xref: /freebsd/sys/netinet/tcp_output.c (revision 4e99f45480598189d49d45a825533a6c9e12f02c)
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 		cc_after_idle(tp);
265 	tp->t_flags &= ~TF_LASTIDLE;
266 	if (idle) {
267 		if (tp->t_flags & TF_MORETOCOME) {
268 			tp->t_flags |= TF_LASTIDLE;
269 			idle = 0;
270 		}
271 	}
272 again:
273 	/*
274 	 * If we've recently taken a timeout, snd_max will be greater than
275 	 * snd_nxt.  There may be SACK information that allows us to avoid
276 	 * resending already delivered data.  Adjust snd_nxt accordingly.
277 	 */
278 	if ((tp->t_flags & TF_SACK_PERMIT) &&
279 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
280 		tcp_sack_adjust(tp);
281 	sendalot = 0;
282 	tso = 0;
283 	mtu = 0;
284 	off = tp->snd_nxt - tp->snd_una;
285 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
286 
287 	flags = tcp_outflags[tp->t_state];
288 	/*
289 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
290 	 * to send out new data (when sendalot is 1), bypass this function.
291 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
292 	 * we're replacing a (future) new transmission with a retransmission
293 	 * now, and we previously incremented snd_cwnd in tcp_input().
294 	 */
295 	/*
296 	 * Still in sack recovery , reset rxmit flag to zero.
297 	 */
298 	sack_rxmit = 0;
299 	sack_bytes_rxmt = 0;
300 	len = 0;
301 	p = NULL;
302 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
303 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
304 		uint32_t cwin;
305 
306 		cwin =
307 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
308 		/* Do not retransmit SACK segments beyond snd_recover */
309 		if (SEQ_GT(p->end, tp->snd_recover)) {
310 			/*
311 			 * (At least) part of sack hole extends beyond
312 			 * snd_recover. Check to see if we can rexmit data
313 			 * for this hole.
314 			 */
315 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
316 				/*
317 				 * Can't rexmit any more data for this hole.
318 				 * That data will be rexmitted in the next
319 				 * sack recovery episode, when snd_recover
320 				 * moves past p->rxmit.
321 				 */
322 				p = NULL;
323 				goto after_sack_rexmit;
324 			} else
325 				/* Can rexmit part of the current hole */
326 				len = ((int32_t)ulmin(cwin,
327 						   tp->snd_recover - p->rxmit));
328 		} else
329 			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
330 		off = p->rxmit - tp->snd_una;
331 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
332 		    __func__, off));
333 		if (len > 0) {
334 			sack_rxmit = 1;
335 			sendalot = 1;
336 			TCPSTAT_INC(tcps_sack_rexmits);
337 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
338 			    min(len, tp->t_maxseg));
339 		}
340 	}
341 after_sack_rexmit:
342 	/*
343 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
344 	 * state flags.
345 	 */
346 	if (tp->t_flags & TF_NEEDFIN)
347 		flags |= TH_FIN;
348 	if (tp->t_flags & TF_NEEDSYN)
349 		flags |= TH_SYN;
350 
351 	SOCKBUF_LOCK(&so->so_snd);
352 	/*
353 	 * If in persist timeout with window of 0, send 1 byte.
354 	 * Otherwise, if window is small but nonzero
355 	 * and timer expired, we will send what we can
356 	 * and go to transmit state.
357 	 */
358 	if (tp->t_flags & TF_FORCEDATA) {
359 		if (sendwin == 0) {
360 			/*
361 			 * If we still have some data to send, then
362 			 * clear the FIN bit.  Usually this would
363 			 * happen below when it realizes that we
364 			 * aren't sending all the data.  However,
365 			 * if we have exactly 1 byte of unsent data,
366 			 * then it won't clear the FIN bit below,
367 			 * and if we are in persist state, we wind
368 			 * up sending the packet without recording
369 			 * that we sent the FIN bit.
370 			 *
371 			 * We can't just blindly clear the FIN bit,
372 			 * because if we don't have any more data
373 			 * to send then the probe will be the FIN
374 			 * itself.
375 			 */
376 			if (off < sbused(&so->so_snd))
377 				flags &= ~TH_FIN;
378 			sendwin = 1;
379 		} else {
380 			tcp_timer_activate(tp, TT_PERSIST, 0);
381 			tp->t_rxtshift = 0;
382 		}
383 	}
384 
385 	/*
386 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
387 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
388 	 * a negative length.  This can also occur when TCP opens up
389 	 * its congestion window while receiving additional duplicate
390 	 * acks after fast-retransmit because TCP will reset snd_nxt
391 	 * to snd_max after the fast-retransmit.
392 	 *
393 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
394 	 * be set to snd_una, the offset will be 0, and the length may
395 	 * wind up 0.
396 	 *
397 	 * If sack_rxmit is true we are retransmitting from the scoreboard
398 	 * in which case len is already set.
399 	 */
400 	if (sack_rxmit == 0) {
401 		if (sack_bytes_rxmt == 0)
402 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
403 			    off);
404 		else {
405 			int32_t cwin;
406 
407                         /*
408 			 * We are inside of a SACK recovery episode and are
409 			 * sending new data, having retransmitted all the
410 			 * data possible in the scoreboard.
411 			 */
412 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
413 			    off);
414 			/*
415 			 * Don't remove this (len > 0) check !
416 			 * We explicitly check for len > 0 here (although it
417 			 * isn't really necessary), to work around a gcc
418 			 * optimization issue - to force gcc to compute
419 			 * len above. Without this check, the computation
420 			 * of len is bungled by the optimizer.
421 			 */
422 			if (len > 0) {
423 				cwin = tp->snd_cwnd -
424 					(tp->snd_nxt - tp->snd_recover) -
425 					sack_bytes_rxmt;
426 				if (cwin < 0)
427 					cwin = 0;
428 				len = imin(len, cwin);
429 			}
430 		}
431 	}
432 
433 	/*
434 	 * Lop off SYN bit if it has already been sent.  However, if this
435 	 * is SYN-SENT state and if segment contains data and if we don't
436 	 * know that foreign host supports TAO, suppress sending segment.
437 	 */
438 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
439 		if (tp->t_state != TCPS_SYN_RECEIVED)
440 			flags &= ~TH_SYN;
441 		/*
442 		 * When sending additional segments following a TFO SYN|ACK,
443 		 * do not include the SYN bit.
444 		 */
445 		if (IS_FASTOPEN(tp->t_flags) &&
446 		    (tp->t_state == TCPS_SYN_RECEIVED))
447 			flags &= ~TH_SYN;
448 		off--, len++;
449 	}
450 
451 	/*
452 	 * Be careful not to send data and/or FIN on SYN segments.
453 	 * This measure is needed to prevent interoperability problems
454 	 * with not fully conformant TCP implementations.
455 	 */
456 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
457 		len = 0;
458 		flags &= ~TH_FIN;
459 	}
460 
461 	/*
462 	 * On TFO sockets, ensure no data is sent in the following cases:
463 	 *
464 	 *  - When retransmitting SYN|ACK on a passively-created socket
465 	 *
466 	 *  - When retransmitting SYN on an actively created socket
467 	 *
468 	 *  - When sending a zero-length cookie (cookie request) on an
469 	 *    actively created socket
470 	 *
471 	 *  - When the socket is in the CLOSED state (RST is being sent)
472 	 */
473 	if (IS_FASTOPEN(tp->t_flags) &&
474 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
475 	     ((tp->t_state == TCPS_SYN_SENT) &&
476 	      (tp->t_tfo_client_cookie_len == 0)) ||
477 	     (flags & TH_RST)))
478 		len = 0;
479 	if (len <= 0) {
480 		/*
481 		 * If FIN has been sent but not acked,
482 		 * but we haven't been called to retransmit,
483 		 * len will be < 0.  Otherwise, window shrank
484 		 * after we sent into it.  If window shrank to 0,
485 		 * cancel pending retransmit, pull snd_nxt back
486 		 * to (closed) window, and set the persist timer
487 		 * if it isn't already going.  If the window didn't
488 		 * close completely, just wait for an ACK.
