xref: /freebsd/sys/netinet/tcp_output.c (revision 3fc9e2c36555140de248a0b4def91bbfa44d7c2c)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_kdtrace.h"
39 #include "opt_tcpdebug.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/hhook.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/mbuf.h>
48 #include <sys/mutex.h>
49 #include <sys/protosw.h>
50 #include <sys/sdt.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
54 
55 #include <net/if.h>
56 #include <net/route.h>
57 #include <net/vnet.h>
58 
59 #include <netinet/cc.h>
60 #include <netinet/in.h>
61 #include <netinet/in_kdtrace.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/ip_options.h>
67 #ifdef INET6
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/ip6_var.h>
71 #endif
72 #define	TCPOUTFLAGS
73 #include <netinet/tcp_fsm.h>
74 #include <netinet/tcp_seq.h>
75 #include <netinet/tcp_timer.h>
76 #include <netinet/tcp_var.h>
77 #include <netinet/tcpip.h>
78 #ifdef TCPDEBUG
79 #include <netinet/tcp_debug.h>
80 #endif
81 #ifdef TCP_OFFLOAD
82 #include <netinet/tcp_offload.h>
83 #endif
84 
85 #ifdef IPSEC
86 #include <netipsec/ipsec.h>
87 #endif /*IPSEC*/
88 
89 #include <machine/in_cksum.h>
90 
91 #include <security/mac/mac_framework.h>
92 
93 VNET_DEFINE(int, path_mtu_discovery) = 1;
94 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
95 	&VNET_NAME(path_mtu_discovery), 1,
96 	"Enable Path MTU Discovery");
97 
98 VNET_DEFINE(int, tcp_do_tso) = 1;
99 #define	V_tcp_do_tso		VNET(tcp_do_tso)
100 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
101 	&VNET_NAME(tcp_do_tso), 0,
102 	"Enable TCP Segmentation Offload");
103 
104 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
105 #define	V_tcp_sendspace	VNET(tcp_sendspace)
106 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
107 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
108 
109 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
110 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
111 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
112 	&VNET_NAME(tcp_do_autosndbuf), 0,
113 	"Enable automatic send buffer sizing");
114 
115 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
116 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
117 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
118 	&VNET_NAME(tcp_autosndbuf_inc), 0,
119 	"Incrementor step size of automatic send buffer");
120 
121 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
122 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
123 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
124 	&VNET_NAME(tcp_autosndbuf_max), 0,
125 	"Max size of automatic send buffer");
126 
127 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
128 			    struct tcphdr *th, struct tcpopt *to,
129 			    long len, int tso);
130 static void inline	cc_after_idle(struct tcpcb *tp);
131 
132 /*
133  * Wrapper for the TCP established output helper hook.
134  */
135 static void inline
136 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
137     struct tcpopt *to, long len, int tso)
138 {
139 	struct tcp_hhook_data hhook_data;
140 
141 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
142 		hhook_data.tp = tp;
143 		hhook_data.th = th;
144 		hhook_data.to = to;
145 		hhook_data.len = len;
146 		hhook_data.tso = tso;
147 
148 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
149 		    tp->osd);
150 	}
151 }
152 
153 /*
154  * CC wrapper hook functions
155  */
156 static void inline
157 cc_after_idle(struct tcpcb *tp)
158 {
159 	INP_WLOCK_ASSERT(tp->t_inpcb);
160 
161 	if (CC_ALGO(tp)->after_idle != NULL)
162 		CC_ALGO(tp)->after_idle(tp->ccv);
163 }
164 
165 /*
166  * Tcp output routine: figure out what should be sent and send it.
167  */
168 int
169 tcp_output(struct tcpcb *tp)
170 {
171 	struct socket *so = tp->t_inpcb->inp_socket;
172 	long len, recwin, sendwin;
173 	int off, flags, error = 0;	/* Keep compiler happy */
174 	struct mbuf *m;
175 	struct ip *ip = NULL;
176 	struct ipovly *ipov = NULL;
177 	struct tcphdr *th;
178 	u_char opt[TCP_MAXOLEN];
179 	unsigned ipoptlen, optlen, hdrlen;
180 #ifdef IPSEC
181 	unsigned ipsec_optlen = 0;
182 #endif
183 	int idle, sendalot;
184 	int sack_rxmit, sack_bytes_rxmt;
185 	struct sackhole *p;
186 	int tso, mtu;
187 	struct tcpopt to;
188 #if 0
189 	int maxburst = TCP_MAXBURST;
190 #endif
191 #ifdef INET6
192 	struct ip6_hdr *ip6 = NULL;
193 	int isipv6;
194 
195 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
196 #endif
197 
198 	INP_WLOCK_ASSERT(tp->t_inpcb);
199 
200 #ifdef TCP_OFFLOAD
201 	if (tp->t_flags & TF_TOE)
202 		return (tcp_offload_output(tp));
203 #endif
204 
205 	/*
206 	 * Determine length of data that should be transmitted,
207 	 * and flags that will be used.
208 	 * If there is some data or critical controls (SYN, RST)
209 	 * to send, then transmit; otherwise, investigate further.
210 	 */
211 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
212 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
213 		cc_after_idle(tp);
214 	tp->t_flags &= ~TF_LASTIDLE;
215 	if (idle) {
216 		if (tp->t_flags & TF_MORETOCOME) {
217 			tp->t_flags |= TF_LASTIDLE;
218 			idle = 0;
219 		}
220 	}
221 again:
222 	/*
223 	 * If we've recently taken a timeout, snd_max will be greater than
224 	 * snd_nxt.  There may be SACK information that allows us to avoid
225 	 * resending already delivered data.  Adjust snd_nxt accordingly.
226 	 */
227 	if ((tp->t_flags & TF_SACK_PERMIT) &&
228 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
229 		tcp_sack_adjust(tp);
230 	sendalot = 0;
231 	tso = 0;
232 	mtu = 0;
233 	off = tp->snd_nxt - tp->snd_una;
234 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
235 
236 	flags = tcp_outflags[tp->t_state];
237 	/*
238 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
239 	 * to send out new data (when sendalot is 1), bypass this function.
240 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
241 	 * we're replacing a (future) new transmission with a retransmission
242 	 * now, and we previously incremented snd_cwnd in tcp_input().
243 	 */
244 	/*
245 	 * Still in sack recovery , reset rxmit flag to zero.
