xref: /freebsd/sys/netinet/tcp_output.c (revision fedf5b965fd623957a9b851418ba3301ac122e27)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_kern_tls.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/arb.h>
41 #include <sys/domain.h>
42 #ifdef TCP_HHOOK
43 #include <sys/hhook.h>
44 #endif
45 #include <sys/kernel.h>
46 #ifdef KERN_TLS
47 #include <sys/ktls.h>
48 #endif
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/qmath.h>
54 #include <sys/sdt.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/stats.h>
59 
60 #include <net/if.h>
61 #include <net/route.h>
62 #include <net/route/nhop.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_kdtrace.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_pcb.h>
70 #include <netinet/ip_var.h>
71 #include <netinet/ip_options.h>
72 #ifdef INET6
73 #include <netinet6/in6_pcb.h>
74 #include <netinet/ip6.h>
75 #include <netinet6/ip6_var.h>
76 #endif
77 #include <netinet/tcp.h>
78 #define	TCPOUTFLAGS
79 #include <netinet/tcp_fsm.h>
80 #include <netinet/tcp_seq.h>
81 #include <netinet/tcp_var.h>
82 #include <netinet/tcp_log_buf.h>
83 #include <netinet/tcp_syncache.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcpip.h>
86 #include <netinet/cc/cc.h>
87 #include <netinet/tcp_fastopen.h>
88 #ifdef TCPPCAP
89 #include <netinet/tcp_pcap.h>
90 #endif
91 #ifdef TCP_OFFLOAD
92 #include <netinet/tcp_offload.h>
93 #endif
94 #include <netinet/tcp_ecn.h>
95 
96 #include <netipsec/ipsec_support.h>
97 
98 #include <netinet/udp.h>
99 #include <netinet/udp_var.h>
100 #include <machine/in_cksum.h>
101 
102 #include <security/mac/mac_framework.h>
103 
104 VNET_DEFINE(int, path_mtu_discovery) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
106 	&VNET_NAME(path_mtu_discovery), 1,
107 	"Enable Path MTU Discovery");
108 
109 VNET_DEFINE(int, tcp_do_tso) = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
111 	&VNET_NAME(tcp_do_tso), 0,
112 	"Enable TCP Segmentation Offload");
113 
114 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
115 #define	V_tcp_sendspace	VNET(tcp_sendspace)
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
117 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
118 
119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
121 	&VNET_NAME(tcp_do_autosndbuf), 0,
122 	"Enable automatic send buffer sizing");
123 
124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
126 	&VNET_NAME(tcp_autosndbuf_inc), 0,
127 	"Incrementor step size of automatic send buffer");
128 
129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
131 	&VNET_NAME(tcp_autosndbuf_max), 0,
132 	"Max size of automatic send buffer");
133 
134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
135 #define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
137 	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
138 	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
139 
140 /*
141  * Make sure that either retransmit or persist timer is set for SYN, FIN and
142  * non-ACK.
143  */
144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
145 	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
146 	    tcp_timer_active((tp), TT_REXMT) ||				\
147 	    tcp_timer_active((tp), TT_PERSIST),				\
148 	    ("neither rexmt nor persist timer is set"))
149 
150 #ifdef TCP_HHOOK
151 /*
152  * Wrapper for the TCP established output helper hook.
153  */
154 void
155 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
156     struct tcpopt *to, uint32_t len, int tso)
157 {
158 	struct tcp_hhook_data hhook_data;
159 
160 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
161 		hhook_data.tp = tp;
162 		hhook_data.th = th;
163 		hhook_data.to = to;
164 		hhook_data.len = len;
165 		hhook_data.tso = tso;
166 
167 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
168 		    &tp->t_osd);
169 	}
170 }
171 #endif
172 
173 /*
174  * CC wrapper hook functions
175  */
176 void
177 cc_after_idle(struct tcpcb *tp)
178 {
179 	INP_WLOCK_ASSERT(tptoinpcb(tp));
180 
181 	if (CC_ALGO(tp)->after_idle != NULL)
182 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
183 }
184 
185 /*
186  * Tcp output routine: figure out what should be sent and send it.
187  */
188 int
189 tcp_default_output(struct tcpcb *tp)
190 {
191 	struct socket *so = tptosocket(tp);
192 	struct inpcb *inp = tptoinpcb(tp);
193 	int32_t len;
194 	uint32_t recwin, sendwin;
195 	uint16_t flags;
196 	int off, error = 0;	/* Keep compiler happy */
197 	u_int if_hw_tsomaxsegcount = 0;
198 	u_int if_hw_tsomaxsegsize = 0;
199 	struct mbuf *m;
200 	struct ip *ip = NULL;
201 	struct tcphdr *th;
202 	u_char opt[TCP_MAXOLEN];
203 	unsigned ipoptlen, optlen, hdrlen, ulen;
204 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
205 	unsigned ipsec_optlen = 0;
206 #endif
207 	int idle, sendalot, curticks;
208 	int sack_rxmit, sack_bytes_rxmt;
209 	struct sackhole *p;
210 	int tso, mtu;
211 	struct tcpopt to;
212 	struct udphdr *udp = NULL;
213 	struct tcp_log_buffer *lgb;
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 	const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
222 #endif
223 #ifdef KERN_TLS
224 	const bool hw_tls = tp->t_nic_ktls_xmit != 0;
225 #else
226 	const bool hw_tls = false;
227 #endif
228 
229 	NET_EPOCH_ASSERT();
230 	INP_WLOCK_ASSERT(inp);
231 
232 #ifdef TCP_OFFLOAD
233 	if (tp->t_flags & TF_TOE)
234 		return (tcp_offload_output(tp));
235 #endif
236 
237 	/*
238 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
239 	 * only allow the initial SYN or SYN|ACK and those sent
240 	 * by the retransmit timer.
241 	 */
242 	if (IS_FASTOPEN(tp->t_flags) &&
243 	    ((tp->t_state == TCPS_SYN_SENT) ||
244 	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
245 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
246 	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
247 		return (0);
248 
249 	/*
250 	 * Determine length of data that should be transmitted,
251 	 * and flags that will be used.
252 	 * If there is some data or critical controls (SYN, RST)
253 	 * to send, then transmit; otherwise, investigate further.
254 	 */
255 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
256 	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
257 	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
258 		cc_after_idle(tp);
259 	tp->t_flags &= ~TF_LASTIDLE;
260 	if (idle) {
261 		if (tp->t_flags & TF_MORETOCOME) {
262 			tp->t_flags |= TF_LASTIDLE;
263 			idle = 0;
264 		}
265 	}
266 again:
267 	/*
268 	 * If we've recently taken a timeout, snd_max will be greater than
269 	 * snd_nxt.  There may be SACK information that allows us to avoid
270 	 * resending already delivered data.  Adjust snd_nxt accordingly.
271 	 */
272 	if ((tp->t_flags & TF_SACK_PERMIT) &&
273 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
274 		tcp_sack_adjust(tp);
275 	sendalot = 0;
276 	tso = 0;
277 	mtu = 0;
278 	off = tp->snd_nxt - tp->snd_una;
279 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
280 
281 	flags = tcp_outflags[tp->t_state];
282 	/*
283 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
284 	 * to send out new data (when sendalot is 1), bypass this function.
285 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
286 	 * we're replacing a (future) new transmission with a retransmission
287 	 * now, and we previously incremented snd_cwnd in tcp_input().
288 	 */
289 	/*
290 	 * Still in sack recovery , reset rxmit flag to zero.
291 	 */
292 	sack_rxmit = 0;
293 	sack_bytes_rxmt = 0;
294 	len = 0;
295 	p = NULL;
296 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
297 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
298 		uint32_t cwin;
299 
300 		cwin =
301 		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
302 		/* Do not retransmit SACK segments beyond snd_recover */
303 		if (SEQ_GT(p->end, tp->snd_recover)) {
304 			/*
305 			 * (At least) part of sack hole extends beyond
306 			 * snd_recover. Check to see if we can rexmit data
307 			 * for this hole.
308 			 */
309 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
310 				/*
311 				 * Can't rexmit any more data for this hole.
312 				 * That data will be rexmitted in the next
313 				 * sack recovery episode, when snd_recover
314 				 * moves past p->rxmit.
315 				 */
316 				p = NULL;
317 				goto after_sack_rexmit;
318 			} else {
319 				/* Can rexmit part of the current hole */
320 				len = ((int32_t)ulmin(cwin,
321 				    SEQ_SUB(tp->snd_recover, p->rxmit)));
322 			}
323 		} else {
324 			len = ((int32_t)ulmin(cwin,
325 			    SEQ_SUB(p->end, p->rxmit)));
326 		}
327 		if (len > 0) {
328 			off = SEQ_SUB(p->rxmit, tp->snd_una);
329 			KASSERT(off >= 0,("%s: sack block to the left of una : %d",
330 			    __func__, off));
331 			sack_rxmit = 1;
332 			sendalot = 1;
333 			TCPSTAT_INC(tcps_sack_rexmits);
334 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
335 			    min(len, tcp_maxseg(tp)));
336 		}
337 	}
338 after_sack_rexmit:
339 	/*
340 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
341 	 * state flags.