489 		 *
490 		 * We also do a general check here to ensure that
491 		 * we will set the persist timer when we have data
492 		 * to send, but a 0-byte window. This makes sure
493 		 * the persist timer is set even if the packet
494 		 * hits one of the "goto send" lines below.
495 		 */
496 		len = 0;
497 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
498 			(off < (int) sbavail(&so->so_snd))) {
499 			tcp_timer_activate(tp, TT_REXMT, 0);
500 			tp->t_rxtshift = 0;
501 			tp->snd_nxt = tp->snd_una;
502 			if (!tcp_timer_active(tp, TT_PERSIST))
503 				tcp_setpersist(tp);
504 		}
505 	}
506 
507 	/* len will be >= 0 after this point. */
508 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
509 
510 	tcp_sndbuf_autoscale(tp, so, sendwin);
511 
512 	/*
513 	 * Decide if we can use TCP Segmentation Offloading (if supported by
514 	 * hardware).
515 	 *
516 	 * TSO may only be used if we are in a pure bulk sending state.  The
517 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
518 	 * IP options prevent using TSO.  With TSO the TCP header is the same
519 	 * (except for the sequence number) for all generated packets.  This
520 	 * makes it impossible to transmit any options which vary per generated
521 	 * segment or packet.
522 	 *
523 	 * IPv4 handling has a clear separation of ip options and ip header
524 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
525 	 * the right thing below to provide length of just ip options and thus
526 	 * checking for ipoptlen is enough to decide if ip options are present.
527 	 */
528 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
529 	/*
530 	 * Pre-calculate here as we save another lookup into the darknesses
531 	 * of IPsec that way and can actually decide if TSO is ok.
532 	 */
533 #ifdef INET6
534 	if (isipv6 && IPSEC_ENABLED(ipv6))
535 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
536 #ifdef INET
537 	else
538 #endif
539 #endif /* INET6 */
540 #ifdef INET
541 	if (IPSEC_ENABLED(ipv4))
542 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
543 #endif /* INET */
544 #endif /* IPSEC */
545 #ifdef INET6
546 	if (isipv6)
547 		ipoptlen = ip6_optlen(tp->t_inpcb);
548 	else
549 #endif
550 	if (tp->t_inpcb->inp_options)
551 		ipoptlen = tp->t_inpcb->inp_options->m_len -
552 				offsetof(struct ipoption, ipopt_list);
553 	else
554 		ipoptlen = 0;
555 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
556 	ipoptlen += ipsec_optlen;
557 #endif
558 
559 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
560 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
561 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
562 	    ipoptlen == 0 && !(flags & TH_SYN))
563 		tso = 1;
564 
565 	if (sack_rxmit) {
566 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
567 			flags &= ~TH_FIN;
568 	} else {
569 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
570 		    sbused(&so->so_snd)))
571 			flags &= ~TH_FIN;
572 	}
573 
574 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
575 	    (long)TCP_MAXWIN << tp->rcv_scale);
576 
577 	/*
578 	 * Sender silly window avoidance.   We transmit under the following
579 	 * conditions when len is non-zero:
580 	 *
581 	 *	- We have a full segment (or more with TSO)
582 	 *	- This is the last buffer in a write()/send() and we are
583 	 *	  either idle or running NODELAY
584 	 *	- we've timed out (e.g. persist timer)
585 	 *	- we have more then 1/2 the maximum send window's worth of
586 	 *	  data (receiver may be limited the window size)
587 	 *	- we need to retransmit
588 	 */
589 	if (len) {
590 		if (len >= tp->t_maxseg)
591 			goto send;
592 		/*
593 		 * NOTE! on localhost connections an 'ack' from the remote
594 		 * end may occur synchronously with the output and cause
595 		 * us to flush a buffer queued with moretocome.  XXX
596 		 *
597 		 * note: the len + off check is almost certainly unnecessary.
598 		 */
599 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
600 		    (idle || (tp->t_flags & TF_NODELAY)) &&
601 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
602 		    (tp->t_flags & TF_NOPUSH) == 0) {
603 			goto send;
604 		}
605 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
606 			goto send;
607 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
608 			goto send;
609 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
610 			goto send;
611 		if (sack_rxmit)
612 			goto send;
613 	}
614 
615 	/*
616 	 * Sending of standalone window updates.
617 	 *
618 	 * Window updates are important when we close our window due to a
619 	 * full socket buffer and are opening it again after the application
620 	 * reads data from it.  Once the window has opened again and the
621 	 * remote end starts to send again the ACK clock takes over and
622 	 * provides the most current window information.
623 	 *
624 	 * We must avoid the silly window syndrome whereas every read
625 	 * from the receive buffer, no matter how small, causes a window
626 	 * update to be sent.  We also should avoid sending a flurry of
627 	 * window updates when the socket buffer had queued a lot of data
628 	 * and the application is doing small reads.
629 	 *
630 	 * Prevent a flurry of pointless window updates by only sending
631 	 * an update when we can increase the advertized window by more
632 	 * than 1/4th of the socket buffer capacity.  When the buffer is
633 	 * getting full or is very small be more aggressive and send an
634 	 * update whenever we can increase by two mss sized segments.
635 	 * In all other situations the ACK's to new incoming data will
636 	 * carry further window increases.
637 	 *
638 	 * Don't send an independent window update if a delayed
639 	 * ACK is pending (it will get piggy-backed on it) or the
640 	 * remote side already has done a half-close and won't send
641 	 * more data.  Skip this if the connection is in T/TCP
642 	 * half-open state.
643 	 */
644 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
645 	    !(tp->t_flags & TF_DELACK) &&
646 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
647 		/*
648 		 * "adv" is the amount we could increase the window,
649 		 * taking into account that we are limited by
650 		 * TCP_MAXWIN << tp->rcv_scale.
651 		 */
652 		int32_t adv;
653 		int oldwin;
654 
655 		adv = recwin;
656 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
657 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
658 			adv -= oldwin;
659 		} else
660 			oldwin = 0;
661 
662 		/*
663 		 * If the new window size ends up being the same as or less
664 		 * than the old size when it is scaled, then don't force
665 		 * a window update.
666 		 */
667 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
668 			goto dontupdate;
669 
670 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
671 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
672 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
673 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
674 		     adv >= TCP_MAXWIN << tp->rcv_scale))
675 			goto send;
676 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
677 			goto send;
678 	}
679 dontupdate:
680 
681 	/*
682 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
683 	 * is also a catch-all for the retransmit timer timeout case.
684 	 */
685 	if (tp->t_flags & TF_ACKNOW)
686 		goto send;
687 	if ((flags & TH_RST) ||
688 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
689 		goto send;
690 	if (SEQ_GT(tp->snd_up, tp->snd_una))
691 		goto send;
692 	/*
693 	 * If our state indicates that FIN should be sent
694 	 * and we have not yet done so, then we need to send.
695 	 */
696 	if (flags & TH_FIN &&
697 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
698 		goto send;
699 	/*
700 	 * In SACK, it is possible for tcp_output to fail to send a segment
701 	 * after the retransmission timer has been turned off.  Make sure
702 	 * that the retransmission timer is set.
703 	 */
704 	if ((tp->t_flags & TF_SACK_PERMIT) &&
705 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
706 	    !tcp_timer_active(tp, TT_REXMT) &&
707 	    !tcp_timer_active(tp, TT_PERSIST)) {
708 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
709 		goto just_return;
710 	}
711 	/*
712 	 * TCP window updates are not reliable, rather a polling protocol
713 	 * using ``persist'' packets is used to insure receipt of window
714 	 * updates.  The three ``states'' for the output side are:
715 	 *	idle			not doing retransmits or persists
716 	 *	persisting		to move a small or zero window
717 	 *	(re)transmitting	and thereby not persisting
718 	 *
719 	 * tcp_timer_active(tp, TT_PERSIST)
720 	 *	is true when we are in persist state.