246 	 */
247 	sack_rxmit = 0;
248 	sack_bytes_rxmt = 0;
249 	len = 0;
250 	p = NULL;
251 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
252 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
253 		long cwin;
254 
255 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
256 		if (cwin < 0)
257 			cwin = 0;
258 		/* Do not retransmit SACK segments beyond snd_recover */
259 		if (SEQ_GT(p->end, tp->snd_recover)) {
260 			/*
261 			 * (At least) part of sack hole extends beyond
262 			 * snd_recover. Check to see if we can rexmit data
263 			 * for this hole.
264 			 */
265 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
266 				/*
267 				 * Can't rexmit any more data for this hole.
268 				 * That data will be rexmitted in the next
269 				 * sack recovery episode, when snd_recover
270 				 * moves past p->rxmit.
271 				 */
272 				p = NULL;
273 				goto after_sack_rexmit;
274 			} else
275 				/* Can rexmit part of the current hole */
276 				len = ((long)ulmin(cwin,
277 						   tp->snd_recover - p->rxmit));
278 		} else
279 			len = ((long)ulmin(cwin, p->end - p->rxmit));
280 		off = p->rxmit - tp->snd_una;
281 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
282 		    __func__, off));
283 		if (len > 0) {
284 			sack_rxmit = 1;
285 			sendalot = 1;
286 			TCPSTAT_INC(tcps_sack_rexmits);
287 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
288 			    min(len, tp->t_maxseg));
289 		}
290 	}
291 after_sack_rexmit:
292 	/*
293 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
294 	 * state flags.
295 	 */
296 	if (tp->t_flags & TF_NEEDFIN)
297 		flags |= TH_FIN;
298 	if (tp->t_flags & TF_NEEDSYN)
299 		flags |= TH_SYN;
300 
301 	SOCKBUF_LOCK(&so->so_snd);
302 	/*
303 	 * If in persist timeout with window of 0, send 1 byte.
304 	 * Otherwise, if window is small but nonzero
305 	 * and timer expired, we will send what we can
306 	 * and go to transmit state.
307 	 */
308 	if (tp->t_flags & TF_FORCEDATA) {
309 		if (sendwin == 0) {
310 			/*
311 			 * If we still have some data to send, then
312 			 * clear the FIN bit.  Usually this would
313 			 * happen below when it realizes that we
314 			 * aren't sending all the data.  However,
315 			 * if we have exactly 1 byte of unsent data,
316 			 * then it won't clear the FIN bit below,
317 			 * and if we are in persist state, we wind
318 			 * up sending the packet without recording
319 			 * that we sent the FIN bit.
320 			 *
321 			 * We can't just blindly clear the FIN bit,
322 			 * because if we don't have any more data
323 			 * to send then the probe will be the FIN
324 			 * itself.
325 			 */
326 			if (off < so->so_snd.sb_cc)
327 				flags &= ~TH_FIN;
328 			sendwin = 1;
329 		} else {
330 			tcp_timer_activate(tp, TT_PERSIST, 0);
331 			tp->t_rxtshift = 0;
332 		}
333 	}
334 
335 	/*
336 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
337 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
338 	 * a negative length.  This can also occur when TCP opens up
339 	 * its congestion window while receiving additional duplicate
340 	 * acks after fast-retransmit because TCP will reset snd_nxt
341 	 * to snd_max after the fast-retransmit.
342 	 *
343 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
344 	 * be set to snd_una, the offset will be 0, and the length may
345 	 * wind up 0.
346 	 *
347 	 * If sack_rxmit is true we are retransmitting from the scoreboard
348 	 * in which case len is already set.
349 	 */
350 	if (sack_rxmit == 0) {
351 		if (sack_bytes_rxmt == 0)
352 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
353 		else {
354 			long cwin;
355 
356                         /*
357 			 * We are inside of a SACK recovery episode and are
358 			 * sending new data, having retransmitted all the
359 			 * data possible in the scoreboard.
360 			 */
361 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
362 			       - off);
363 			/*
364 			 * Don't remove this (len > 0) check !
365 			 * We explicitly check for len > 0 here (although it
366 			 * isn't really necessary), to work around a gcc
367 			 * optimization issue - to force gcc to compute
368 			 * len above. Without this check, the computation
369 			 * of len is bungled by the optimizer.
370 			 */
371 			if (len > 0) {
372 				cwin = tp->snd_cwnd -
373 					(tp->snd_nxt - tp->sack_newdata) -
374 					sack_bytes_rxmt;
375 				if (cwin < 0)
376 					cwin = 0;
377 				len = lmin(len, cwin);
378 			}
379 		}
380 	}
381 
382 	/*
383 	 * Lop off SYN bit if it has already been sent.  However, if this
384 	 * is SYN-SENT state and if segment contains data and if we don't
385 	 * know that foreign host supports TAO, suppress sending segment.
386 	 */
387 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
388 		if (tp->t_state != TCPS_SYN_RECEIVED)
389 			flags &= ~TH_SYN;
390 		off--, len++;
391 	}
392 
393 	/*
394 	 * Be careful not to send data and/or FIN on SYN segments.
395 	 * This measure is needed to prevent interoperability problems
396 	 * with not fully conformant TCP implementations.
397 	 */
398 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
399 		len = 0;
400 		flags &= ~TH_FIN;
401 	}
402 
403 	if (len < 0) {
404 		/*
405 		 * If FIN has been sent but not acked,
406 		 * but we haven't been called to retransmit,
407 		 * len will be < 0.  Otherwise, window shrank
408 		 * after we sent into it.  If window shrank to 0,
409 		 * cancel pending retransmit, pull snd_nxt back
410 		 * to (closed) window, and set the persist timer
411 		 * if it isn't already going.  If the window didn't
412 		 * close completely, just wait for an ACK.
413 		 */
414 		len = 0;
415 		if (sendwin == 0) {
416 			tcp_timer_activate(tp, TT_REXMT, 0);
417 			tp->t_rxtshift = 0;
418 			tp->snd_nxt = tp->snd_una;
419 			if (!tcp_timer_active(tp, TT_PERSIST))
420 				tcp_setpersist(tp);
421 		}
422 	}
423 
424 	/* len will be >= 0 after this point. */
425 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
426 
427 	/*
428 	 * Automatic sizing of send socket buffer.  Often the send buffer
429 	 * size is not optimally adjusted to the actual network conditions
430 	 * at hand (delay bandwidth product).  Setting the buffer size too
431 	 * small limits throughput on links with high bandwidth and high
432 	 * delay (eg. trans-continental/oceanic links).  Setting the
433 	 * buffer size too big consumes too much real kernel memory,
434 	 * especially with many connections on busy servers.
