xref: /netbsd-src/sys/net/if_tun.c (revision 3816d47b2c42fcd6e549e3407f842a5b1a1d23ad)
1 /*	$NetBSD: if_tun.c,v 1.111 2009/05/08 11:09:44 elad Exp $	*/
2 
3 /*
4  * Copyright (c) 1988, Julian Onions <jpo@cs.nott.ac.uk>
5  * Nottingham University 1987.
6  *
7  * This source may be freely distributed, however I would be interested
8  * in any changes that are made.
9  *
10  * This driver takes packets off the IP i/f and hands them up to a
11  * user process to have its wicked way with. This driver has its
12  * roots in a similar driver written by Phil Cockcroft (formerly) at
13  * UCL. This driver is based much more on read/write/poll mode of
14  * operation though.
15  */
16 
17 #include <sys/cdefs.h>
18 __KERNEL_RCSID(0, "$NetBSD: if_tun.c,v 1.111 2009/05/08 11:09:44 elad Exp $");
19 
20 #include "opt_inet.h"
21 
22 #include <sys/param.h>
23 #include <sys/proc.h>
24 #include <sys/systm.h>
25 #include <sys/mbuf.h>
26 #include <sys/buf.h>
27 #include <sys/protosw.h>
28 #include <sys/socket.h>
29 #include <sys/ioctl.h>
30 #include <sys/errno.h>
31 #include <sys/syslog.h>
32 #include <sys/select.h>
33 #include <sys/poll.h>
34 #include <sys/file.h>
35 #include <sys/signalvar.h>
36 #include <sys/conf.h>
37 #include <sys/kauth.h>
38 #include <sys/simplelock.h>
39 #include <sys/cpu.h>
40 
41 #include <net/if.h>
42 #include <net/if_types.h>
43 #include <net/netisr.h>
44 #include <net/route.h>
45 
46 
47 #ifdef INET
48 #include <netinet/in.h>
49 #include <netinet/in_systm.h>
50 #include <netinet/in_var.h>
51 #include <netinet/ip.h>
52 #include <netinet/if_inarp.h>
53 #endif
54 
55 
56 #include "bpfilter.h"
57 #if NBPFILTER > 0
58 #include <sys/time.h>
59 #include <net/bpf.h>
60 #endif
61 
62 #include <net/if_tun.h>
63 
64 #define TUNDEBUG	if (tundebug) printf
65 int	tundebug = 0;
66 
67 extern int ifqmaxlen;
68 void	tunattach(int);
69 
70 static LIST_HEAD(, tun_softc) tun_softc_list;
71 static LIST_HEAD(, tun_softc) tunz_softc_list;
72 static struct simplelock tun_softc_lock;
73 
74 static int	tun_ioctl(struct ifnet *, u_long, void *);
75 static int	tun_output(struct ifnet *, struct mbuf *,
76 			const struct sockaddr *, struct rtentry *rt);
77 static int	tun_clone_create(struct if_clone *, int);
78 static int	tun_clone_destroy(struct ifnet *);
79 
80 static struct if_clone tun_cloner =
81     IF_CLONE_INITIALIZER("tun", tun_clone_create, tun_clone_destroy);
82 
83 static void tunattach0(struct tun_softc *);
84 static void tuninit(struct tun_softc *);
85 static void tun_i_softintr(void *);
86 static void tun_o_softintr(void *);
87 #ifdef ALTQ
88 static void tunstart(struct ifnet *);
89 #endif
90 static struct tun_softc *tun_find_unit(dev_t);
91 static struct tun_softc *tun_find_zunit(int);
92 
93 static dev_type_open(tunopen);
94 static dev_type_close(tunclose);
95 static dev_type_read(tunread);
96 static dev_type_write(tunwrite);
97 static dev_type_ioctl(tunioctl);
98 static dev_type_poll(tunpoll);
99 static dev_type_kqfilter(tunkqfilter);
100 
101 const struct cdevsw tun_cdevsw = {
102 	tunopen, tunclose, tunread, tunwrite, tunioctl,
103 	nostop, notty, tunpoll, nommap, tunkqfilter, D_OTHER,
104 };
105 
106 void
107 tunattach(int unused)
108 {
109 
110 	simple_lock_init(&tun_softc_lock);
111 	LIST_INIT(&tun_softc_list);
112 	LIST_INIT(&tunz_softc_list);
113 	if_clone_attach(&tun_cloner);
114 }
115 
116 /*
117  * Find driver instance from dev_t.
118  * Call at splnet().
119  * Returns with tp locked (if found).
120  */
121 static struct tun_softc *
122 tun_find_unit(dev_t dev)
123 {
124 	struct tun_softc *tp;
125 	int unit = minor(dev);
126 
127 	simple_lock(&tun_softc_lock);
128 	LIST_FOREACH(tp, &tun_softc_list, tun_list)
129 		if (unit == tp->tun_unit)
130 			break;
131 	if (tp)
132 		simple_lock(&tp->tun_lock);
133 	simple_unlock(&tun_softc_lock);
134 
135 	return (tp);
136 }
137 
138 /*
139  * Find zombie driver instance by unit number.
