xref: /netbsd-src/sys/net/if_tun.c (revision c2f76ff004a2cb67efe5b12d97bd3ef7fe89e18d)
1 /*	$NetBSD: if_tun.c,v 1.113 2010/04/05 07:22:24 joerg 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 1.113 2010/04/05 07:22:24 joerg 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 <sys/time.h>
57 #include <net/bpf.h>
58 
59 #include <net/if_tun.h>
60 
61 #define TUNDEBUG	if (tundebug) printf
62 int	tundebug = 0;
63 
64 extern int ifqmaxlen;
65 void	tunattach(int);
66 
67 static LIST_HEAD(, tun_softc) tun_softc_list;
68 static LIST_HEAD(, tun_softc) tunz_softc_list;
69 static struct simplelock tun_softc_lock;
70 
71 static int	tun_ioctl(struct ifnet *, u_long, void *);
72 static int	tun_output(struct ifnet *, struct mbuf *,
73 			const struct sockaddr *, struct rtentry *rt);
74 static int	tun_clone_create(struct if_clone *, int);
75 static int	tun_clone_destroy(struct ifnet *);
76 
77 static struct if_clone tun_cloner =
78     IF_CLONE_INITIALIZER("tun", tun_clone_create, tun_clone_destroy);
79 
80 static void tunattach0(struct tun_softc *);
81 static void tuninit(struct tun_softc *);
82 static void tun_i_softintr(void *);
83 static void tun_o_softintr(void *);
84 #ifdef ALTQ
85 static void tunstart(struct ifnet *);
86 #endif
87 static struct tun_softc *tun_find_unit(dev_t);
88 static struct tun_softc *tun_find_zunit(int);
89 
90 static dev_type_open(tunopen);
91 static dev_type_close(tunclose);
92 static dev_type_read(tunread);
93 static dev_type_write(tunwrite);
94 static dev_type_ioctl(tunioctl);
95 static dev_type_poll(tunpoll);
96 static dev_type_kqfilter(tunkqfilter);
97 
98 const struct cdevsw tun_cdevsw = {
99 	tunopen, tunclose, tunread, tunwrite, tunioctl,
100 	nostop, notty, tunpoll, nommap, tunkqfilter, D_OTHER,
101 };
102 
103 void
104 tunattach(int unused)
105 {
106 
107 	simple_lock_init(&tun_softc_lock);
108 	LIST_INIT(&tun_softc_list);
109 	LIST_INIT(&tunz_softc_list);
110 	if_clone_attach(&tun_cloner);
111 }
112 
113 /*
114  * Find driver instance from dev_t.
115  * Call at splnet().
116  * Returns with tp locked (if found).
117  */
118 static struct tun_softc *
119 tun_find_unit(dev_t dev)
120 {
121 	struct tun_softc *tp;
122 	int unit = minor(dev);
123 
124 	simple_lock(&tun_softc_lock);
125 	LIST_FOREACH(tp, &tun_softc_list, tun_list)
126 		if (unit == tp->tun_unit)
127 			break;
128 	if (tp)
129 		simple_lock(&tp->tun_lock);
130 	simple_unlock(&tun_softc_lock);
131 
132 	return (tp);
133 }
134 
135 /*
136  * Find zombie driver instance by unit number.
137  * Call at splnet().
138  * Remove tp from list and return it unlocked (if found).