721 	 * (tp->t_flags & TF_FORCEDATA)
722 	 *	is set when we are called to send a persist packet.
723 	 * tcp_timer_active(tp, TT_REXMT)
724 	 *	is set when we are retransmitting
725 	 * The output side is idle when both timers are zero.
726 	 *
727 	 * If send window is too small, there is data to transmit, and no
728 	 * retransmit or persist is pending, then go to persist state.
729 	 * If nothing happens soon, send when timer expires:
730 	 * if window is nonzero, transmit what we can,
731 	 * otherwise force out a byte.
732 	 */
733 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
734 	    !tcp_timer_active(tp, TT_PERSIST)) {
735 		tp->t_rxtshift = 0;
736 		tcp_setpersist(tp);
737 	}
738 
739 	/*
740 	 * No reason to send a segment, just return.
741 	 */
742 just_return:
743 	SOCKBUF_UNLOCK(&so->so_snd);
744 	return (0);
745 
746 send:
747 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
748 	if (len > 0) {
749 		if (len >= tp->t_maxseg)
750 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
751 		else
752 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
753 	}
754 	/*
755 	 * Before ESTABLISHED, force sending of initial options
756 	 * unless TCP set not to do any options.
757 	 * NOTE: we assume that the IP/TCP header plus TCP options
758 	 * always fit in a single mbuf, leaving room for a maximum
759 	 * link header, i.e.
760 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
761 	 */
762 	optlen = 0;
763 #ifdef INET6
764 	if (isipv6)
765 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
766 	else
767 #endif
768 		hdrlen = sizeof (struct tcpiphdr);
769 
770 	/*
771 	 * Compute options for segment.
772 	 * We only have to care about SYN and established connection
773 	 * segments.  Options for SYN-ACK segments are handled in TCP
774 	 * syncache.
775 	 */
776 	to.to_flags = 0;
777 	if ((tp->t_flags & TF_NOOPT) == 0) {
778 		/* Maximum segment size. */
779 		if (flags & TH_SYN) {
780 			tp->snd_nxt = tp->iss;
781 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
782 			to.to_flags |= TOF_MSS;
783 
784 			/*
785 			 * On SYN or SYN|ACK transmits on TFO connections,
786 			 * only include the TFO option if it is not a
787 			 * retransmit, as the presence of the TFO option may
788 			 * have caused the original SYN or SYN|ACK to have
789 			 * been dropped by a middlebox.
790 			 */
791 			if (IS_FASTOPEN(tp->t_flags) &&
792 			    (tp->t_rxtshift == 0)) {
793 				if (tp->t_state == TCPS_SYN_RECEIVED) {
794 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
795 					to.to_tfo_cookie =
796 					    (u_int8_t *)&tp->t_tfo_cookie.server;
797 					to.to_flags |= TOF_FASTOPEN;
798 					wanted_cookie = 1;
799 				} else if (tp->t_state == TCPS_SYN_SENT) {
800 					to.to_tfo_len =
801 					    tp->t_tfo_client_cookie_len;
802 					to.to_tfo_cookie =
803 					    tp->t_tfo_cookie.client;
804 					to.to_flags |= TOF_FASTOPEN;
805 					wanted_cookie = 1;
806 					/*
807 					 * If we wind up having more data to
808 					 * send with the SYN than can fit in
809 					 * one segment, don't send any more
810 					 * until the SYN|ACK comes back from
811 					 * the other end.
812 					 */
813 					dont_sendalot = 1;
814 				}
815 			}
816 		}
817 		/* Window scaling. */
818 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
819 			to.to_wscale = tp->request_r_scale;
820 			to.to_flags |= TOF_SCALE;
821 		}
822 		/* Timestamps. */
823 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
824 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
825 			curticks = tcp_ts_getticks();
826 			to.to_tsval = curticks + tp->ts_offset;
827 			to.to_tsecr = tp->ts_recent;
828 			to.to_flags |= TOF_TS;
829 			if (tp->t_rxtshift == 1)
830 				tp->t_badrxtwin = curticks;
831 		}
832 
833 		/* Set receive buffer autosizing timestamp. */
834 		if (tp->rfbuf_ts == 0 &&
835 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
836 			tp->rfbuf_ts = tcp_ts_getticks();
837 
838 		/* Selective ACK's. */
839 		if (tp->t_flags & TF_SACK_PERMIT) {
840 			if (flags & TH_SYN)
841 				to.to_flags |= TOF_SACKPERM;
842 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
843 			    (tp->t_flags & TF_SACK_PERMIT) &&
844 			    tp->rcv_numsacks > 0) {
845 				to.to_flags |= TOF_SACK;
846 				to.to_nsacks = tp->rcv_numsacks;
847 				to.to_sacks = (u_char *)tp->sackblks;
848 			}
849 		}
850 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
851 		/* TCP-MD5 (RFC2385). */
852 		/*
853 		 * Check that TCP_MD5SIG is enabled in tcpcb to
854 		 * account the size needed to set this TCP option.
855 		 */
856 		if (tp->t_flags & TF_SIGNATURE)
857 			to.to_flags |= TOF_SIGNATURE;
858 #endif /* TCP_SIGNATURE */
859 
860 		/* Processing the options. */
861 		hdrlen += optlen = tcp_addoptions(&to, opt);
862 		/*
863 		 * If we wanted a TFO option to be added, but it was unable
864 		 * to fit, ensure no data is sent.
865 		 */
866 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
867 		    !(to.to_flags & TOF_FASTOPEN))
868 			len = 0;
869 	}
870 
871 	/*
872 	 * Adjust data length if insertion of options will
873 	 * bump the packet length beyond the t_maxseg length.
874 	 * Clear the FIN bit because we cut off the tail of
875 	 * the segment.
876 	 */
877 	if (len + optlen + ipoptlen > tp->t_maxseg) {
878 		flags &= ~TH_FIN;
879 
880 		if (tso) {
881 			u_int if_hw_tsomax;
882 			u_int moff;
883 			int max_len;
884 
885 			/* extract TSO information */
886 			if_hw_tsomax = tp->t_tsomax;
887 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
888 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
889 
890 			/*
891 			 * Limit a TSO burst to prevent it from
892 			 * overflowing or exceeding the maximum length
893 			 * allowed by the network interface:
894 			 */
895 			KASSERT(ipoptlen == 0,
896 			    ("%s: TSO can't do IP options", __func__));
897 
898 			/*
899 			 * Check if we should limit by maximum payload
900 			 * length:
901 			 */
902 			if (if_hw_tsomax != 0) {
903 				/* compute maximum TSO length */
904 				max_len = (if_hw_tsomax - hdrlen -
905 				    max_linkhdr);
906 				if (max_len <= 0) {
907 					len = 0;
908 				} else if (len > max_len) {
909 					sendalot = 1;
910 					len = max_len;
911 				}
912 			}
913 
914 			/*
915 			 * Prevent the last segment from being
916 			 * fractional unless the send sockbuf can be
917 			 * emptied:
918 			 */
919 			max_len = (tp->t_maxseg - optlen);
920 			if (((uint32_t)off + (uint32_t)len) <
921 			    sbavail(&so->so_snd)) {
922 				moff = len % max_len;
923 				if (moff != 0) {
924 					len -= moff;
925 					sendalot = 1;
926 				}
927 			}
928 
929 			/*
930 			 * In case there are too many small fragments
931 			 * don't use TSO:
932 			 */
933 			if (len <= max_len) {
934 				len = max_len;
935 				sendalot = 1;
936 				tso = 0;
937 			}
938 
939 			/*
940 			 * Send the FIN in a separate segment
941 			 * after the bulk sending is done.
942 			 * We don't trust the TSO implementations
943 			 * to clear the FIN flag on all but the
944 			 * last segment.