342 	 */
343 	if (tp->t_flags & TF_NEEDFIN)
344 		flags |= TH_FIN;
345 	if (tp->t_flags & TF_NEEDSYN)
346 		flags |= TH_SYN;
347 
348 	SOCKBUF_LOCK(&so->so_snd);
349 	/*
350 	 * If in persist timeout with window of 0, send 1 byte.
351 	 * Otherwise, if window is small but nonzero
352 	 * and timer expired, we will send what we can
353 	 * and go to transmit state.
354 	 */
355 	if (tp->t_flags & TF_FORCEDATA) {
356 		if (sendwin == 0) {
357 			/*
358 			 * If we still have some data to send, then
359 			 * clear the FIN bit.  Usually this would
360 			 * happen below when it realizes that we
361 			 * aren't sending all the data.  However,
362 			 * if we have exactly 1 byte of unsent data,
363 			 * then it won't clear the FIN bit below,
364 			 * and if we are in persist state, we wind
365 			 * up sending the packet without recording
366 			 * that we sent the FIN bit.
367 			 *
368 			 * We can't just blindly clear the FIN bit,
369 			 * because if we don't have any more data
370 			 * to send then the probe will be the FIN
371 			 * itself.
372 			 */
373 			if (off < sbused(&so->so_snd))
374 				flags &= ~TH_FIN;
375 			sendwin = 1;
376 		} else {
377 			tcp_timer_activate(tp, TT_PERSIST, 0);
378 			tp->t_rxtshift = 0;
379 		}
380 	}
381 
382 	/*
383 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
384 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
385 	 * a negative length.  This can also occur when TCP opens up
386 	 * its congestion window while receiving additional duplicate
387 	 * acks after fast-retransmit because TCP will reset snd_nxt
388 	 * to snd_max after the fast-retransmit.
389 	 *
390 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
391 	 * be set to snd_una, the offset will be 0, and the length may
392 	 * wind up 0.
393 	 *
394 	 * If sack_rxmit is true we are retransmitting from the scoreboard
395 	 * in which case len is already set.
396 	 */
397 	if (sack_rxmit == 0) {
398 		if (sack_bytes_rxmt == 0)
399 			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
400 			    off);
401 		else {
402 			int32_t cwin;
403 
404 			/*
405 			 * We are inside of a SACK recovery episode and are
406 			 * sending new data, having retransmitted all the
407 			 * data possible in the scoreboard.
408 			 */
409 			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
410 			    off);
411 			/*
412 			 * Don't remove this (len > 0) check !
413 			 * We explicitly check for len > 0 here (although it
414 			 * isn't really necessary), to work around a gcc
415 			 * optimization issue - to force gcc to compute
416 			 * len above. Without this check, the computation
417 			 * of len is bungled by the optimizer.
418 			 */
419 			if (len > 0) {
420 				cwin = tp->snd_cwnd - imax(0, (int32_t)
421 					(tp->snd_nxt - tp->snd_recover)) -
422 					sack_bytes_rxmt;
423 				if (cwin < 0)
424 					cwin = 0;
425 				len = imin(len, cwin);
426 			}
427 		}
428 	}
429 
430 	/*
431 	 * Lop off SYN bit if it has already been sent.  However, if this
432 	 * is SYN-SENT state and if segment contains data and if we don't
433 	 * know that foreign host supports TAO, suppress sending segment.
434 	 */
435 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
436 		if (tp->t_state != TCPS_SYN_RECEIVED)
437 			flags &= ~TH_SYN;
438 		/*
439 		 * When sending additional segments following a TFO SYN|ACK,
440 		 * do not include the SYN bit.
441 		 */
442 		if (IS_FASTOPEN(tp->t_flags) &&
443 		    (tp->t_state == TCPS_SYN_RECEIVED))
444 			flags &= ~TH_SYN;
445 		off--, len++;
446 	}
447 
448 	/*
449 	 * Be careful not to send data and/or FIN on SYN segments.
450 	 * This measure is needed to prevent interoperability problems
451 	 * with not fully conformant TCP implementations.
452 	 */
453 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
454 		len = 0;
455 		flags &= ~TH_FIN;
456 	}
457 
458 	/*
459 	 * On TFO sockets, ensure no data is sent in the following cases:
460 	 *
461 	 *  - When retransmitting SYN|ACK on a passively-created socket
462 	 *
463 	 *  - When retransmitting SYN on an actively created socket
464 	 *
465 	 *  - When sending a zero-length cookie (cookie request) on an
466 	 *    actively created socket
467 	 *
468 	 *  - When the socket is in the CLOSED state (RST is being sent)
469 	 */
470 	if (IS_FASTOPEN(tp->t_flags) &&
471 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
472 	     ((tp->t_state == TCPS_SYN_SENT) &&
473 	      (tp->t_tfo_client_cookie_len == 0)) ||
474 	     (flags & TH_RST)))
475 		len = 0;
476 	if (len <= 0) {
477 		/*
478 		 * If FIN has been sent but not acked,
479 		 * but we haven't been called to retransmit,
480 		 * len will be < 0.  Otherwise, window shrank
481 		 * after we sent into it.  If window shrank to 0,
482 		 * cancel pending retransmit, pull snd_nxt back
483 		 * to (closed) window, and set the persist timer
484 		 * if it isn't already going.  If the window didn't
485 		 * close completely, just wait for an ACK.
486 		 *
487 		 * We also do a general check here to ensure that
488 		 * we will set the persist timer when we have data
489 		 * to send, but a 0-byte window. This makes sure
490 		 * the persist timer is set even if the packet
491 		 * hits one of the "goto send" lines below.
492 		 */
493 		len = 0;
494 		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
495 		    (off < (int) sbavail(&so->so_snd)) &&
496 		    !tcp_timer_active(tp, TT_PERSIST)) {
497 			tcp_timer_activate(tp, TT_REXMT, 0);
498 			tp->t_rxtshift = 0;
499 			tp->snd_nxt = tp->snd_una;
500 			if (!tcp_timer_active(tp, TT_PERSIST))
501 				tcp_setpersist(tp);
502 		}
503 	}
504 
505 	/* len will be >= 0 after this point. */
506 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
507 
508 	tcp_sndbuf_autoscale(tp, so, sendwin);
509 
510 	/*
511 	 * Decide if we can use TCP Segmentation Offloading (if supported by
512 	 * hardware).
513 	 *
514 	 * TSO may only be used if we are in a pure bulk sending state.  The
515 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
516 	 * IP options prevent using TSO.  With TSO the TCP header is the same
517 	 * (except for the sequence number) for all generated packets.  This
518 	 * makes it impossible to transmit any options which vary per generated
519 	 * segment or packet.
520 	 *
521 	 * IPv4 handling has a clear separation of ip options and ip header
522 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
523 	 * the right thing below to provide length of just ip options and thus
524 	 * checking for ipoptlen is enough to decide if ip options are present.
525 	 */
526 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
527 	/*
528 	 * Pre-calculate here as we save another lookup into the darknesses
529 	 * of IPsec that way and can actually decide if TSO is ok.
530 	 */
531 #ifdef INET6
532 	if (isipv6 && IPSEC_ENABLED(ipv6))
533 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
534 #ifdef INET
535 	else
536 #endif
537 #endif /* INET6 */
538 #ifdef INET
539 	if (IPSEC_ENABLED(ipv4))
540 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
541 #endif /* INET */
542 #endif /* IPSEC */
543 #ifdef INET6
544 	if (isipv6)
545 		ipoptlen = ip6_optlen(inp);
546 	else
547 #endif
548 	if (inp->inp_options)
549 		ipoptlen = inp->inp_options->m_len -
550 				offsetof(struct ipoption, ipopt_list);
551 	else
552 		ipoptlen = 0;
553 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
554 	ipoptlen += ipsec_optlen;
555 #endif
556 
557 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
558 	    (tp->t_port == 0) &&
559 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
560 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
561 	    ipoptlen == 0 && !(flags & TH_SYN))
562 		tso = 1;
563 
564 	if (sack_rxmit) {
565 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
566 			flags &= ~TH_FIN;
567 	} else {
568 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
569 		    sbused(&so->so_snd)))
570 			flags &= ~TH_FIN;
571 	}
572 
573 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
574 	    (long)TCP_MAXWIN << tp->rcv_scale);
575 
576 	/*
577 	 * Sender silly window avoidance.   We transmit under the following
578 	 * conditions when len is non-zero:
579 	 *
580 	 *	- We have a full segment (or more with TSO)
581 	 *	- This is the last buffer in a write()/send() and we are
582 	 *	  either idle or running NODELAY
583 	 *	- we've timed out (e.g. persist timer)
584 	 *	- we have more then 1/2 the maximum send window's worth of
585 	 *	  data (receiver may be limited the window size)
586 	 *	- we need to retransmit
587 	 */
588 	if (len) {
589 		if (len >= tp->t_maxseg)
590 			goto send;
591 		/*
592 		 * As the TCP header options are now
593 		 * considered when setting up the initial
594 		 * window, we would not send the last segment
595 		 * if we skip considering the option length here.