140  * Call at splnet().
141  * Remove tp from list and return it unlocked (if found).
142  */
143 static struct tun_softc *
144 tun_find_zunit(int unit)
145 {
146 	struct tun_softc *tp;
147 
148 	simple_lock(&tun_softc_lock);
149 	LIST_FOREACH(tp, &tunz_softc_list, tun_list)
150 		if (unit == tp->tun_unit)
151 			break;
152 	if (tp)
153 		LIST_REMOVE(tp, tun_list);
154 	simple_unlock(&tun_softc_lock);
155 #ifdef DIAGNOSTIC
156 	if (tp != NULL && (tp->tun_flags & (TUN_INITED|TUN_OPEN)) != TUN_OPEN)
157 		printf("tun%d: inconsistent flags: %x\n", unit, tp->tun_flags);
158 #endif
159 
160 	return (tp);
161 }
162 
163 static int
164 tun_clone_create(struct if_clone *ifc, int unit)
165 {
166 	struct tun_softc *tp;
167 
168 	if ((tp = tun_find_zunit(unit)) == NULL) {
169 		/* Allocate a new instance */
170 		tp = malloc(sizeof(*tp), M_DEVBUF, M_WAITOK|M_ZERO);
171 
172 		tp->tun_unit = unit;
173 		simple_lock_init(&tp->tun_lock);
174 		selinit(&tp->tun_rsel);
175 		selinit(&tp->tun_wsel);
176 	} else {
177 		/* Revive tunnel instance; clear ifp part */
178 		(void)memset(&tp->tun_if, 0, sizeof(struct ifnet));
179 	}
180 
181 	if_initname(&tp->tun_if, ifc->ifc_name, unit);
182 	tunattach0(tp);
183 	tp->tun_flags |= TUN_INITED;
184 	tp->tun_osih = softint_establish(SOFTINT_CLOCK, tun_o_softintr, tp);
185 	tp->tun_isih = softint_establish(SOFTINT_CLOCK, tun_i_softintr, tp);
186 
187 	simple_lock(&tun_softc_lock);
188 	LIST_INSERT_HEAD(&tun_softc_list, tp, tun_list);
189 	simple_unlock(&tun_softc_lock);
190 
191 	return (0);
192 }
193 
194 static void
195 tunattach0(struct tun_softc *tp)
196 {
197 	struct ifnet *ifp;
198 
199 	ifp = &tp->tun_if;
200 	ifp->if_softc = tp;
201 	ifp->if_mtu = TUNMTU;
202 	ifp->if_ioctl = tun_ioctl;
203 	ifp->if_output = tun_output;
204 #ifdef ALTQ
205 	ifp->if_start = tunstart;
206 #endif
207 	ifp->if_flags = IFF_POINTOPOINT;
208 	ifp->if_type = IFT_TUNNEL;
209 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
210 	ifp->if_collisions = 0;
211 	ifp->if_ierrors = 0;
212 	ifp->if_oerrors = 0;
213 	ifp->if_ipackets = 0;
214 	ifp->if_opackets = 0;
215 	ifp->if_ibytes   = 0;
216 	ifp->if_obytes   = 0;
217 	ifp->if_dlt = DLT_NULL;
218 	IFQ_SET_READY(&ifp->if_snd);
219 	if_attach(ifp);
220 	if_alloc_sadl(ifp);
221 #if NBPFILTER > 0
222 	bpfattach(ifp, DLT_NULL, sizeof(uint32_t));
223 #endif
224 }
225 
226 static int
227 tun_clone_destroy(struct ifnet *ifp)
228 {
229 	struct tun_softc *tp = (void *)ifp;
230 	int s, zombie = 0;
231 
232 	s = splnet();
233 	simple_lock(&tun_softc_lock);
234 	simple_lock(&tp->tun_lock);
235 	LIST_REMOVE(tp, tun_list);
236 	if (tp->tun_flags & TUN_OPEN) {
237 		/* Hang on to storage until last close */
238 		zombie = 1;
239 		tp->tun_flags &= ~TUN_INITED;
240 		LIST_INSERT_HEAD(&tunz_softc_list, tp, tun_list);
241 	}
242 	simple_unlock(&tun_softc_lock);
243 
244 	IF_PURGE(&ifp->if_snd);
245 	ifp->if_flags &= ~IFF_RUNNING;
246 
247 	if (tp->tun_flags & TUN_RWAIT) {
248 		tp->tun_flags &= ~TUN_RWAIT;
249 		wakeup((void *)tp);
250 	}
251 	selnotify(&tp->tun_rsel, 0, 0);
