xref: /freebsd/sys/netinet/tcp_output.c (revision 0629b152762b06325dd75a41bcb0a2789514141b)
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_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
45 #ifdef TCP_HHOOK
46 #include <sys/hhook.h>
47 #endif
48 #include <sys/kernel.h>
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/sdt.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
57 
58 #include <net/if.h>
59 #include <net/route.h>
60 #include <net/vnet.h>
61 
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
69 #ifdef INET6
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
73 #endif
74 #include <netinet/tcp.h>
75 #define	TCPOUTFLAGS
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/tcp_log_buf.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcpip.h>
82 #include <netinet/cc/cc.h>
83 #include <netinet/tcp_fastopen.h>
84 #ifdef TCPPCAP
85 #include <netinet/tcp_pcap.h>
86 #endif
87 #ifdef TCPDEBUG
88 #include <netinet/tcp_debug.h>
89 #endif
90 #ifdef TCP_OFFLOAD
91 #include <netinet/tcp_offload.h>
92 #endif
93 
94 #include <netipsec/ipsec_support.h>
95 
96 #include <machine/in_cksum.h>
97 
98 #include <security/mac/mac_framework.h>
99 
100 VNET_DEFINE(int, path_mtu_discovery) = 1;
101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
102 	&VNET_NAME(path_mtu_discovery), 1,
103 	"Enable Path MTU Discovery");
104 
105 VNET_DEFINE(int, tcp_do_tso) = 1;
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
107 	&VNET_NAME(tcp_do_tso), 0,
108 	"Enable TCP Segmentation Offload");
109 
110 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
111 #define	V_tcp_sendspace	VNET(tcp_sendspace)
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
113 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114 
115 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
117 	&VNET_NAME(tcp_do_autosndbuf), 0,
118 	"Enable automatic send buffer sizing");
119 
120 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
122 	&VNET_NAME(tcp_autosndbuf_inc), 0,
123 	"Incrementor step size of automatic send buffer");
124 
125 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
127 	&VNET_NAME(tcp_autosndbuf_max), 0,
128 	"Max size of automatic send buffer");
129 
130 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
131 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
133 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
134 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
135 
136 /*
137  * Make sure that either retransmit or persist timer is set for SYN, FIN and
138  * non-ACK.
139  */
140 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
141 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
142 	    tcp_timer_active((tp), TT_REXMT) ||				\
143 	    tcp_timer_active((tp), TT_PERSIST),				\
144 	    ("neither rexmt nor persist timer is set"))
145 
146 #ifdef TCP_HHOOK
147 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
148 			    struct tcphdr *th, struct tcpopt *to,
149 			    uint32_t len, int tso);
150 #endif
151 static void inline	cc_after_idle(struct tcpcb *tp);
152 
153 #ifdef TCP_HHOOK
154 /*
155  * Wrapper for the TCP established output helper hook.
156  */
157 static void inline
158 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
159     struct tcpopt *to, uint32_t len, int tso)
160 {
161 	struct tcp_hhook_data hhook_data;
162 
163 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
164 		hhook_data.tp = tp;
165 		hhook_data.th = th;
166 		hhook_data.to = to;
167 		hhook_data.len = len;
168 		hhook_data.tso = tso;
169 
170 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
171 		    tp->osd);
172 	}
173 }
174 #endif
175 
176 /*
177  * CC wrapper hook functions
178  */
179 static void inline
180 cc_after_idle(struct tcpcb *tp)
181 {
182 	INP_WLOCK_ASSERT(tp->t_inpcb);
183 
184 	if (CC_ALGO(tp)->after_idle != NULL)
185 		CC_ALGO(tp)->after_idle(tp->ccv);
186 }
187 
188 /*
189  * Tcp output routine: figure out what should be sent and send it.
190  */
191 int
192 tcp_output(struct tcpcb *tp)
193 {
194 	struct socket *so = tp->t_inpcb->inp_socket;
195 	int32_t len;
196 	uint32_t recwin, sendwin;
197 	int off, flags, error = 0;	/* Keep compiler happy */
198 	struct mbuf *m;
199 	struct ip *ip = NULL;
200 #ifdef TCPDEBUG
201 	struct ipovly *ipov = NULL;
202 #endif
203 	struct tcphdr *th;
204 	u_char opt[TCP_MAXOLEN];
205 	unsigned ipoptlen, optlen, hdrlen;
206 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
207 	unsigned ipsec_optlen = 0;
208 #endif
209 	int idle, sendalot;
210 	int sack_rxmit, sack_bytes_rxmt;
211 	struct sackhole *p;
212 	int tso, mtu;
213 	struct tcpopt to;
214 	unsigned int wanted_cookie = 0;
215 	unsigned int dont_sendalot = 0;
216 #if 0
217 	int maxburst = TCP_MAXBURST;
218 #endif
219 #ifdef INET6
220 	struct ip6_hdr *ip6 = NULL;
221 	int isipv6;
222 
223 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
224 #endif
225 
226 	INP_WLOCK_ASSERT(tp->t_inpcb);
227 
228 #ifdef TCP_OFFLOAD
229 	if (tp->t_flags & TF_TOE)
230 		return (tcp_offload_output(tp));
231 #endif
232 
233 	/*
234 	 * For TFO connections in SYN_RECEIVED, only allow the initial
235 	 * SYN|ACK and those sent by the retransmit timer.
236 	 */
237 	if (IS_FASTOPEN(tp->t_flags) &&
238 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
239 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
240 	    (tp->snd_nxt != tp->snd_una))         /* not a retransmit */
241 		return (0);
242 
243 	/*
244 	 * Determine length of data that should be transmitted,
245 	 * and flags that will be used.
246 	 * If there is some data or critical controls (SYN, RST)
247 	 * to send, then transmit; otherwise, investigate further.
248 	 */
249 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
250 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
251 		cc_after_idle(tp);
252 	tp->t_flags &= ~TF_LASTIDLE;
253 	if (idle) {
254 		if (tp->t_flags & TF_MORETOCOME) {
255 			tp->t_flags |= TF_LASTIDLE;
256 			idle = 0;
257 		}
258 	}
259 again:
260 	/*
261 	 * If we've recently taken a timeout, snd_max will be greater than
262 	 * snd_nxt.  There may be SACK information that allows us to avoid
263 	 * resending already delivered data.  Adjust snd_nxt accordingly.
264 	 */
265 	if ((tp->t_flags & TF_SACK_PERMIT) &&
266 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
267 		tcp_sack_adjust(tp);
268 	sendalot = 0;
269 	tso = 0;
270 	mtu = 0;
271 	off = tp->snd_nxt - tp->snd_una;
272 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
273 
274 	flags = tcp_outflags[tp->t_state];
275 	/*
276 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
277 	 * to send out new data (when sendalot is 1), bypass this function.
278 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
279 	 * we're replacing a (future) new transmission with a retransmission
280 	 * now, and we previously incremented snd_cwnd in tcp_input().
281 	 */
282 	/*
283 	 * Still in sack recovery , reset rxmit flag to zero.