596 		 * Note: this may not work when tcp headers change
597 		 * very dynamically in the future.
598 		 */
599 		if ((((tp->t_flags & TF_SIGNATURE) ?
600 			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
601 		    ((tp->t_flags & TF_RCVD_TSTMP) ?
602 			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
603 		    len) >= tp->t_maxseg)
604 			goto send;
605 		/*
606 		 * NOTE! on localhost connections an 'ack' from the remote
607 		 * end may occur synchronously with the output and cause
608 		 * us to flush a buffer queued with moretocome.  XXX
609 		 *
610 		 * note: the len + off check is almost certainly unnecessary.
611 		 */
612 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
613 		    (idle || (tp->t_flags & TF_NODELAY)) &&
614 		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
615 		    (tp->t_flags & TF_NOPUSH) == 0) {
616 			goto send;
617 		}
618 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
619 			goto send;
620 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
621 			goto send;
622 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
623 			goto send;
624 		if (sack_rxmit)
625 			goto send;
626 	}
627 
628 	/*
629 	 * Sending of standalone window updates.
630 	 *
631 	 * Window updates are important when we close our window due to a
632 	 * full socket buffer and are opening it again after the application
633 	 * reads data from it.  Once the window has opened again and the
634 	 * remote end starts to send again the ACK clock takes over and
635 	 * provides the most current window information.
636 	 *
637 	 * We must avoid the silly window syndrome whereas every read
638 	 * from the receive buffer, no matter how small, causes a window
639 	 * update to be sent.  We also should avoid sending a flurry of
640 	 * window updates when the socket buffer had queued a lot of data
641 	 * and the application is doing small reads.
642 	 *
643 	 * Prevent a flurry of pointless window updates by only sending
644 	 * an update when we can increase the advertized window by more
645 	 * than 1/4th of the socket buffer capacity.  When the buffer is
646 	 * getting full or is very small be more aggressive and send an
647 	 * update whenever we can increase by two mss sized segments.
648 	 * In all other situations the ACK's to new incoming data will
649 	 * carry further window increases.
650 	 *
651 	 * Don't send an independent window update if a delayed
652 	 * ACK is pending (it will get piggy-backed on it) or the
653 	 * remote side already has done a half-close and won't send
654 	 * more data.
655 	 */
656 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
657 	    !(tp->t_flags & TF_DELACK) &&
658 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
659 		/*
660 		 * "adv" is the amount we could increase the window,
661 		 * taking into account that we are limited by
662 		 * TCP_MAXWIN << tp->rcv_scale.
663 		 */
664 		int32_t adv;
665 		int oldwin;
666 
667 		adv = recwin;
668 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
669 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
670 			if (adv > oldwin)
671 				adv -= oldwin;
672 			else
673 				adv = 0;
674 		} else
675 			oldwin = 0;
676 
677 		/*
678 		 * If the new window size ends up being the same as or less
679 		 * than the old size when it is scaled, then don't force
680 		 * a window update.
681 		 */
682 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
683 			goto dontupdate;
684 
685 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
686 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
687 		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
688 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
689 		     adv >= TCP_MAXWIN << tp->rcv_scale))
690 			goto send;
691 		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
692 			goto send;
693 	}
694 dontupdate:
695 
696 	/*
697 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
698 	 * is also a catch-all for the retransmit timer timeout case.
699 	 */
700 	if (tp->t_flags & TF_ACKNOW)
701 		goto send;
702 	if ((flags & TH_RST) ||
703 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
704 		goto send;
705 	if (SEQ_GT(tp->snd_up, tp->snd_una))
706 		goto send;
707 	/*
708 	 * If our state indicates that FIN should be sent
709 	 * and we have not yet done so, then we need to send.
710 	 */
711 	if (flags & TH_FIN &&
712 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
713 		goto send;
714 	/*
715 	 * In SACK, it is possible for tcp_output to fail to send a segment
716 	 * after the retransmission timer has been turned off.  Make sure
717 	 * that the retransmission timer is set.
718 	 */
719 	if ((tp->t_flags & TF_SACK_PERMIT) &&
720 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
721 	    !tcp_timer_active(tp, TT_REXMT) &&
722 	    !tcp_timer_active(tp, TT_PERSIST)) {
723 		tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
724 		goto just_return;
725 	}
726 	/*
727 	 * TCP window updates are not reliable, rather a polling protocol
728 	 * using ``persist'' packets is used to insure receipt of window
729 	 * updates.  The three ``states'' for the output side are:
730 	 *	idle			not doing retransmits or persists
731 	 *	persisting		to move a small or zero window
732 	 *	(re)transmitting	and thereby not persisting
733 	 *
734 	 * tcp_timer_active(tp, TT_PERSIST)
735 	 *	is true when we are in persist state.
736 	 * (tp->t_flags & TF_FORCEDATA)
737 	 *	is set when we are called to send a persist packet.
738 	 * tcp_timer_active(tp, TT_REXMT)
739 	 *	is set when we are retransmitting
740 	 * The output side is idle when both timers are zero.
741 	 *
742 	 * If send window is too small, there is data to transmit, and no
743 	 * retransmit or persist is pending, then go to persist state.
744 	 * If nothing happens soon, send when timer expires:
745 	 * if window is nonzero, transmit what we can,
746 	 * otherwise force out a byte.
747 	 */
748 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
749 	    !tcp_timer_active(tp, TT_PERSIST)) {
750 		tp->t_rxtshift = 0;
751 		tcp_setpersist(tp);
752 	}
753 
754 	/*
755 	 * No reason to send a segment, just return.
756 	 */
757 just_return:
758 	SOCKBUF_UNLOCK(&so->so_snd);
759 	return (0);
760 
761 send:
762 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
763 	if (len > 0) {
764 		if (len >= tp->t_maxseg)
765 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
766 		else
767 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
768 	}
769 	/*
770 	 * Before ESTABLISHED, force sending of initial options
771 	 * unless TCP set not to do any options.
772 	 * NOTE: we assume that the IP/TCP header plus TCP options
773 	 * always fit in a single mbuf, leaving room for a maximum
774 	 * link header, i.e.
775 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
776 	 */
777 	optlen = 0;
778 #ifdef INET6
779 	if (isipv6)
780 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
781 	else
782 #endif
783 		hdrlen = sizeof (struct tcpiphdr);
784 
785 	if (flags & TH_SYN) {
786 		tp->snd_nxt = tp->iss;
787 	}
788 
789 	/*
790 	 * Compute options for segment.
791 	 * We only have to care about SYN and established connection
792 	 * segments.  Options for SYN-ACK segments are handled in TCP
793 	 * syncache.
794 	 */
795 	to.to_flags = 0;
796 	if ((tp->t_flags & TF_NOOPT) == 0) {
797 		/* Maximum segment size. */
798 		if (flags & TH_SYN) {
799 			to.to_mss = tcp_mssopt(&inp->inp_inc);
800 			if (tp->t_port)
801 				to.to_mss -= V_tcp_udp_tunneling_overhead;
802 			to.to_flags |= TOF_MSS;
803 
804 			/*
805 			 * On SYN or SYN|ACK transmits on TFO connections,
806 			 * only include the TFO option if it is not a
807 			 * retransmit, as the presence of the TFO option may
808 			 * have caused the original SYN or SYN|ACK to have
809 			 * been dropped by a middlebox.
810 			 */
811 			if (IS_FASTOPEN(tp->t_flags) &&
812 			    (tp->t_rxtshift == 0)) {
813 				if (tp->t_state == TCPS_SYN_RECEIVED) {
814 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
815 					to.to_tfo_cookie =
816 					    (u_int8_t *)&tp->t_tfo_cookie.server;
817 					to.to_flags |= TOF_FASTOPEN;
818 					wanted_cookie = 1;
819 				} else if (tp->t_state == TCPS_SYN_SENT) {
820 					to.to_tfo_len =
821 					    tp->t_tfo_client_cookie_len;
822 					to.to_tfo_cookie =
823 					    tp->t_tfo_cookie.client;
824 					to.to_flags |= TOF_FASTOPEN;
825 					wanted_cookie = 1;
826 					/*
827 					 * If we wind up having more data to
828 					 * send with the SYN than can fit in
829 					 * one segment, don't send any more
830 					 * until the SYN|ACK comes back from
831 					 * the other end.
