flags & TH_SYN))
1105 			flags |= TH_PUSH;
1106 		SOCKBUF_UNLOCK(&so->so_snd);
1107 	} else {
1108 		SOCKBUF_UNLOCK(&so->so_snd);
1109 		if (tp->t_flags & TF_ACKNOW)
1110 			TCPSTAT_INC(tcps_sndacks);
1111 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1112 			TCPSTAT_INC(tcps_sndctrl);
1113 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1114 			TCPSTAT_INC(tcps_sndurg);
1115 		else
1116 			TCPSTAT_INC(tcps_sndwinup);
1117 
1118 		m = m_gethdr(M_NOWAIT, MT_DATA);
1119 		if (m == NULL) {
1120 			error = ENOBUFS;
1121 			sack_rxmit = 0;
1122 			goto out;
1123 		}
1124 #ifdef INET6
1125 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1126 		    MHLEN >= hdrlen) {
1127 			M_ALIGN(m, hdrlen);
1128 		} else
1129 #endif
1130 		m->m_data += max_linkhdr;
1131 		m->m_len = hdrlen;
1132 	}
1133 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1134 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1135 #ifdef MAC
1136 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1137 #endif
1138 #ifdef INET6
1139 	if (isipv6) {
1140 		ip6 = mtod(m, struct ip6_hdr *);
1141 		th = (struct tcphdr *)(ip6 + 1);
1142 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1143 	} else
1144 #endif /* INET6 */
1145 	{
1146 		ip = mtod(m, struct ip *);
1147 #ifdef TCPDEBUG
1148 		ipov = (struct ipovly *)ip;
1149 #endif
1150 		th = (struct tcphdr *)(ip + 1);
1151 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1152 	}
1153 
1154 	/*
1155 	 * Fill in fields, remembering maximum advertised
1156 	 * window for use in delaying messages about window sizes.
1157 	 * If resending a FIN, be sure not to use a new sequence number.
1158 	 */
1159 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1160 	    tp->snd_nxt == tp->snd_max)
1161 		tp->snd_nxt--;
1162 	/*
1163 	 * If we are starting a connection, send ECN setup
1164 	 * SYN packet. If we are on a retransmit, we may
1165 	 * resend those bits a number of times as per
1166 	 * RFC 3168.
1167 	 */
1168 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1169 		if (tp->t_rxtshift >= 1) {
1170 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1171 				flags |= TH_ECE|TH_CWR;
1172 		} else
1173 			flags |= TH_ECE|TH_CWR;
1174 	}
1175 	/* Handle parallel SYN for ECN */
1176 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1177 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1178 			flags |= TH_ECE;
1179 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1180 	}
1181 
1182 	if (tp->t_state == TCPS_ESTABLISHED &&
1183 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
1184 		/*
1185 		 * If the peer has ECN, mark data packets with
1186 		 * ECN capable transmission (ECT).
1187 		 * Ignore pure ack packets, retransmissions and window probes.
1188 		 */
1189 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1190 		    (sack_rxmit == 0) &&
1191 		    !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1192 		    SEQ_LT(tp->snd_una, tp->snd_max))) {
1193 #ifdef INET6
1194 			if (isipv6)
1195 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1196 			else
1197 #endif
1198 				ip->ip_tos |= IPTOS_ECN_ECT0;
1199 			TCPSTAT_INC(tcps_ecn_ect0);
1200 			/*
1201 			 * Reply with proper ECN notifications.
1202 			 * Only set CWR on new data segments.
1203 			 */
1204 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1205 				flags |= TH_CWR;
1206 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1207 			}
1208 		}
1209 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
1210 			flags |= TH_ECE;
1211 	}
1212 
1213 	/*
1214 	 * If we are doing retransmissions, then snd_nxt will
1215 	 * not reflect the first unsent octet.  For ACK only
1216 	 * packets, we do not want the sequence number of the
1217 	 * retransmitted packet, we want the sequence number
1218 	 * of the next unsent octet.  So, if there is no data
1219 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1220 	 * when filling in ti_seq.  But if we are in persist
1221 	 * state, snd_max might reflect one byte beyond the
1222 	 * right edge of the window, so use snd_nxt in that
1223 	 * case, since we know we aren't doing a retransmission.
1224 	 * (retransmit and persist are mutually exclusive...)