284 	 */
285 	sack_rxmit = 0;
286 	sack_bytes_rxmt = 0;
287 	len = 0;
288 	p = NULL;
289 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
290 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
291 		uint32_t cwin;
292 
293 		cwin =
294 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
295 		/* Do not retransmit SACK segments beyond snd_recover */
296 		if (SEQ_GT(p->end, tp->snd_recover)) {
297 			/*
298 			 * (At least) part of sack hole extends beyond
299 			 * snd_recover. Check to see if we can rexmit data
300 			 * for this hole.
301 			 */
302 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
303 				/*
304 				 * Can't rexmit any more data for this hole.
305 				 * That data will be rexmitted in the next
306 				 * sack recovery episode, when snd_recover
307 				 * moves past p->rxmit.
308 				 */
309 				p = NULL;
310 				goto after_sack_rexmit;
311 			} else
312 				/* Can rexmit part of the current hole */
313 				len = ((int32_t)ulmin(cwin,
314 						   tp->snd_recover - p->rxmit));
315 		} else
316 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
317 		off = p->rxmit - tp->snd_una;
318 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
319 		    __func__, off));
320 		if (len > 0) {
321 			sack_rxmit = 1;
322 			sendalot = 1;
323 			TCPSTAT_INC(tcps_sack_rexmits);
324 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
325 			    min(len, tp->t_maxseg));
326 		}
327 	}
328 after_sack_rexmit:
329 	/*
330 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
331 	 * state flags.
332 	 */
333 	if (tp->t_flags & TF_NEEDFIN)
334 		flags |= TH_FIN;
335 	if (tp->t_flags & TF_NEEDSYN)
336 		flags |= TH_SYN;
337 
338 	SOCKBUF_LOCK(&so->so_snd);
339 	/*
340 	 * If in persist timeout with window of 0, send 1 byte.
341 	 * Otherwise, if window is small but nonzero
342 	 * and timer expired, we will send what we can
343 	 * and go to transmit state.
344 	 */
345 	if (tp->t_flags & TF_FORCEDATA) {
346 		if (sendwin == 0) {
347 			/*
348 			 * If we still have some data to send, then
349 			 * clear the FIN bit.  Usually this would
350 			 * happen below when it realizes that we
351 			 * aren't sending all the data.  However,
352 			 * if we have exactly 1 byte of unsent data,
353 			 * then it won't clear the FIN bit below,
354 			 * and if we are in persist state, we wind
355 			 * up sending the packet without recording
356 			 * that we sent the FIN bit.
357 			 *
358 			 * We can't just blindly clear the FIN bit,
359 			 * because if we don't have any more data
360 			 * to send then the probe will be the FIN
361 			 * itself.
362 			 */
363 			if (off < sbused(&so->so_snd))
364 				flags &= ~TH_FIN;
365 			sendwin = 1;
366 		} else {
367 			tcp_timer_activate(tp, TT_PERSIST, 0);
368 			tp->t_rxtshift = 0;
369 		}
370 	}
371 
372 	/*
373 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
374 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
375 	 * a negative length.  This can also occur when TCP opens up
376 	 * its congestion window while receiving additional duplicate
377 	 * acks after fast-retransmit because TCP will reset snd_nxt
378 	 * to snd_max after the fast-retransmit.
379 	 *
380 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
381 	 * be set to snd_una, the offset will be 0, and the length may
382 	 * wind up 0.
383 	 *
384 	 * If sack_rxmit is true we are retransmitting from the scoreboard
385 	 * in which case len is already set.
386 	 */
387 	if (sack_rxmit == 0) {
388 		if (sack_bytes_rxmt == 0)
389 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
390 			    off);
391 		else {
392 			int32_t cwin;
393 
394                         /*
395 			 * We are inside of a SACK recovery episode and are
396 			 * sending new data, having retransmitted all the
397 			 * data possible in the scoreboard.
398 			 */
399 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
400 			    off);
401 			/*
402 			 * Don't remove this (len > 0) check !
403 			 * We explicitly check for len > 0 here (although it
404 			 * isn't really necessary), to work around a gcc
405 			 * optimization issue - to force gcc to compute
406 			 * len above. Without this check, the computation
407 			 * of len is bungled by the optimizer.
408 			 */
409 			if (len > 0) {
410 				cwin = tp->snd_cwnd -
411 					(tp->snd_nxt - tp->sack_newdata) -
412 					sack_bytes_rxmt;
413 				if (cwin < 0)
414 					cwin = 0;
415 				len = imin(len, cwin);
416 			}
417 		}
418 	}
419 
420 	/*
421 	 * Lop off SYN bit if it has already been sent.  However, if this
422 	 * is SYN-SENT state and if segment contains data and if we don't
423 	 * know that foreign host supports TAO, suppress sending segment.
424 	 */
425 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
426 		if (tp->t_state != TCPS_SYN_RECEIVED)
427 			flags &= ~TH_SYN;
428 		/*
429 		 * When sending additional segments following a TFO SYN|ACK,
430 		 * do not include the SYN bit.
431 		 */
432 		if (IS_FASTOPEN(tp->t_flags) &&
433 		    (tp->t_state == TCPS_SYN_RECEIVED))
434 			flags &= ~TH_SYN;
435 		off--, len++;
436 	}
437 
438 	/*
439 	 * Be careful not to send data and/or FIN on SYN segments.
440 	 * This measure is needed to prevent interoperability problems
441 	 * with not fully conformant TCP implementations.
442 	 */
443 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
444 		len = 0;
445 		flags &= ~TH_FIN;
446 	}
447 
448 	/*
449 	 * On TFO sockets, ensure no data is sent in the following cases:
450 	 *
451 	 *  - When retransmitting SYN|ACK on a passively-created socket
452 	 *
453 	 *  - When retransmitting SYN on an actively created socket
454 	 *
455 	 *  - When sending a zero-length cookie (cookie request) on an
456 	 *    actively created socket
457 	 *
458 	 *  - When the socket is in the CLOSED state (RST is being sent)
459 	 */
460 	if (IS_FASTOPEN(tp->t_flags) &&
461 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
462 	     ((tp->t_state == TCPS_SYN_SENT) &&
463 	      (tp->t_tfo_client_cookie_len == 0)) ||
464 	     (flags & TH_RST)))
465 		len = 0;
466 	if (len <= 0) {
467 		/*
468 		 * If FIN has been sent but not acked,
469 		 * but we haven't been called to retransmit,
470 		 * len will be < 0.  Otherwise, window shrank
471 		 * after we sent into it.  If window shrank to 0,
472 		 * cancel pending retransmit, pull snd_nxt back
473 		 * to (closed) window, and set the persist timer
474 		 * if it isn't already going.  If the window didn't
475 		 * close completely, just wait for an ACK.
476 		 *
477 		 * We also do a general check here to ensure that
478 		 * we will set the persist timer when we have data
479 		 * to send, but a 0-byte window. This makes sure
480 		 * the persist timer is set even if the packet
481 		 * hits one of the "goto send" lines below.
482 		 */
483 		len = 0;
484 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
485 			(off < (int) sbavail(&so->so_snd))) {
486 			tcp_timer_activate(tp, TT_REXMT, 0);
487 			tp->t_rxtshift = 0;
488 			tp->snd_nxt = tp->snd_una;
489 			if (!tcp_timer_active(tp, TT_PERSIST))
490 				tcp_setpersist(tp);
491 		}
492 	}
493 
494 	/* len will be >= 0 after this point. */
495 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
496 
497 	tcp_sndbuf_autoscale(tp, so, sendwin);
498 
499 	/*
500 	 * Decide if we can use TCP Segmentation Offloading (if supported by
501 	 * hardware).