139  */
140 static struct tun_softc *
141 tun_find_zunit(int unit)
142 {
143 	struct tun_softc *tp;
144 
145 	simple_lock(&tun_softc_lock);
146 	LIST_FOREACH(tp, &tunz_softc_list, tun_list)
147 		if (unit == tp->tun_unit)
148 			break;
149 	if (tp)
150 		LIST_REMOVE(tp, tun_list);
151 	simple_unlock(&tun_softc_lock);
152 #ifdef DIAGNOSTIC
153 	if (tp != NULL && (tp->tun_flags & (TUN_INITED|TUN_OPEN)) != TUN_OPEN)
154 		printf("tun%d: inconsistent flags: %x\n", unit, tp->tun_flags);
155 #endif
156 
157 	return (tp);
158 }
159 
160 static int
161 tun_clone_create(struct if_clone *ifc, int unit)
162 {
163 	struct tun_softc *tp;
164 
165 	if ((tp = tun_find_zunit(unit)) == NULL) {
166 		/* Allocate a new instance */
167 		tp = malloc(sizeof(*tp), M_DEVBUF, M_WAITOK|M_ZERO);
168 
169 		tp->tun_unit = unit;
170 		simple_lock_init(&tp->tun_lock);
171 		selinit(&tp->tun_rsel);
172 		selinit(&tp->tun_wsel);
173 	} else {
174 		/* Revive tunnel instance; clear ifp part */
175 		(void)memset(&tp->tun_if, 0, sizeof(struct ifnet));
176 	}
177 
178 	if_initname(&tp->tun_if, ifc->ifc_name, unit);
179 	tunattach0(tp);
180 	tp->tun_flags |= TUN_INITED;
181 	tp->tun_osih = softint_establish(SOFTINT_CLOCK, tun_o_softintr, tp);
182 	tp->tun_isih = softint_establish(SOFTINT_CLOCK, tun_i_softintr, tp);
183 
184 	simple_lock(&tun_softc_lock);
185 	LIST_INSERT_HEAD(&tun_softc_list, tp, tun_list);
186 	simple_unlock(&tun_softc_lock);
187 
188 	return (0);
189 }
190 
191 static void
192 tunattach0(struct tun_softc *tp)
193 {
194 	struct ifnet *ifp;
195 
196 	ifp = &tp->tun_if;
197 	ifp->if_softc = tp;
198 	ifp->if_mtu = TUNMTU;
199 	ifp->if_ioctl = tun_ioctl;
200 	ifp->if_output = tun_output;
201 #ifdef ALTQ
202 	ifp->if_start = tunstart;
203 #endif
204 	ifp->if_flags = IFF_POINTOPOINT;
205 	ifp->if_type = IFT_TUNNEL;
206 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
207 	ifp->if_collisions = 0;
208 	ifp->if_ierrors = 0;
209 	ifp->if_oerrors = 0;
210 	ifp->if_ipackets = 0;
211 	ifp->if_opackets = 0;
212 	ifp->if_ibytes   = 0;
213 	ifp->if_obytes   = 0;
214 	ifp->if_dlt = DLT_NULL;
215 	IFQ_SET_READY(&ifp->if_snd);
216 	if_attach(ifp);
217 	if_alloc_sadl(ifp);
218 	bpf_attach(ifp, DLT_NULL, sizeof(uint32_t));
219 }
220 
221 static int
222 tun_clone_destroy(struct ifnet *ifp)
223 {
224 	struct tun_softc *tp = (void *)ifp;
225 	int s, zombie = 0;
226 
227 	s = splnet();
228 	simple_lock(&tun_softc_lock);
229 	simple_lock(&tp->tun_lock);
230 	LIST_REMOVE(tp, tun_list);
231 	if (tp->tun_flags & TUN_OPEN) {
232 		/* Hang on to storage until last close */
233 		zombie = 1;
234 		tp->tun_flags &= ~TUN_INITED;
235 		LIST_INSERT_HEAD(&tunz_softc_list, tp, tun_list);
236 	}
237 	simple_unlock(&tun_softc_lock);
238 
239 	IF_PURGE(&ifp->if_snd);
240 	ifp->if_flags &= ~IFF_RUNNING;
241 
242 	if (tp->tun_flags & TUN_RWAIT) {
243 		tp->tun_flags &= ~TUN_RWAIT;
244 		wakeup((void *)tp);
245 	}
246 	selnotify(&tp->tun_rsel, 0, 0);
247 
248 	simple_unlock(&tp->tun_lock);