832 					 */
833 					dont_sendalot = 1;
834 				}
835 			}
836 		}
837 		/* Window scaling. */
838 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
839 			to.to_wscale = tp->request_r_scale;
840 			to.to_flags |= TOF_SCALE;
841 		}
842 		/* Timestamps. */
843 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
844 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
845 			curticks = tcp_ts_getticks();
846 			to.to_tsval = curticks + tp->ts_offset;
847 			to.to_tsecr = tp->ts_recent;
848 			to.to_flags |= TOF_TS;
849 			if (tp->t_rxtshift == 1)
850 				tp->t_badrxtwin = curticks;
851 		}
852 
853 		/* Set receive buffer autosizing timestamp. */
854 		if (tp->rfbuf_ts == 0 &&
855 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
856 			tp->rfbuf_ts = tcp_ts_getticks();
857 
858 		/* Selective ACK's. */
859 		if (tp->t_flags & TF_SACK_PERMIT) {
860 			if (flags & TH_SYN)
861 				to.to_flags |= TOF_SACKPERM;
862 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
863 			    tp->rcv_numsacks > 0) {
864 				to.to_flags |= TOF_SACK;
865 				to.to_nsacks = tp->rcv_numsacks;
866 				to.to_sacks = (u_char *)tp->sackblks;
867 			}
868 		}
869 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
870 		/* TCP-MD5 (RFC2385). */
871 		/*
872 		 * Check that TCP_MD5SIG is enabled in tcpcb to
873 		 * account the size needed to set this TCP option.
874 		 */
875 		if (tp->t_flags & TF_SIGNATURE)
876 			to.to_flags |= TOF_SIGNATURE;
877 #endif /* TCP_SIGNATURE */
878 
879 		/* Processing the options. */
880 		hdrlen += optlen = tcp_addoptions(&to, opt);
881 		/*
882 		 * If we wanted a TFO option to be added, but it was unable
883 		 * to fit, ensure no data is sent.
884 		 */
885 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
886 		    !(to.to_flags & TOF_FASTOPEN))
887 			len = 0;
888 	}
889 	if (tp->t_port) {
890 		if (V_tcp_udp_tunneling_port == 0) {
891 			/* The port was removed?? */
892 			SOCKBUF_UNLOCK(&so->so_snd);
893 			return (EHOSTUNREACH);
894 		}
895 		hdrlen += sizeof(struct udphdr);
896 	}
897 	/*
898 	 * Adjust data length if insertion of options will
899 	 * bump the packet length beyond the t_maxseg length.
900 	 * Clear the FIN bit because we cut off the tail of
901 	 * the segment.
902 	 */
903 	if (len + optlen + ipoptlen > tp->t_maxseg) {
904 		flags &= ~TH_FIN;
905 
906 		if (tso) {
907 			u_int if_hw_tsomax;
908 			u_int moff;
909 			int max_len;
910 
911 			/* extract TSO information */
912 			if_hw_tsomax = tp->t_tsomax;
913 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
914 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
915 
916 			/*
917 			 * Limit a TSO burst to prevent it from
918 			 * overflowing or exceeding the maximum length
919 			 * allowed by the network interface:
920 			 */
921 			KASSERT(ipoptlen == 0,
922 			    ("%s: TSO can't do IP options", __func__));
923 
924 			/*
925 			 * Check if we should limit by maximum payload
926 			 * length:
927 			 */
928 			if (if_hw_tsomax != 0) {
929 				/* compute maximum TSO length */
930 				max_len = (if_hw_tsomax - hdrlen -
931 				    max_linkhdr);
932 				if (max_len <= 0) {
933 					len = 0;
934 				} else if (len > max_len) {
935 					sendalot = 1;
936 					len = max_len;
937 				}
938 			}
939 
940 			/*
941 			 * Prevent the last segment from being
942 			 * fractional unless the send sockbuf can be
943 			 * emptied:
944 			 */
945 			max_len = (tp->t_maxseg - optlen);
946 			if (((uint32_t)off + (uint32_t)len) <
947 			    sbavail(&so->so_snd)) {
948 				moff = len % max_len;
949 				if (moff != 0) {
950 					len -= moff;
951 					sendalot = 1;
952 				}
953 			}
954 
955 			/*
956 			 * In case there are too many small fragments
957 			 * don't use TSO:
958 			 */
959 			if (len <= max_len) {
960 				len = max_len;
961 				sendalot = 1;
962 				tso = 0;
963 			}
964 
965 			/*
966 			 * Send the FIN in a separate segment
967 			 * after the bulk sending is done.
968 			 * We don't trust the TSO implementations
969 			 * to clear the FIN flag on all but the
970 			 * last segment.
971 			 */
972 			if (tp->t_flags & TF_NEEDFIN)
973 				sendalot = 1;
974 		} else {
975 			if (optlen + ipoptlen >= tp->t_maxseg) {
976 				/*
977 				 * Since we don't have enough space to put
978 				 * the IP header chain and the TCP header in
979 				 * one packet as required by RFC 7112, don't
980 				 * send it. Also ensure that at least one
981 				 * byte of the payload can be put into the
982 				 * TCP segment.
983 				 */
984 				SOCKBUF_UNLOCK(&so->so_snd);
985 				error = EMSGSIZE;
986 				sack_rxmit = 0;
987 				goto out;
988 			}
989 			len = tp->t_maxseg - optlen - ipoptlen;
990 			sendalot = 1;
991 			if (dont_sendalot)
992 				sendalot = 0;
993 		}
994 	} else
995 		tso = 0;
996 
997 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
998 	    ("%s: len > IP_MAXPACKET", __func__));
999 
1000 /*#ifdef DIAGNOSTIC*/
1001 #ifdef INET6
1002 	if (max_linkhdr + hdrlen > MCLBYTES)
1003 #else
1004 	if (max_linkhdr + hdrlen > MHLEN)
1005 #endif
1006 		panic("tcphdr too big");
1007 /*#endif*/
1008 
1009 	/*
1010 	 * This KASSERT is here to catch edge cases at a well defined place.
1011 	 * Before, those had triggered (random) panic conditions further down.
1012 	 */
1013 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1014 
1015 	/*
1016 	 * Grab a header mbuf, attaching a copy of data to
1017 	 * be transmitted, and initialize the header from
1018 	 * the template for sends on this connection.
1019 	 */
1020 	if (len) {
1021 		struct mbuf *mb;
1022 		struct sockbuf *msb;
1023 		u_int moff;
1024 
1025 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1026 			TCPSTAT_INC(tcps_sndprobe);
1027 #ifdef STATS
1028 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1029 				stats_voi_update_abs_u32(tp->t_stats,
1030 				VOI_TCP_RETXPB, len);
1031 			else
1032 				stats_voi_update_abs_u64(tp->t_stats,
1033 				    VOI_TCP_TXPB, len);
1034 #endif /* STATS */
1035 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1036 			tp->t_sndrexmitpack++;
1037 			TCPSTAT_INC(tcps_sndrexmitpack);
1038 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1039 #ifdef STATS
1040 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1041 			    len);
1042 #endif /* STATS */
1043 		} else {
1044 			TCPSTAT_INC(tcps_sndpack);
1045 			TCPSTAT_ADD(tcps_sndbyte, len);
1046 #ifdef STATS
1047 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1048 			    len);
1049 #endif /* STATS */
1050 		}
1051 #ifdef INET6
1052 		if (MHLEN < hdrlen + max_linkhdr)
1053 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1054 		else
1055 #endif
1056 			m = m_gethdr(M_NOWAIT, MT_DATA);
1057 
1058 		if (m == NULL) {
1059 			SOCKBUF_UNLOCK(&so->so_snd);
1060 			error = ENOBUFS;
1061 			sack_rxmit = 0;
1062 			goto out;
1063 		}
1064 
1065 		m->m_data += max_linkhdr;
1066 		m->m_len = hdrlen;
1067 
1068 		/*
1069 		 * Start the m_copy functions from the closest mbuf
1070 		 * to the offset in the socket buffer chain.
1071 		 */
1072 		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1073 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1074 			m_copydata(mb, moff, len,
1075 			    mtod(m, caddr_t) + hdrlen);
1076 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1077 				sbsndptr_adv(&so->so_snd, mb, len);
1078 			m->m_len += len;
1079 		} else {
1080 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1081 				msb = NULL;
1082 			else
1083 				msb = &so->so_snd;
1084 			m->m_next = tcp_m_copym(mb, moff,
1085 			    &len, if_hw_tsomaxsegcount,
1086 			    if_hw_tsomaxsegsize, msb, hw_tls);
1087 			if (len <= (tp->t_maxseg - optlen)) {
1088 				/*
1089 				 * Must have ran out of mbufs for the copy
1090 				 * shorten it to no longer need tso. Lets
1091 				 * not put on sendalot since we are low on
1092 				 * mbufs.
1093 				 */
1094 				tso = 0;
1095 			}
1096 			if (m->m_next == NULL) {
1097 				SOCKBUF_UNLOCK(&so->so_snd);
1098 				(void) m_free(m);
1099 				error = ENOBUFS;
1100 				sack_rxmit = 0;
1101 				goto out;
1102 			}
1103 		}
1104 
1105 		/*
1106 		 * If we're sending everything we've got, set PUSH.
1107 		 * (This will keep happy those implementations which only
1108 		 * give data to the user when a buffer fills or
1109 		 * a PUSH comes in.)