1225 	 */
1226 	if (sack_rxmit == 0) {
1227 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1228 		    tcp_timer_active(tp, TT_PERSIST))
1229 			th->th_seq = htonl(tp->snd_nxt);
1230 		else
1231 			th->th_seq = htonl(tp->snd_max);
1232 	} else {
1233 		th->th_seq = htonl(p->rxmit);
1234 		p->rxmit += len;
1235 		tp->sackhint.sack_bytes_rexmit += len;
1236 	}
1237 	th->th_ack = htonl(tp->rcv_nxt);
1238 	if (optlen) {
1239 		bcopy(opt, th + 1, optlen);
1240 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1241 	}
1242 	th->th_flags = flags;
1243 	/*
1244 	 * Calculate receive window.  Don't shrink window,
1245 	 * but avoid silly window syndrome.
1246 	 * If a RST segment is sent, advertise a window of zero.
1247 	 */
1248 	if (flags & TH_RST) {
1249 		recwin = 0;
1250 	} else {
1251 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1252 		    recwin < tp->t_maxseg)
1253 			recwin = 0;
1254 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1255 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1256 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1257 	}
1258 	/*
1259 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1260 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1261 	 * case is handled in syncache.
1262 	 */
1263 	if (flags & TH_SYN)
1264 		th->th_win = htons((u_short)
1265 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1266 	else {
1267 		/* Avoid shrinking window with window scaling. */
1268 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1269 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1270 	}
1271 
1272 	/*
1273 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1274 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1275 	 * flag tells soreceive() to disable delayed acknowledgements when
1276 	 * draining the buffer.  This can occur if the receiver is attempting
1277 	 * to read more data than can be buffered prior to transmitting on
1278 	 * the connection.
1279 	 */
1280 	if (th->th_win == 0) {
1281 		tp->t_sndzerowin++;
1282 		tp->t_flags |= TF_RXWIN0SENT;
1283 	} else
1284 		tp->t_flags &= ~TF_RXWIN0SENT;
1285 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1286 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1287 		th->th_flags |= TH_URG;
1288 	} else
1289 		/*
1290 		 * If no urgent pointer to send, then we pull
1291 		 * the urgent pointer to the left edge of the send window
1292 		 * so that it doesn't drift into the send window on sequence
1293 		 * number wraparound.
1294 		 */
1295 		tp->snd_up = tp->snd_una;		/* drag it along */
1296 
1297 	/*
1298 	 * Put TCP length in extended header, and then
1299 	 * checksum extended header and data.
1300 	 */
1301 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1302 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1303 
1304 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1305 	if (to.to_flags & TOF_SIGNATURE) {
1306 		/*
1307 		 * Calculate MD5 signature and put it into the place
1308 		 * determined before.
1309 		 * NOTE: since TCP options buffer doesn't point into
1310 		 * mbuf's data, calculate offset and use it.
1311 		 */
1312 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1313 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1314 			/*
1315 			 * Do not send segment if the calculation of MD5
1316 			 * digest has failed.
1317 			 */
1318 			m_freem(m);
1319 			goto out;
1320 		}
1321 	}
1322 #endif
1323 #ifdef INET6
1324 	if (isipv6) {
1325 		/*
1326 		 * There is no need to fill in ip6_plen right now.
1327 		 * It will be filled later by ip6_output.
1328 		 */
1329 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1330 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1331 		    optlen + len, IPPROTO_TCP, 0);
1332 	}
1333 #endif
1334 #if defined(INET6) && defined(INET)
1335 	else
1336 #endif
1337 #ifdef INET
1338 	{
1339 		m->m_pkthdr.csum_flags = CSUM_TCP;
1340 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1341 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1342 
1343 		/* IP version must be set here for ipv4/ipv6 checking later */
1344 		KASSERT(ip->ip_v == IPVERSION,
1345 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1346 	}
1347 #endif
1348 
1349 	/*
1350 	 * Enable TSO and specify the size of the segments.
1351 	 * The TCP pseudo header checksum is always provided.
1352 	 */
1353 	if (tso) {
1354 		KASSERT(len > tp->t_maxseg - optlen,
1355 		    ("%s: len <= tso_segsz", __func__));
1356 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1357 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1358 	}
1359 
1360 	KASSERT(len + hdrlen == m_length(m, NULL),
1361 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1362 	    __func__, len, hdrlen, m_length(m, NULL)));
1363 
1364 #ifdef TCP_HHOOK
1365 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1366 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1367 #endif
1368 
1369 #ifdef TCPDEBUG
1370 	/*
1371 	 * Trace.