502 	 *
503 	 * TSO may only be used if we are in a pure bulk sending state.  The
504 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
505 	 * IP options prevent using TSO.  With TSO the TCP header is the same
506 	 * (except for the sequence number) for all generated packets.  This
507 	 * makes it impossible to transmit any options which vary per generated
508 	 * segment or packet.
509 	 *
510 	 * IPv4 handling has a clear separation of ip options and ip header
511 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
512 	 * the right thing below to provide length of just ip options and thus
513 	 * checking for ipoptlen is enough to decide if ip options are present.
514 	 */
515 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
516 	/*
517 	 * Pre-calculate here as we save another lookup into the darknesses
518 	 * of IPsec that way and can actually decide if TSO is ok.
519 	 */
520 #ifdef INET6
521 	if (isipv6 && IPSEC_ENABLED(ipv6))
522 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
523 #ifdef INET
524 	else
525 #endif
526 #endif /* INET6 */
527 #ifdef INET
528 	if (IPSEC_ENABLED(ipv4))
529 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
530 #endif /* INET */
531 #endif /* IPSEC */
532 #ifdef INET6
533 	if (isipv6)
534 		ipoptlen = ip6_optlen(tp->t_inpcb);
535 	else
536 #endif
537 	if (tp->t_inpcb->inp_options)
538 		ipoptlen = tp->t_inpcb->inp_options->m_len -
539 				offsetof(struct ipoption, ipopt_list);
540 	else
541 		ipoptlen = 0;
542 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
543 	ipoptlen += ipsec_optlen;
544 #endif
545 
546 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
547 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
548 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
549 	    ipoptlen == 0 && !(flags & TH_SYN))
550 		tso = 1;
551 
552 	if (sack_rxmit) {
553 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
554 			flags &= ~TH_FIN;
555 	} else {
556 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
557 		    sbused(&so->so_snd)))
558 			flags &= ~TH_FIN;
559 	}
560 
561 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
562 	    (long)TCP_MAXWIN << tp->rcv_scale);
563 
564 	/*
565 	 * Sender silly window avoidance.   We transmit under the following
566 	 * conditions when len is non-zero:
567 	 *
568 	 *	- We have a full segment (or more with TSO)
569 	 *	- This is the last buffer in a write()/send() and we are
570 	 *	  either idle or running NODELAY
571 	 *	- we've timed out (e.g. persist timer)
572 	 *	- we have more then 1/2 the maximum send window's worth of
573 	 *	  data (receiver may be limited the window size)
574 	 *	- we need to retransmit
575 	 */
576 	if (len) {
577 		if (len >= tp->t_maxseg)
578 			goto send;
579 		/*
580 		 * NOTE! on localhost connections an 'ack' from the remote
581 		 * end may occur synchronously with the output and cause
582 		 * us to flush a buffer queued with moretocome.  XXX
583 		 *
584 		 * note: the len + off check is almost certainly unnecessary.
585 		 */
586 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
587 		    (idle || (tp->t_flags & TF_NODELAY)) &&
588 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
589 		    (tp->t_flags & TF_NOPUSH) == 0) {
590 			goto send;
591 		}
592 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
593 			goto send;
594 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
595 			goto send;
596 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
597 			goto send;
598 		if (sack_rxmit)
599 			goto send;
600 	}
601 
602 	/*
603 	 * Sending of standalone window updates.
604 	 *
605 	 * Window updates are important when we close our window due to a
606 	 * full socket buffer and are opening it again after the application
607 	 * reads data from it.  Once the window has opened again and the
608 	 * remote end starts to send again the ACK clock takes over and
609 	 * provides the most current window information.
610 	 *
611 	 * We must avoid the silly window syndrome whereas every read
612 	 * from the receive buffer, no matter how small, causes a window
613 	 * update to be sent.  We also should avoid sending a flurry of
614 	 * window updates when the socket buffer had queued a lot of data
615 	 * and the application is doing small reads.
616 	 *
617 	 * Prevent a flurry of pointless window updates by only sending
618 	 * an update when we can increase the advertized window by more
619 	 * than 1/4th of the socket buffer capacity.  When the buffer is
620 	 * getting full or is very small be more aggressive and send an
621 	 * update whenever we can increase by two mss sized segments.
622 	 * In all other situations the ACK's to new incoming data will
623 	 * carry further window increases.
624 	 *
625 	 * Don't send an independent window update if a delayed
626 	 * ACK is pending (it will get piggy-backed on it) or the
627 	 * remote side already has done a half-close and won't send
628 	 * more data.  Skip this if the connection is in T/TCP
629 	 * half-open state.
630 	 */
631 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
632 	    !(tp->t_flags & TF_DELACK) &&
633 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
634 		/*
635 		 * "adv" is the amount we could increase the window,
636 		 * taking into account that we are limited by
637 		 * TCP_MAXWIN << tp->rcv_scale.
638 		 */
639 		int32_t adv;
640 		int oldwin;
641 
642 		adv = recwin;
643 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
644 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
645 			adv -= oldwin;
646 		} else
647 			oldwin = 0;
648 
649 		/*
650 		 * If the new window size ends up being the same as or less
651 		 * than the old size when it is scaled, then don't force
652 		 * a window update.
653 		 */
654 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
655 			goto dontupdate;
656 
657 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
658 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
659 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
660 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
661 			goto send;
662 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
663 			goto send;
664 	}
665 dontupdate:
666 
667 	/*
668 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
669 	 * is also a catch-all for the retransmit timer timeout case.
670 	 */
671 	if (tp->t_flags & TF_ACKNOW)
672 		goto send;
673 	if ((flags & TH_RST) ||
674 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
675 		goto send;
676 	if (SEQ_GT(tp->snd_up, tp->snd_una))
677 		goto send;
678 	/*
679 	 * If our state indicates that FIN should be sent
680 	 * and we have not yet done so, then we need to send.
681 	 */
682 	if (flags & TH_FIN &&
683 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
684 		goto send;
685 	/*
686 	 * In SACK, it is possible for tcp_output to fail to send a segment
687 	 * after the retransmission timer has been turned off.  Make sure
688 	 * that the retransmission timer is set.
689 	 */
690 	if ((tp->t_flags & TF_SACK_PERMIT) &&
691 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
692 	    !tcp_timer_active(tp, TT_REXMT) &&
693 	    !tcp_timer_active(tp, TT_PERSIST)) {
694 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
695 		goto just_return;
696 	}
697 	/*
698 	 * TCP window updates are not reliable, rather a polling protocol
699 	 * using ``persist'' packets is used to insure receipt of window
700 	 * updates.  The three ``states'' for the output side are:
701 	 *	idle			not doing retransmits or persists
702 	 *	persisting		to move a small or zero window
703 	 *	(re)transmitting	and thereby not persisting
704 	 *
705 	 * tcp_timer_active(tp, TT_PERSIST)
706 	 *	is true when we are in persist state.
707 	 * (tp->t_flags & TF_FORCEDATA)
708 	 *	is set when we are called to send a persist packet.