1110 		 */
1111 		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1112 		    !(flags & TH_SYN))
1113 			flags |= TH_PUSH;
1114 		SOCKBUF_UNLOCK(&so->so_snd);
1115 	} else {
1116 		SOCKBUF_UNLOCK(&so->so_snd);
1117 		if (tp->t_flags & TF_ACKNOW)
1118 			TCPSTAT_INC(tcps_sndacks);
1119 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1120 			TCPSTAT_INC(tcps_sndctrl);
1121 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1122 			TCPSTAT_INC(tcps_sndurg);
1123 		else
1124 			TCPSTAT_INC(tcps_sndwinup);
1125 
1126 		m = m_gethdr(M_NOWAIT, MT_DATA);
1127 		if (m == NULL) {
1128 			error = ENOBUFS;
1129 			sack_rxmit = 0;
1130 			goto out;
1131 		}
1132 #ifdef INET6
1133 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1134 		    MHLEN >= hdrlen) {
1135 			M_ALIGN(m, hdrlen);
1136 		} else
1137 #endif
1138 		m->m_data += max_linkhdr;
1139 		m->m_len = hdrlen;
1140 	}
1141 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1142 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1143 #ifdef MAC
1144 	mac_inpcb_create_mbuf(inp, m);
1145 #endif
1146 #ifdef INET6
1147 	if (isipv6) {
1148 		ip6 = mtod(m, struct ip6_hdr *);
1149 		if (tp->t_port) {
1150 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1151 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1152 			udp->uh_dport = tp->t_port;
1153 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1154 			udp->uh_ulen = htons(ulen);
1155 			th = (struct tcphdr *)(udp + 1);
1156 		} else {
1157 			th = (struct tcphdr *)(ip6 + 1);
1158 		}
1159 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
1160 	} else
1161 #endif /* INET6 */
1162 	{
1163 		ip = mtod(m, struct ip *);
1164 		if (tp->t_port) {
1165 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1166 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1167 			udp->uh_dport = tp->t_port;
1168 			ulen = hdrlen + len - sizeof(struct ip);
1169 			udp->uh_ulen = htons(ulen);
1170 			th = (struct tcphdr *)(udp + 1);
1171 		} else
1172 			th = (struct tcphdr *)(ip + 1);
1173 		tcpip_fillheaders(inp, tp->t_port, ip, th);
1174 	}
1175 
1176 	/*
1177 	 * Fill in fields, remembering maximum advertised
1178 	 * window for use in delaying messages about window sizes.
1179 	 * If resending a FIN, be sure not to use a new sequence number.
1180 	 */
1181 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1182 	    tp->snd_nxt == tp->snd_max)
1183 		tp->snd_nxt--;
1184 	/*
1185 	 * If we are starting a connection, send ECN setup
1186 	 * SYN packet. If we are on a retransmit, we may
1187 	 * resend those bits a number of times as per
1188 	 * RFC 3168.
1189 	 */
1190 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1191 		flags |= tcp_ecn_output_syn_sent(tp);
1192 	}
1193 	/* Also handle parallel SYN for ECN */
1194 	if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1195 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1196 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1197 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1198 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
1199 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1200 #ifdef INET6
1201 		if (isipv6) {
1202 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1203 			ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1204 		}
1205 		else
1206 #endif
1207 		{
1208 			ip->ip_tos &= ~IPTOS_ECN_MASK;
1209 			ip->ip_tos |= ect;
1210 		}
1211 	}
1212 
1213 	/*
1214 	 * If we are doing retransmissions, then snd_nxt will
1215 	 * not reflect the first unsent octet.  For ACK only
1216 	 * packets, we do not want the sequence number of the
1217 	 * retransmitted packet, we want the sequence number
1218 	 * of the next unsent octet.  So, if there is no data
1219 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1220 	 * when filling in ti_seq.  But if we are in persist
1221 	 * state, snd_max might reflect one byte beyond the
1222 	 * right edge of the window, so use snd_nxt in that
1223 	 * case, since we know we aren't doing a retransmission.
1224 	 * (retransmit and persist are mutually exclusive...)
1225 	 */
1226 	if (sack_rxmit == 0) {
1227 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1228 		    tcp_timer_active(tp, TT_PERSIST))
1229 			th->th_seq = htonl(tp->snd_nxt);
1230 		else
1231 			th->th_seq = htonl(tp->snd_max);
1232 	} else {
1233 		th->th_seq = htonl(p->rxmit);
1234 		p->rxmit += len;
1235 		/*
1236 		 * Lost Retransmission Detection
1237 		 * trigger resending of a (then
1238 		 * still existing) hole, when
1239 		 * fack acks recoverypoint.
1240 		 */
1241 		if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1242 			p->rxmit = tp->snd_recover;
1243 		tp->sackhint.sack_bytes_rexmit += len;
1244 	}
1245 	if (IN_RECOVERY(tp->t_flags)) {
1246 		/*
1247 		 * Account all bytes transmitted while
1248 		 * IN_RECOVERY, simplifying PRR and
1249 		 * Lost Retransmit Detection
1250 		 */
1251 		tp->sackhint.prr_out += len;
1252 	}
1253 	th->th_ack = htonl(tp->rcv_nxt);
1254 	if (optlen) {
1255 		bcopy(opt, th + 1, optlen);
1256 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1257 	}
1258 	tcp_set_flags(th, flags);
1259 	/*
1260 	 * Calculate receive window.  Don't shrink window,
1261 	 * but avoid silly window syndrome.
1262 	 * If a RST segment is sent, advertise a window of zero.
1263 	 */
1264 	if (flags & TH_RST) {
1265 		recwin = 0;
1266 	} else {
1267 		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1268 		    recwin < tp->t_maxseg)
1269 			recwin = 0;
1270 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1271 		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1272 			recwin = (tp->rcv_adv - tp->rcv_nxt);
1273 	}
1274 	/*
1275 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1276 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1277 	 * case is handled in syncache.
1278 	 */
1279 	if (flags & TH_SYN)
1280 		th->th_win = htons((u_short)
1281 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1282 	else {
1283 		/* Avoid shrinking window with window scaling. */
1284 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1285 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1286 	}
1287 
1288 	/*
1289 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1290 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1291 	 * flag tells soreceive() to disable delayed acknowledgements when
1292 	 * draining the buffer.  This can occur if the receiver is attempting
1293 	 * to read more data than can be buffered prior to transmitting on
1294 	 * the connection.
1295 	 */
1296 	if (th->th_win == 0) {
1297 		tp->t_sndzerowin++;
1298 		tp->t_flags |= TF_RXWIN0SENT;
1299 	} else
1300 		tp->t_flags &= ~TF_RXWIN0SENT;
1301 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1302 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1303 		th->th_flags |= TH_URG;
1304 	} else
1305 		/*
1306 		 * If no urgent pointer to send, then we pull
1307 		 * the urgent pointer to the left edge of the send window
1308 		 * so that it doesn't drift into the send window on sequence
1309 		 * number wraparound.
1310 		 */
1311 		tp->snd_up = tp->snd_una;		/* drag it along */
1312 
1313 	/*
1314 	 * Put TCP length in extended header, and then
1315 	 * checksum extended header and data.
1316 	 */
1317 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1318 
1319 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1320 	if (to.to_flags & TOF_SIGNATURE) {
1321 		/*
1322 		 * Calculate MD5 signature and put it into the place
1323 		 * determined before.
1324 		 * NOTE: since TCP options buffer doesn't point into
1325 		 * mbuf's data, calculate offset and use it.
1326 		 */
1327 		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1328 		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1329 			/*
1330 			 * Do not send segment if the calculation of MD5
1331 			 * digest has failed.
1332 			 */
1333 			m_freem(m);
1334 			goto out;
1335 		}
1336 	}
1337 #endif
1338 #ifdef INET6
1339 	if (isipv6) {
1340 		/*
1341 		 * There is no need to fill in ip6_plen right now.
1342 		 * It will be filled later by ip6_output.
1343 		 */
1344 		if (tp->t_port) {
1345 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1346 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1347 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1348 			th->th_sum = htons(0);
1349 			UDPSTAT_INC(udps_opackets);
1350 		} else {
1351 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1352 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1353 			th->th_sum = in6_cksum_pseudo(ip6,
1354 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1355 			    0);
1356 		}
1357 	}
1358 #endif
1359 #if defined(INET6) && defined(INET)
1360 	else
1361 #endif
1362 #ifdef INET
1363 	{
1364 		if (tp->t_port) {
1365 			m->m_pkthdr.csum_flags = CSUM_UDP;
1366 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1367 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1368 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1369 			th->th_sum = htons(0);
1370 			UDPSTAT_INC(udps_opackets);
1371 		} else {
1372 			m->m_pkthdr.csum_flags = CSUM_TCP;
1373 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1374 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1375 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1376 			    IPPROTO_TCP + len + optlen));
1377 		}
1378 
1379 		/* IP version must be set here for ipv4/ipv6 checking later */
1380 		KASSERT(ip->ip_v == IPVERSION,
1381 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1382 	}
1383 #endif
1384 
1385 	/*
1386 	 * Enable TSO and specify the size of the segments.
1387 	 * The TCP pseudo header checksum is always provided.