249 	splx(s);
250 
251 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
252 		fownsignal(tp->tun_pgid, SIGIO, POLL_HUP, 0, NULL);
253 
254 	bpf_detach(ifp);
255 	if_detach(ifp);
256 
257 	if (!zombie) {
258 		seldestroy(&tp->tun_rsel);
259 		seldestroy(&tp->tun_wsel);
260 		softint_disestablish(tp->tun_osih);
261 		softint_disestablish(tp->tun_isih);
262 		free(tp, M_DEVBUF);
263 	}
264 
265 	return (0);
266 }
267 
268 /*
269  * tunnel open - must be superuser & the device must be
270  * configured in
271  */
272 static int
273 tunopen(dev_t dev, int flag, int mode, struct lwp *l)
274 {
275 	struct ifnet	*ifp;
276 	struct tun_softc *tp;
277 	int	s, error;
278 
279 	error = kauth_authorize_network(l->l_cred, KAUTH_NETWORK_INTERFACE_TUN,
280 	    KAUTH_REQ_NETWORK_INTERFACE_TUN_ADD, NULL, NULL, NULL);
281 	if (error)
282 		return (error);
283 
284 	s = splnet();
285 	tp = tun_find_unit(dev);
286 
287 	if (tp == NULL) {
288 		(void)tun_clone_create(&tun_cloner, minor(dev));
289 		tp = tun_find_unit(dev);
290 		if (tp == NULL) {
291 			error = ENXIO;
292 			goto out_nolock;
293 		}
294 	}
295 
296 	if (tp->tun_flags & TUN_OPEN) {
297 		error = EBUSY;
298 		goto out;
299 	}
300 
301 	ifp = &tp->tun_if;
302 	tp->tun_flags |= TUN_OPEN;
303 	TUNDEBUG("%s: open\n", ifp->if_xname);
304 out:
305 	simple_unlock(&tp->tun_lock);
306 out_nolock:
307 	splx(s);
308 	return (error);
309 }
310 
311 /*
312  * tunclose - close the device - mark i/f down & delete
313  * routing info
314  */
315 int
316 tunclose(dev_t dev, int flag, int mode,
317     struct lwp *l)
318 {
319 	int	s;
320 	struct tun_softc *tp;
321 	struct ifnet	*ifp;
322 
323 	s = splnet();
324 	if ((tp = tun_find_zunit(minor(dev))) != NULL) {
325 		/* interface was "destroyed" before the close */
326 		seldestroy(&tp->tun_rsel);
327 		seldestroy(&tp->tun_wsel);
328 		softint_disestablish(tp->tun_osih);
329 		softint_disestablish(tp->tun_isih);
330 		free(tp, M_DEVBUF);
331 		goto out_nolock;
332 	}
333 
334 	if ((tp = tun_find_unit(dev)) == NULL)
335 		goto out_nolock;
336 
337 	ifp = &tp->tun_if;
338 
339 	tp->tun_flags &= ~TUN_OPEN;
340 
341 	/*
342 	 * junk all pending output
343 	 */
344 	IFQ_PURGE(&ifp->if_snd);
345 
346 	if (ifp->if_flags & IFF_UP) {
347 		if_down(ifp);
348 		if (ifp->if_flags & IFF_RUNNING) {
349 			/* find internet addresses and delete routes */
350 			struct ifaddr *ifa;
351 			IFADDR_FOREACH(ifa, ifp) {
352 #if defined(INET) || defined(INET6)
353 				if (ifa->ifa_addr->sa_family == AF_INET ||
354 				    ifa->ifa_addr->sa_family == AF_INET6) {
355 					rtinit(ifa, (int)RTM_DELETE,
356 					       tp->tun_flags & TUN_DSTADDR
357 							? RTF_HOST
358 							: 0);
359 				}
360 #endif
361 			}
362 		}
363 	}
364 	tp->tun_pgid = 0;
365 	selnotify(&tp->tun_rsel, 0, 0);
366 
367 	TUNDEBUG ("%s: closed\n", ifp->if_xname);
368 	simple_unlock(&tp->tun_lock);
369 out_nolock:
370 	splx(s);
371 	return (0);
372 }
373 
374 /*
375  * Call at splnet().