435 	 *
436 	 * The criteria to step up the send buffer one notch are:
437 	 *  1. receive window of remote host is larger than send buffer
438 	 *     (with a fudge factor of 5/4th);
439 	 *  2. send buffer is filled to 7/8th with data (so we actually
440 	 *     have data to make use of it);
441 	 *  3. send buffer fill has not hit maximal automatic size;
442 	 *  4. our send window (slow start and cogestion controlled) is
443 	 *     larger than sent but unacknowledged data in send buffer.
444 	 *
445 	 * The remote host receive window scaling factor may limit the
446 	 * growing of the send buffer before it reaches its allowed
447 	 * maximum.
448 	 *
449 	 * It scales directly with slow start or congestion window
450 	 * and does at most one step per received ACK.  This fast
451 	 * scaling has the drawback of growing the send buffer beyond
452 	 * what is strictly necessary to make full use of a given
453 	 * delay*bandwith product.  However testing has shown this not
454 	 * to be much of an problem.  At worst we are trading wasting
455 	 * of available bandwith (the non-use of it) for wasting some
456 	 * socket buffer memory.
457 	 *
458 	 * TODO: Shrink send buffer during idle periods together
459 	 * with congestion window.  Requires another timer.  Has to
460 	 * wait for upcoming tcp timer rewrite.
461 	 */
462 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
463 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
464 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
465 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
466 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
467 			if (!sbreserve_locked(&so->so_snd,
468 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
469 			     V_tcp_autosndbuf_max), so, curthread))
470 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
471 		}
472 	}
473 
474 	/*
475 	 * Decide if we can use TCP Segmentation Offloading (if supported by
476 	 * hardware).
477 	 *
478 	 * TSO may only be used if we are in a pure bulk sending state.  The
479 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
480 	 * IP options prevent using TSO.  With TSO the TCP header is the same
481 	 * (except for the sequence number) for all generated packets.  This
482 	 * makes it impossible to transmit any options which vary per generated
483 	 * segment or packet.
484 	 */
485 #ifdef IPSEC
486 	/*
487 	 * Pre-calculate here as we save another lookup into the darknesses
488 	 * of IPsec that way and can actually decide if TSO is ok.
489 	 */
490 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
491 #endif
492 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
493 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
494 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
495 #ifdef IPSEC
496 	    ipsec_optlen == 0 &&
497 #endif
498 	    tp->t_inpcb->inp_options == NULL &&
499 	    tp->t_inpcb->in6p_options == NULL)
500 		tso = 1;
501 
502 	if (sack_rxmit) {
503 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
504 			flags &= ~TH_FIN;
505 	} else {
506 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
507 			flags &= ~TH_FIN;
508 	}
509 
510 	recwin = sbspace(&so->so_rcv);
511 
512 	/*
513 	 * Sender silly window avoidance.   We transmit under the following
514 	 * conditions when len is non-zero:
515 	 *
516 	 *	- We have a full segment (or more with TSO)
517 	 *	- This is the last buffer in a write()/send() and we are
518 	 *	  either idle or running NODELAY
519 	 *	- we've timed out (e.g. persist timer)
520 	 *	- we have more then 1/2 the maximum send window's worth of
521 	 *	  data (receiver may be limited the window size)
522 	 *	- we need to retransmit
523 	 */
524 	if (len) {
525 		if (len >= tp->t_maxseg)
526 			goto send;
527 		/*
528 		 * NOTE! on localhost connections an 'ack' from the remote
529 		 * end may occur synchronously with the output and cause
530 		 * us to flush a buffer queued with moretocome.  XXX
531 		 *
532 		 * note: the len + off check is almost certainly unnecessary.
533 		 */
534 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
535 		    (idle || (tp->t_flags & TF_NODELAY)) &&
536 		    len + off >= so->so_snd.sb_cc &&
537 		    (tp->t_flags & TF_NOPUSH) == 0) {
538 			goto send;
539 		}
540 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
541 			goto send;
542 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
543 			goto send;
544 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
545 			goto send;
546 		if (sack_rxmit)
547 			goto send;
548 	}
549 
550 	/*
551 	 * Sending of standalone window updates.
552 	 *
553 	 * Window updates are important when we close our window due to a
554 	 * full socket buffer and are opening it again after the application
555 	 * reads data from it.  Once the window has opened again and the
556 	 * remote end starts to send again the ACK clock takes over and
557 	 * provides the most current window information.
558 	 *
559 	 * We must avoid the silly window syndrome whereas every read
560 	 * from the receive buffer, no matter how small, causes a window
561 	 * update to be sent.  We also should avoid sending a flurry of
562 	 * window updates when the socket buffer had queued a lot of data
563 	 * and the application is doing small reads.
564 	 *
565 	 * Prevent a flurry of pointless window updates by only sending
566 	 * an update when we can increase the advertized window by more
567 	 * than 1/4th of the socket buffer capacity.  When the buffer is
568 	 * getting full or is very small be more aggressive and send an
569 	 * update whenever we can increase by two mss sized segments.
570 	 * In all other situations the ACK's to new incoming data will
571 	 * carry further window increases.
572 	 *
573 	 * Don't send an independent window update if a delayed
574 	 * ACK is pending (it will get piggy-backed on it) or the
575 	 * remote side already has done a half-close and won't send
576 	 * more data.  Skip this if the connection is in T/TCP
577 	 * half-open state.
578 	 */
579 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
580 	    !(tp->t_flags & TF_DELACK) &&
581 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
582 		/*
583 		 * "adv" is the amount we could increase the window,
584 		 * taking into account that we are limited by
585 		 * TCP_MAXWIN << tp->rcv_scale.
586 		 */
587 		long adv;
588 		int oldwin;
589 
590 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
591 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
592 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
593 			adv -= oldwin;
594 		} else
595 			oldwin = 0;
596 
597 		/*
598 		 * If the new window size ends up being the same as the old
599 		 * size when it is scaled, then don't force a window update.
600 		 */
601 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
602 			goto dontupdate;
603 
604 		if (adv >= (long)(2 * tp->t_maxseg) &&
605 		    (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
606 		     recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
607 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
608 			goto send;
609 	}
610 dontupdate:
611 
612 	/*
613 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
614 	 * is also a catch-all for the retransmit timer timeout case.
615 	 */
616 	if (tp->t_flags & TF_ACKNOW)
617 		goto send;
618 	if ((flags & TH_RST) ||
619 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
620 		goto send;
621 	if (SEQ_GT(tp->snd_up, tp->snd_una))
622 		goto send;
623 	/*
624 	 * If our state indicates that FIN should be sent
625 	 * and we have not yet done so, then we need to send.