252 
253 	simple_unlock(&tp->tun_lock);
254 	splx(s);
255 
256 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
257 		fownsignal(tp->tun_pgid, SIGIO, POLL_HUP, 0, NULL);
258 
259 #if NBPFILTER > 0
260 	bpfdetach(ifp);
261 #endif
262 	if_detach(ifp);
263 
264 	if (!zombie) {
265 		seldestroy(&tp->tun_rsel);
266 		seldestroy(&tp->tun_wsel);
267 		softint_disestablish(tp->tun_osih);
268 		softint_disestablish(tp->tun_isih);
269 		free(tp, M_DEVBUF);
270 	}
271 
272 	return (0);
273 }
274 
275 /*
276  * tunnel open - must be superuser & the device must be
277  * configured in
278  */
279 static int
280 tunopen(dev_t dev, int flag, int mode, struct lwp *l)
281 {
282 	struct ifnet	*ifp;
283 	struct tun_softc *tp;
284 	int	s, error;
285 
286 	error = kauth_authorize_network(l->l_cred, KAUTH_NETWORK_INTERFACE_TUN,
287 	    KAUTH_REQ_NETWORK_INTERFACE_TUN_ADD, NULL, NULL, NULL);
288 	if (error)
289 		return (error);
290 
291 	s = splnet();
292 	tp = tun_find_unit(dev);
293 
294 	if (tp == NULL) {
295 		(void)tun_clone_create(&tun_cloner, minor(dev));
296 		tp = tun_find_unit(dev);
297 		if (tp == NULL) {
298 			error = ENXIO;
299 			goto out_nolock;
300 		}
301 	}
302 
303 	if (tp->tun_flags & TUN_OPEN) {
304 		error = EBUSY;
305 		goto out;
306 	}
307 
308 	ifp = &tp->tun_if;
309 	tp->tun_flags |= TUN_OPEN;
310 	TUNDEBUG("%s: open\n", ifp->if_xname);
311 out:
312 	simple_unlock(&tp->tun_lock);
313 out_nolock:
314 	splx(s);
315 	return (error);
316 }
317 
318 /*
319  * tunclose - close the device - mark i/f down & delete
320  * routing info
321  */
322 int
323 tunclose(dev_t dev, int flag, int mode,
324     struct lwp *l)
325 {
326 	int	s;
327 	struct tun_softc *tp;
328 	struct ifnet	*ifp;
329 
330 	s = splnet();
331 	if ((tp = tun_find_zunit(minor(dev))) != NULL) {
332 		/* interface was "destroyed" before the close */
333 		seldestroy(&tp->tun_rsel);
334 		seldestroy(&tp->tun_wsel);
335 		softint_disestablish(tp->tun_osih);
336 		softint_disestablish(tp->tun_isih);
337 		free(tp, M_DEVBUF);
338 		goto out_nolock;
339 	}
340 
341 	if ((tp = tun_find_unit(dev)) == NULL)
342 		goto out_nolock;
343 
344 	ifp = &tp->tun_if;
345 
346 	tp->tun_flags &= ~TUN_OPEN;
347 
348 	/*
349 	 * junk all pending output
350 	 */
351 	IFQ_PURGE(&ifp->if_snd);
352 
353 	if (ifp->if_flags & IFF_UP) {
354 		if_down(ifp);
355 		if (ifp->if_flags & IFF_RUNNING) {
356 			/* find internet addresses and delete routes */
357 			struct ifaddr *ifa;
358 			IFADDR_FOREACH(ifa, ifp) {
359 #if defined(INET) || defined(INET6)
360 				if (ifa->ifa_addr->sa_family == AF_INET ||
361 				    ifa->ifa_addr->sa_family == AF_INET6) {
362 					rtinit(ifa, (int)RTM_DELETE,
363 					       tp->tun_flags & TUN_DSTADDR
364 							? RTF_HOST
365 							: 0);
366 				}
367 #endif
368 			}
369 		}
370 	}
371 	tp->tun_pgid = 0;
372 	selnotify(&tp->tun_rsel, 0, 0);
373 
374 	TUNDEBUG ("%s: closed\n", ifp->if_xname);
375 	simple_unlock(&tp->tun_lock);
376 out_nolock:
377 	splx(s);
378 	return (0);
379 }
380 
381 /*
382  * Call at splnet().