376  */
377 static void
378 tuninit(struct tun_softc *tp)
379 {
380 	struct ifnet	*ifp = &tp->tun_if;
381 	struct ifaddr	*ifa;
382 
383 	TUNDEBUG("%s: tuninit\n", ifp->if_xname);
384 
385 	simple_lock(&tp->tun_lock);
386 	ifp->if_flags |= IFF_UP | IFF_RUNNING;
387 
388 	tp->tun_flags &= ~(TUN_IASET|TUN_DSTADDR);
389 	IFADDR_FOREACH(ifa, ifp) {
390 #ifdef INET
391 		if (ifa->ifa_addr->sa_family == AF_INET) {
392 			struct sockaddr_in *sin;
393 
394 			sin = satosin(ifa->ifa_addr);
395 			if (sin && sin->sin_addr.s_addr)
396 				tp->tun_flags |= TUN_IASET;
397 
398 			if (ifp->if_flags & IFF_POINTOPOINT) {
399 				sin = satosin(ifa->ifa_dstaddr);
400 				if (sin && sin->sin_addr.s_addr)
401 					tp->tun_flags |= TUN_DSTADDR;
402 			}
403 		}
404 #endif
405 #ifdef INET6
406 		if (ifa->ifa_addr->sa_family == AF_INET6) {
407 			struct sockaddr_in6 *sin;
408 
409 			sin = (struct sockaddr_in6 *)ifa->ifa_addr;
410 			if (!IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
411 				tp->tun_flags |= TUN_IASET;
412 
413 			if (ifp->if_flags & IFF_POINTOPOINT) {
414 				sin = (struct sockaddr_in6 *)ifa->ifa_dstaddr;
415 				if (sin &&
416 				    !IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
417 					tp->tun_flags |= TUN_DSTADDR;
418 			} else
419 				tp->tun_flags &= ~TUN_DSTADDR;
420 		}
421 #endif /* INET6 */
422 	}
423 
424 	simple_unlock(&tp->tun_lock);
425 	return;
426 }
427 
428 /*
429  * Process an ioctl request.
430  */
431 static int
432 tun_ioctl(struct ifnet *ifp, u_long cmd, void *data)
433 {
434 	int		error = 0, s;
435 	struct tun_softc *tp = (struct tun_softc *)(ifp->if_softc);
436 	struct ifreq *ifr = data;
437 
438 	s = splnet();
439 
440 	switch (cmd) {
441 	case SIOCINITIFADDR:
442 		tuninit(tp);
443 		TUNDEBUG("%s: address set\n", ifp->if_xname);
444 		break;
445 	case SIOCSIFDSTADDR:
446 		tuninit(tp);
447 		TUNDEBUG("%s: destination address set\n", ifp->if_xname);
448 		break;
449 	case SIOCSIFBRDADDR:
450 		TUNDEBUG("%s: broadcast address set\n", ifp->if_xname);
451 		break;
452 	case SIOCSIFMTU:
453 		if (ifr->ifr_mtu > TUNMTU || ifr->ifr_mtu < 576) {
454 			error = EINVAL;
455 			break;
456 		}
457 		TUNDEBUG("%s: interface mtu set\n", ifp->if_xname);
458 		if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET)
459 			error = 0;
460 		break;
461 	case SIOCADDMULTI:
462 	case SIOCDELMULTI:
463 		if (ifr == NULL) {
464 	        	error = EAFNOSUPPORT;           /* XXX */
465 			break;
466 		}
467 		switch (ifreq_getaddr(cmd, ifr)->sa_family) {
468 #ifdef INET
469 		case AF_INET:
470 			break;
471 #endif
472 #ifdef INET6
473 		case AF_INET6:
474 			break;
475 #endif
476 		default:
477 			error = EAFNOSUPPORT;
478 			break;
479 		}
480 		break;
481 	default:
482 		error = ifioctl_common(ifp, cmd, data);
483 	}
484 
485 	splx(s);
486 	return (error);
487 }
488 
489 /*
490  * tun_output - queue packets from higher level ready to put out.
491  */
492 static int
493 tun_output(struct ifnet *ifp, struct mbuf *m0, const struct sockaddr *dst,
494     struct rtentry *rt)
495 {
496 	struct tun_softc *tp = ifp->if_softc;
497 	int		s;
498 	int		error;
499 #if defined(INET) || defined(INET6)
500 	int		mlen;
501 	uint32_t	*af;
502 #endif
503 	ALTQ_DECL(struct altq_pktattr pktattr;)
504 
505 	s = splnet();
506 	simple_lock(&tp->tun_lock);
507 	TUNDEBUG ("%s: tun_output\n", ifp->if_xname);
508 
509 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
510 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname,
511 			  tp->tun_flags);
512 		m_freem (m0);
513 		error = EHOSTDOWN;
514 		goto out;
515 	}
516 
517 	/*
518 	 * if the queueing discipline needs packet classification,
519 	 * do it before prepending link headers.