626 	 */
627 	if (flags & TH_FIN &&
628 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
629 		goto send;
630 	/*
631 	 * In SACK, it is possible for tcp_output to fail to send a segment
632 	 * after the retransmission timer has been turned off.  Make sure
633 	 * that the retransmission timer is set.
634 	 */
635 	if ((tp->t_flags & TF_SACK_PERMIT) &&
636 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
637 	    !tcp_timer_active(tp, TT_REXMT) &&
638 	    !tcp_timer_active(tp, TT_PERSIST)) {
639 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
640 		goto just_return;
641 	}
642 	/*
643 	 * TCP window updates are not reliable, rather a polling protocol
644 	 * using ``persist'' packets is used to insure receipt of window
645 	 * updates.  The three ``states'' for the output side are:
646 	 *	idle			not doing retransmits or persists
647 	 *	persisting		to move a small or zero window
648 	 *	(re)transmitting	and thereby not persisting
649 	 *
650 	 * tcp_timer_active(tp, TT_PERSIST)
651 	 *	is true when we are in persist state.
652 	 * (tp->t_flags & TF_FORCEDATA)
653 	 *	is set when we are called to send a persist packet.
654 	 * tcp_timer_active(tp, TT_REXMT)
655 	 *	is set when we are retransmitting
656 	 * The output side is idle when both timers are zero.
657 	 *
658 	 * If send window is too small, there is data to transmit, and no
659 	 * retransmit or persist is pending, then go to persist state.
660 	 * If nothing happens soon, send when timer expires:
661 	 * if window is nonzero, transmit what we can,
662 	 * otherwise force out a byte.
663 	 */
664 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
665 	    !tcp_timer_active(tp, TT_PERSIST)) {
666 		tp->t_rxtshift = 0;
667 		tcp_setpersist(tp);
668 	}
669 
670 	/*
671 	 * No reason to send a segment, just return.
672 	 */
673 just_return:
674 	SOCKBUF_UNLOCK(&so->so_snd);
675 	return (0);
676 
677 send:
678 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
679 	/*
680 	 * Before ESTABLISHED, force sending of initial options
681 	 * unless TCP set not to do any options.
682 	 * NOTE: we assume that the IP/TCP header plus TCP options
683 	 * always fit in a single mbuf, leaving room for a maximum
684 	 * link header, i.e.
685 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
686 	 */
687 	optlen = 0;
688 #ifdef INET6
689 	if (isipv6)
690 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
691 	else
692 #endif
693 		hdrlen = sizeof (struct tcpiphdr);
694 
695 	/*
696 	 * Compute options for segment.
697 	 * We only have to care about SYN and established connection
698 	 * segments.  Options for SYN-ACK segments are handled in TCP
699 	 * syncache.
700 	 */
701 	if ((tp->t_flags & TF_NOOPT) == 0) {
702 		to.to_flags = 0;
703 		/* Maximum segment size. */
704 		if (flags & TH_SYN) {
705 			tp->snd_nxt = tp->iss;
706 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
707 			to.to_flags |= TOF_MSS;
708 		}
709 		/* Window scaling. */
710 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
711 			to.to_wscale = tp->request_r_scale;
712 			to.to_flags |= TOF_SCALE;
713 		}
714 		/* Timestamps. */
715 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
716 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
717 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
718 			to.to_tsecr = tp->ts_recent;
719 			to.to_flags |= TOF_TS;
720 			/* Set receive buffer autosizing timestamp. */
721 			if (tp->rfbuf_ts == 0 &&
722 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
723 				tp->rfbuf_ts = tcp_ts_getticks();
724 		}
725 		/* Selective ACK's. */
726 		if (tp->t_flags & TF_SACK_PERMIT) {
727 			if (flags & TH_SYN)
728 				to.to_flags |= TOF_SACKPERM;
729 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
730 			    (tp->t_flags & TF_SACK_PERMIT) &&
731 			    tp->rcv_numsacks > 0) {
732 				to.to_flags |= TOF_SACK;
733 				to.to_nsacks = tp->rcv_numsacks;
734 				to.to_sacks = (u_char *)tp->sackblks;
735 			}
736 		}
737 #ifdef TCP_SIGNATURE
738 		/* TCP-MD5 (RFC2385). */
739 		if (tp->t_flags & TF_SIGNATURE)
740 			to.to_flags |= TOF_SIGNATURE;
741 #endif /* TCP_SIGNATURE */
742 
743 		/* Processing the options. */
744 		hdrlen += optlen = tcp_addoptions(&to, opt);
745 	}
746 
747 #ifdef INET6
748 	if (isipv6)
749 		ipoptlen = ip6_optlen(tp->t_inpcb);
750 	else
751 #endif
752 	if (tp->t_inpcb->inp_options)
753 		ipoptlen = tp->t_inpcb->inp_options->m_len -
754 				offsetof(struct ipoption, ipopt_list);
755 	else
756 		ipoptlen = 0;
757 #ifdef IPSEC
758 	ipoptlen += ipsec_optlen;
759 #endif
760 
761 	/*
762 	 * Adjust data length if insertion of options will
763 	 * bump the packet length beyond the t_maxopd length.
764 	 * Clear the FIN bit because we cut off the tail of
765 	 * the segment.
766 	 */
767 	if (len + optlen + ipoptlen > tp->t_maxopd) {
768 		flags &= ~TH_FIN;
769 
770 		if (tso) {
771 			KASSERT(ipoptlen == 0,
772 			    ("%s: TSO can't do IP options", __func__));
773 
774 			/*
775 			 * Limit a burst to t_tsomax minus IP,
776 			 * TCP and options length to keep ip->ip_len
777 			 * from overflowing or exceeding the maximum
778 			 * length allowed by the network interface.
779 			 */
780 			if (len > tp->t_tsomax - hdrlen) {
781 				len = tp->t_tsomax - hdrlen;
782 				sendalot = 1;
783 			}
784 
785 			/*
786 			 * Prevent the last segment from being
787 			 * fractional unless the send sockbuf can
788 			 * be emptied.
789 			 */
790 			if (sendalot && off + len < so->so_snd.sb_cc) {
791 				len -= len % (tp->t_maxopd - optlen);
792 				sendalot = 1;
793 			}
794 
795 			/*
796 			 * Send the FIN in a separate segment
797 			 * after the bulk sending is done.
798 			 * We don't trust the TSO implementations
799 			 * to clear the FIN flag on all but the
800 			 * last segment.