1372 	 */
1373 	if (so->so_options & SO_DEBUG) {
1374 		u_short save = 0;
1375 #ifdef INET6
1376 		if (!isipv6)
1377 #endif
1378 		{
1379 			save = ipov->ih_len;
1380 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1381 		}
1382 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1383 #ifdef INET6
1384 		if (!isipv6)
1385 #endif
1386 		ipov->ih_len = save;
1387 	}
1388 #endif /* TCPDEBUG */
1389 	TCP_PROBE3(debug__output, tp, th, m);
1390 
1391 	/* We're getting ready to send; log now. */
1392 	TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1393 	    len, NULL, false);
1394 
1395 	/*
1396 	 * Fill in IP length and desired time to live and
1397 	 * send to IP level.  There should be a better way
1398 	 * to handle ttl and tos; we could keep them in
1399 	 * the template, but need a way to checksum without them.
1400 	 */
1401 	/*
1402 	 * m->m_pkthdr.len should have been set before checksum calculation,
1403 	 * because in6_cksum() need it.
1404 	 */
1405 #ifdef INET6
1406 	if (isipv6) {
1407 		/*
1408 		 * we separately set hoplimit for every segment, since the
1409 		 * user might want to change the value via setsockopt.
1410 		 * Also, desired default hop limit might be changed via
1411 		 * Neighbor Discovery.
1412 		 */
1413 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1414 
1415 		/*
1416 		 * Set the packet size here for the benefit of DTrace probes.
1417 		 * ip6_output() will set it properly; it's supposed to include
1418 		 * the option header lengths as well.
1419 		 */
1420 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1421 
1422 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1423 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1424 		else
1425 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1426 
1427 		if (tp->t_state == TCPS_SYN_SENT)
1428 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1429 
1430 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1431 
1432 #ifdef TCPPCAP
1433 		/* Save packet, if requested. */
1434 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1435 #endif
1436 
1437 		/* TODO: IPv6 IP6TOS_ECT bit on */
1438 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1439 		    &tp->t_inpcb->inp_route6,
1440 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1441 		    NULL, NULL, tp->t_inpcb);
1442 
1443 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1444 			mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1445 	}
1446 #endif /* INET6 */
1447 #if defined(INET) && defined(INET6)
1448 	else
1449 #endif
1450 #ifdef INET
1451     {
1452 	ip->ip_len = htons(m->m_pkthdr.len);
1453 #ifdef INET6
1454 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1455 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1456 #endif /* INET6 */
1457 	/*
1458 	 * If we do path MTU discovery, then we set DF on every packet.
1459 	 * This might not be the best thing to do according to RFC3390
1460 	 * Section 2. However the tcp hostcache migitates the problem
1461 	 * so it affects only the first tcp connection with a host.
1462 	 *
1463 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1464 	 */
1465 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1466 		ip->ip_off |= htons(IP_DF);
1467 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1468 	} else {
1469 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1470 	}
1471 
1472 	if (tp->t_state == TCPS_SYN_SENT)
1473 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1474 
1475 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1476 
1477 #ifdef TCPPCAP
1478 	/* Save packet, if requested. */
1479 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1480 #endif
1481 
1482 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1483 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1484 	    tp->t_inpcb);
1485 
1486 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1487 		mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1488     }
1489 #endif /* INET */
1490 
1491 out:
1492 	/*
1493 	 * In transmit state, time the transmission and arrange for
1494 	 * the retransmit.  In persist state, just set snd_max.
1495 	 */
1496 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1497 	    !tcp_timer_active(tp, TT_PERSIST)) {
1498 		tcp_seq startseq = tp->snd_nxt;
1499 
1500 		/*
1501 		 * Advance snd_nxt over sequence space of this segment.
1502 		 */
1503 		if (flags & (TH_SYN|TH_FIN)) {
1504 			if (flags & TH_SYN)
1505 				tp->snd_nxt++;
1506 			if (flags & TH_FIN) {
1507 				tp->snd_nxt++;
1508 				tp->t_flags |= TF_SENTFIN;
1509 			}
1510 		}
1511 		if (sack_rxmit)
1512 			goto timer;
1513 		tp->snd_nxt += len;
1514 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1515 			tp->snd_max = tp->snd_nxt;
1516 			/*
1517 			 * Time this transmission if not a retransmission and
1518 			 * not currently timing anything.