520 	 */
521 	IFQ_CLASSIFY(&ifp->if_snd, m0, dst->sa_family, &pktattr);
522 
523 	bpf_mtap_af(ifp, dst->sa_family, m0);
524 
525 	switch(dst->sa_family) {
526 #ifdef INET6
527 	case AF_INET6:
528 #endif
529 #ifdef INET
530 	case AF_INET:
531 #endif
532 #if defined(INET) || defined(INET6)
533 		if (tp->tun_flags & TUN_PREPADDR) {
534 			/* Simple link-layer header */
535 			M_PREPEND(m0, dst->sa_len, M_DONTWAIT);
536 			if (m0 == NULL) {
537 				IF_DROP(&ifp->if_snd);
538 				error = ENOBUFS;
539 				goto out;
540 			}
541 			bcopy(dst, mtod(m0, char *), dst->sa_len);
542 		}
543 
544 		if (tp->tun_flags & TUN_IFHEAD) {
545 			/* Prepend the address family */
546 			M_PREPEND(m0, sizeof(*af), M_DONTWAIT);
547 			if (m0 == NULL) {
548 				IF_DROP(&ifp->if_snd);
549 				error = ENOBUFS;
550 				goto out;
551 			}
552 			af = mtod(m0,uint32_t *);
553 			*af = htonl(dst->sa_family);
554 		} else {
555 #ifdef INET
556 			if (dst->sa_family != AF_INET)
557 #endif
558 			{
559 				m_freem(m0);
560 				error = EAFNOSUPPORT;
561 				goto out;
562 			}
563 		}
564 		/* FALLTHROUGH */
565 	case AF_UNSPEC:
566 		IFQ_ENQUEUE(&ifp->if_snd, m0, &pktattr, error);
567 		if (error) {
568 			ifp->if_collisions++;
569 			error = EAFNOSUPPORT;
570 			goto out;
571 		}
572 		mlen = m0->m_pkthdr.len;
573 		ifp->if_opackets++;
574 		ifp->if_obytes += mlen;
575 		break;
576 #endif
577 	default:
578 		m_freem(m0);
579 		error = EAFNOSUPPORT;
580 		goto out;
581 	}
582 
583 	if (tp->tun_flags & TUN_RWAIT) {
584 		tp->tun_flags &= ~TUN_RWAIT;
585 		wakeup((void *)tp);
586 	}
587 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
588 		softint_schedule(tp->tun_isih);
589 
590 	selnotify(&tp->tun_rsel, 0, 0);
591 out:
592 	simple_unlock(&tp->tun_lock);
593 	splx(s);
594 	return (0);
595 }
596 
597 static void
598 tun_i_softintr(void *cookie)
599 {
600 	struct tun_softc *tp = cookie;
601 
602 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
603 		fownsignal(tp->tun_pgid, SIGIO, POLL_IN, POLLIN|POLLRDNORM,
604 		    NULL);
605 }
606 
607 static void
608 tun_o_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_OUT, POLLOUT|POLLWRNORM,
614 		    NULL);
615 }
616 
617 /*
618  * the cdevsw interface is now pretty minimal.