801 			 */
802 			if (tp->t_flags & TF_NEEDFIN)
803 				sendalot = 1;
804 
805 		} else {
806 			len = tp->t_maxopd - optlen - ipoptlen;
807 			sendalot = 1;
808 		}
809 	} else
810 		tso = 0;
811 
812 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
813 	    ("%s: len > IP_MAXPACKET", __func__));
814 
815 /*#ifdef DIAGNOSTIC*/
816 #ifdef INET6
817 	if (max_linkhdr + hdrlen > MCLBYTES)
818 #else
819 	if (max_linkhdr + hdrlen > MHLEN)
820 #endif
821 		panic("tcphdr too big");
822 /*#endif*/
823 
824 	/*
825 	 * This KASSERT is here to catch edge cases at a well defined place.
826 	 * Before, those had triggered (random) panic conditions further down.
827 	 */
828 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
829 
830 	/*
831 	 * Grab a header mbuf, attaching a copy of data to
832 	 * be transmitted, and initialize the header from
833 	 * the template for sends on this connection.
834 	 */
835 	if (len) {
836 		struct mbuf *mb;
837 		u_int moff;
838 
839 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
840 			TCPSTAT_INC(tcps_sndprobe);
841 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
842 			tp->t_sndrexmitpack++;
843 			TCPSTAT_INC(tcps_sndrexmitpack);
844 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
845 		} else {
846 			TCPSTAT_INC(tcps_sndpack);
847 			TCPSTAT_ADD(tcps_sndbyte, len);
848 		}
849 #ifdef INET6
850 		if (MHLEN < hdrlen + max_linkhdr)
851 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
852 		else
853 #endif
854 			m = m_gethdr(M_NOWAIT, MT_DATA);
855 
856 		if (m == NULL) {
857 			SOCKBUF_UNLOCK(&so->so_snd);
858 			error = ENOBUFS;
859 			sack_rxmit = 0;
860 			goto out;
861 		}
862 
863 		m->m_data += max_linkhdr;
864 		m->m_len = hdrlen;
865 
866 		/*
867 		 * Start the m_copy functions from the closest mbuf
868 		 * to the offset in the socket buffer chain.
869 		 */
870 		mb = sbsndptr(&so->so_snd, off, len, &moff);
871 
872 		if (len <= MHLEN - hdrlen - max_linkhdr) {
873 			m_copydata(mb, moff, (int)len,
874 			    mtod(m, caddr_t) + hdrlen);
875 			m->m_len += len;
876 		} else {
877 			m->m_next = m_copy(mb, moff, (int)len);
878 			if (m->m_next == NULL) {
879 				SOCKBUF_UNLOCK(&so->so_snd);
880 				(void) m_free(m);
881 				error = ENOBUFS;
882 				sack_rxmit = 0;
883 				goto out;
884 			}
885 		}
886 
887 		/*
888 		 * If we're sending everything we've got, set PUSH.
889 		 * (This will keep happy those implementations which only
890 		 * give data to the user when a buffer fills or
891 		 * a PUSH comes in.)
892 		 */
893 		if (off + len == so->so_snd.sb_cc)
894 			flags |= TH_PUSH;
895 		SOCKBUF_UNLOCK(&so->so_snd);
896 	} else {
897 		SOCKBUF_UNLOCK(&so->so_snd);
898 		if (tp->t_flags & TF_ACKNOW)
899 			TCPSTAT_INC(tcps_sndacks);
900 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
901 			TCPSTAT_INC(tcps_sndctrl);
902 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
903 			TCPSTAT_INC(tcps_sndurg);
904 		else
905 			TCPSTAT_INC(tcps_sndwinup);
906 
907 		m = m_gethdr(M_NOWAIT, MT_DATA);
908 		if (m == NULL) {
909 			error = ENOBUFS;
910 			sack_rxmit = 0;
911 			goto out;
912 		}
913 #ifdef INET6
914 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
915 		    MHLEN >= hdrlen) {
916 			MH_ALIGN(m, hdrlen);
917 		} else
918 #endif
919 		m->m_data += max_linkhdr;
920 		m->m_len = hdrlen;
921 	}
922 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
923 	m->m_pkthdr.rcvif = (struct ifnet *)0;
924 #ifdef MAC
925 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
926 #endif
927 #ifdef INET6
928 	if (isipv6) {
929 		ip6 = mtod(m, struct ip6_hdr *);
930 		th = (struct tcphdr *)(ip6 + 1);
931 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
932 	} else
933 #endif /* INET6 */
934 	{
935 		ip = mtod(m, struct ip *);
936 		ipov = (struct ipovly *)ip;
937 		th = (struct tcphdr *)(ip + 1);
938 		tcpip_fillheaders(tp->t_inpcb, ip, th);
939 	}
940 
941 	/*
942 	 * Fill in fields, remembering maximum advertised
943 	 * window for use in delaying messages about window sizes.
944 	 * If resending a FIN, be sure not to use a new sequence number.
945 	 */
946 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
947 	    tp->snd_nxt == tp->snd_max)
948 		tp->snd_nxt--;
949 	/*
950 	 * If we are starting a connection, send ECN setup
951 	 * SYN packet. If we are on a retransmit, we may
952 	 * resend those bits a number of times as per
953 	 * RFC 3168.
954 	 */
955 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
956 		if (tp->t_rxtshift >= 1) {
957 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
958 				flags |= TH_ECE|TH_CWR;
959 		} else
960 			flags |= TH_ECE|TH_CWR;
961 	}
962 
963 	if (tp->t_state == TCPS_ESTABLISHED &&
964 	    (tp->t_flags & TF_ECN_PERMIT)) {
965 		/*
966 		 * If the peer has ECN, mark data packets with
967 		 * ECN capable transmission (ECT).
968 		 * Ignore pure ack packets, retransmissions and window probes.
969 		 */
970 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
971 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
972 #ifdef INET6
973 			if (isipv6)
974 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
975 			else
976 #endif
977 				ip->ip_tos |= IPTOS_ECN_ECT0;
978 			TCPSTAT_INC(tcps_ecn_ect0);
979 		}
980 
981 		/*
982 		 * Reply with proper ECN notifications.
983 		 */
984 		if (tp->t_flags & TF_ECN_SND_CWR) {
985 			flags |= TH_CWR;
986 			tp->t_flags &= ~TF_ECN_SND_CWR;
987 		}
988 		if (tp->t_flags & TF_ECN_SND_ECE)
989 			flags |= TH_ECE;
990 	}
991 
992 	/*
993 	 * If we are doing retransmissions, then snd_nxt will
994 	 * not reflect the first unsent octet.  For ACK only
995 	 * packets, we do not want the sequence number of the
996 	 * retransmitted packet, we want the sequence number
997 	 * of the next unsent octet.  So, if there is no data
998 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
999 	 * when filling in ti_seq.  But if we are in persist
1000 	 * state, snd_max might reflect one byte beyond the
1001 	 * right edge of the window, so use snd_nxt in that
1002 	 * case, since we know we aren't doing a retransmission.