1519 			 */
1520 			tp->t_sndtime = ticks;
1521 			if (tp->t_rtttime == 0) {
1522 				tp->t_rtttime = ticks;
1523 				tp->t_rtseq = startseq;
1524 				TCPSTAT_INC(tcps_segstimed);
1525 			}
1526 #ifdef STATS
1527 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1528 				tp->t_flags |= TF_GPUTINPROG;
1529 				tp->gput_seq = startseq;
1530 				tp->gput_ack = startseq +
1531 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1532 				tp->gput_ts = tcp_ts_getticks();
1533 			}
1534 #endif /* STATS */
1535 		}
1536 
1537 		/*
1538 		 * Set retransmit timer if not currently set,
1539 		 * and not doing a pure ack or a keep-alive probe.
1540 		 * Initial value for retransmit timer is smoothed
1541 		 * round-trip time + 2 * round-trip time variance.
1542 		 * Initialize shift counter which is used for backoff
1543 		 * of retransmit time.
1544 		 */
1545 timer:
1546 		if (!tcp_timer_active(tp, TT_REXMT) &&
1547 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1548 		     (tp->snd_nxt != tp->snd_una))) {
1549 			if (tcp_timer_active(tp, TT_PERSIST)) {
1550 				tcp_timer_activate(tp, TT_PERSIST, 0);
1551 				tp->t_rxtshift = 0;
1552 			}
1553 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1554 		} else if (len == 0 && sbavail(&so->so_snd) &&
1555 		    !tcp_timer_active(tp, TT_REXMT) &&
1556 		    !tcp_timer_active(tp, TT_PERSIST)) {
1557 			/*
1558 			 * Avoid a situation where we do not set persist timer
1559 			 * after a zero window condition. For example:
1560 			 * 1) A -> B: packet with enough data to fill the window
1561 			 * 2) B -> A: ACK for #1 + new data (0 window
1562 			 *    advertisement)
1563 			 * 3) A -> B: ACK for #2, 0 len packet
1564 			 *
1565 			 * In this case, A will not activate the persist timer,
1566 			 * because it chose to send a packet. Unless tcp_output
1567 			 * is called for some other reason (delayed ack timer,
1568 			 * another input packet from B, socket syscall), A will
1569 			 * not send zero window probes.
1570 			 *
1571 			 * So, if you send a 0-length packet, but there is data
1572 			 * in the socket buffer, and neither the rexmt or
1573 			 * persist timer is already set, then activate the
1574 			 * persist timer.
1575 			 */
1576 			tp->t_rxtshift = 0;
1577 			tcp_setpersist(tp);
1578 		}
1579 	} else {
1580 		/*
1581 		 * Persist case, update snd_max but since we are in
1582 		 * persist mode (no window) we do not update snd_nxt.
1583 		 */
1584 		int xlen = len;
1585 		if (flags & TH_SYN)
1586 			++xlen;
1587 		if (flags & TH_FIN) {
1588 			++xlen;
1589 			tp->t_flags |= TF_SENTFIN;
1590 		}
1591 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1592 			tp->snd_max = tp->snd_nxt + xlen;
1593 	}
1594 	if ((error == 0) &&
1595 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1596 	     (tp->t_flags & TF_SACK_PERMIT) &&
1597 	     tp->rcv_numsacks > 0)) {
1598 		    /* Clean up any DSACK's sent */
1599 		    tcp_clean_dsack_blocks(tp);
1600 	}
1601 	if (error) {
1602 		/* Record the error. */
1603 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1604 		    error, 0, NULL, false);
1605 
1606 		/*
1607 		 * We know that the packet was lost, so back out the
1608 		 * sequence number advance, if any.
1609 		 *
1610 		 * If the error is EPERM the packet got blocked by the
1611 		 * local firewall.  Normally we should terminate the
1612 		 * connection but the blocking may have been spurious
1613 		 * due to a firewall reconfiguration cycle.  So we treat
1614 		 * it like a packet loss and let the retransmit timer and
1615 		 * timeouts do their work over time.
1616 		 * XXX: It is a POLA question whether calling tcp_drop right
1617 		 * away would be the really correct behavior instead.