383  */
384 static void
385 tuninit(struct tun_softc *tp)
386 {
387 	struct ifnet	*ifp = &tp->tun_if;
388 	struct ifaddr	*ifa;
389 
390 	TUNDEBUG("%s: tuninit\n", ifp->if_xname);
391 
392 	simple_lock(&tp->tun_lock);
393 	ifp->if_flags |= IFF_UP | IFF_RUNNING;
394 
395 	tp->tun_flags &= ~(TUN_IASET|TUN_DSTADDR);
396 	IFADDR_FOREACH(ifa, ifp) {
397 #ifdef INET
398 		if (ifa->ifa_addr->sa_family == AF_INET) {
399 			struct sockaddr_in *sin;
400 
401 			sin = satosin(ifa->ifa_addr);
402 			if (sin && sin->sin_addr.s_addr)
403 				tp->tun_flags |= TUN_IASET;
404 
405 			if (ifp->if_flags & IFF_POINTOPOINT) {
406 				sin = satosin(ifa->ifa_dstaddr);
407 				if (sin && sin->sin_addr.s_addr)
408 					tp->tun_flags |= TUN_DSTADDR;
409 			}
410 		}
411 #endif
412 #ifdef INET6
413 		if (ifa->ifa_addr->sa_family == AF_INET6) {
414 			struct sockaddr_in6 *sin;
415 
416 			sin = (struct sockaddr_in6 *)ifa->ifa_addr;
417 			if (!IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
418 				tp->tun_flags |= TUN_IASET;
419 
420 			if (ifp->if_flags & IFF_POINTOPOINT) {
421 				sin = (struct sockaddr_in6 *)ifa->ifa_dstaddr;
422 				if (sin &&
423 				    !IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
424 					tp->tun_flags |= TUN_DSTADDR;
425 			} else
426 				tp->tun_flags &= ~TUN_DSTADDR;
427 		}
428 #endif /* INET6 */
429 	}
430 
431 	simple_unlock(&tp->tun_lock);
432 	return;
433 }
434 
435 /*
436  * Process an ioctl request.
437  */
438 static int
439 tun_ioctl(struct ifnet *ifp, u_long cmd, void *data)
440 {
441 	int		error = 0, s;
442 	struct tun_softc *tp = (struct tun_softc *)(ifp->if_softc);
443 	struct ifreq *ifr = data;
444 
445 	s = splnet();
446 
447 	switch (cmd) {
448 	case SIOCINITIFADDR:
449 		tuninit(tp);
450 		TUNDEBUG("%s: address set\n", ifp->if_xname);
451 		break;
452 	case SIOCSIFDSTADDR:
453 		tuninit(tp);
454 		TUNDEBUG("%s: destination address set\n", ifp->if_xname);
455 		break;
456 	case SIOCSIFBRDADDR:
457 		TUNDEBUG("%s: broadcast address set\n", ifp->if_xname);
458 		break;
459 	case SIOCSIFMTU:
460 		if (ifr->ifr_mtu > TUNMTU || ifr->ifr_mtu < 576) {
461 			error = EINVAL;
462 			break;
463 		}
464 		TUNDEBUG("%s: interface mtu set\n", ifp->if_xname);
465 		if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET)
466 			error = 0;
467 		break;
468 	case SIOCADDMULTI:
469 	case SIOCDELMULTI:
470 		if (ifr == NULL) {
471 	        	error = EAFNOSUPPORT;           /* XXX */
472 			break;
473 		}
474 		switch (ifreq_getaddr(cmd, ifr)->sa_family) {
475 #ifdef INET
476 		case AF_INET:
477 			break;
478 #endif
479 #ifdef INET6
480 		case AF_INET6:
481 			break;
482 #endif
483 		default:
484 			error = EAFNOSUPPORT;
485 			break;
486 		}
487 		break;
488 	default:
489 		error = ifioctl_common(ifp, cmd, data);
490 	}
491 
492 	splx(s);
493 	return (error);
494 }
495 
496 /*
497  * tun_output - queue packets from higher level ready to put out.
498  */
499 static int
500 tun_output(struct ifnet *ifp, struct mbuf *m0, const struct sockaddr *dst,
501     struct rtentry *rt)
502 {
503 	struct tun_softc *tp = ifp->if_softc;
504 	int		s;
505 	int		error;
506 #if defined(INET) || defined(INET6)
507 	int		mlen;
508 	uint32_t	*af;
509 #endif
510 	ALTQ_DECL(struct altq_pktattr pktattr;)
511 
512 	s = splnet();
513 	simple_lock(&tp->tun_lock);
514 	TUNDEBUG ("%s: tun_output\n", ifp->if_xname);
515 
516 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
517 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname,
518 			  tp->tun_flags);
519 		m_freem (m0);
520 		error = EHOSTDOWN;
521 		goto out;
522 	}
523 
524 	/*
525 	 * if the queueing discipline needs packet classification,
526 	 * do it before prepending link headers.