1003 	 * (retransmit and persist are mutually exclusive...)
1004 	 */
1005 	if (sack_rxmit == 0) {
1006 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1007 		    tcp_timer_active(tp, TT_PERSIST))
1008 			th->th_seq = htonl(tp->snd_nxt);
1009 		else
1010 			th->th_seq = htonl(tp->snd_max);
1011 	} else {
1012 		th->th_seq = htonl(p->rxmit);
1013 		p->rxmit += len;
1014 		tp->sackhint.sack_bytes_rexmit += len;
1015 	}
1016 	th->th_ack = htonl(tp->rcv_nxt);
1017 	if (optlen) {
1018 		bcopy(opt, th + 1, optlen);
1019 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1020 	}
1021 	th->th_flags = flags;
1022 	/*
1023 	 * Calculate receive window.  Don't shrink window,
1024 	 * but avoid silly window syndrome.
1025 	 */
1026 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1027 	    recwin < (long)tp->t_maxseg)
1028 		recwin = 0;
1029 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1030 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1031 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1032 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1033 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1034 
1035 	/*
1036 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1037 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1038 	 * case is handled in syncache.
1039 	 */
1040 	if (flags & TH_SYN)
1041 		th->th_win = htons((u_short)
1042 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1043 	else
1044 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1045 
1046 	/*
1047 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1048 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1049 	 * flag tells soreceive() to disable delayed acknowledgements when
1050 	 * draining the buffer.  This can occur if the receiver is attempting
1051 	 * to read more data than can be buffered prior to transmitting on
1052 	 * the connection.
1053 	 */
1054 	if (th->th_win == 0) {
1055 		tp->t_sndzerowin++;
1056 		tp->t_flags |= TF_RXWIN0SENT;
1057 	} else
1058 		tp->t_flags &= ~TF_RXWIN0SENT;
1059 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1060 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1061 		th->th_flags |= TH_URG;
1062 	} else
1063 		/*
1064 		 * If no urgent pointer to send, then we pull
1065 		 * the urgent pointer to the left edge of the send window
1066 		 * so that it doesn't drift into the send window on sequence
1067 		 * number wraparound.
1068 		 */
1069 		tp->snd_up = tp->snd_una;		/* drag it along */
1070 
1071 #ifdef TCP_SIGNATURE
1072 	if (tp->t_flags & TF_SIGNATURE) {
1073 		int sigoff = to.to_signature - opt;
1074 		tcp_signature_compute(m, 0, len, optlen,
1075 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1076 	}
1077 #endif
1078 
1079 	/*
1080 	 * Put TCP length in extended header, and then
1081 	 * checksum extended header and data.
1082 	 */
1083 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1084 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1085 #ifdef INET6
1086 	if (isipv6) {
1087 		/*
1088 		 * ip6_plen is not need to be filled now, and will be filled
1089 		 * in ip6_output.
1090 		 */
1091 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1092 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1093 		    optlen + len, IPPROTO_TCP, 0);
1094 	}
1095 #endif
1096 #if defined(INET6) && defined(INET)
1097 	else
1098 #endif
1099 #ifdef INET
1100 	{
1101 		m->m_pkthdr.csum_flags = CSUM_TCP;
1102 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1103 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1104 
1105 		/* IP version must be set here for ipv4/ipv6 checking later */
1106 		KASSERT(ip->ip_v == IPVERSION,
1107 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1108 	}
1109 #endif
1110 
1111 	/*
1112 	 * Enable TSO and specify the size of the segments.
1113 	 * The TCP pseudo header checksum is always provided.
1114 	 * XXX: Fixme: This is currently not the case for IPv6.
1115 	 */
1116 	if (tso) {
1117 		KASSERT(len > tp->t_maxopd - optlen,
1118 		    ("%s: len <= tso_segsz", __func__));
1119 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1120 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1121 	}
1122 
1123 #ifdef IPSEC
1124 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1125 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1126 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1127 #else
1128 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1129 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1130 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1131 #endif
1132 
1133 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1134 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1135 
1136 #ifdef TCPDEBUG
1137 	/*
1138 	 * Trace.
1139 	 */
1140 	if (so->so_options & SO_DEBUG) {
1141 		u_short save = 0;
1142 #ifdef INET6
1143 		if (!isipv6)
1144 #endif
1145 		{
1146 			save = ipov->ih_len;
1147 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1148 		}
1149 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1150 #ifdef INET6
1151 		if (!isipv6)
1152 #endif
1153 		ipov->ih_len = save;
1154 	}
1155 #endif /* TCPDEBUG */
1156 
1157 	/*
1158 	 * Fill in IP length and desired time to live and
1159 	 * send to IP level.  There should be a better way
1160 	 * to handle ttl and tos; we could keep them in
1161 	 * the template, but need a way to checksum without them.
1162 	 */
1163 	/*
1164 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1165 	 * because in6_cksum() need it.
1166 	 */
1167 #ifdef INET6
1168 	if (isipv6) {
1169 		struct route_in6 ro;
1170 
1171 		bzero(&ro, sizeof(ro));
1172 		/*
1173 		 * we separately set hoplimit for every segment, since the
1174 		 * user might want to change the value via setsockopt.
1175 		 * Also, desired default hop limit might be changed via
1176 		 * Neighbor Discovery.
1177 		 */
1178 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1179 
1180 		/*
1181 		 * Set the packet size here for the benefit of DTrace probes.
1182 		 * ip6_output() will set it properly; it's supposed to include
1183 		 * the option header lengths as well.
1184 		 */
1185 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1186 
1187 		if (tp->t_state == TCPS_SYN_SENT)
1188 			TCP_PROBE5(connect_request, NULL, tp, ip6, tp, th);
1189 
1190 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1191 
1192 		/* TODO: IPv6 IP6TOS_ECT bit on */
1193 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1194 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1195 		    NULL, NULL, tp->t_inpcb);
1196 
1197 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1198 			mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1199 		RO_RTFREE(&ro);
1200 	}
1201 #endif /* INET6 */
1202 #if defined(INET) && defined(INET6)
1203 	else
1204 #endif
1205 #ifdef INET
1206     {
1207 	struct route ro;
1208 
1209 	bzero(&ro, sizeof(ro));
1210 	ip->ip_len = htons(m->m_pkthdr.len);
1211 #ifdef INET6
1212 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1213 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1214 #endif /* INET6 */
1215 	/*
1216 	 * If we do path MTU discovery, then we set DF on every packet.