709 	 * tcp_timer_active(tp, TT_REXMT)
710 	 *	is set when we are retransmitting
711 	 * The output side is idle when both timers are zero.
712 	 *
713 	 * If send window is too small, there is data to transmit, and no
714 	 * retransmit or persist is pending, then go to persist state.
715 	 * If nothing happens soon, send when timer expires:
716 	 * if window is nonzero, transmit what we can,
717 	 * otherwise force out a byte.
718 	 */
719 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
720 	    !tcp_timer_active(tp, TT_PERSIST)) {
721 		tp->t_rxtshift = 0;
722 		tcp_setpersist(tp);
723 	}
724 
725 	/*
726 	 * No reason to send a segment, just return.
727 	 */
728 just_return:
729 	SOCKBUF_UNLOCK(&so->so_snd);
730 	return (0);
731 
732 send:
733 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
734 	if (len > 0) {
735 		if (len >= tp->t_maxseg)
736 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
737 		else
738 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
739 	}
740 	/*
741 	 * Before ESTABLISHED, force sending of initial options
742 	 * unless TCP set not to do any options.
743 	 * NOTE: we assume that the IP/TCP header plus TCP options
744 	 * always fit in a single mbuf, leaving room for a maximum
745 	 * link header, i.e.
746 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
747 	 */
748 	optlen = 0;
749 #ifdef INET6
750 	if (isipv6)
751 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
752 	else
753 #endif
754 		hdrlen = sizeof (struct tcpiphdr);
755 
756 	/*
757 	 * Compute options for segment.
758 	 * We only have to care about SYN and established connection
759 	 * segments.  Options for SYN-ACK segments are handled in TCP
760 	 * syncache.
761 	 */
762 	to.to_flags = 0;
763 	if ((tp->t_flags & TF_NOOPT) == 0) {
764 		/* Maximum segment size. */
765 		if (flags & TH_SYN) {
766 			tp->snd_nxt = tp->iss;
767 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
768 			to.to_flags |= TOF_MSS;
769 
770 			/*
771 			 * On SYN or SYN|ACK transmits on TFO connections,
772 			 * only include the TFO option if it is not a
773 			 * retransmit, as the presence of the TFO option may
774 			 * have caused the original SYN or SYN|ACK to have
775 			 * been dropped by a middlebox.
776 			 */
777 			if (IS_FASTOPEN(tp->t_flags) &&
778 			    (tp->t_rxtshift == 0)) {
779 				if (tp->t_state == TCPS_SYN_RECEIVED) {
780 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
781 					to.to_tfo_cookie =
782 					    (u_int8_t *)&tp->t_tfo_cookie.server;
783 					to.to_flags |= TOF_FASTOPEN;
784 					wanted_cookie = 1;
785 				} else if (tp->t_state == TCPS_SYN_SENT) {
786 					to.to_tfo_len =
787 					    tp->t_tfo_client_cookie_len;
788 					to.to_tfo_cookie =
789 					    tp->t_tfo_cookie.client;
790 					to.to_flags |= TOF_FASTOPEN;
791 					wanted_cookie = 1;
792 					/*
793 					 * If we wind up having more data to
794 					 * send with the SYN than can fit in
795 					 * one segment, don't send any more
796 					 * until the SYN|ACK comes back from
797 					 * the other end.
798 					 */
799 					dont_sendalot = 1;
800 				}
801 			}
802 		}
803 		/* Window scaling. */
804 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
805 			to.to_wscale = tp->request_r_scale;
806 			to.to_flags |= TOF_SCALE;
807 		}
808 		/* Timestamps. */
809 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
810 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
811 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
812 			to.to_tsecr = tp->ts_recent;
813 			to.to_flags |= TOF_TS;
814 		}
815 
816 		/* Set receive buffer autosizing timestamp. */
817 		if (tp->rfbuf_ts == 0 &&
818 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
819 			tp->rfbuf_ts = tcp_ts_getticks();
820 
821 		/* Selective ACK's. */
822 		if (tp->t_flags & TF_SACK_PERMIT) {
823 			if (flags & TH_SYN)
824 				to.to_flags |= TOF_SACKPERM;
825 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
826 			    (tp->t_flags & TF_SACK_PERMIT) &&
827 			    tp->rcv_numsacks > 0) {
828 				to.to_flags |= TOF_SACK;
829 				to.to_nsacks = tp->rcv_numsacks;
830 				to.to_sacks = (u_char *)tp->sackblks;
831 			}
832 		}
833 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
834 		/* TCP-MD5 (RFC2385). */
835 		/*
836 		 * Check that TCP_MD5SIG is enabled in tcpcb to
837 		 * account the size needed to set this TCP option.
838 		 */
839 		if (tp->t_flags & TF_SIGNATURE)
840 			to.to_flags |= TOF_SIGNATURE;
841 #endif /* TCP_SIGNATURE */
842 
843 		/* Processing the options. */
844 		hdrlen += optlen = tcp_addoptions(&to, opt);
845 		/*
846 		 * If we wanted a TFO option to be added, but it was unable
847 		 * to fit, ensure no data is sent.
848 		 */
849 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
850 		    !(to.to_flags & TOF_FASTOPEN))
851 			len = 0;
852 	}
853 
854 	/*
855 	 * Adjust data length if insertion of options will
856 	 * bump the packet length beyond the t_maxseg length.
857 	 * Clear the FIN bit because we cut off the tail of
858 	 * the segment.
859 	 */
860 	if (len + optlen + ipoptlen > tp->t_maxseg) {
861 		flags &= ~TH_FIN;
862 
863 		if (tso) {
864 			u_int if_hw_tsomax;
865 			u_int if_hw_tsomaxsegcount;
866 			u_int if_hw_tsomaxsegsize;
867 			struct mbuf *mb;
868 			u_int moff;
869 			int max_len;
870 
871 			/* extract TSO information */
872 			if_hw_tsomax = tp->t_tsomax;
873 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
874 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
875 
876 			/*
877 			 * Limit a TSO burst to prevent it from
878 			 * overflowing or exceeding the maximum length
879 			 * allowed by the network interface:
880 			 */
881 			KASSERT(ipoptlen == 0,
882 			    ("%s: TSO can't do IP options", __func__));
883 
884 			/*
885 			 * Check if we should limit by maximum payload
886 			 * length:
887 			 */
888 			if (if_hw_tsomax != 0) {
889 				/* compute maximum TSO length */
890 				max_len = (if_hw_tsomax - hdrlen -
891 				    max_linkhdr);
892 				if (max_len <= 0) {
893 					len = 0;
894 				} else if (len > max_len) {
895 					sendalot = 1;
896 					len = max_len;
897 				}
898 			}
899 
900 			/*
901 			 * Check if we should limit by maximum segment
902 			 * size and count:
903 			 */
904 			if (if_hw_tsomaxsegcount != 0 &&
905 			    if_hw_tsomaxsegsize != 0) {
906 				/*
907 				 * Subtract one segment for the LINK
908 				 * and TCP/IP headers mbuf that will
909 				 * be prepended to this mbuf chain
910 				 * after the code in this section
911 				 * limits the number of mbufs in the
912 				 * chain to if_hw_tsomaxsegcount.