619  */
620 int
621 tunioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
622 {
623 	struct tun_softc *tp;
624 	int s, error = 0;
625 
626 	s = splnet();
627 	tp = tun_find_unit(dev);
628 
629 	/* interface was "destroyed" already */
630 	if (tp == NULL) {
631 		error = ENXIO;
632 		goto out_nolock;
633 	}
634 
635 	switch (cmd) {
636 	case TUNSDEBUG:
637 		tundebug = *(int *)data;
638 		break;
639 
640 	case TUNGDEBUG:
641 		*(int *)data = tundebug;
642 		break;
643 
644 	case TUNSIFMODE:
645 		switch (*(int *)data & (IFF_POINTOPOINT|IFF_BROADCAST)) {
646 		case IFF_POINTOPOINT:
647 		case IFF_BROADCAST:
648 			if (tp->tun_if.if_flags & IFF_UP) {
649 				error = EBUSY;
650 				goto out;
651 			}
652 			tp->tun_if.if_flags &=
653 				~(IFF_BROADCAST|IFF_POINTOPOINT|IFF_MULTICAST);
654 			tp->tun_if.if_flags |= *(int *)data;
655 			break;
656 		default:
657 			error = EINVAL;
658 			goto out;
659 		}
660 		break;
661 
662 	case TUNSLMODE:
663 		if (*(int *)data) {
664 			tp->tun_flags |= TUN_PREPADDR;
665 			tp->tun_flags &= ~TUN_IFHEAD;
666 		} else
667 			tp->tun_flags &= ~TUN_PREPADDR;
668 		break;
669 
670 	case TUNSIFHEAD:
671 		if (*(int *)data) {
672 			tp->tun_flags |= TUN_IFHEAD;
673 			tp->tun_flags &= ~TUN_PREPADDR;
674 		} else
675 			tp->tun_flags &= ~TUN_IFHEAD;
676 		break;
677 
678 	case TUNGIFHEAD:
679 		*(int *)data = (tp->tun_flags & TUN_IFHEAD);
680 		break;
681 
682 	case FIONBIO:
683 		if (*(int *)data)
684 			tp->tun_flags |= TUN_NBIO;
685 		else
686 			tp->tun_flags &= ~TUN_NBIO;
687 		break;
688 
689 	case FIOASYNC:
690 		if (*(int *)data)
691 			tp->tun_flags |= TUN_ASYNC;
692 		else
693 			tp->tun_flags &= ~TUN_ASYNC;
694 		break;
695 
696 	case FIONREAD:
697 		if (tp->tun_if.if_snd.ifq_head)
698 			*(int *)data = tp->tun_if.if_snd.ifq_head->m_pkthdr.len;
699 		else
700 			*(int *)data = 0;
701 		break;
702 
703 	case TIOCSPGRP:
704 	case FIOSETOWN:
705 		error = fsetown(&tp->tun_pgid, cmd, data);
706 		break;
707 
708 	case TIOCGPGRP:
709 	case FIOGETOWN:
710 		error = fgetown(tp->tun_pgid, cmd, data);
711 		break;
712 
713 	default:
714 		error = ENOTTY;
715 	}
716 
717 out:
718 	simple_unlock(&tp->tun_lock);
719 out_nolock:
720 	splx(s);
721 	return (error);
722 }
723 
724 /*
725  * The cdevsw read interface - reads a packet at a time, or at
726  * least as much of a packet as can be read.
727  */
728 int
729 tunread(dev_t dev, struct uio *uio, int ioflag)
730 {
731 	struct tun_softc *tp;
732 	struct ifnet	*ifp;
733 	struct mbuf	*m, *m0;
734 	int		error = 0, len, s, index;
735 
736 	s = splnet();
737 	tp = tun_find_unit(dev);
738 
739 	/* interface was "destroyed" already */
740 	if (tp == NULL) {
741 		error = ENXIO;
742 		goto out_nolock;
743 	}
744 
745 	index = tp->tun_if.if_index;
746 	ifp = &tp->tun_if;
747 
748 	TUNDEBUG ("%s: read\n", ifp->if_xname);
749 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
750 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname, tp->tun_flags);
751 		error = EHOSTDOWN;
752 		goto out;
753 	}
754 
755 	tp->tun_flags &= ~TUN_RWAIT;
756 
757 	do {
758 		IFQ_DEQUEUE(&ifp->if_snd, m0);
759 		if (m0 == 0) {
760 			if (tp->tun_flags & TUN_NBIO) {
761 				error = EWOULDBLOCK;
762 				goto out;
763 			}
764 			tp->tun_flags |= TUN_RWAIT;
765 			if (ltsleep((void *)tp, PZERO|PCATCH|PNORELOCK,
766 					"tunread", 0, &tp->tun_lock) != 0) {
767 				error = EINTR;
768 				goto out_nolock;
769 			} else {
770 				/*
771 				 * Maybe the interface was destroyed while
772 				 * we were sleeping, so let's ensure that
773 				 * we're looking at the same (valid) tun
774 				 * interface before looping.