527 	 */
528 	IFQ_CLASSIFY(&ifp->if_snd, m0, dst->sa_family, &pktattr);
529 
530 #if NBPFILTER > 0
531 	if (ifp->if_bpf)
532 		bpf_mtap_af(ifp->if_bpf, dst->sa_family, m0);
533 #endif
534 
535 	switch(dst->sa_family) {
536 #ifdef INET6
537 	case AF_INET6:
538 #endif
539 #ifdef INET
540 	case AF_INET:
541 #endif
542 #if defined(INET) || defined(INET6)
543 		if (tp->tun_flags & TUN_PREPADDR) {
544 			/* Simple link-layer header */
545 			M_PREPEND(m0, dst->sa_len, M_DONTWAIT);
546 			if (m0 == NULL) {
547 				IF_DROP(&ifp->if_snd);
548 				error = ENOBUFS;
549 				goto out;
550 			}
551 			bcopy(dst, mtod(m0, char *), dst->sa_len);
552 		}
553 
554 		if (tp->tun_flags & TUN_IFHEAD) {
555 			/* Prepend the address family */
556 			M_PREPEND(m0, sizeof(*af), M_DONTWAIT);
557 			if (m0 == NULL) {
558 				IF_DROP(&ifp->if_snd);
559 				error = ENOBUFS;
560 				goto out;
561 			}
562 			af = mtod(m0,uint32_t *);
563 			*af = htonl(dst->sa_family);
564 		} else {
565 #ifdef INET
566 			if (dst->sa_family != AF_INET)
567 #endif
568 			{
569 				m_freem(m0);
570 				error = EAFNOSUPPORT;
571 				goto out;
572 			}
573 		}
574 		/* FALLTHROUGH */
575 	case AF_UNSPEC:
576 		IFQ_ENQUEUE(&ifp->if_snd, m0, &pktattr, error);
577 		if (error) {
578 			ifp->if_collisions++;
579 			error = EAFNOSUPPORT;
580 			goto out;
581 		}
582 		mlen = m0->m_pkthdr.len;
583 		ifp->if_opackets++;
584 		ifp->if_obytes += mlen;
585 		break;
586 #endif
587 	default:
588 		m_freem(m0);
589 		error = EAFNOSUPPORT;
590 		goto out;
591 	}
592 
593 	if (tp->tun_flags & TUN_RWAIT) {
594 		tp->tun_flags &= ~TUN_RWAIT;
595 		wakeup((void *)tp);
596 	}
597 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
598 		softint_schedule(tp->tun_isih);
599 
600 	selnotify(&tp->tun_rsel, 0, 0);
601 out:
602 	simple_unlock(&tp->tun_lock);
603 	splx(s);
604 	return (0);
605 }
606 
607 static void
608 tun_i_softintr(void *cookie)
609 {
610 	struct tun_softc *tp = cookie;
611 
612 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
613 		fownsignal(tp->tun_pgid, SIGIO, POLL_IN, POLLIN|POLLRDNORM,
614 		    NULL);
615 }
616 
617 static void
618 tun_o_softintr(void *cookie)
619 {
620 	struct tun_softc *tp = cookie;
621 
622 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
623 		fownsignal(tp->tun_pgid, SIGIO, POLL_OUT, POLLOUT|POLLWRNORM,
624 		    NULL);
625 }
626 
627 /*
628  * the cdevsw interface is now pretty minimal.
629  */
630 int
631 tunioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
632 {
633 	struct tun_softc *tp;
634 	int s, error = 0;
635 
636 	s = splnet();
637 	tp = tun_find_unit(dev);
638 
639 	/* interface was "destroyed" already */
640 	if (tp == NULL) {
641 		error = ENXIO;
642 		goto out_nolock;
643 	}
644 
645 	switch (cmd) {
646 	case TUNSDEBUG:
647 		tundebug = *(int *)data;
648 		break;
649 
650 	case TUNGDEBUG:
651 		*(int *)data = tundebug;
652 		break;
653 
654 	case TUNSIFMODE:
655 		switch (*(int *)data & (IFF_POINTOPOINT|IFF_BROADCAST)) {
656 		case IFF_POINTOPOINT:
657 		case IFF_BROADCAST:
658 			if (tp->tun_if.if_flags & IFF_UP) {
659 				error = EBUSY;
660 				goto out;
661 			}
662 			tp->tun_if.if_flags &=
663 				~(IFF_BROADCAST|IFF_POINTOPOINT|IFF_MULTICAST);
664 			tp->tun_if.if_flags |= *(int *)data;
665 			break;
666 		default:
667 			error = EINVAL;
668 			goto out;
669 		}
670 		break;
671 
672 	case TUNSLMODE:
673 		if (*(int *)data) {
674 			tp->tun_flags |= TUN_PREPADDR;
675 			tp->tun_flags &= ~TUN_IFHEAD;