1217 	 * This might not be the best thing to do according to RFC3390
1218 	 * Section 2. However the tcp hostcache migitates the problem
1219 	 * so it affects only the first tcp connection with a host.
1220 	 *
1221 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1222 	 */
1223 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1224 		ip->ip_off |= htons(IP_DF);
1225 
1226 	if (tp->t_state == TCPS_SYN_SENT)
1227 		TCP_PROBE5(connect_request, NULL, tp, ip, tp, th);
1228 
1229 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1230 
1231 	error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1232 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1233 	    tp->t_inpcb);
1234 
1235 	if (error == EMSGSIZE && ro.ro_rt != NULL)
1236 		mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1237 	RO_RTFREE(&ro);
1238     }
1239 #endif /* INET */
1240 
1241 out:
1242 	/*
1243 	 * In transmit state, time the transmission and arrange for
1244 	 * the retransmit.  In persist state, just set snd_max.
1245 	 */
1246 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1247 	    !tcp_timer_active(tp, TT_PERSIST)) {
1248 		tcp_seq startseq = tp->snd_nxt;
1249 
1250 		/*
1251 		 * Advance snd_nxt over sequence space of this segment.
1252 		 */
1253 		if (flags & (TH_SYN|TH_FIN)) {
1254 			if (flags & TH_SYN)
1255 				tp->snd_nxt++;
1256 			if (flags & TH_FIN) {
1257 				tp->snd_nxt++;
1258 				tp->t_flags |= TF_SENTFIN;
1259 			}
1260 		}
1261 		if (sack_rxmit)
1262 			goto timer;
1263 		tp->snd_nxt += len;
1264 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1265 			tp->snd_max = tp->snd_nxt;
1266 			/*
1267 			 * Time this transmission if not a retransmission and
1268 			 * not currently timing anything.
1269 			 */
1270 			if (tp->t_rtttime == 0) {
1271 				tp->t_rtttime = ticks;
1272 				tp->t_rtseq = startseq;
1273 				TCPSTAT_INC(tcps_segstimed);
1274 			}
1275 		}
1276 
1277 		/*
1278 		 * Set retransmit timer if not currently set,
1279 		 * and not doing a pure ack or a keep-alive probe.
1280 		 * Initial value for retransmit timer is smoothed
1281 		 * round-trip time + 2 * round-trip time variance.
1282 		 * Initialize shift counter which is used for backoff
1283 		 * of retransmit time.
1284 		 */
1285 timer:
1286 		if (!tcp_timer_active(tp, TT_REXMT) &&
1287 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1288 		     (tp->snd_nxt != tp->snd_una))) {
1289 			if (tcp_timer_active(tp, TT_PERSIST)) {
1290 				tcp_timer_activate(tp, TT_PERSIST, 0);
1291 				tp->t_rxtshift = 0;
1292 			}
1293 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1294 		}
1295 	} else {
1296 		/*
1297 		 * Persist case, update snd_max but since we are in
1298 		 * persist mode (no window) we do not update snd_nxt.
1299 		 */
1300 		int xlen = len;
1301 		if (flags & TH_SYN)
1302 			++xlen;
1303 		if (flags & TH_FIN) {
1304 			++xlen;
1305 			tp->t_flags |= TF_SENTFIN;
1306 		}
1307 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1308 			tp->snd_max = tp->snd_nxt + len;
1309 	}
1310 
1311 	if (error) {
1312 
1313 		/*
1314 		 * We know that the packet was lost, so back out the
1315 		 * sequence number advance, if any.
1316 		 *
1317 		 * If the error is EPERM the packet got blocked by the
1318 		 * local firewall.  Normally we should terminate the
1319 		 * connection but the blocking may have been spurious
1320 		 * due to a firewall reconfiguration cycle.  So we treat
1321 		 * it like a packet loss and let the retransmit timer and
1322 		 * timeouts do their work over time.
1323 		 * XXX: It is a POLA question whether calling tcp_drop right
1324 		 * away would be the really correct behavior instead.
1325 		 */
1326 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1327 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1328 		    ((flags & TH_SYN) == 0) &&
1329 		    (error != EPERM)) {
1330 			if (sack_rxmit) {
1331 				p->rxmit -= len;
1332 				tp->sackhint.sack_bytes_rexmit -= len;
1333 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1334 				    ("sackhint bytes rtx >= 0"));
1335 			} else
1336 				tp->snd_nxt -= len;
1337 		}
1338 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1339 		switch (error) {
1340 		case EPERM:
1341 			tp->t_softerror = error;
1342 			return (error);
1343 		case ENOBUFS:
1344 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1345 			    !tcp_timer_active(tp, TT_PERSIST))
1346 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1347 			tp->snd_cwnd = tp->t_maxseg;
1348 			return (0);
1349 		case EMSGSIZE:
1350 			/*
1351 			 * For some reason the interface we used initially
1352 			 * to send segments changed to another or lowered
1353 			 * its MTU.
1354 			 * If TSO was active we either got an interface
1355 			 * without TSO capabilits or TSO was turned off.
1356 			 * If we obtained mtu from ip_output() then update
1357 			 * it and try again.
1358 			 */
1359 			if (tso)
1360 				tp->t_flags &= ~TF_TSO;
1361 			if (mtu != 0) {
1362 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1363 				goto again;
1364 			}
1365 			return (error);
1366 		case EHOSTDOWN:
1367 		case EHOSTUNREACH:
1368 		case ENETDOWN:
1369 		case ENETUNREACH:
1370 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1371 				tp->t_softerror = error;
1372 				return (0);
1373 			}
1374 			/* FALLTHROUGH */
1375 		default:
1376 			return (error);
1377 		}
1378 	}
1379 	TCPSTAT_INC(tcps_sndtotal);
1380 
1381 	/*
1382 	 * Data sent (as far as we can tell).
1383 	 * If this advertises a larger window than any other segment,
1384 	 * then remember the size of the advertised window.
1385 	 * Any pending ACK has now been sent.
1386 	 */
1387 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1388 		tp->rcv_adv = tp->rcv_nxt + recwin;
1389 	tp->last_ack_sent = tp->rcv_nxt;
1390 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1391 	if (tcp_timer_active(tp, TT_DELACK))
1392 		tcp_timer_activate(tp, TT_DELACK, 0);
1393 #if 0
1394 	/*
1395 	 * This completely breaks TCP if newreno is turned on.  What happens
1396 	 * is that if delayed-acks are turned on on the receiver, this code
1397 	 * on the transmitter effectively destroys the TCP window, forcing
1398 	 * it to four packets (1.5Kx4 = 6K window).