945 			 */
946 			if (tp->t_flags & TF_NEEDFIN)
947 				sendalot = 1;
948 		} else {
949 			if (optlen + ipoptlen >= tp->t_maxseg) {
950 				/*
951 				 * Since we don't have enough space to put
952 				 * the IP header chain and the TCP header in
953 				 * one packet as required by RFC 7112, don't
954 				 * send it. Also ensure that at least one
955 				 * byte of the payload can be put into the
956 				 * TCP segment.
957 				 */
958 				SOCKBUF_UNLOCK(&so->so_snd);
959 				error = EMSGSIZE;
960 				sack_rxmit = 0;
961 				goto out;
962 			}
963 			len = tp->t_maxseg - optlen - ipoptlen;
964 			sendalot = 1;
965 			if (dont_sendalot)
966 				sendalot = 0;
967 		}
968 	} else
969 		tso = 0;
970 
971 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
972 	    ("%s: len > IP_MAXPACKET", __func__));
973 
974 /*#ifdef DIAGNOSTIC*/
975 #ifdef INET6
976 	if (max_linkhdr + hdrlen > MCLBYTES)
977 #else
978 	if (max_linkhdr + hdrlen > MHLEN)
979 #endif
980 		panic("tcphdr too big");
981 /*#endif*/
982 
983 	/*
984 	 * This KASSERT is here to catch edge cases at a well defined place.
985 	 * Before, those had triggered (random) panic conditions further down.
986 	 */
987 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
988 
989 	/*
990 	 * Grab a header mbuf, attaching a copy of data to
991 	 * be transmitted, and initialize the header from
992 	 * the template for sends on this connection.
993 	 */
994 	if (len) {
995 		struct mbuf *mb;
996 		struct sockbuf *msb;
997 		u_int moff;
998 
999 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1000 			TCPSTAT_INC(tcps_sndprobe);
1001 #ifdef STATS
1002 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1003 				stats_voi_update_abs_u32(tp->t_stats,
1004 				VOI_TCP_RETXPB, len);
1005 			else
1006 				stats_voi_update_abs_u64(tp->t_stats,
1007 				    VOI_TCP_TXPB, len);
1008 #endif /* STATS */
1009 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1010 			tp->t_sndrexmitpack++;
1011 			TCPSTAT_INC(tcps_sndrexmitpack);
1012 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1013 #ifdef STATS
1014 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1015 			    len);
1016 #endif /* STATS */
1017 		} else {
1018 			TCPSTAT_INC(tcps_sndpack);
1019 			TCPSTAT_ADD(tcps_sndbyte, len);
1020 #ifdef STATS
1021 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1022 			    len);
1023 #endif /* STATS */
1024 		}
1025 #ifdef INET6
1026 		if (MHLEN < hdrlen + max_linkhdr)
1027 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1028 		else
1029 #endif
1030 			m = m_gethdr(M_NOWAIT, MT_DATA);
1031 
1032 		if (m == NULL) {
1033 			SOCKBUF_UNLOCK(&so->so_snd);
1034 			error = ENOBUFS;
1035 			sack_rxmit = 0;
1036 			goto out;
1037 		}
1038 
1039 		m->m_data += max_linkhdr;
1040 		m->m_len = hdrlen;
1041 
1042 		/*
1043 		 * Start the m_copy functions from the closest mbuf
1044 		 * to the offset in the socket buffer chain.
1045 		 */
1046 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1047 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1048 			m_copydata(mb, moff, len,
1049 			    mtod(m, caddr_t) + hdrlen);
1050 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1051 				sbsndptr_adv(&so->so_snd, mb, len);
1052 			m->m_len += len;
1053 		} else {
1054 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1055 				msb = NULL;
1056 			else
1057 				msb = &so->so_snd;
1058 			m->m_next = tcp_m_copym(mb, moff,
1059 			    &len, if_hw_tsomaxsegcount,
1060 			    if_hw_tsomaxsegsize, msb, hw_tls);
1061 			if (len <= (tp->t_maxseg - optlen)) {
1062 				/*
1063 				 * Must have ran out of mbufs for the copy
1064 				 * shorten it to no longer need tso. Lets
1065 				 * not put on sendalot since we are low on
1066 				 * mbufs.
1067 				 */
1068 				tso = 0;
1069 			}
1070 			if (m->m_next == NULL) {
1071 				SOCKBUF_UNLOCK(&so->so_snd);
1072 				(void) m_free(m);
1073 				error = ENOBUFS;
1074 				sack_rxmit = 0;
1075 				goto out;
1076 			}
1077 		}
1078 
1079 		/*
1080 		 * If we're sending everything we've got, set PUSH.
1081 		 * (This will keep happy those implementations which only
1082 		 * give data to the user when a buffer fills or
1083 		 * a PUSH comes in.)
1084 		 */
1085 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1086 		    !(flags & TH_SYN))
1087 			flags |= TH_PUSH;
1088 		SOCKBUF_UNLOCK(&so->so_snd);
1089 	} else {
1090 		SOCKBUF_UNLOCK(&so->so_snd);
1091 		if (tp->t_flags & TF_ACKNOW)
1092 			TCPSTAT_INC(tcps_sndacks);
1093 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1094 			TCPSTAT_INC(tcps_sndctrl);
1095 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1096 			TCPSTAT_INC(tcps_sndurg);
1097 		else
1098 			TCPSTAT_INC(tcps_sndwinup);
1099 
1100 		m = m_gethdr(M_NOWAIT, MT_DATA);
1101 		if (m == NULL) {
1102 			error = ENOBUFS;
1103 			sack_rxmit = 0;
1104 			goto out;
1105 		}
1106 #ifdef INET6
1107 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1108 		    MHLEN >= hdrlen) {
1109 			M_ALIGN(m, hdrlen);
1110 		} else
1111 #endif
1112 		m->m_data += max_linkhdr;
1113 		m->m_len = hdrlen;
1114 	}
1115 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1116 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1117 #ifdef MAC
1118 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1119 #endif
1120 #ifdef INET6
1121 	if (isipv6) {
1122 		ip6 = mtod(m, struct ip6_hdr *);
1123 		th = (struct tcphdr *)(ip6 + 1);
1124 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1125 	} else
1126 #endif /* INET6 */
1127 	{
1128 		ip = mtod(m, struct ip *);
1129 #ifdef TCPDEBUG
1130 		ipov = (struct ipovly *)ip;
1131 #endif
1132 		th = (struct tcphdr *)(ip + 1);
1133 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1134 	}
1135 
1136 	/*
1137 	 * Fill in fields, remembering maximum advertised
1138 	 * window for use in delaying messages about window sizes.
1139 	 * If resending a FIN, be sure not to use a new sequence number.
1140 	 */
1141 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1142 	    tp->snd_nxt == tp->snd_max)
1143 		tp->snd_nxt--;
1144 	/*
1145 	 * If we are starting a connection, send ECN setup
1146 	 * SYN packet. If we are on a retransmit, we may
1147 	 * resend those bits a number of times as per
1148 	 * RFC 3168.
1149 	 */
1150 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1151 		if (tp->t_rxtshift >= 1) {
1152 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1153 				flags |= TH_ECE|TH_CWR;
1154 		} else
1155 			flags |= TH_ECE|TH_CWR;
1156 	}
1157 	/* Handle parallel SYN for ECN */
1158 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1159 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1160 			flags |= TH_ECE;
1161 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1162 	}
1163 
1164 	if (tp->t_state == TCPS_ESTABLISHED &&
1165 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
1166 		/*
1167 		 * If the peer has ECN, mark data packets with
1168 		 * ECN capable transmission (ECT).
1169 		 * Ignore pure ack packets, retransmissions and window probes.
1170 		 */
1171 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1172 		    (sack_rxmit == 0) &&
1173 		    !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1174 		    SEQ_LT(tp->snd_una, tp->snd_max))) {
1175 #ifdef INET6
1176 			if (isipv6)
1177 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1178 			else
1179 #endif
1180 				ip->ip_tos |= IPTOS_ECN_ECT0;
1181 			TCPSTAT_INC(tcps_ecn_ect0);
1182 			/*
1183 			 * Reply with proper ECN notifications.
1184 			 * Only set CWR on new data segments.