913 				 */
914 				if_hw_tsomaxsegcount -= 1;
915 				max_len = 0;
916 				mb = sbsndmbuf(&so->so_snd, off, &moff);
917 
918 				while (mb != NULL && max_len < len) {
919 					u_int mlen;
920 					u_int frags;
921 
922 					/*
923 					 * Get length of mbuf fragment
924 					 * and how many hardware frags,
925 					 * rounded up, it would use:
926 					 */
927 					mlen = (mb->m_len - moff);
928 					frags = howmany(mlen,
929 					    if_hw_tsomaxsegsize);
930 
931 					/* Handle special case: Zero Length Mbuf */
932 					if (frags == 0)
933 						frags = 1;
934 
935 					/*
936 					 * Check if the fragment limit
937 					 * will be reached or exceeded:
938 					 */
939 					if (frags >= if_hw_tsomaxsegcount) {
940 						max_len += min(mlen,
941 						    if_hw_tsomaxsegcount *
942 						    if_hw_tsomaxsegsize);
943 						break;
944 					}
945 					max_len += mlen;
946 					if_hw_tsomaxsegcount -= frags;
947 					moff = 0;
948 					mb = mb->m_next;
949 				}
950 				if (max_len <= 0) {
951 					len = 0;
952 				} else if (len > max_len) {
953 					sendalot = 1;
954 					len = max_len;
955 				}
956 			}
957 
958 			/*
959 			 * Prevent the last segment from being
960 			 * fractional unless the send sockbuf can be
961 			 * emptied:
962 			 */
963 			max_len = (tp->t_maxseg - optlen);
964 			if (((uint32_t)off + (uint32_t)len) <
965 			    sbavail(&so->so_snd)) {
966 				moff = len % max_len;
967 				if (moff != 0) {
968 					len -= moff;
969 					sendalot = 1;
970 				}
971 			}
972 
973 			/*
974 			 * In case there are too many small fragments
975 			 * don't use TSO:
976 			 */
977 			if (len <= max_len) {
978 				len = max_len;
979 				sendalot = 1;
980 				tso = 0;
981 			}
982 
983 			/*
984 			 * Send the FIN in a separate segment
985 			 * after the bulk sending is done.
986 			 * We don't trust the TSO implementations
987 			 * to clear the FIN flag on all but the
988 			 * last segment.
989 			 */
990 			if (tp->t_flags & TF_NEEDFIN)
991 				sendalot = 1;
992 
993 		} else {
994 			len = tp->t_maxseg - optlen - ipoptlen;
995 			sendalot = 1;
996 			if (dont_sendalot)
997 				sendalot = 0;
998 		}
999 	} else
1000 		tso = 0;
1001 
1002 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1003 	    ("%s: len > IP_MAXPACKET", __func__));
1004 
1005 /*#ifdef DIAGNOSTIC*/
1006 #ifdef INET6
1007 	if (max_linkhdr + hdrlen > MCLBYTES)
1008 #else
1009 	if (max_linkhdr + hdrlen > MHLEN)
1010 #endif
1011 		panic("tcphdr too big");
1012 /*#endif*/
1013 
1014 	/*
1015 	 * This KASSERT is here to catch edge cases at a well defined place.
1016 	 * Before, those had triggered (random) panic conditions further down.
1017 	 */
1018 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1019 
1020 	/*
1021 	 * Grab a header mbuf, attaching a copy of data to
1022 	 * be transmitted, and initialize the header from
1023 	 * the template for sends on this connection.
1024 	 */
1025 	if (len) {
1026 		struct mbuf *mb;
1027 		u_int moff;
1028 
1029 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1030 			TCPSTAT_INC(tcps_sndprobe);
1031 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1032 			tp->t_sndrexmitpack++;
1033 			TCPSTAT_INC(tcps_sndrexmitpack);
1034 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1035 		} else {
1036 			TCPSTAT_INC(tcps_sndpack);
1037 			TCPSTAT_ADD(tcps_sndbyte, len);
1038 		}
1039 #ifdef INET6
1040 		if (MHLEN < hdrlen + max_linkhdr)
1041 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1042 		else
1043 #endif
1044 			m = m_gethdr(M_NOWAIT, MT_DATA);
1045 
1046 		if (m == NULL) {
1047 			SOCKBUF_UNLOCK(&so->so_snd);
1048 			error = ENOBUFS;
1049 			sack_rxmit = 0;
1050 			goto out;
1051 		}
1052 
1053 		m->m_data += max_linkhdr;
1054 		m->m_len = hdrlen;
1055 
1056 		/*
1057 		 * Start the m_copy functions from the closest mbuf
1058 		 * to the offset in the socket buffer chain.
1059 		 */
1060 		mb = sbsndptr(&so->so_snd, off, len, &moff);
1061 
1062 		if (len <= MHLEN - hdrlen - max_linkhdr) {
1063 			m_copydata(mb, moff, len,
1064 			    mtod(m, caddr_t) + hdrlen);
1065 			m->m_len += len;
1066 		} else {
1067 			m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1068 			if (m->m_next == NULL) {
1069 				SOCKBUF_UNLOCK(&so->so_snd);
1070 				(void) m_free(m);
1071 				error = ENOBUFS;
1072 				sack_rxmit = 0;
1073 				goto out;
1074 			}
1075 		}
1076 
1077 		/*
1078 		 * If we're sending everything we've got, set PUSH.
1079 		 * (This will keep happy those implementations which only
1080 		 * give data to the user when a buffer fills or
1081 		 * a PUSH comes in.)
1082 		 */
1083 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1084 		    !(flags & TH_SYN))
1085 			flags |= TH_PUSH;
1086 		SOCKBUF_UNLOCK(&so->so_snd);
1087 	} else {
1088 		SOCKBUF_UNLOCK(&so->so_snd);
1089 		if (tp->t_flags & TF_ACKNOW)
1090 			TCPSTAT_INC(tcps_sndacks);
1091 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1092 			TCPSTAT_INC(tcps_sndctrl);
1093 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1094 			TCPSTAT_INC(tcps_sndurg);
1095 		else
1096 			TCPSTAT_INC(tcps_sndwinup);
1097 
1098 		m = m_gethdr(M_NOWAIT, MT_DATA);
1099 		if (m == NULL) {
1100 			error = ENOBUFS;
1101 			sack_rxmit = 0;
1102 			goto out;
1103 		}
1104 #ifdef INET6
1105 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1106 		    MHLEN >= hdrlen) {
1107 			M_ALIGN(m, hdrlen);
1108 		} else
1109 #endif
1110 		m->m_data += max_linkhdr;
1111 		m->m_len = hdrlen;
1112 	}
1113 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1114 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1115 #ifdef MAC
1116 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1117 #endif
1118 #ifdef INET6
1119 	if (isipv6) {
1120 		ip6 = mtod(m, struct ip6_hdr *);
1121 		th = (struct tcphdr *)(ip6 + 1);
1122 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1123 	} else
1124 #endif /* INET6 */
1125 	{
1126 		ip = mtod(m, struct ip *);
1127 #ifdef TCPDEBUG
1128 		ipov = (struct ipovly *)ip;
1129 #endif
1130 		th = (struct tcphdr *)(ip + 1);
1131 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1132 	}
1133 
1134 	/*
1135 	 * Fill in fields, remembering maximum advertised
1136 	 * window for use in delaying messages about window sizes.
1137 	 * If resending a FIN, be sure not to use a new sequence number.