1388 	 */
1389 	if (tso) {
1390 		KASSERT(len > tp->t_maxseg - optlen,
1391 		    ("%s: len <= tso_segsz", __func__));
1392 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1393 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1394 	}
1395 
1396 	KASSERT(len + hdrlen == m_length(m, NULL),
1397 	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1398 	    __func__, len, hdrlen, m_length(m, NULL)));
1399 
1400 #ifdef TCP_HHOOK
1401 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1402 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1403 #endif
1404 
1405 	TCP_PROBE3(debug__output, tp, th, m);
1406 
1407 	/* We're getting ready to send; log now. */
1408 	/* XXXMT: We are not honoring verbose logging. */
1409 
1410 	if (tcp_bblogging_on(tp))
1411 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1412 		    TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1413 		    NULL);
1414 	else
1415 		lgb = NULL;
1416 
1417 	/*
1418 	 * Fill in IP length and desired time to live and
1419 	 * send to IP level.  There should be a better way
1420 	 * to handle ttl and tos; we could keep them in
1421 	 * the template, but need a way to checksum without them.
1422 	 */
1423 	/*
1424 	 * m->m_pkthdr.len should have been set before checksum calculation,
1425 	 * because in6_cksum() need it.
1426 	 */
1427 #ifdef INET6
1428 	if (isipv6) {
1429 		/*
1430 		 * we separately set hoplimit for every segment, since the
1431 		 * user might want to change the value via setsockopt.
1432 		 * Also, desired default hop limit might be changed via
1433 		 * Neighbor Discovery.
1434 		 */
1435 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1436 
1437 		/*
1438 		 * Set the packet size here for the benefit of DTrace probes.
1439 		 * ip6_output() will set it properly; it's supposed to include
1440 		 * the option header lengths as well.
1441 		 */
1442 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1443 
1444 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1445 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1446 		else
1447 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1448 
1449 		if (tp->t_state == TCPS_SYN_SENT)
1450 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1451 
1452 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1453 
1454 #ifdef TCPPCAP
1455 		/* Save packet, if requested. */
1456 		tcp_pcap_add(th, m, &(tp->t_outpkts));
1457 #endif
1458 
1459 		/* TODO: IPv6 IP6TOS_ECT bit on */
1460 		error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1461 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1462 		    NULL, NULL, inp);
1463 
1464 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1465 			mtu = inp->inp_route6.ro_nh->nh_mtu;
1466 	}
1467 #endif /* INET6 */
1468 #if defined(INET) && defined(INET6)
1469 	else
1470 #endif
1471 #ifdef INET
1472     {
1473 	ip->ip_len = htons(m->m_pkthdr.len);
1474 #ifdef INET6
1475 	if (inp->inp_vflag & INP_IPV6PROTO)
1476 		ip->ip_ttl = in6_selecthlim(inp, NULL);
1477 #endif /* INET6 */
1478 	/*
1479 	 * If we do path MTU discovery, then we set DF on every packet.
1480 	 * This might not be the best thing to do according to RFC3390
1481 	 * Section 2. However the tcp hostcache migitates the problem
1482 	 * so it affects only the first tcp connection with a host.
1483 	 *
1484 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1485 	 */
1486 	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1487 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1488 		if (tp->t_port == 0 || len < V_tcp_minmss) {
1489 			ip->ip_off |= htons(IP_DF);
1490 		}
1491 	} else {
1492 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1493 	}
1494 
1495 	if (tp->t_state == TCPS_SYN_SENT)
1496 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1497 
1498 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1499 
1500 #ifdef TCPPCAP
1501 	/* Save packet, if requested. */
1502 	tcp_pcap_add(th, m, &(tp->t_outpkts));
1503 #endif
1504 
1505 	error = ip_output(m, inp->inp_options, &inp->inp_route,
1506 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1507 
1508 	if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1509 		mtu = inp->inp_route.ro_nh->nh_mtu;
1510     }
1511 #endif /* INET */
1512 
1513 	if (lgb != NULL) {
1514 		lgb->tlb_errno = error;
1515 		lgb = NULL;
1516 	}
1517 out:
1518 	if (error == 0)
1519 		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1520 	/*
1521 	 * In transmit state, time the transmission and arrange for
1522 	 * the retransmit.  In persist state, just set snd_max.
1523 	 */
1524 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1525 	    !tcp_timer_active(tp, TT_PERSIST)) {
1526 		tcp_seq startseq = tp->snd_nxt;
1527 
1528 		/*
1529 		 * Advance snd_nxt over sequence space of this segment.
1530 		 */
1531 		if (flags & (TH_SYN|TH_FIN)) {
1532 			if (flags & TH_SYN)
1533 				tp->snd_nxt++;
1534 			if (flags & TH_FIN) {
1535 				tp->snd_nxt++;
1536 				tp->t_flags |= TF_SENTFIN;
1537 			}
1538 		}
1539 		if (sack_rxmit)
1540 			goto timer;
1541 		tp->snd_nxt += len;
1542 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1543 			/*
1544 			 * Update "made progress" indication if we just
1545 			 * added new data to an empty socket buffer.
1546 			 */
1547 			if (tp->snd_una == tp->snd_max)
1548 				tp->t_acktime = ticks;
1549 			tp->snd_max = tp->snd_nxt;
1550 			/*
1551 			 * Time this transmission if not a retransmission and
1552 			 * not currently timing anything.
1553 			 */
1554 			tp->t_sndtime = ticks;
1555 			if (tp->t_rtttime == 0) {
1556 				tp->t_rtttime = ticks;
1557 				tp->t_rtseq = startseq;
1558 				TCPSTAT_INC(tcps_segstimed);
1559 			}
1560 #ifdef STATS
1561 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1562 				tp->t_flags |= TF_GPUTINPROG;
1563 				tp->gput_seq = startseq;
1564 				tp->gput_ack = startseq +
1565 				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1566 				tp->gput_ts = tcp_ts_getticks();
1567 			}
1568 #endif /* STATS */
1569 		}
1570 
1571 		/*
1572 		 * Set retransmit timer if not currently set,
1573 		 * and not doing a pure ack or a keep-alive probe.
1574 		 * Initial value for retransmit timer is smoothed
1575 		 * round-trip time + 2 * round-trip time variance.
1576 		 * Initialize shift counter which is used for backoff
1577 		 * of retransmit time.
1578 		 */
1579 timer:
1580 		if (!tcp_timer_active(tp, TT_REXMT) &&
1581 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1582 		     (tp->snd_nxt != tp->snd_una))) {
1583 			if (tcp_timer_active(tp, TT_PERSIST)) {
1584 				tcp_timer_activate(tp, TT_PERSIST, 0);
1585 				tp->t_rxtshift = 0;
1586 			}
1587 			tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1588 		} else if (len == 0 && sbavail(&so->so_snd) &&
1589 		    !tcp_timer_active(tp, TT_REXMT) &&
1590 		    !tcp_timer_active(tp, TT_PERSIST)) {
1591 			/*
1592 			 * Avoid a situation where we do not set persist timer
1593 			 * after a zero window condition. For example:
1594 			 * 1) A -> B: packet with enough data to fill the window
1595 			 * 2) B -> A: ACK for #1 + new data (0 window
1596 			 *    advertisement)
1597 			 * 3) A -> B: ACK for #2, 0 len packet
1598 			 *
1599 			 * In this case, A will not activate the persist timer,
1600 			 * because it chose to send a packet. Unless tcp_output
1601 			 * is called for some other reason (delayed ack timer,
1602 			 * another input packet from B, socket syscall), A will
1603 			 * not send zero window probes.
1604 			 *
1605 			 * So, if you send a 0-length packet, but there is data
1606 			 * in the socket buffer, and neither the rexmt or
1607 			 * persist timer is already set, then activate the
1608 			 * persist timer.
1609 			 */
1610 			tp->t_rxtshift = 0;
1611 			tcp_setpersist(tp);
1612 		}
1613 	} else {
1614 		/*
1615 		 * Persist case, update snd_max but since we are in
1616 		 * persist mode (no window) we do not update snd_nxt.
1617 		 */
1618 		int xlen = len;
1619 		if (flags & TH_SYN)
1620 			++xlen;
1621 		if (flags & TH_FIN) {
1622 			++xlen;
1623 			tp->t_flags |= TF_SENTFIN;
1624 		}
1625 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1626 			tp->snd_max = tp->snd_nxt + xlen;
1627 	}
1628 	if ((error == 0) &&
1629 	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1630 	     (tp->t_flags & TF_SACK_PERMIT) &&
1631 	     tp->rcv_numsacks > 0)) {
1632 		    /* Clean up any DSACK's sent */
1633 		    tcp_clean_dsack_blocks(tp);
1634 	}
1635 	if (error) {
1636 		/*
1637 		 * We know that the packet was lost, so back out the
1638 		 * sequence number advance, if any.
1639 		 *
1640 		 * If the error is EPERM the packet got blocked by the
1641 		 * local firewall.  Normally we should terminate the
1642 		 * connection but the blocking may have been spurious
1643 		 * due to a firewall reconfiguration cycle.  So we treat
1644 		 * it like a packet loss and let the retransmit timer and
1645 		 * timeouts do their work over time.
1646 		 * XXX: It is a POLA question whether calling tcp_drop right
1647 		 * away would be the really correct behavior instead.