1618 		 */
1619 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1620 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1621 		    ((flags & TH_SYN) == 0) &&
1622 		    (error != EPERM)) {
1623 			if (sack_rxmit) {
1624 				p->rxmit -= len;
1625 				tp->sackhint.sack_bytes_rexmit -= len;
1626 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1627 				    ("sackhint bytes rtx >= 0"));
1628 			} else
1629 				tp->snd_nxt -= len;
1630 		}
1631 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1632 		switch (error) {
1633 		case EACCES:
1634 		case EPERM:
1635 			tp->t_softerror = error;
1636 			return (error);
1637 		case ENOBUFS:
1638 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1639 			tp->snd_cwnd = tp->t_maxseg;
1640 			return (0);
1641 		case EMSGSIZE:
1642 			/*
1643 			 * For some reason the interface we used initially
1644 			 * to send segments changed to another or lowered
1645 			 * its MTU.
1646 			 * If TSO was active we either got an interface
1647 			 * without TSO capabilits or TSO was turned off.
1648 			 * If we obtained mtu from ip_output() then update
1649 			 * it and try again.
1650 			 */
1651 			if (tso)
1652 				tp->t_flags &= ~TF_TSO;
1653 			if (mtu != 0) {
1654 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1655 				goto again;
1656 			}
1657 			return (error);
1658 		case EHOSTDOWN:
1659 		case EHOSTUNREACH:
1660 		case ENETDOWN:
1661 		case ENETUNREACH:
1662 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1663 				tp->t_softerror = error;
1664 				return (0);
1665 			}
1666 			/* FALLTHROUGH */
1667 		default:
1668 			return (error);
1669 		}
1670 	}
1671 	TCPSTAT_INC(tcps_sndtotal);
1672 
1673 	/*
1674 	 * Data sent (as far as we can tell).
1675 	 * If this advertises a larger window than any other segment,
1676 	 * then remember the size of the advertised window.
1677 	 * Any pending ACK has now been sent.
1678 	 */
1679 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1680 		tp->rcv_adv = tp->rcv_nxt + recwin;
1681 	tp->last_ack_sent = tp->rcv_nxt;
1682 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1683 	if (tcp_timer_active(tp, TT_DELACK))
1684 		tcp_timer_activate(tp, TT_DELACK, 0);
1685 #if 0
1686 	/*
1687 	 * This completely breaks TCP if newreno is turned on.  What happens
1688 	 * is that if delayed-acks are turned on on the receiver, this code
1689 	 * on the transmitter effectively destroys the TCP window, forcing
1690 	 * it to four packets (1.5Kx4 = 6K window).
1691 	 */
1692 	if (sendalot && --maxburst)
1693 		goto again;
1694 #endif
1695 	if (sendalot)
1696 		goto again;
1697 	return (0);
1698 }
1699 
1700 void
1701 tcp_setpersist(struct tcpcb *tp)
1702 {
1703 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1704 	int tt;
1705 
1706 	tp->t_flags &= ~TF_PREVVALID;
1707 	if (tcp_timer_active(tp, TT_REXMT))
1708 		panic("tcp_setpersist: retransmit pending");
1709 	/*
1710 	 * Start/restart persistence timer.
1711 	 */
1712 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1713 		      tcp_persmin, tcp_persmax);
1714 	tcp_timer_activate(tp, TT_PERSIST, tt);
1715 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1716 		tp->t_rxtshift++;
1717 }
1718 
1719 /*
1720  * Insert TCP options according to the supplied parameters to the place
1721  * optp in a consistent way.  Can handle unaligned destinations.
1722  *
1723  * The order of the option processing is crucial for optimal packing and
1724  * alignment for the scarce option space.