1138 	 */
1139 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1140 	    tp->snd_nxt == tp->snd_max)
1141 		tp->snd_nxt--;
1142 	/*
1143 	 * If we are starting a connection, send ECN setup
1144 	 * SYN packet. If we are on a retransmit, we may
1145 	 * resend those bits a number of times as per
1146 	 * RFC 3168.
1147 	 */
1148 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1149 		if (tp->t_rxtshift >= 1) {
1150 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1151 				flags |= TH_ECE|TH_CWR;
1152 		} else
1153 			flags |= TH_ECE|TH_CWR;
1154 	}
1155 
1156 	if (tp->t_state == TCPS_ESTABLISHED &&
1157 	    (tp->t_flags & TF_ECN_PERMIT)) {
1158 		/*
1159 		 * If the peer has ECN, mark data packets with
1160 		 * ECN capable transmission (ECT).
1161 		 * Ignore pure ack packets, retransmissions and window probes.
1162 		 */
1163 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1164 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1165 #ifdef INET6
1166 			if (isipv6)
1167 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1168 			else
1169 #endif
1170 				ip->ip_tos |= IPTOS_ECN_ECT0;
1171 			TCPSTAT_INC(tcps_ecn_ect0);
1172 		}
1173 
1174 		/*
1175 		 * Reply with proper ECN notifications.
1176 		 */
1177 		if (tp->t_flags & TF_ECN_SND_CWR) {
1178 			flags |= TH_CWR;
1179 			tp->t_flags &= ~TF_ECN_SND_CWR;
1180 		}
1181 		if (tp->t_flags & TF_ECN_SND_ECE)
1182 			flags |= TH_ECE;
1183 	}
1184 
1185 	/*
1186 	 * If we are doing retransmissions, then snd_nxt will
1187 	 * not reflect the first unsent octet.  For ACK only
1188 	 * packets, we do not want the sequence number of the
1189 	 * retransmitted packet, we want the sequence number
1190 	 * of the next unsent octet.  So, if there is no data
1191 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1192 	 * when filling in ti_seq.  But if we are in persist
1193 	 * state, snd_max might reflect one byte beyond the
1194 	 * right edge of the window, so use snd_nxt in that
1195 	 * case, since we know we aren't doing a retransmission.
1196 	 * (retransmit and persist are mutually exclusive...)
1197 	 */
1198 	if (sack_rxmit == 0) {
1199 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1200 		    tcp_timer_active(tp, TT_PERSIST))
1201 			th->th_seq = htonl(tp->snd_nxt);
1202 		else
1203 			th->th_seq = htonl(tp->snd_max);
1204 	} else {
1205 		th->th_seq = htonl(p->rxmit);
1206 		p->rxmit += len;
1207 		tp->sackhint.sack_bytes_rexmit += len;
1208 	}
1209 	th->th_ack = htonl(tp->rcv_nxt);
1210 	if (optlen) {
1211 		bcopy(opt, th + 1, optlen);
1212 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1213 	}
1214 	th->th_flags = flags;
1215 	/*
1216 	 * Calculate receive window.  Don't shrink window,
1217 	 * but avoid silly window syndrome.
1218 	 */
1219 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1220 	    recwin < tp->t_maxseg)
1221 		recwin = 0;
1222 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1223 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
1224 		recwin = (tp->rcv_adv - tp->rcv_nxt);
1225 
1226 	/*
1227 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1228 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1229 	 * case is handled in syncache.
1230 	 */
1231 	if (flags & TH_SYN)
1232 		th->th_win = htons((u_short)
1233 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1234 	else
1235 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1236 
1237 	/*
1238 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1239 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1240 	 * flag tells soreceive() to disable delayed acknowledgements when
1241 	 * draining the buffer.  This can occur if the receiver is attempting
1242 	 * to read more data than can be buffered prior to transmitting on
1243 	 * the connection.
1244 	 */
1245 	if (th->th_win == 0) {
1246 		tp->t_sndzerowin++;
1247 		tp->t_flags |= TF_RXWIN0SENT;
1248 	} else
1249 		tp->t_flags &= ~TF_RXWIN0SENT;
1250 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1251 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1252 		th->th_flags |= TH_URG;
1253 	} else
1254 		/*
1255 		 * If no urgent pointer to send, then we pull
1256 		 * the urgent pointer to the left edge of the send window
1257 		 * so that it doesn't drift into the send window on sequence
1258 		 * number wraparound.
1259 		 */
1260 		tp->snd_up = tp->snd_una;		/* drag it along */
1261 
1262 	/*
1263 	 * Put TCP length in extended header, and then
1264 	 * checksum extended header and data.
1265 	 */
1266 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1267 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1268 
1269 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1270 	if (to.to_flags & TOF_SIGNATURE) {
1271 		/*
1272 		 * Calculate MD5 signature and put it into the place
1273 		 * determined before.
1274 		 * NOTE: since TCP options buffer doesn't point into
1275 		 * mbuf's data, calculate offset and use it.
1276 		 */
1277 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1278 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1279 			/*
1280 			 * Do not send segment if the calculation of MD5
1281 			 * digest has failed.
1282 			 */
1283 			m_freem(m);
1284 			goto out;
1285 		}
1286 	}
1287 #endif
1288 #ifdef INET6
1289 	if (isipv6) {
1290 		/*
1291 		 * There is no need to fill in ip6_plen right now.
1292 		 * It will be filled later by ip6_output.
1293 		 */
1294 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1295 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1296 		    optlen + len, IPPROTO_TCP, 0);
1297 	}
1298 #endif
1299 #if defined(INET6) && defined(INET)
1300 	else
1301 #endif
1302 #ifdef INET
1303 	{
1304 		m->m_pkthdr.csum_flags = CSUM_TCP;
1305 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1306 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1307 
1308 		/* IP version must be set here for ipv4/ipv6 checking later */
1309 		KASSERT(ip->ip_v == IPVERSION,
1310 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1311 	}
1312 #endif
1313 
1314 	/*
1315 	 * Enable TSO and specify the size of the segments.
1316 	 * The TCP pseudo header checksum is always provided.
1317 	 */
1318 	if (tso) {
1319 		KASSERT(len > tp->t_maxseg - optlen,
1320 		    ("%s: len <= tso_segsz", __func__));
1321 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1322 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1323 	}
1324 
1325 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1326 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1327 	    ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1328 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1329 #else
1330 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1331 	    ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1332 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1333 #endif
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 		}
1497 
1498 		/*
1499 		 * Set retransmit timer if not currently set,
1500 		 * and not doing a pure ack or a keep-alive probe.
1501 		 * Initial value for retransmit timer is smoothed
1502 		 * round-trip time + 2 * round-trip time variance.
1503 		 * Initialize shift counter which is used for backoff
1504 		 * of retransmit time.