1399 	 */
1400 	if (sendalot && --maxburst)
1401 		goto again;
1402 #endif
1403 	if (sendalot)
1404 		goto again;
1405 	return (0);
1406 }
1407 
1408 void
1409 tcp_setpersist(struct tcpcb *tp)
1410 {
1411 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1412 	int tt;
1413 
1414 	tp->t_flags &= ~TF_PREVVALID;
1415 	if (tcp_timer_active(tp, TT_REXMT))
1416 		panic("tcp_setpersist: retransmit pending");
1417 	/*
1418 	 * Start/restart persistance timer.
1419 	 */
1420 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1421 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1422 	tcp_timer_activate(tp, TT_PERSIST, tt);
1423 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1424 		tp->t_rxtshift++;
1425 }
1426 
1427 /*
1428  * Insert TCP options according to the supplied parameters to the place
1429  * optp in a consistent way.  Can handle unaligned destinations.
1430  *
1431  * The order of the option processing is crucial for optimal packing and
1432  * alignment for the scarce option space.
1433  *
1434  * The optimal order for a SYN/SYN-ACK segment is:
1435  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1436  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1437  *
1438  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1439  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1440  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1441  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1442  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1443  */
1444 int
1445 tcp_addoptions(struct tcpopt *to, u_char *optp)
1446 {
1447 	u_int mask, optlen = 0;
1448 
1449 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1450 		if ((to->to_flags & mask) != mask)
1451 			continue;
1452 		if (optlen == TCP_MAXOLEN)
1453 			break;
1454 		switch (to->to_flags & mask) {
1455 		case TOF_MSS:
1456 			while (optlen % 4) {
1457 				optlen += TCPOLEN_NOP;
1458 				*optp++ = TCPOPT_NOP;
1459 			}
1460 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1461 				continue;
1462 			optlen += TCPOLEN_MAXSEG;
1463 			*optp++ = TCPOPT_MAXSEG;
1464 			*optp++ = TCPOLEN_MAXSEG;
1465 			to->to_mss = htons(to->to_mss);
1466 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1467 			optp += sizeof(to->to_mss);
1468 			break;
1469 		case TOF_SCALE:
1470 			while (!optlen || optlen % 2 != 1) {
1471 				optlen += TCPOLEN_NOP;
1472 				*optp++ = TCPOPT_NOP;
1473 			}
1474 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1475 				continue;
1476 			optlen += TCPOLEN_WINDOW;
1477 			*optp++ = TCPOPT_WINDOW;
1478 			*optp++ = TCPOLEN_WINDOW;
1479 			*optp++ = to->to_wscale;
1480 			break;
1481 		case TOF_SACKPERM:
1482 			while (optlen % 2) {
1483 				optlen += TCPOLEN_NOP;
1484 				*optp++ = TCPOPT_NOP;
1485 			}
1486 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1487 				continue;
1488 			optlen += TCPOLEN_SACK_PERMITTED;
1489 			*optp++ = TCPOPT_SACK_PERMITTED;
1490 			*optp++ = TCPOLEN_SACK_PERMITTED;
1491 			break;
1492 		case TOF_TS:
1493 			while (!optlen || optlen % 4 != 2) {
1494 				optlen += TCPOLEN_NOP;
1495 				*optp++ = TCPOPT_NOP;
1496 			}
1497 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1498 				continue;
1499 			optlen += TCPOLEN_TIMESTAMP;
1500 			*optp++ = TCPOPT_TIMESTAMP;
1501 			*optp++ = TCPOLEN_TIMESTAMP;
1502 			to->to_tsval = htonl(to->to_tsval);
1503 			to->to_tsecr = htonl(to->to_tsecr);
1504 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1505 			optp += sizeof(to->to_tsval);
1506 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1507 			optp += sizeof(to->to_tsecr);
1508 			break;
1509 		case TOF_SIGNATURE:
1510 			{
1511 			int siglen = TCPOLEN_SIGNATURE - 2;
1512 
1513 			while (!optlen || optlen % 4 != 2) {
1514 				optlen += TCPOLEN_NOP;
1515 				*optp++ = TCPOPT_NOP;
1516 			}
1517 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1518 				continue;
1519 			optlen += TCPOLEN_SIGNATURE;
1520 			*optp++ = TCPOPT_SIGNATURE;
1521 			*optp++ = TCPOLEN_SIGNATURE;
1522 			to->to_signature = optp;
1523 			while (siglen--)
1524 				 *optp++ = 0;
1525 			break;
1526 			}
1527 		case TOF_SACK:
1528 			{
1529 			int sackblks = 0;
1530 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1531 			tcp_seq sack_seq;
1532 
1533 			while (!optlen || optlen % 4 != 2) {
1534 				optlen += TCPOLEN_NOP;
1535 				*optp++ = TCPOPT_NOP;
1536 			}
1537 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1538 				continue;
1539 			optlen += TCPOLEN_SACKHDR;
1540 			*optp++ = TCPOPT_SACK;
1541 			sackblks = min(to->to_nsacks,
1542 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1543 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1544 			while (sackblks--) {
1545 				sack_seq = htonl(sack->start);
1546 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1547 				optp += sizeof(sack_seq);
1548 				sack_seq = htonl(sack->end);
1549 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1550 				optp += sizeof(sack_seq);
1551 				optlen += TCPOLEN_SACK;
1552 				sack++;
1553 			}
1554 			TCPSTAT_INC(tcps_sack_send_blocks);
1555 			break;
1556 			}
1557 		default:
1558 			panic("%s: unknown TCP option type", __func__);
1559 			break;
1560 		}
1561 	}
1562 
1563 	/* Terminate and pad TCP options to a 4 byte boundary. */
1564 	if (optlen % 4) {
1565 		optlen += TCPOLEN_EOL;
1566 		*optp++ = TCPOPT_EOL;
1567 	}
1568 	/*
1569 	 * According to RFC 793 (STD0007):
1570 	 *   "The content of the header beyond the End-of-Option option
1571 	 *    must be header padding (i.e., zero)."
1572 	 *   and later: "The padding is composed of zeros."
1573 	 */
1574 	while (optlen % 4) {
1575 		optlen += TCPOLEN_PAD;
1576 		*optp++ = TCPOPT_PAD;
1577 	}
1578 
1579 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1580 	return (optlen);
1581 }
1582