1185 			 */
1186 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1187 				flags |= TH_CWR;
1188 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1189 			}
1190 		}
1191 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
1192 			flags |= TH_ECE;
1193 	}
1194 
1195 	/*
1196 	 * If we are doing retransmissions, then snd_nxt will
1197 	 * not reflect the first unsent octet.  For ACK only
1198 	 * packets, we do not want the sequence number of the
1199 	 * retransmitted packet, we want the sequence number
1200 	 * of the next unsent octet.  So, if there is no data
1201 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1202 	 * when filling in ti_seq.  But if we are in persist
1203 	 * state, snd_max might reflect one byte beyond the
1204 	 * right edge of the window, so use snd_nxt in that
1205 	 * case, since we know we aren't doing a retransmission.
1206 	 * (retransmit and persist are mutually exclusive...)
1207 	 */
1208 	if (sack_rxmit == 0) {
1209 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1210 		    tcp_timer_active(tp, TT_PERSIST))
1211 			th->th_seq = htonl(tp->snd_nxt);
1212 		else
1213 			th->th_seq = htonl(tp->snd_max);
1214 	} else {
1215 		th->th_seq = htonl(p->rxmit);
1216 		p->rxmit += len;
1217 		tp->sackhint.sack_bytes_rexmit += len;
1218 	}
1219 	th->th_ack = htonl(tp->rcv_nxt);
1220 	if (optlen) {
1221 		bcopy(opt, th + 1, optlen);
1222 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1223 	}
1224 	th->th_flags = flags;
1225 	/*
1226 	 * Calculate receive window.  Don't shrink window,
1227 	 * but avoid silly window syndrome.
1228 	 * If a RST segment is sent, advertise a window of zero.
1229 	 */
1230 	if (flags & TH_RST) {
1231 		recwin = 0;
1232 	} else {
1233 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1234 		    recwin < tp->t_maxseg)
1235 			recwin = 0;
1236 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1237 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1238 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1239 	}
1240 	/*
1241 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1242 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1243 	 * case is handled in syncache.
1244 	 */
1245 	if (flags & TH_SYN)
1246 		th->th_win = htons((u_short)
1247 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1248 	else {
1249 		/* Avoid shrinking window with window scaling. */
1250 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1251 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1252 	}
1253 
1254 	/*
1255 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1256 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1257 	 * flag tells soreceive() to disable delayed acknowledgements when
1258 	 * draining the buffer.  This can occur if the receiver is attempting
1259 	 * to read more data than can be buffered prior to transmitting on
1260 	 * the connection.
1261 	 */
1262 	if (th->th_win == 0) {
1263 		tp->t_sndzerowin++;
1264 		tp->t_flags |= TF_RXWIN0SENT;
1265 	} else
1266 		tp->t_flags &= ~TF_RXWIN0SENT;
1267 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1268 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1269 		th->th_flags |= TH_URG;
1270 	} else
1271 		/*
1272 		 * If no urgent pointer to send, then we pull
1273 		 * the urgent pointer to the left edge of the send window
1274 		 * so that it doesn't drift into the send window on sequence
1275 		 * number wraparound.
1276 		 */
1277 		tp->snd_up = tp->snd_una;		/* drag it along */
1278 
1279 	/*
1280 	 * Put TCP length in extended header, and then
1281 	 * checksum extended header and data.
1282 	 */
1283 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1284 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1285 
1286 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1287 	if (to.to_flags & TOF_SIGNATURE) {
1288 		/*
1289 		 * Calculate MD5 signature and put it into the place
1290 		 * determined before.
1291 		 * NOTE: since TCP options buffer doesn't point into
1292 		 * mbuf's data, calculate offset and use it.
1293 		 */
1294 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1295 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1296 			/*
1297 			 * Do not send segment if the calculation of MD5
1298 			 * digest has failed.
1299 			 */
1300 			m_freem(m);
1301 			goto out;
1302 		}
1303 	}
1304 #endif
1305 #ifdef INET6
1306 	if (isipv6) {
1307 		/*
1308 		 * There is no need to fill in ip6_plen right now.
1309 		 * It will be filled later by ip6_output.
1310 		 */
1311 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1312 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1313 		    optlen + len, IPPROTO_TCP, 0);
1314 	}
1315 #endif
1316 #if defined(INET6) && defined(INET)
1317 	else
1318 #endif
1319 #ifdef INET
1320 	{
1321 		m->m_pkthdr.csum_flags = CSUM_TCP;
1322 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1323 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1324 
1325 		/* IP version must be set here for ipv4/ipv6 checking later */
1326 		KASSERT(ip->ip_v == IPVERSION,
1327 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1328 	}
1329 #endif
1330 
1331 	/*
1332 	 * Enable TSO and specify the size of the segments.
1333 	 * The TCP pseudo header checksum is always provided.
1334 	 */
1335 	if (tso) {
1336 		KASSERT(len > tp->t_maxseg - optlen,
1337 		    ("%s: len <= tso_segsz", __func__));
1338 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1339 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1340 	}
1341 
1342 	KASSERT(len + hdrlen == m_length(m, NULL),
1343 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1344 	    __func__, len, hdrlen, m_length(m, NULL)));
1345 
1346 #ifdef TCP_HHOOK
1347 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1348 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1349 #endif
1350 
1351 #ifdef TCPDEBUG
1352 	/*
1353 	 * Trace.
1354 	 */
1355 	if (so->so_options & SO_DEBUG) {
1356 		u_short save = 0;
1357 #ifdef INET6
1358 		if (!isipv6)
1359 #endif
1360 		{
1361 			save = ipov->ih_len;
1362 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1363 		}
1364 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1365 #ifdef INET6
1366 		if (!isipv6)
1367 #endif
1368 		ipov->ih_len = save;
1369 	}
1370 #endif /* TCPDEBUG */
1371 	TCP_PROBE3(debug__output, tp, th, m);
1372 
1373 	/* We're getting ready to send; log now. */
1374 	TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1375 	    len, NULL, false);
1376 
1377 	/*
1378 	 * Fill in IP length and desired time to live and
1379 	 * send to IP level.  There should be a better way
1380 	 * to handle ttl and tos; we could keep them in
1381 	 * the template, but need a way to checksum without them.
1382 	 */
1383 	/*
1384 	 * m->m_pkthdr.len should have been set before checksum calculation,
1385 	 * because in6_cksum() need it.
1386 	 */
1387 #ifdef INET6
1388 	if (isipv6) {
1389 		/*
1390 		 * we separately set hoplimit for every segment, since the
1391 		 * user might want to change the value via setsockopt.
1392 		 * Also, desired default hop limit might be changed via
1393 		 * Neighbor Discovery.
1394 		 */
1395 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1396 
1397 		/*
1398 		 * Set the packet size here for the benefit of DTrace probes.
1399 		 * ip6_output() will set it properly; it's supposed to include
1400 		 * the option header lengths as well.
1401 		 */
1402 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1403 
1404 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1405 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1406 		else
1407 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1408 
1409 		if (tp->t_state == TCPS_SYN_SENT)
1410 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1411 
1412 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1413 
1414 #ifdef TCPPCAP
1415 		/* Save packet, if requested. */
1416 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1417 #endif
1418 
1419 		/* TODO: IPv6 IP6TOS_ECT bit on */
1420 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1421 		    &tp->t_inpcb->inp_route6,
1422 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1423 		    NULL, NULL, tp->t_inpcb);
1424 
1425 		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1426 			mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1427 	}
1428 #endif /* INET6 */
1429 #if defined(INET) && defined(INET6)
1430 	else
1431 #endif
1432 #ifdef INET
1433     {
1434 	ip->ip_len = htons(m->m_pkthdr.len);
1435 #ifdef INET6
1436 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1437 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1438 #endif /* INET6 */
1439 	/*
1440 	 * If we do path MTU discovery, then we set DF on every packet.
1441 	 * This might not be the best thing to do according to RFC3390
1442 	 * Section 2. However the tcp hostcache migitates the problem
1443 	 * so it affects only the first tcp connection with a host.