1505 		 */
1506 timer:
1507 		if (!tcp_timer_active(tp, TT_REXMT) &&
1508 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1509 		     (tp->snd_nxt != tp->snd_una))) {
1510 			if (tcp_timer_active(tp, TT_PERSIST)) {
1511 				tcp_timer_activate(tp, TT_PERSIST, 0);
1512 				tp->t_rxtshift = 0;
1513 			}
1514 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1515 		} else if (len == 0 && sbavail(&so->so_snd) &&
1516 		    !tcp_timer_active(tp, TT_REXMT) &&
1517 		    !tcp_timer_active(tp, TT_PERSIST)) {
1518 			/*
1519 			 * Avoid a situation where we do not set persist timer
1520 			 * after a zero window condition. For example:
1521 			 * 1) A -> B: packet with enough data to fill the window
1522 			 * 2) B -> A: ACK for #1 + new data (0 window
1523 			 *    advertisement)
1524 			 * 3) A -> B: ACK for #2, 0 len packet
1525 			 *
1526 			 * In this case, A will not activate the persist timer,
1527 			 * because it chose to send a packet. Unless tcp_output
1528 			 * is called for some other reason (delayed ack timer,
1529 			 * another input packet from B, socket syscall), A will
1530 			 * not send zero window probes.
1531 			 *
1532 			 * So, if you send a 0-length packet, but there is data
1533 			 * in the socket buffer, and neither the rexmt or
1534 			 * persist timer is already set, then activate the
1535 			 * persist timer.
1536 			 */
1537 			tp->t_rxtshift = 0;
1538 			tcp_setpersist(tp);
1539 		}
1540 	} else {
1541 		/*
1542 		 * Persist case, update snd_max but since we are in
1543 		 * persist mode (no window) we do not update snd_nxt.
1544 		 */
1545 		int xlen = len;
1546 		if (flags & TH_SYN)
1547 			++xlen;
1548 		if (flags & TH_FIN) {
1549 			++xlen;
1550 			tp->t_flags |= TF_SENTFIN;
1551 		}
1552 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1553 			tp->snd_max = tp->snd_nxt + xlen;
1554 	}
1555 
1556 	if (error) {
1557 		/* Record the error. */
1558 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1559 		    error, 0, NULL, false);
1560 
1561 		/*
1562 		 * We know that the packet was lost, so back out the
1563 		 * sequence number advance, if any.
1564 		 *
1565 		 * If the error is EPERM the packet got blocked by the
1566 		 * local firewall.  Normally we should terminate the
1567 		 * connection but the blocking may have been spurious
1568 		 * due to a firewall reconfiguration cycle.  So we treat
1569 		 * it like a packet loss and let the retransmit timer and
1570 		 * timeouts do their work over time.
1571 		 * XXX: It is a POLA question whether calling tcp_drop right
1572 		 * away would be the really correct behavior instead.
1573 		 */
1574 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1575 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1576 		    ((flags & TH_SYN) == 0) &&
1577 		    (error != EPERM)) {
1578 			if (sack_rxmit) {
1579 				p->rxmit -= len;
1580 				tp->sackhint.sack_bytes_rexmit -= len;
1581 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1582 				    ("sackhint bytes rtx >= 0"));
1583 			} else
1584 				tp->snd_nxt -= len;
1585 		}
1586 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1587 		switch (error) {
1588 		case EACCES:
1589 			tp->t_softerror = error;
1590 			return (0);
1591 		case EPERM:
1592 			tp->t_softerror = error;
1593 			return (error);
1594 		case ENOBUFS:
1595 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1596 			tp->snd_cwnd = tp->t_maxseg;
1597 			return (0);
1598 		case EMSGSIZE:
1599 			/*
1600 			 * For some reason the interface we used initially
1601 			 * to send segments changed to another or lowered
1602 			 * its MTU.
1603 			 * If TSO was active we either got an interface
1604 			 * without TSO capabilits or TSO was turned off.
1605 			 * If we obtained mtu from ip_output() then update
1606 			 * it and try again.
1607 			 */
1608 			if (tso)
1609 				tp->t_flags &= ~TF_TSO;
1610 			if (mtu != 0) {
1611 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1612 				goto again;
1613 			}
1614 			return (error);
1615 		case EHOSTDOWN:
1616 		case EHOSTUNREACH:
1617 		case ENETDOWN:
1618 		case ENETUNREACH:
1619 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1620 				tp->t_softerror = error;
1621 				return (0);
1622 			}
1623 			/* FALLTHROUGH */
1624 		default:
1625 			return (error);
1626 		}
1627 	}
1628 	TCPSTAT_INC(tcps_sndtotal);
1629 
1630 	/*
1631 	 * Data sent (as far as we can tell).
1632 	 * If this advertises a larger window than any other segment,
1633 	 * then remember the size of the advertised window.
1634 	 * Any pending ACK has now been sent.
1635 	 */
1636 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1637 		tp->rcv_adv = tp->rcv_nxt + recwin;
1638 	tp->last_ack_sent = tp->rcv_nxt;
1639 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1640 	if (tcp_timer_active(tp, TT_DELACK))
1641 		tcp_timer_activate(tp, TT_DELACK, 0);
1642 #if 0
1643 	/*
1644 	 * This completely breaks TCP if newreno is turned on.  What happens
1645 	 * is that if delayed-acks are turned on on the receiver, this code
1646 	 * on the transmitter effectively destroys the TCP window, forcing
1647 	 * it to four packets (1.5Kx4 = 6K window).
1648 	 */
1649 	if (sendalot && --maxburst)
1650 		goto again;
1651 #endif
1652 	if (sendalot)
1653 		goto again;
1654 	return (0);
1655 }
1656 
1657 void
1658 tcp_setpersist(struct tcpcb *tp)
1659 {
1660 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1661 	int tt;
1662 
1663 	tp->t_flags &= ~TF_PREVVALID;
1664 	if (tcp_timer_active(tp, TT_REXMT))
1665 		panic("tcp_setpersist: retransmit pending");
1666 	/*
1667 	 * Start/restart persistence timer.
1668 	 */
1669 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1670 		      tcp_persmin, tcp_persmax);
1671 	tcp_timer_activate(tp, TT_PERSIST, tt);
1672 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1673 		tp->t_rxtshift++;
1674 }
1675 
1676 /*
1677  * Insert TCP options according to the supplied parameters to the place
1678  * optp in a consistent way.  Can handle unaligned destinations.
1679  *
1680  * The order of the option processing is crucial for optimal packing and
1681  * alignment for the scarce option space.