1725  *
1726  * The optimal order for a SYN/SYN-ACK segment is:
1727  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1728  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1729  *
1730  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1731  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1732  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1733  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1734  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1735  */
1736 int
1737 tcp_addoptions(struct tcpopt *to, u_char *optp)
1738 {
1739 	u_int32_t mask, optlen = 0;
1740 
1741 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1742 		if ((to->to_flags & mask) != mask)
1743 			continue;
1744 		if (optlen == TCP_MAXOLEN)
1745 			break;
1746 		switch (to->to_flags & mask) {
1747 		case TOF_MSS:
1748 			while (optlen % 4) {
1749 				optlen += TCPOLEN_NOP;
1750 				*optp++ = TCPOPT_NOP;
1751 			}
1752 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1753 				continue;
1754 			optlen += TCPOLEN_MAXSEG;
1755 			*optp++ = TCPOPT_MAXSEG;
1756 			*optp++ = TCPOLEN_MAXSEG;
1757 			to->to_mss = htons(to->to_mss);
1758 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1759 			optp += sizeof(to->to_mss);
1760 			break;
1761 		case TOF_SCALE:
1762 			while (!optlen || optlen % 2 != 1) {
1763 				optlen += TCPOLEN_NOP;
1764 				*optp++ = TCPOPT_NOP;
1765 			}
1766 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1767 				continue;
1768 			optlen += TCPOLEN_WINDOW;
1769 			*optp++ = TCPOPT_WINDOW;
1770 			*optp++ = TCPOLEN_WINDOW;
1771 			*optp++ = to->to_wscale;
1772 			break;
1773 		case TOF_SACKPERM:
1774 			while (optlen % 2) {
1775 				optlen += TCPOLEN_NOP;
1776 				*optp++ = TCPOPT_NOP;
1777 			}
1778 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1779 				continue;
1780 			optlen += TCPOLEN_SACK_PERMITTED;
1781 			*optp++ = TCPOPT_SACK_PERMITTED;
1782 			*optp++ = TCPOLEN_SACK_PERMITTED;
1783 			break;
1784 		case TOF_TS:
1785 			while (!optlen || optlen % 4 != 2) {
1786 				optlen += TCPOLEN_NOP;
1787 				*optp++ = TCPOPT_NOP;
1788 			}
1789 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1790 				continue;
1791 			optlen += TCPOLEN_TIMESTAMP;
1792 			*optp++ = TCPOPT_TIMESTAMP;
1793 			*optp++ = TCPOLEN_TIMESTAMP;
1794 			to->to_tsval = htonl(to->to_tsval);
1795 			to->to_tsecr = htonl(to->to_tsecr);
1796 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1797 			optp += sizeof(to->to_tsval);
1798 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1799 			optp += sizeof(to->to_tsecr);
1800 			break;
1801 		case TOF_SIGNATURE:
1802 			{
1803 			int siglen = TCPOLEN_SIGNATURE - 2;
1804 
1805 			while (!optlen || optlen % 4 != 2) {
1806 				optlen += TCPOLEN_NOP;
1807 				*optp++ = TCPOPT_NOP;
1808 			}
1809 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1810 				to->to_flags &= ~TOF_SIGNATURE;
1811 				continue;
1812 			}
1813 			optlen += TCPOLEN_SIGNATURE;
1814 			*optp++ = TCPOPT_SIGNATURE;
1815 			*optp++ = TCPOLEN_SIGNATURE;
1816 			to->to_signature = optp;
1817 			while (siglen--)
1818 				 *optp++ = 0;
1819 			break;
1820 			}
1821 		case TOF_SACK:
1822 			{
1823 			int sackblks = 0;
1824 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1825 			tcp_seq sack_seq;
1826 
1827 			while (!optlen || optlen % 4 != 2) {
1828 				optlen += TCPOLEN_NOP;
1829 				*optp++ = TCPOPT_NOP;
1830 			}
1831 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1832 				continue;
1833 			optlen += TCPOLEN_SACKHDR;
1834 			*optp++ = TCPOPT_SACK;
1835 			sackblks = min(to->to_nsacks,
1836 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1837 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1838 			while (sackblks--) {
1839 				sack_seq = htonl(sack->start);
1840 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1841 				optp += sizeof(sack_seq);
1842 				sack_seq = htonl(sack->end);
1843 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1844 				optp += sizeof(sack_seq);
1845 				optlen += TCPOLEN_SACK;
1846 				sack++;
1847 			}
1848 			TCPSTAT_INC(tcps_sack_send_blocks);
1849 			break;
1850 			}
1851 		case TOF_FASTOPEN:
1852 			{
1853 			int total_len;
1854 
1855 			/* XXX is there any point to aligning this option? */
1856 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1857 			if (TCP_MAXOLEN - optlen < total_len) {
1858 				to->to_flags &= ~TOF_FASTOPEN;
1859 				continue;
1860 			}
1861 			*optp++ = TCPOPT_FAST_OPEN;
1862 			*optp++ = total_len;
1863 			if (to->to_tfo_len > 0) {
1864 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1865 				optp += to->to_tfo_len;
1866 			}
1867 			optlen += total_len;
1868 			break;
1869 			}
1870 		default:
1871 			panic("%s: unknown TCP option type", __func__);
1872 			break;
1873 		}
1874 	}
1875 
1876 	/* Terminate and pad TCP options to a 4 byte boundary. */
1877 	if (optlen % 4) {
1878 		optlen += TCPOLEN_EOL;
1879 		*optp++ = TCPOPT_EOL;
1880 	}
1881 	/*
1882 	 * According to RFC 793 (STD0007):
1883 	 *   "The content of the header beyond the End-of-Option option
1884 	 *    must be header padding (i.e., zero)."