1444 	 *
1445 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1446 	 */
1447 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1448 		ip->ip_off |= htons(IP_DF);
1449 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1450 	} else {
1451 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1452 	}
1453 
1454 	if (tp->t_state == TCPS_SYN_SENT)
1455 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1456 
1457 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1458 
1459 #ifdef TCPPCAP
1460 	/* Save packet, if requested. */
1461 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1462 #endif
1463 
1464 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1465 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1466 	    tp->t_inpcb);
1467 
1468 	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1469 		mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1470     }
1471 #endif /* INET */
1472 
1473 out:
1474 	/*
1475 	 * In transmit state, time the transmission and arrange for
1476 	 * the retransmit.  In persist state, just set snd_max.
1477 	 */
1478 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1479 	    !tcp_timer_active(tp, TT_PERSIST)) {
1480 		tcp_seq startseq = tp->snd_nxt;
1481 
1482 		/*
1483 		 * Advance snd_nxt over sequence space of this segment.
1484 		 */
1485 		if (flags & (TH_SYN|TH_FIN)) {
1486 			if (flags & TH_SYN)
1487 				tp->snd_nxt++;
1488 			if (flags & TH_FIN) {
1489 				tp->snd_nxt++;
1490 				tp->t_flags |= TF_SENTFIN;
1491 			}
1492 		}
1493 		if (sack_rxmit)
1494 			goto timer;
1495 		tp->snd_nxt += len;
1496 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1497 			tp->snd_max = tp->snd_nxt;
1498 			/*
1499 			 * Time this transmission if not a retransmission and
1500 			 * not currently timing anything.
1501 			 */
1502 			if (tp->t_rtttime == 0) {
1503 				tp->t_rtttime = ticks;
1504 				tp->t_rtseq = startseq;
1505 				TCPSTAT_INC(tcps_segstimed);
1506 			}
1507 #ifdef STATS
1508 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1509 				tp->t_flags |= TF_GPUTINPROG;
1510 				tp->gput_seq = startseq;
1511 				tp->gput_ack = startseq +
1512 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1513 				tp->gput_ts = tcp_ts_getticks();
1514 			}
1515 #endif /* STATS */
1516 		}
1517 
1518 		/*
1519 		 * Set retransmit timer if not currently set,
1520 		 * and not doing a pure ack or a keep-alive probe.
1521 		 * Initial value for retransmit timer is smoothed
1522 		 * round-trip time + 2 * round-trip time variance.
1523 		 * Initialize shift counter which is used for backoff
1524 		 * of retransmit time.
1525 		 */
1526 timer:
1527 		if (!tcp_timer_active(tp, TT_REXMT) &&
1528 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1529 		     (tp->snd_nxt != tp->snd_una))) {
1530 			if (tcp_timer_active(tp, TT_PERSIST)) {
1531 				tcp_timer_activate(tp, TT_PERSIST, 0);
1532 				tp->t_rxtshift = 0;
1533 			}
1534 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1535 		} else if (len == 0 && sbavail(&so->so_snd) &&
1536 		    !tcp_timer_active(tp, TT_REXMT) &&
1537 		    !tcp_timer_active(tp, TT_PERSIST)) {
1538 			/*
1539 			 * Avoid a situation where we do not set persist timer
1540 			 * after a zero window condition. For example:
1541 			 * 1) A -> B: packet with enough data to fill the window
1542 			 * 2) B -> A: ACK for #1 + new data (0 window
1543 			 *    advertisement)
1544 			 * 3) A -> B: ACK for #2, 0 len packet
1545 			 *
1546 			 * In this case, A will not activate the persist timer,
1547 			 * because it chose to send a packet. Unless tcp_output
1548 			 * is called for some other reason (delayed ack timer,
1549 			 * another input packet from B, socket syscall), A will
1550 			 * not send zero window probes.
1551 			 *
1552 			 * So, if you send a 0-length packet, but there is data
1553 			 * in the socket buffer, and neither the rexmt or
1554 			 * persist timer is already set, then activate the
1555 			 * persist timer.
1556 			 */
1557 			tp->t_rxtshift = 0;
1558 			tcp_setpersist(tp);
1559 		}
1560 	} else {
1561 		/*
1562 		 * Persist case, update snd_max but since we are in
1563 		 * persist mode (no window) we do not update snd_nxt.
1564 		 */
1565 		int xlen = len;
1566 		if (flags & TH_SYN)
1567 			++xlen;
1568 		if (flags & TH_FIN) {
1569 			++xlen;
1570 			tp->t_flags |= TF_SENTFIN;
1571 		}
1572 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1573 			tp->snd_max = tp->snd_nxt + xlen;
1574 	}
1575 	if ((error == 0) &&
1576 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1577 	     (tp->t_flags & TF_SACK_PERMIT) &&
1578 	     tp->rcv_numsacks > 0)) {
1579 		    /* Clean up any DSACK's sent */
1580 		    tcp_clean_dsack_blocks(tp);
1581 	}
1582 	if (error) {
1583 		/* Record the error. */
1584 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1585 		    error, 0, NULL, false);
1586 
1587 		/*
1588 		 * We know that the packet was lost, so back out the
1589 		 * sequence number advance, if any.
1590 		 *
1591 		 * If the error is EPERM the packet got blocked by the
1592 		 * local firewall.  Normally we should terminate the
1593 		 * connection but the blocking may have been spurious
1594 		 * due to a firewall reconfiguration cycle.  So we treat
1595 		 * it like a packet loss and let the retransmit timer and
1596 		 * timeouts do their work over time.
1597 		 * XXX: It is a POLA question whether calling tcp_drop right
1598 		 * away would be the really correct behavior instead.
1599 		 */
1600 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1601 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1602 		    ((flags & TH_SYN) == 0) &&
1603 		    (error != EPERM)) {
1604 			if (sack_rxmit) {
1605 				p->rxmit -= len;
1606 				tp->sackhint.sack_bytes_rexmit -= len;
1607 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1608 				    ("sackhint bytes rtx >= 0"));
1609 			} else
1610 				tp->snd_nxt -= len;
1611 		}
1612 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1613 		switch (error) {
1614 		case EACCES:
1615 		case EPERM:
1616 			tp->t_softerror = error;
1617 			return (error);
1618 		case ENOBUFS:
1619 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1620 			tp->snd_cwnd = tp->t_maxseg;
1621 			return (0);
1622 		case EMSGSIZE:
1623 			/*
1624 			 * For some reason the interface we used initially
1625 			 * to send segments changed to another or lowered
1626 			 * its MTU.
1627 			 * If TSO was active we either got an interface
1628 			 * without TSO capabilits or TSO was turned off.
1629 			 * If we obtained mtu from ip_output() then update
1630 			 * it and try again.
1631 			 */
1632 			if (tso)
1633 				tp->t_flags &= ~TF_TSO;
1634 			if (mtu != 0) {
1635 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1636 				goto again;
1637 			}
1638 			return (error);
1639 		case EHOSTDOWN:
1640 		case EHOSTUNREACH:
1641 		case ENETDOWN:
1642 		case ENETUNREACH:
1643 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1644 				tp->t_softerror = error;
1645 				return (0);
1646 			}
1647 			/* FALLTHROUGH */
1648 		default:
1649 			return (error);
1650 		}
1651 	}
1652 	TCPSTAT_INC(tcps_sndtotal);
1653 
1654 	/*
1655 	 * Data sent (as far as we can tell).
1656 	 * If this advertises a larger window than any other segment,
1657 	 * then remember the size of the advertised window.
1658 	 * Any pending ACK has now been sent.
1659 	 */
1660 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1661 		tp->rcv_adv = tp->rcv_nxt + recwin;
1662 	tp->last_ack_sent = tp->rcv_nxt;
1663 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1664 	if (tcp_timer_active(tp, TT_DELACK))
1665 		tcp_timer_activate(tp, TT_DELACK, 0);
1666 #if 0
1667 	/*
1668 	 * This completely breaks TCP if newreno is turned on.  What happens
1669 	 * is that if delayed-acks are turned on on the receiver, this code
1670 	 * on the transmitter effectively destroys the TCP window, forcing
1671 	 * it to four packets (1.5Kx4 = 6K window).