1682  *
1683  * The optimal order for a SYN/SYN-ACK segment is:
1684  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1685  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1686  *
1687  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1688  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1689  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1690  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1691  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1692  */
1693 int
1694 tcp_addoptions(struct tcpopt *to, u_char *optp)
1695 {
1696 	u_int32_t mask, optlen = 0;
1697 
1698 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1699 		if ((to->to_flags & mask) != mask)
1700 			continue;
1701 		if (optlen == TCP_MAXOLEN)
1702 			break;
1703 		switch (to->to_flags & mask) {
1704 		case TOF_MSS:
1705 			while (optlen % 4) {
1706 				optlen += TCPOLEN_NOP;
1707 				*optp++ = TCPOPT_NOP;
1708 			}
1709 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1710 				continue;
1711 			optlen += TCPOLEN_MAXSEG;
1712 			*optp++ = TCPOPT_MAXSEG;
1713 			*optp++ = TCPOLEN_MAXSEG;
1714 			to->to_mss = htons(to->to_mss);
1715 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1716 			optp += sizeof(to->to_mss);
1717 			break;
1718 		case TOF_SCALE:
1719 			while (!optlen || optlen % 2 != 1) {
1720 				optlen += TCPOLEN_NOP;
1721 				*optp++ = TCPOPT_NOP;
1722 			}
1723 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1724 				continue;
1725 			optlen += TCPOLEN_WINDOW;
1726 			*optp++ = TCPOPT_WINDOW;
1727 			*optp++ = TCPOLEN_WINDOW;
1728 			*optp++ = to->to_wscale;
1729 			break;
1730 		case TOF_SACKPERM:
1731 			while (optlen % 2) {
1732 				optlen += TCPOLEN_NOP;
1733 				*optp++ = TCPOPT_NOP;
1734 			}
1735 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1736 				continue;
1737 			optlen += TCPOLEN_SACK_PERMITTED;
1738 			*optp++ = TCPOPT_SACK_PERMITTED;
1739 			*optp++ = TCPOLEN_SACK_PERMITTED;
1740 			break;
1741 		case TOF_TS:
1742 			while (!optlen || optlen % 4 != 2) {
1743 				optlen += TCPOLEN_NOP;
1744 				*optp++ = TCPOPT_NOP;
1745 			}
1746 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1747 				continue;
1748 			optlen += TCPOLEN_TIMESTAMP;
1749 			*optp++ = TCPOPT_TIMESTAMP;
1750 			*optp++ = TCPOLEN_TIMESTAMP;
1751 			to->to_tsval = htonl(to->to_tsval);
1752 			to->to_tsecr = htonl(to->to_tsecr);
1753 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1754 			optp += sizeof(to->to_tsval);
1755 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1756 			optp += sizeof(to->to_tsecr);
1757 			break;
1758 		case TOF_SIGNATURE:
1759 			{
1760 			int siglen = TCPOLEN_SIGNATURE - 2;
1761 
1762 			while (!optlen || optlen % 4 != 2) {
1763 				optlen += TCPOLEN_NOP;
1764 				*optp++ = TCPOPT_NOP;
1765 			}
1766 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1767 				to->to_flags &= ~TOF_SIGNATURE;
1768 				continue;
1769 			}
1770 			optlen += TCPOLEN_SIGNATURE;
1771 			*optp++ = TCPOPT_SIGNATURE;
1772 			*optp++ = TCPOLEN_SIGNATURE;
1773 			to->to_signature = optp;
1774 			while (siglen--)
1775 				 *optp++ = 0;
1776 			break;
1777 			}
1778 		case TOF_SACK:
1779 			{
1780 			int sackblks = 0;
1781 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1782 			tcp_seq sack_seq;
1783 
1784 			while (!optlen || optlen % 4 != 2) {
1785 				optlen += TCPOLEN_NOP;
1786 				*optp++ = TCPOPT_NOP;
1787 			}
1788 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1789 				continue;
1790 			optlen += TCPOLEN_SACKHDR;
1791 			*optp++ = TCPOPT_SACK;
1792 			sackblks = min(to->to_nsacks,
1793 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1794 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1795 			while (sackblks--) {
1796 				sack_seq = htonl(sack->start);
1797 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1798 				optp += sizeof(sack_seq);
1799 				sack_seq = htonl(sack->end);
1800 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1801 				optp += sizeof(sack_seq);
1802 				optlen += TCPOLEN_SACK;
1803 				sack++;
1804 			}
1805 			TCPSTAT_INC(tcps_sack_send_blocks);
1806 			break;
1807 			}
1808 		case TOF_FASTOPEN:
1809 			{
1810 			int total_len;
1811 
1812 			/* XXX is there any point to aligning this option? */
1813 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1814 			if (TCP_MAXOLEN - optlen < total_len) {
1815 				to->to_flags &= ~TOF_FASTOPEN;
1816 				continue;
1817 			}
1818 			*optp++ = TCPOPT_FAST_OPEN;
1819 			*optp++ = total_len;
1820 			if (to->to_tfo_len > 0) {
1821 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1822 				optp += to->to_tfo_len;
1823 			}
1824 			optlen += total_len;
1825 			break;
1826 			}
1827 		default:
1828 			panic("%s: unknown TCP option type", __func__);
1829 			break;
1830 		}
1831 	}
1832 
1833 	/* Terminate and pad TCP options to a 4 byte boundary. */
1834 	if (optlen % 4) {
1835 		optlen += TCPOLEN_EOL;
1836 		*optp++ = TCPOPT_EOL;
1837 	}
1838 	/*
1839 	 * According to RFC 793 (STD0007):
1840 	 *   "The content of the header beyond the End-of-Option option
1841 	 *    must be header padding (i.e., zero)."
1842 	 *   and later: "The padding is composed of zeros."
1843 	 */
1844 	while (optlen % 4) {
1845 		optlen += TCPOLEN_PAD;
1846 		*optp++ = TCPOPT_PAD;
1847 	}
1848 
1849 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1850 	return (optlen);
1851 }
1852 
1853 void
1854 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1855 {
1856 
1857 	/*
1858 	 * Automatic sizing of send socket buffer.  Often the send buffer
1859 	 * size is not optimally adjusted to the actual network conditions
1860 	 * at hand (delay bandwidth product).  Setting the buffer size too
1861 	 * small limits throughput on links with high bandwidth and high
1862 	 * delay (eg. trans-continental/oceanic links).  Setting the
1863 	 * buffer size too big consumes too much real kernel memory,
1864 	 * especially with many connections on busy servers.
1865 	 *
1866 	 * The criteria to step up the send buffer one notch are:
1867 	 *  1. receive window of remote host is larger than send buffer
1868 	 *     (with a fudge factor of 5/4th);
1869 	 *  2. send buffer is filled to 7/8th with data (so we actually
1870 	 *     have data to make use of it);
1871 	 *  3. send buffer fill has not hit maximal automatic size;
1872 	 *  4. our send window (slow start and cogestion controlled) is
1873 	 *     larger than sent but unacknowledged data in send buffer.
1874 	 *
1875 	 * The remote host receive window scaling factor may limit the
1876 	 * growing of the send buffer before it reaches its allowed
1877 	 * maximum.
1878 	 *
1879 	 * It scales directly with slow start or congestion window
1880 	 * and does at most one step per received ACK.  This fast
1881 	 * scaling has the drawback of growing the send buffer beyond
1882 	 * what is strictly necessary to make full use of a given
1883 	 * delay*bandwidth product.  However testing has shown this not
1884 	 * to be much of an problem.  At worst we are trading wasting
1885 	 * of available bandwidth (the non-use of it) for wasting some
1886 	 * socket buffer memory.
1887 	 *
1888 	 * TODO: Shrink send buffer during idle periods together
1889 	 * with congestion window.  Requires another timer.  Has to
1890 	 * wait for upcoming tcp timer rewrite.
1891 	 *
1892 	 * XXXGL: should there be used sbused() or sbavail()?
1893 	 */
1894 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
1895 		int lowat;
1896 
1897 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
1898 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
1899 		    sbused(&so->so_snd) >=
1900 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
1901 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
1902 		    sendwin >= (sbused(&so->so_snd) -
1903 		    (tp->snd_nxt - tp->snd_una))) {
1904 			if (!sbreserve_locked(&so->so_snd,
1905 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
1906 			     V_tcp_autosndbuf_max), so, curthread))
1907 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
1908 		}
1909 	}
1910 }
1911