1648 		 */
1649 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1650 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1651 		    ((flags & TH_SYN) == 0) &&
1652 		    (error != EPERM)) {
1653 			if (sack_rxmit) {
1654 				p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1655 				tp->sackhint.sack_bytes_rexmit -= len;
1656 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1657 				    ("sackhint bytes rtx >= 0"));
1658 				KASSERT((flags & TH_FIN) == 0,
1659 				    ("error while FIN with SACK rxmit"));
1660 			} else {
1661 				tp->snd_nxt -= len;
1662 				if (flags & TH_FIN)
1663 					tp->snd_nxt--;
1664 			}
1665 			if (IN_RECOVERY(tp->t_flags))
1666 				tp->sackhint.prr_out -= len;
1667 		}
1668 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1669 		switch (error) {
1670 		case EACCES:
1671 		case EPERM:
1672 			tp->t_softerror = error;
1673 			return (error);
1674 		case ENOBUFS:
1675 			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1676 			tp->snd_cwnd = tp->t_maxseg;
1677 			return (0);
1678 		case EMSGSIZE:
1679 			/*
1680 			 * For some reason the interface we used initially
1681 			 * to send segments changed to another or lowered
1682 			 * its MTU.
1683 			 * If TSO was active we either got an interface
1684 			 * without TSO capabilits or TSO was turned off.
1685 			 * If we obtained mtu from ip_output() then update
1686 			 * it and try again.
1687 			 */
1688 			if (tso)
1689 				tp->t_flags &= ~TF_TSO;
1690 			if (mtu != 0) {
1691 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1692 				goto again;
1693 			}
1694 			return (error);
1695 		case EHOSTDOWN:
1696 		case EHOSTUNREACH:
1697 		case ENETDOWN:
1698 		case ENETUNREACH:
1699 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1700 				tp->t_softerror = error;
1701 				return (0);
1702 			}
1703 			/* FALLTHROUGH */
1704 		default:
1705 			return (error);
1706 		}
1707 	}
1708 	TCPSTAT_INC(tcps_sndtotal);
1709 
1710 	/*
1711 	 * Data sent (as far as we can tell).
1712 	 * If this advertises a larger window than any other segment,
1713 	 * then remember the size of the advertised window.
1714 	 * Any pending ACK has now been sent.
1715 	 */
1716 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1717 		tp->rcv_adv = tp->rcv_nxt + recwin;
1718 	tp->last_ack_sent = tp->rcv_nxt;
1719 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1720 	if (tcp_timer_active(tp, TT_DELACK))
1721 		tcp_timer_activate(tp, TT_DELACK, 0);
1722 	if (sendalot)
1723 		goto again;
1724 	return (0);
1725 }
1726 
1727 void
1728 tcp_setpersist(struct tcpcb *tp)
1729 {
1730 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1731 	int tt;
1732 	int maxunacktime;
1733 
1734 	tp->t_flags &= ~TF_PREVVALID;
1735 	if (tcp_timer_active(tp, TT_REXMT))
1736 		panic("tcp_setpersist: retransmit pending");
1737 	/*
1738 	 * If the state is already closed, don't bother.
1739 	 */
1740 	if (tp->t_state == TCPS_CLOSED)
1741 		return;
1742 
1743 	/*
1744 	 * Start/restart persistence timer.
1745 	 */
1746 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1747 		      tcp_persmin, tcp_persmax);
1748 	if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1749 		maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1750 		if (maxunacktime < 1)
1751 			maxunacktime = 1;
1752 		if (maxunacktime < tt)
1753 			tt = maxunacktime;
1754 	}
1755 	tcp_timer_activate(tp, TT_PERSIST, tt);
1756 	if (tp->t_rxtshift < V_tcp_retries)
1757 		tp->t_rxtshift++;
1758 }
1759 
1760 /*
1761  * Insert TCP options according to the supplied parameters to the place
1762  * optp in a consistent way.  Can handle unaligned destinations.
1763  *
1764  * The order of the option processing is crucial for optimal packing and
1765  * alignment for the scarce option space.
1766  *
1767  * The optimal order for a SYN/SYN-ACK segment is:
1768  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1769  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1770  *
1771  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1772  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1773  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1774  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1775  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1776  */
1777 int
1778 tcp_addoptions(struct tcpopt *to, u_char *optp)
1779 {
1780 	u_int32_t mask, optlen = 0;
1781 
1782 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1783 		if ((to->to_flags & mask) != mask)
1784 			continue;
1785 		if (optlen == TCP_MAXOLEN)
1786 			break;
1787 		switch (to->to_flags & mask) {
1788 		case TOF_MSS:
1789 			while (optlen % 4) {
1790 				optlen += TCPOLEN_NOP;
1791 				*optp++ = TCPOPT_NOP;
1792 			}
1793 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1794 				continue;
1795 			optlen += TCPOLEN_MAXSEG;
1796 			*optp++ = TCPOPT_MAXSEG;
1797 			*optp++ = TCPOLEN_MAXSEG;
1798 			to->to_mss = htons(to->to_mss);
1799 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1800 			optp += sizeof(to->to_mss);
1801 			break;
1802 		case TOF_SCALE:
1803 			while (!optlen || optlen % 2 != 1) {
1804 				optlen += TCPOLEN_NOP;
1805 				*optp++ = TCPOPT_NOP;
1806 			}
1807 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1808 				continue;
1809 			optlen += TCPOLEN_WINDOW;
1810 			*optp++ = TCPOPT_WINDOW;
1811 			*optp++ = TCPOLEN_WINDOW;
1812 			*optp++ = to->to_wscale;
1813 			break;
1814 		case TOF_SACKPERM:
1815 			while (optlen % 2) {
1816 				optlen += TCPOLEN_NOP;
1817 				*optp++ = TCPOPT_NOP;
1818 			}
1819 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1820 				continue;
1821 			optlen += TCPOLEN_SACK_PERMITTED;
1822 			*optp++ = TCPOPT_SACK_PERMITTED;
1823 			*optp++ = TCPOLEN_SACK_PERMITTED;
1824 			break;
1825 		case TOF_TS:
1826 			while (!optlen || optlen % 4 != 2) {
1827 				optlen += TCPOLEN_NOP;
1828 				*optp++ = TCPOPT_NOP;
1829 			}
1830 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1831 				continue;
1832 			optlen += TCPOLEN_TIMESTAMP;
1833 			*optp++ = TCPOPT_TIMESTAMP;
1834 			*optp++ = TCPOLEN_TIMESTAMP;
1835 			to->to_tsval = htonl(to->to_tsval);
1836 			to->to_tsecr = htonl(to->to_tsecr);
1837 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1838 			optp += sizeof(to->to_tsval);
1839 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1840 			optp += sizeof(to->to_tsecr);
1841 			break;
1842 		case TOF_SIGNATURE:
1843 			{
1844 			int siglen = TCPOLEN_SIGNATURE - 2;
1845 
1846 			while (!optlen || optlen % 4 != 2) {
1847 				optlen += TCPOLEN_NOP;
1848 				*optp++ = TCPOPT_NOP;
1849 			}
1850 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1851 				to->to_flags &= ~TOF_SIGNATURE;
1852 				continue;
1853 			}
1854 			optlen += TCPOLEN_SIGNATURE;
1855 			*optp++ = TCPOPT_SIGNATURE;
1856 			*optp++ = TCPOLEN_SIGNATURE;
1857 			to->to_signature = optp;
1858 			while (siglen--)
1859 				 *optp++ = 0;
1860 			break;
1861 			}
1862 		case TOF_SACK:
1863 			{
1864 			int sackblks = 0;
1865 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1866 			tcp_seq sack_seq;
1867 
1868 			while (!optlen || optlen % 4 != 2) {
1869 				optlen += TCPOLEN_NOP;
1870 				*optp++ = TCPOPT_NOP;
1871 			}
1872 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1873 				continue;
1874 			optlen += TCPOLEN_SACKHDR;
1875 			*optp++ = TCPOPT_SACK;
1876 			sackblks = min(to->to_nsacks,
1877 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1878 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1879 			while (sackblks--) {
1880 				sack_seq = htonl(sack->start);
1881 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1882 				optp += sizeof(sack_seq);
1883 				sack_seq = htonl(sack->end);
1884 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1885 				optp += sizeof(sack_seq);
1886 				optlen += TCPOLEN_SACK;
1887 				sack++;
1888 			}
1889 			TCPSTAT_INC(tcps_sack_send_blocks);
1890 			break;
1891 			}
1892 		case TOF_FASTOPEN:
1893 			{
1894 			int total_len;
1895 
1896 			/* XXX is there any point to aligning this option? */
1897 			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1898 			if (TCP_MAXOLEN - optlen < total_len) {
1899 				to->to_flags &= ~TOF_FASTOPEN;
1900 				continue;
1901 			}
1902 			*optp++ = TCPOPT_FAST_OPEN;
1903 			*optp++ = total_len;
1904 			if (to->to_tfo_len > 0) {
1905 				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1906 				optp += to->to_tfo_len;
1907 			}
1908 			optlen += total_len;
1909 			break;
1910 			}
1911 		default:
1912 			panic("%s: unknown TCP option type", __func__);
1913 			break;
1914 		}
1915 	}
1916 
1917 	/* Terminate and pad TCP options to a 4 byte boundary. */
1918 	if (optlen % 4) {
1919 		optlen += TCPOLEN_EOL;
1920 		*optp++ = TCPOPT_EOL;
1921 	}
1922 	/*
1923 	 * According to RFC 793 (STD0007):
1924 	 *   "The content of the header beyond the End-of-Option option
1925 	 *    must be header padding (i.e., zero)."