1885 	 *   and later: "The padding is composed of zeros."
1886 	 */
1887 	while (optlen % 4) {
1888 		optlen += TCPOLEN_PAD;
1889 		*optp++ = TCPOPT_PAD;
1890 	}
1891 
1892 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1893 	return (optlen);
1894 }
1895 
1896 /*
1897  * This is a copy of m_copym(), taking the TSO segment size/limit
1898  * constraints into account, and advancing the sndptr as it goes.
1899  */
1900 struct mbuf *
1901 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1902     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1903 {
1904 #ifdef KERN_TLS
1905 	struct ktls_session *tls, *ntls;
1906 	struct mbuf *start;
1907 #endif
1908 	struct mbuf *n, **np;
1909 	struct mbuf *top;
1910 	int32_t off = off0;
1911 	int32_t len = *plen;
1912 	int32_t fragsize;
1913 	int32_t len_cp = 0;
1914 	int32_t *pkthdrlen;
1915 	uint32_t mlen, frags;
1916 	bool copyhdr;
1917 
1918 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1919 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1920 	if (off == 0 && m->m_flags & M_PKTHDR)
1921 		copyhdr = true;
1922 	else
1923 		copyhdr = false;
1924 	while (off > 0) {
1925 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1926 		if (off < m->m_len)
1927 			break;
1928 		off -= m->m_len;
1929 		if ((sb) && (m == sb->sb_sndptr)) {
1930 			sb->sb_sndptroff += m->m_len;
1931 			sb->sb_sndptr = m->m_next;
1932 		}
1933 		m = m->m_next;
1934 	}
1935 	np = &top;
1936 	top = NULL;
1937 	pkthdrlen = NULL;
1938 #ifdef KERN_TLS
1939 	if (hw_tls && (m->m_flags & M_EXTPG))
1940 		tls = m->m_epg_tls;
1941 	else
1942 		tls = NULL;
1943 	start = m;
1944 #endif
1945 	while (len > 0) {
1946 		if (m == NULL) {
1947 			KASSERT(len == M_COPYALL,
1948 			    ("tcp_m_copym, length > size of mbuf chain"));
1949 			*plen = len_cp;
1950 			if (pkthdrlen != NULL)
1951 				*pkthdrlen = len_cp;
1952 			break;
1953 		}
1954 #ifdef KERN_TLS
1955 		if (hw_tls) {
1956 			if (m->m_flags & M_EXTPG)
1957 				ntls = m->m_epg_tls;
1958 			else
1959 				ntls = NULL;
1960 
1961 			/*
1962 			 * Avoid mixing TLS records with handshake
1963 			 * data or TLS records from different
1964 			 * sessions.
1965 			 */
1966 			if (tls != ntls) {
1967 				MPASS(m != start);
1968 				*plen = len_cp;
1969 				if (pkthdrlen != NULL)
1970 					*pkthdrlen = len_cp;
1971 				break;
1972 			}
1973 		}
1974 #endif
1975 		mlen = min(len, m->m_len - off);
1976 		if (seglimit) {
1977 			/*
1978 			 * For M_EXTPG mbufs, add 3 segments
1979 			 * + 1 in case we are crossing page boundaries
1980 			 * + 2 in case the TLS hdr/trailer are used
1981 			 * It is cheaper to just add the segments
1982 			 * than it is to take the cache miss to look
1983 			 * at the mbuf ext_pgs state in detail.
1984 			 */
1985 			if (m->m_flags & M_EXTPG) {
1986 				fragsize = min(segsize, PAGE_SIZE);
1987 				frags = 3;
1988 			} else {
1989 				fragsize = segsize;
1990 				frags = 0;
1991 			}
1992 
1993 			/* Break if we really can't fit anymore. */
1994 			if ((frags + 1) >= seglimit) {
1995 				*plen =	len_cp;
1996 				if (pkthdrlen != NULL)
1997 					*pkthdrlen = len_cp;
1998 				break;
1999 			}
2000 
2001 			/*
2002 			 * Reduce size if you can't copy the whole
2003 			 * mbuf. If we can't copy the whole mbuf, also
2004 			 * adjust len so the loop will end after this
2005 			 * mbuf.