676 		} else
677 			tp->tun_flags &= ~TUN_PREPADDR;
678 		break;
679 
680 	case TUNSIFHEAD:
681 		if (*(int *)data) {
682 			tp->tun_flags |= TUN_IFHEAD;
683 			tp->tun_flags &= ~TUN_PREPADDR;
684 		} else
685 			tp->tun_flags &= ~TUN_IFHEAD;
686 		break;
687 
688 	case TUNGIFHEAD:
689 		*(int *)data = (tp->tun_flags & TUN_IFHEAD);
690 		break;
691 
692 	case FIONBIO:
693 		if (*(int *)data)
694 			tp->tun_flags |= TUN_NBIO;
695 		else
696 			tp->tun_flags &= ~TUN_NBIO;
697 		break;
698 
699 	case FIOASYNC:
700 		if (*(int *)data)
701 			tp->tun_flags |= TUN_ASYNC;
702 		else
703 			tp->tun_flags &= ~TUN_ASYNC;
704 		break;
705 
706 	case FIONREAD:
707 		if (tp->tun_if.if_snd.ifq_head)
708 			*(int *)data = tp->tun_if.if_snd.ifq_head->m_pkthdr.len;
709 		else
710 			*(int *)data = 0;
711 		break;
712 
713 	case TIOCSPGRP:
714 	case FIOSETOWN:
715 		error = fsetown(&tp->tun_pgid, cmd, data);
716 		break;
717 
718 	case TIOCGPGRP:
719 	case FIOGETOWN:
720 		error = fgetown(tp->tun_pgid, cmd, data);
721 		break;
722 
723 	default:
724 		error = ENOTTY;
725 	}
726 
727 out:
728 	simple_unlock(&tp->tun_lock);
729 out_nolock:
730 	splx(s);
731 	return (error);
732 }
733 
734 /*
735  * The cdevsw read interface - reads a packet at a time, or at
736  * least as much of a packet as can be read.
737  */
738 int
739 tunread(dev_t dev, struct uio *uio, int ioflag)
740 {
741 	struct tun_softc *tp;
742 	struct ifnet	*ifp;
743 	struct mbuf	*m, *m0;
744 	int		error = 0, len, s, index;
745 
746 	s = splnet();
747 	tp = tun_find_unit(dev);
748 
749 	/* interface was "destroyed" already */
750 	if (tp == NULL) {
751 		error = ENXIO;
752 		goto out_nolock;
753 	}
754 
755 	index = tp->tun_if.if_index;
756 	ifp = &tp->tun_if;
757 
758 	TUNDEBUG ("%s: read\n", ifp->if_xname);
759 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
760 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname, tp->tun_flags);
761 		error = EHOSTDOWN;
762 		goto out;
763 	}
764 
765 	tp->tun_flags &= ~TUN_RWAIT;
766 
767 	do {
768 		IFQ_DEQUEUE(&ifp->if_snd, m0);
769 		if (m0 == 0) {
770 			if (tp->tun_flags & TUN_NBIO) {
771 				error = EWOULDBLOCK;
772 				goto out;
773 			}
774 			tp->tun_flags |= TUN_RWAIT;
775 			if (ltsleep((void *)tp, PZERO|PCATCH|PNORELOCK,
776 					"tunread", 0, &tp->tun_lock) != 0) {
777 				error = EINTR;
778 				goto out_nolock;
779 			} else {
780 				/*
781 				 * Maybe the interface was destroyed while
782 				 * we were sleeping, so let's ensure that
783 				 * we're looking at the same (valid) tun
784 				 * interface before looping.
785 				 */
786 				tp = tun_find_unit(dev);
787 				if (tp == NULL) {
788 					error = ENXIO;
789 					goto out_nolock;
790 				}
791 				if (tp->tun_if.if_index != index) {
792 					error = ENXIO;
793 					goto out;
794 				}
795 			}
796 		}
797 	} while (m0 == 0);
798 
799 	simple_unlock(&tp->tun_lock);
800 	splx(s);
801 
802 	/* Copy the mbuf chain */
803 	while (m0 && uio->uio_resid > 0 && error == 0) {
804 		len = min(uio->uio_resid, m0->m_len);
805 		if (len != 0)
806 			error = uiomove(mtod(m0, void *), len, uio);
807 		MFREE(m0, m);
808 		m0 = m;
809 	}
810 
811 	if (m0) {
812 		TUNDEBUG("Dropping mbuf\n");
813 		m_freem(m0);
814 	}
815 	if (error)
816 		ifp->if_ierrors++;
817 
818 	return (error);
819 
820 out:
821 	simple_unlock(&tp->tun_lock);
822 out_nolock:
823 	splx(s);
824 	return (error);
825 }
826 
827 /*
828  * the cdevsw write interface - an atomic write is a packet - or else!