1672 	 */
1673 	if (sendalot && --maxburst)
1674 		goto again;
1675 #endif
1676 	if (sendalot)
1677 		goto again;
1678 	return (0);
1679 }
1680 
1681 void
1682 tcp_setpersist(struct tcpcb *tp)
1683 {
1684 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1685 	int tt;
1686 
1687 	tp->t_flags &= ~TF_PREVVALID;
1688 	if (tcp_timer_active(tp, TT_REXMT))
1689 		panic("tcp_setpersist: retransmit pending");
1690 	/*
1691 	 * Start/restart persistence timer.
1692 	 */
1693 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1694 		      tcp_persmin, tcp_persmax);
1695 	tcp_timer_activate(tp, TT_PERSIST, tt);
1696 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1697 		tp->t_rxtshift++;
1698 }
1699 
1700 /*
1701  * Insert TCP options according to the supplied parameters to the place
1702  * optp in a consistent way.  Can handle unaligned destinations.
1703  *
1704  * The order of the option processing is crucial for optimal packing and
1705  * alignment for the scarce option space.
1706  *
1707  * The optimal order for a SYN/SYN-ACK segment is:
1708  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1709  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1710  *
1711  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1712  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1713  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1714  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1715  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1716  */
1717 int
1718 tcp_addoptions(struct tcpopt *to, u_char *optp)
1719 {
1720 	u_int32_t mask, optlen = 0;
1721 
1722 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1723 		if ((to->to_flags & mask) != mask)
1724 			continue;
1725 		if (optlen == TCP_MAXOLEN)
1726 			break;
1727 		switch (to->to_flags & mask) {
1728 		case TOF_MSS:
1729 			while (optlen % 4) {
1730 				optlen += TCPOLEN_NOP;
1731 				*optp++ = TCPOPT_NOP;
1732 			}
1733 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1734 				continue;
1735 			optlen += TCPOLEN_MAXSEG;
1736 			*optp++ = TCPOPT_MAXSEG;
1737 			*optp++ = TCPOLEN_MAXSEG;
1738 			to->to_mss = htons(to->to_mss);
1739 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1740 			optp += sizeof(to->to_mss);
1741 			break;
1742 		case TOF_SCALE:
1743 			while (!optlen || optlen % 2 != 1) {
1744 				optlen += TCPOLEN_NOP;
1745 				*optp++ = TCPOPT_NOP;
1746 			}
1747 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1748 				continue;
1749 			optlen += TCPOLEN_WINDOW;
1750 			*optp++ = TCPOPT_WINDOW;
1751 			*optp++ = TCPOLEN_WINDOW;
1752 			*optp++ = to->to_wscale;
1753 			break;
1754 		case TOF_SACKPERM:
1755 			while (optlen % 2) {
1756 				optlen += TCPOLEN_NOP;
1757 				*optp++ = TCPOPT_NOP;
1758 			}
1759 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1760 				continue;
1761 			optlen += TCPOLEN_SACK_PERMITTED;
1762 			*optp++ = TCPOPT_SACK_PERMITTED;
1763 			*optp++ = TCPOLEN_SACK_PERMITTED;
1764 			break;
1765 		case TOF_TS:
1766 			while (!optlen || optlen % 4 != 2) {
1767 				optlen += TCPOLEN_NOP;
1768 				*optp++ = TCPOPT_NOP;
1769 			}
1770 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1771 				continue;
1772 			optlen += TCPOLEN_TIMESTAMP;
1773 			*optp++ = TCPOPT_TIMESTAMP;
1774 			*optp++ = TCPOLEN_TIMESTAMP;
1775 			to->to_tsval = htonl(to->to_tsval);
1776 			to->to_tsecr = htonl(to->to_tsecr);
1777 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1778 			optp += sizeof(to->to_tsval);
1779 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1780 			optp += sizeof(to->to_tsecr);
1781 			break;
1782 		case TOF_SIGNATURE:
1783 			{
1784 			int siglen = TCPOLEN_SIGNATURE - 2;
1785 
1786 			while (!optlen || optlen % 4 != 2) {
1787 				optlen += TCPOLEN_NOP;
1788 				*optp++ = TCPOPT_NOP;
1789 			}
1790 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1791 				to->to_flags &= ~TOF_SIGNATURE;
1792 				continue;
1793 			}
1794 			optlen += TCPOLEN_SIGNATURE;
1795 			*optp++ = TCPOPT_SIGNATURE;
1796 			*optp++ = TCPOLEN_SIGNATURE;
1797 			to->to_signature = optp;
1798 			while (siglen--)
1799 				 *optp++ = 0;
1800 			break;
1801 			}
1802 		case TOF_SACK:
1803 			{
1804 			int sackblks = 0;
1805 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1806 			tcp_seq sack_seq;
1807 
1808 			while (!optlen || optlen % 4 != 2) {
1809 				optlen += TCPOLEN_NOP;
1810 				*optp++ = TCPOPT_NOP;
1811 			}
1812 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1813 				continue;
1814 			optlen += TCPOLEN_SACKHDR;
1815 			*optp++ = TCPOPT_SACK;
1816 			sackblks = min(to->to_nsacks,
1817 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1818 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1819 			while (sackblks--) {
1820 				sack_seq = htonl(sack->start);
1821 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1822 				optp += sizeof(sack_seq);
1823 				sack_seq = htonl(sack->end);
1824 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1825 				optp += sizeof(sack_seq);
1826 				optlen += TCPOLEN_SACK;
1827 				sack++;
1828 			}
1829 			TCPSTAT_INC(tcps_sack_send_blocks);
1830 			break;
1831 			}
1832 		case TOF_FASTOPEN:
1833 			{
1834 			int total_len;
1835 
1836 			/* XXX is there any point to aligning this option? */
1837 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1838 			if (TCP_MAXOLEN - optlen < total_len) {
1839 				to->to_flags &= ~TOF_FASTOPEN;
1840 				continue;
1841 			}
1842 			*optp++ = TCPOPT_FAST_OPEN;
1843 			*optp++ = total_len;
1844 			if (to->to_tfo_len > 0) {
1845 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1846 				optp += to->to_tfo_len;
1847 			}
1848 			optlen += total_len;
1849 			break;
1850 			}
1851 		default:
1852 			panic("%s: unknown TCP option type", __func__);
1853 			break;
1854 		}
1855 	}
1856 
1857 	/* Terminate and pad TCP options to a 4 byte boundary. */
1858 	if (optlen % 4) {
1859 		optlen += TCPOLEN_EOL;
1860 		*optp++ = TCPOPT_EOL;
1861 	}
1862 	/*
1863 	 * According to RFC 793 (STD0007):
1864 	 *   "The content of the header beyond the End-of-Option option
1865 	 *    must be header padding (i.e., zero)."
1866 	 *   and later: "The padding is composed of zeros."
1867 	 */
1868 	while (optlen % 4) {
1869 		optlen += TCPOLEN_PAD;
1870 		*optp++ = TCPOPT_PAD;
1871 	}
1872 
1873 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1874 	return (optlen);
1875 }
1876 
1877 /*
1878  * This is a copy of m_copym(), taking the TSO segment size/limit
1879  * constraints into account, and advancing the sndptr as it goes.