2006 			 */
2007 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2008 				mlen = (seglimit - frags - 1) * fragsize;
2009 				len = mlen;
2010 				*plen = len_cp + len;
2011 				if (pkthdrlen != NULL)
2012 					*pkthdrlen = *plen;
2013 			}
2014 			frags += howmany(mlen, fragsize);
2015 			if (frags == 0)
2016 				frags++;
2017 			seglimit -= frags;
2018 			KASSERT(seglimit > 0,
2019 			    ("%s: seglimit went too low", __func__));
2020 		}
2021 		if (copyhdr)
2022 			n = m_gethdr(M_NOWAIT, m->m_type);
2023 		else
2024 			n = m_get(M_NOWAIT, m->m_type);
2025 		*np = n;
2026 		if (n == NULL)
2027 			goto nospace;
2028 		if (copyhdr) {
2029 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2030 				goto nospace;
2031 			if (len == M_COPYALL)
2032 				n->m_pkthdr.len -= off0;
2033 			else
2034 				n->m_pkthdr.len = len;
2035 			pkthdrlen = &n->m_pkthdr.len;
2036 			copyhdr = false;
2037 		}
2038 		n->m_len = mlen;
2039 		len_cp += n->m_len;
2040 		if (m->m_flags & (M_EXT|M_EXTPG)) {
2041 			n->m_data = m->m_data + off;
2042 			mb_dupcl(n, m);
2043 		} else
2044 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2045 			    (u_int)n->m_len);
2046 
2047 		if (sb && (sb->sb_sndptr == m) &&
2048 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2049 			sb->sb_sndptroff += m->m_len;
2050 			sb->sb_sndptr = m->m_next;
2051 		}
2052 		off = 0;
2053 		if (len != M_COPYALL) {
2054 			len -= n->m_len;
2055 		}
2056 		m = m->m_next;
2057 		np = &n->m_next;
2058 	}
2059 	return (top);
2060 nospace:
2061 	m_freem(top);
2062 	return (NULL);
2063 }
2064 
2065 void
2066 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2067 {
2068 
2069 	/*
2070 	 * Automatic sizing of send socket buffer.  Often the send buffer
2071 	 * size is not optimally adjusted to the actual network conditions
2072 	 * at hand (delay bandwidth product).  Setting the buffer size too
2073 	 * small limits throughput on links with high bandwidth and high
2074 	 * delay (eg. trans-continental/oceanic links).  Setting the
2075 	 * buffer size too big consumes too much real kernel memory,
2076 	 * especially with many connections on busy servers.
2077 	 *
2078 	 * The criteria to step up the send buffer one notch are:
2079 	 *  1. receive window of remote host is larger than send buffer
2080 	 *     (with a fudge factor of 5/4th);
2081 	 *  2. send buffer is filled to 7/8th with data (so we actually
2082 	 *     have data to make use of it);
2083 	 *  3. send buffer fill has not hit maximal automatic size;
2084 	 *  4. our send window (slow start and cogestion controlled) is
2085 	 *     larger than sent but unacknowledged data in send buffer.
2086 	 *
2087 	 * The remote host receive window scaling factor may limit the
2088 	 * growing of the send buffer before it reaches its allowed
2089 	 * maximum.
2090 	 *
2091 	 * It scales directly with slow start or congestion window
2092 	 * and does at most one step per received ACK.  This fast
2093 	 * scaling has the drawback of growing the send buffer beyond
2094 	 * what is strictly necessary to make full use of a given
2095 	 * delay*bandwidth product.  However testing has shown this not
2096 	 * to be much of an problem.  At worst we are trading wasting
2097 	 * of available bandwidth (the non-use of it) for wasting some
2098 	 * socket buffer memory.
2099 	 *
2100 	 * TODO: Shrink send buffer during idle periods together
2101 	 * with congestion window.  Requires another timer.  Has to
2102 	 * wait for upcoming tcp timer rewrite.
2103 	 *
2104 	 * XXXGL: should there be used sbused() or sbavail()?
2105 	 */
2106 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2107 		int lowat;
2108 
2109 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2110 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2111 		    sbused(&so->so_snd) >=
2112 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2113 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2114 		    sendwin >= (sbused(&so->so_snd) -
2115 		    (tp->snd_nxt - tp->snd_una))) {
2116 			if (!sbreserve_locked(&so->so_snd,
2117 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2118 			     V_tcp_autosndbuf_max), so, curthread))
2119 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2120 		}
2121 	}
2122 }
2123