1926 	 *   and later: "The padding is composed of zeros."
1927 	 */
1928 	while (optlen % 4) {
1929 		optlen += TCPOLEN_PAD;
1930 		*optp++ = TCPOPT_PAD;
1931 	}
1932 
1933 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1934 	return (optlen);
1935 }
1936 
1937 /*
1938  * This is a copy of m_copym(), taking the TSO segment size/limit
1939  * constraints into account, and advancing the sndptr as it goes.
1940  */
1941 struct mbuf *
1942 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1943     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1944 {
1945 #ifdef KERN_TLS
1946 	struct ktls_session *tls, *ntls;
1947 	struct mbuf *start __diagused;
1948 #endif
1949 	struct mbuf *n, **np;
1950 	struct mbuf *top;
1951 	int32_t off = off0;
1952 	int32_t len = *plen;
1953 	int32_t fragsize;
1954 	int32_t len_cp = 0;
1955 	int32_t *pkthdrlen;
1956 	uint32_t mlen, frags;
1957 	bool copyhdr;
1958 
1959 	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1960 	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1961 	if (off == 0 && m->m_flags & M_PKTHDR)
1962 		copyhdr = true;
1963 	else
1964 		copyhdr = false;
1965 	while (off > 0) {
1966 		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1967 		if (off < m->m_len)
1968 			break;
1969 		off -= m->m_len;
1970 		if ((sb) && (m == sb->sb_sndptr)) {
1971 			sb->sb_sndptroff += m->m_len;
1972 			sb->sb_sndptr = m->m_next;
1973 		}
1974 		m = m->m_next;
1975 	}
1976 	np = &top;
1977 	top = NULL;
1978 	pkthdrlen = NULL;
1979 #ifdef KERN_TLS
1980 	if (hw_tls && (m->m_flags & M_EXTPG))
1981 		tls = m->m_epg_tls;
1982 	else
1983 		tls = NULL;
1984 	start = m;
1985 #endif
1986 	while (len > 0) {
1987 		if (m == NULL) {
1988 			KASSERT(len == M_COPYALL,
1989 			    ("tcp_m_copym, length > size of mbuf chain"));
1990 			*plen = len_cp;
1991 			if (pkthdrlen != NULL)
1992 				*pkthdrlen = len_cp;
1993 			break;
1994 		}
1995 #ifdef KERN_TLS
1996 		if (hw_tls) {
1997 			if (m->m_flags & M_EXTPG)
1998 				ntls = m->m_epg_tls;
1999 			else
2000 				ntls = NULL;
2001 
2002 			/*
2003 			 * Avoid mixing TLS records with handshake
2004 			 * data or TLS records from different
2005 			 * sessions.
2006 			 */
2007 			if (tls != ntls) {
2008 				MPASS(m != start);
2009 				*plen = len_cp;
2010 				if (pkthdrlen != NULL)
2011 					*pkthdrlen = len_cp;
2012 				break;
2013 			}
2014 		}
2015 #endif
2016 		mlen = min(len, m->m_len - off);
2017 		if (seglimit) {
2018 			/*
2019 			 * For M_EXTPG mbufs, add 3 segments
2020 			 * + 1 in case we are crossing page boundaries
2021 			 * + 2 in case the TLS hdr/trailer are used
2022 			 * It is cheaper to just add the segments
2023 			 * than it is to take the cache miss to look
2024 			 * at the mbuf ext_pgs state in detail.
2025 			 */
2026 			if (m->m_flags & M_EXTPG) {
2027 				fragsize = min(segsize, PAGE_SIZE);
2028 				frags = 3;
2029 			} else {
2030 				fragsize = segsize;
2031 				frags = 0;
2032 			}
2033 
2034 			/* Break if we really can't fit anymore. */
2035 			if ((frags + 1) >= seglimit) {
2036 				*plen =	len_cp;
2037 				if (pkthdrlen != NULL)
2038 					*pkthdrlen = len_cp;
2039 				break;
2040 			}
2041 
2042 			/*
2043 			 * Reduce size if you can't copy the whole
2044 			 * mbuf. If we can't copy the whole mbuf, also
2045 			 * adjust len so the loop will end after this
2046 			 * mbuf.
2047 			 */
2048 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2049 				mlen = (seglimit - frags - 1) * fragsize;
2050 				len = mlen;
2051 				*plen = len_cp + len;
2052 				if (pkthdrlen != NULL)
2053 					*pkthdrlen = *plen;
2054 			}
2055 			frags += howmany(mlen, fragsize);
2056 			if (frags == 0)
2057 				frags++;
2058 			seglimit -= frags;
2059 			KASSERT(seglimit > 0,
2060 			    ("%s: seglimit went too low", __func__));
2061 		}
2062 		if (copyhdr)
2063 			n = m_gethdr(M_NOWAIT, m->m_type);
2064 		else
2065 			n = m_get(M_NOWAIT, m->m_type);
2066 		*np = n;
2067 		if (n == NULL)
2068 			goto nospace;
2069 		if (copyhdr) {
2070 			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2071 				goto nospace;
2072 			if (len == M_COPYALL)
2073 				n->m_pkthdr.len -= off0;
2074 			else
2075 				n->m_pkthdr.len = len;
2076 			pkthdrlen = &n->m_pkthdr.len;
2077 			copyhdr = false;
2078 		}
2079 		n->m_len = mlen;
2080 		len_cp += n->m_len;
2081 		if (m->m_flags & (M_EXT | M_EXTPG)) {
2082 			n->m_data = m->m_data + off;
2083 			mb_dupcl(n, m);
2084 		} else
2085 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2086 			    (u_int)n->m_len);
2087 
2088 		if (sb && (sb->sb_sndptr == m) &&
2089 		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2090 			sb->sb_sndptroff += m->m_len;
2091 			sb->sb_sndptr = m->m_next;
2092 		}
2093 		off = 0;
2094 		if (len != M_COPYALL) {
2095 			len -= n->m_len;
2096 		}
2097 		m = m->m_next;
2098 		np = &n->m_next;
2099 	}
2100 	return (top);
2101 nospace:
2102 	m_freem(top);
2103 	return (NULL);
2104 }
2105 
2106 void
2107 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2108 {
2109 
2110 	/*
2111 	 * Automatic sizing of send socket buffer.  Often the send buffer
2112 	 * size is not optimally adjusted to the actual network conditions
2113 	 * at hand (delay bandwidth product).  Setting the buffer size too
2114 	 * small limits throughput on links with high bandwidth and high
2115 	 * delay (eg. trans-continental/oceanic links).  Setting the
2116 	 * buffer size too big consumes too much real kernel memory,
2117 	 * especially with many connections on busy servers.
2118 	 *
2119 	 * The criteria to step up the send buffer one notch are:
2120 	 *  1. receive window of remote host is larger than send buffer
2121 	 *     (with a fudge factor of 5/4th);
2122 	 *  2. send buffer is filled to 7/8th with data (so we actually
2123 	 *     have data to make use of it);
2124 	 *  3. send buffer fill has not hit maximal automatic size;
2125 	 *  4. our send window (slow start and cogestion controlled) is
2126 	 *     larger than sent but unacknowledged data in send buffer.
2127 	 *
2128 	 * The remote host receive window scaling factor may limit the
2129 	 * growing of the send buffer before it reaches its allowed
2130 	 * maximum.
2131 	 *
2132 	 * It scales directly with slow start or congestion window
2133 	 * and does at most one step per received ACK.  This fast
2134 	 * scaling has the drawback of growing the send buffer beyond
2135 	 * what is strictly necessary to make full use of a given
2136 	 * delay*bandwidth product.  However testing has shown this not
2137 	 * to be much of an problem.  At worst we are trading wasting
2138 	 * of available bandwidth (the non-use of it) for wasting some
2139 	 * socket buffer memory.
2140 	 *
2141 	 * TODO: Shrink send buffer during idle periods together
2142 	 * with congestion window.  Requires another timer.  Has to
2143 	 * wait for upcoming tcp timer rewrite.
2144 	 *
2145 	 * XXXGL: should there be used sbused() or sbavail()?
2146 	 */
2147 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2148 		int lowat;
2149 
2150 		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2151 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2152 		    sbused(&so->so_snd) >=
2153 		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2154 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2155 		    sendwin >= (sbused(&so->so_snd) -
2156 		    (tp->snd_nxt - tp->snd_una))) {
2157 			if (!sbreserve_locked(so, SO_SND,
2158 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2159 			     V_tcp_autosndbuf_max), curthread))
2160 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2161 		}
2162 	}
2163 }
2164