1880  */
1881 struct mbuf *
1882 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1883     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1884 {
1885 #ifdef KERN_TLS
1886 	struct ktls_session *tls, *ntls;
1887 	struct mbuf *start;
1888 #endif
1889 	struct mbuf *n, **np;
1890 	struct mbuf *top;
1891 	int32_t off = off0;
1892 	int32_t len = *plen;
1893 	int32_t fragsize;
1894 	int32_t len_cp = 0;
1895 	int32_t *pkthdrlen;
1896 	uint32_t mlen, frags;
1897 	bool copyhdr;
1898 
1899 
1900 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1901 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1902 	if (off == 0 && m->m_flags & M_PKTHDR)
1903 		copyhdr = true;
1904 	else
1905 		copyhdr = false;
1906 	while (off > 0) {
1907 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1908 		if (off < m->m_len)
1909 			break;
1910 		off -= m->m_len;
1911 		if ((sb) && (m == sb->sb_sndptr)) {
1912 			sb->sb_sndptroff += m->m_len;
1913 			sb->sb_sndptr = m->m_next;
1914 		}
1915 		m = m->m_next;
1916 	}
1917 	np = &top;
1918 	top = NULL;
1919 	pkthdrlen = NULL;
1920 #ifdef KERN_TLS
1921 	if (hw_tls && (m->m_flags & M_EXTPG))
1922 		tls = m->m_epg_tls;
1923 	else
1924 		tls = NULL;
1925 	start = m;
1926 #endif
1927 	while (len > 0) {
1928 		if (m == NULL) {
1929 			KASSERT(len == M_COPYALL,
1930 			    ("tcp_m_copym, length > size of mbuf chain"));
1931 			*plen = len_cp;
1932 			if (pkthdrlen != NULL)
1933 				*pkthdrlen = len_cp;
1934 			break;
1935 		}
1936 #ifdef KERN_TLS
1937 		if (hw_tls) {
1938 			if (m->m_flags & M_EXTPG)
1939 				ntls = m->m_epg_tls;
1940 			else
1941 				ntls = NULL;
1942 
1943 			/*
1944 			 * Avoid mixing TLS records with handshake
1945 			 * data or TLS records from different
1946 			 * sessions.
1947 			 */
1948 			if (tls != ntls) {
1949 				MPASS(m != start);
1950 				*plen = len_cp;
1951 				if (pkthdrlen != NULL)
1952 					*pkthdrlen = len_cp;
1953 				break;
1954 			}
1955 
1956 			/*
1957 			 * Don't end a send in the middle of a TLS
1958 			 * record if it spans multiple TLS records.
1959 			 */
1960 			if (tls != NULL && (m != start) && len < m->m_len) {
1961 				*plen = len_cp;
1962 				if (pkthdrlen != NULL)
1963 					*pkthdrlen = len_cp;
1964 				break;
1965 			}
1966 		}
1967 #endif
1968 		mlen = min(len, m->m_len - off);
1969 		if (seglimit) {
1970 			/*
1971 			 * For M_EXTPG mbufs, add 3 segments
1972 			 * + 1 in case we are crossing page boundaries
1973 			 * + 2 in case the TLS hdr/trailer are used
1974 			 * It is cheaper to just add the segments
1975 			 * than it is to take the cache miss to look
1976 			 * at the mbuf ext_pgs state in detail.
1977 			 */
1978 			if (m->m_flags & M_EXTPG) {
1979 				fragsize = min(segsize, PAGE_SIZE);
1980 				frags = 3;
1981 			} else {
1982 				fragsize = segsize;
1983 				frags = 0;
1984 			}
1985 
1986 			/* Break if we really can't fit anymore. */
1987 			if ((frags + 1) >= seglimit) {
1988 				*plen =	len_cp;
1989 				if (pkthdrlen != NULL)
1990 					*pkthdrlen = len_cp;
1991 				break;
1992 			}
1993 
1994 			/*
1995 			 * Reduce size if you can't copy the whole
1996 			 * mbuf. If we can't copy the whole mbuf, also
1997 			 * adjust len so the loop will end after this
1998 			 * mbuf.
1999 			 */
2000 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2001 				mlen = (seglimit - frags - 1) * fragsize;
2002 				len = mlen;
2003 				*plen = len_cp + len;
2004 				if (pkthdrlen != NULL)
2005 					*pkthdrlen = *plen;
2006 			}
2007 			frags += howmany(mlen, fragsize);
2008 			if (frags == 0)
2009 				frags++;
2010 			seglimit -= frags;
2011 			KASSERT(seglimit > 0,
2012 			    ("%s: seglimit went too low", __func__));
2013 		}
2014 		if (copyhdr)
2015 			n = m_gethdr(M_NOWAIT, m->m_type);
2016 		else
2017 			n = m_get(M_NOWAIT, m->m_type);
2018 		*np = n;
2019 		if (n == NULL)
2020 			goto nospace;
2021 		if (copyhdr) {
2022 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2023 				goto nospace;
2024 			if (len == M_COPYALL)
2025 				n->m_pkthdr.len -= off0;
2026 			else
2027 				n->m_pkthdr.len = len;
2028 			pkthdrlen = &n->m_pkthdr.len;
2029 			copyhdr = false;
2030 		}
2031 		n->m_len = mlen;
2032 		len_cp += n->m_len;
2033 		if (m->m_flags & (M_EXT|M_EXTPG)) {
2034 			n->m_data = m->m_data + off;
2035 			mb_dupcl(n, m);
2036 		} else
2037 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2038 			    (u_int)n->m_len);
2039 
2040 		if (sb && (sb->sb_sndptr == m) &&
2041 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2042 			sb->sb_sndptroff += m->m_len;
2043 			sb->sb_sndptr = m->m_next;
2044 		}
2045 		off = 0;
2046 		if (len != M_COPYALL) {
2047 			len -= n->m_len;
2048 		}
2049 		m = m->m_next;
2050 		np = &n->m_next;
2051 	}
2052 	return (top);
2053 nospace:
2054 	m_freem(top);
2055 	return (NULL);
2056 }
2057 
2058 void
2059 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2060 {
2061 
2062 	/*
2063 	 * Automatic sizing of send socket buffer.  Often the send buffer
2064 	 * size is not optimally adjusted to the actual network conditions
2065 	 * at hand (delay bandwidth product).  Setting the buffer size too
2066 	 * small limits throughput on links with high bandwidth and high
2067 	 * delay (eg. trans-continental/oceanic links).  Setting the
2068 	 * buffer size too big consumes too much real kernel memory,
2069 	 * especially with many connections on busy servers.
2070 	 *
2071 	 * The criteria to step up the send buffer one notch are:
2072 	 *  1. receive window of remote host is larger than send buffer
2073 	 *     (with a fudge factor of 5/4th);
2074 	 *  2. send buffer is filled to 7/8th with data (so we actually
2075 	 *     have data to make use of it);
2076 	 *  3. send buffer fill has not hit maximal automatic size;
2077 	 *  4. our send window (slow start and cogestion controlled) is
2078 	 *     larger than sent but unacknowledged data in send buffer.
2079 	 *
2080 	 * The remote host receive window scaling factor may limit the
2081 	 * growing of the send buffer before it reaches its allowed
2082 	 * maximum.
2083 	 *
2084 	 * It scales directly with slow start or congestion window
2085 	 * and does at most one step per received ACK.  This fast
2086 	 * scaling has the drawback of growing the send buffer beyond
2087 	 * what is strictly necessary to make full use of a given
2088 	 * delay*bandwidth product.  However testing has shown this not
2089 	 * to be much of an problem.  At worst we are trading wasting
2090 	 * of available bandwidth (the non-use of it) for wasting some
2091 	 * socket buffer memory.
2092 	 *
2093 	 * TODO: Shrink send buffer during idle periods together
2094 	 * with congestion window.  Requires another timer.  Has to
2095 	 * wait for upcoming tcp timer rewrite.
2096 	 *
2097 	 * XXXGL: should there be used sbused() or sbavail()?
2098 	 */
2099 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2100 		int lowat;
2101 
2102 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2103 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2104 		    sbused(&so->so_snd) >=
2105 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2106 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2107 		    sendwin >= (sbused(&so->so_snd) -
2108 		    (tp->snd_nxt - tp->snd_una))) {
2109 			if (!sbreserve_locked(&so->so_snd,
2110 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2111 			     V_tcp_autosndbuf_max), so, curthread))
2112 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2113 		}
2114 	}
2115 }
2116