829  */
830 int
831 tunwrite(dev_t dev, struct uio *uio, int ioflag)
832 {
833 	struct tun_softc *tp;
834 	struct ifnet	*ifp;
835 	struct mbuf	*top, **mp, *m;
836 	struct ifqueue	*ifq;
837 	struct sockaddr	dst;
838 	int		isr, error = 0, s, tlen, mlen;
839 	uint32_t	family;
840 
841 	s = splnet();
842 	tp = tun_find_unit(dev);
843 
844 	/* interface was "destroyed" already */
845 	if (tp == NULL) {
846 		error = ENXIO;
847 		goto out_nolock;
848 	}
849 
850 	/* Unlock until we've got the data */
851 	simple_unlock(&tp->tun_lock);
852 	splx(s);
853 
854 	ifp = &tp->tun_if;
855 
856 	TUNDEBUG("%s: tunwrite\n", ifp->if_xname);
857 
858 	if (tp->tun_flags & TUN_PREPADDR) {
859 		if (uio->uio_resid < sizeof(dst)) {
860 			error = EIO;
861 			goto out0;
862 		}
863 		error = uiomove((void *)&dst, sizeof(dst), uio);
864 		if (dst.sa_len > sizeof(dst)) {
865 			/* Duh.. */
866 			char discard;
867 			int n = dst.sa_len - sizeof(dst);
868 			while (n--)
869 				if ((error = uiomove(&discard, 1, uio)) != 0) {
870 					goto out0;
871 				}
872 		}
873 	} else if (tp->tun_flags & TUN_IFHEAD) {
874 		if (uio->uio_resid < sizeof(family)){
875 			error = EIO;
876 			goto out0;
877 		}
878 		error = uiomove((void *)&family, sizeof(family), uio);
879 		dst.sa_family = ntohl(family);
880 	} else {
881 #ifdef INET
882 		dst.sa_family = AF_INET;
883 #endif
884 	}
885 
886 	if (uio->uio_resid > TUNMTU) {
887 		TUNDEBUG("%s: len=%lu!\n", ifp->if_xname,
888 		    (unsigned long)uio->uio_resid);
889 		error = EIO;
890 		goto out0;
891 	}
892 
893 	switch (dst.sa_family) {
894 #ifdef INET
895 	case AF_INET:
896 		ifq = &ipintrq;
897 		isr = NETISR_IP;
898 		break;
899 #endif
900 #ifdef INET6
901 	case AF_INET6:
902 		ifq = &ip6intrq;
903 		isr = NETISR_IPV6;
904 		break;
905 #endif
906 	default:
907 		error = EAFNOSUPPORT;
908 		goto out0;
909 	}
910 
911 	tlen = uio->uio_resid;
912 
913 	/* get a header mbuf */
914 	MGETHDR(m, M_DONTWAIT, MT_DATA);
915 	if (m == NULL) {
916 		error = ENOBUFS;
917 		goto out0;
918 	}
919 	mlen = MHLEN;
920 
921 	top = NULL;
922 	mp = &top;
923 	while (error == 0 && uio->uio_resid > 0) {
924 		m->m_len = min(mlen, uio->uio_resid);
925 		error = uiomove(mtod(m, void *), m->m_len, uio);
926 		*mp = m;
927 		mp = &m->m_next;
928 		if (error == 0 && uio->uio_resid > 0) {
929 			MGET(m, M_DONTWAIT, MT_DATA);
930 			if (m == NULL) {
931 				error = ENOBUFS;
932 				break;
933 			}
934 			mlen = MLEN;
935 		}
936 	}
937 	if (error) {
938 		if (top != NULL)
939 			m_freem (top);
940 		ifp->if_ierrors++;
941 		goto out0;
942 	}
943 
944 	top->m_pkthdr.len = tlen;
945 	top->m_pkthdr.rcvif = ifp;
946 
947 #if NBPFILTER > 0
948 	if (ifp->if_bpf)
949 		bpf_mtap_af(ifp->if_bpf, dst.sa_family, top);
950 #endif
951 
952 	s = splnet();
953 	simple_lock(&tp->tun_lock);
954 	if ((tp->tun_flags & TUN_INITED) == 0) {
955 		/* Interface was destroyed */
956 		error = ENXIO;
957 		goto out;
958 	}
959 	if (IF_QFULL(ifq)) {
960 		IF_DROP(ifq);
961 		ifp->if_collisions++;
962 		m_freem(top);
963 		error = ENOBUFS;
964 		goto out;
965 	}
966 
967 	IF_ENQUEUE(ifq, top);
968 	ifp->if_ipackets++;
969 	ifp->if_ibytes += tlen;
970 	schednetisr(isr);
971 out:
972 	simple_unlock(&tp->tun_lock);
973 out_nolock:
974 	splx(s);
975 out0:
976 	return (error);
977 }
978 
979 #ifdef ALTQ
980 /*
981  * Start packet transmission on the interface.
982  * when the interface queue is rate-limited by ALTQ or TBR,
983  * if_start is needed to drain packets from the queue in order
984  * to notify readers when outgoing packets become ready.
985  *
986  * Should be called at splnet.
987  */
988 static void
989 tunstart(struct ifnet *ifp)
990 {
991 	struct tun_softc *tp = ifp->if_softc;
992 
993 	if (!ALTQ_IS_ENABLED(&ifp->if_snd) && !TBR_IS_ENABLED(&ifp->if_snd))
994 		return;
995 
996 	simple_lock(&tp->tun_lock);
997 	if (!IF_IS_EMPTY(&ifp->if_snd)) {
998 		if (tp->tun_flags & TUN_RWAIT) {
999 			tp->tun_flags &= ~TUN_RWAIT;
1000 			wakeup((void *)tp);
1001 		}
1002 		if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
1003 			softint_schedule(tp->tun_osih);
1004 
1005 		selnotify(&tp->tun_rsel, 0, 0);
1006 	}
1007 	simple_unlock(&tp->tun_lock);
1008 }
1009 #endif /* ALTQ */
1010 /*
1011  * tunpoll - the poll interface, this is only useful on reads
1012  * really. The write detect always returns true, write never blocks
1013  * anyway, it either accepts the packet or drops it.
1014  */
1015 int
1016 tunpoll(dev_t dev, int events, struct lwp *l)
1017 {
1018 	struct tun_softc *tp;
1019 	struct ifnet	*ifp;
1020 	int		s, revents = 0;
1021 
1022 	s = splnet();
1023 	tp = tun_find_unit(dev);
1024 
1025 	/* interface was "destroyed" already */
1026 	if (tp == NULL)
1027 		goto out_nolock;
1028 
1029 	ifp = &tp->tun_if;
1030 
1031 	TUNDEBUG("%s: tunpoll\n", ifp->if_xname);
1032 
1033 	if (events & (POLLIN | POLLRDNORM)) {
1034 		if (!IFQ_IS_EMPTY(&ifp->if_snd)) {
1035 			TUNDEBUG("%s: tunpoll q=%d\n", ifp->if_xname,
1036 			    ifp->if_snd.ifq_len);
1037 			revents |= events & (POLLIN | POLLRDNORM);
1038 		} else {
1039 			TUNDEBUG("%s: tunpoll waiting\n", ifp->if_xname);
1040 			selrecord(l, &tp->tun_rsel);
1041 		}
1042 	}
1043 
1044 	if (events & (POLLOUT | POLLWRNORM))
1045 		revents |= events & (POLLOUT | POLLWRNORM);
1046 
1047 	simple_unlock(&tp->tun_lock);
1048 out_nolock:
1049 	splx(s);
1050 	return (revents);
1051 }
1052 
1053 static void
1054 filt_tunrdetach(struct knote *kn)
1055 {
1056 	struct tun_softc *tp = kn->kn_hook;
1057 	int s;
1058 
1059 	s = splnet();
1060 	SLIST_REMOVE(&tp->tun_rsel.sel_klist, kn, knote, kn_selnext);
1061 	splx(s);
1062 }
1063 
1064 static int
1065 filt_tunread(struct knote *kn, long hint)
1066 {
1067 	struct tun_softc *tp = kn->kn_hook;
1068 	struct ifnet *ifp = &tp->tun_if;
1069 	struct mbuf *m;
1070 	int s;
1071 
1072 	s = splnet();
1073 	IF_POLL(&ifp->if_snd, m);
1074 	if (m == NULL) {
1075 		splx(s);
1076 		return (0);
1077 	}
1078 
1079 	for (kn->kn_data = 0; m != NULL; m = m->m_next)
1080 		kn->kn_data += m->m_len;
1081 
1082 	splx(s);
1083 	return (1);
1084 }
1085 
1086 static const struct filterops tunread_filtops =
1087 	{ 1, NULL, filt_tunrdetach, filt_tunread };
1088 
1089 static const struct filterops tun_seltrue_filtops =
1090 	{ 1, NULL, filt_tunrdetach, filt_seltrue };
1091 
1092 int
1093 tunkqfilter(dev_t dev, struct knote *kn)
1094 {
1095 	struct tun_softc *tp;
1096 	struct klist *klist;
1097 	int rv = 0, s;
1098 
1099 	s = splnet();
1100 	tp = tun_find_unit(dev);
1101 	if (tp == NULL)
1102 		goto out_nolock;
1103 
1104 	switch (kn->kn_filter) {
1105 	case EVFILT_READ:
1106 		klist = &tp->tun_rsel.sel_klist;
1107 		kn->kn_fop = &tunread_filtops;
1108 		break;
1109 
1110 	case EVFILT_WRITE:
1111 		klist = &tp->tun_rsel.sel_klist;
1112 		kn->kn_fop = &tun_seltrue_filtops;
1113 		break;
1114 
1115 	default:
1116 		rv = EINVAL;
1117 		goto out;
1118 	}
1119 
1120 	kn->kn_hook = tp;
1121 
1122 	SLIST_INSERT_HEAD(klist, kn, kn_selnext);
1123 
1124 out:
1125 	simple_unlock(&tp->tun_lock);
1126 out_nolock:
1127 	splx(s);
1128 	return (rv);
1129 }
1130