775 				 */
776 				tp = tun_find_unit(dev);
777 				if (tp == NULL) {
778 					error = ENXIO;
779 					goto out_nolock;
780 				}
781 				if (tp->tun_if.if_index != index) {
782 					error = ENXIO;
783 					goto out;
784 				}
785 			}
786 		}
787 	} while (m0 == 0);
788 
789 	simple_unlock(&tp->tun_lock);
790 	splx(s);
791 
792 	/* Copy the mbuf chain */
793 	while (m0 && uio->uio_resid > 0 && error == 0) {
794 		len = min(uio->uio_resid, m0->m_len);
795 		if (len != 0)
796 			error = uiomove(mtod(m0, void *), len, uio);
797 		MFREE(m0, m);
798 		m0 = m;
799 	}
800 
801 	if (m0) {
802 		TUNDEBUG("Dropping mbuf\n");
803 		m_freem(m0);
804 	}
805 	if (error)
806 		ifp->if_ierrors++;
807 
808 	return (error);
809 
810 out:
811 	simple_unlock(&tp->tun_lock);
812 out_nolock:
813 	splx(s);
814 	return (error);
815 }
816 
817 /*
818  * the cdevsw write interface - an atomic write is a packet - or else!
819  */
820 int
821 tunwrite(dev_t dev, struct uio *uio, int ioflag)
822 {
823 	struct tun_softc *tp;
824 	struct ifnet	*ifp;
825 	struct mbuf	*top, **mp, *m;
826 	struct ifqueue	*ifq;
827 	struct sockaddr	dst;
828 	int		isr, error = 0, s, tlen, mlen;
829 	uint32_t	family;
830 
831 	s = splnet();
832 	tp = tun_find_unit(dev);
833 
834 	/* interface was "destroyed" already */
835 	if (tp == NULL) {
836 		error = ENXIO;
837 		goto out_nolock;
838 	}
839 
840 	/* Unlock until we've got the data */
841 	simple_unlock(&tp->tun_lock);
842 	splx(s);
843 
844 	ifp = &tp->tun_if;
845 
846 	TUNDEBUG("%s: tunwrite\n", ifp->if_xname);
847 
848 	if (tp->tun_flags & TUN_PREPADDR) {
849 		if (uio->uio_resid < sizeof(dst)) {
850 			error = EIO;
851 			goto out0;
852 		}
853 		error = uiomove((void *)&dst, sizeof(dst), uio);
854 		if (dst.sa_len > sizeof(dst)) {
855 			/* Duh.. */
856 			char discard;
857 			int n = dst.sa_len - sizeof(dst);
858 			while (n--)
859 				if ((error = uiomove(&discard, 1, uio)) != 0) {
860 					goto out0;
861 				}
862 		}
863 	} else if (tp->tun_flags & TUN_IFHEAD) {
864 		if (uio->uio_resid < sizeof(family)){
865 			error = EIO;
866 			goto out0;
867 		}
868 		error = uiomove((void *)&family, sizeof(family), uio);
869 		dst.sa_family = ntohl(family);
870 	} else {
871 #ifdef INET
872 		dst.sa_family = AF_INET;
873 #endif
874 	}
875 
876 	if (uio->uio_resid > TUNMTU) {
877 		TUNDEBUG("%s: len=%lu!\n", ifp->if_xname,
878 		    (unsigned long)uio->uio_resid);
879 		error = EIO;
880 		goto out0;
881 	}
882 
883 	switch (dst.sa_family) {
884 #ifdef INET
885 	case AF_INET:
886 		ifq = &ipintrq;
887 		isr = NETISR_IP;
888 		break;
889 #endif
890 #ifdef INET6
891 	case AF_INET6:
892 		ifq = &ip6intrq;
893 		isr = NETISR_IPV6;
894 		break;
895 #endif
896 	default:
897 		error = EAFNOSUPPORT;
898 		goto out0;
899 	}
900 
901 	tlen = uio->uio_resid;
902 
903 	/* get a header mbuf */
904 	MGETHDR(m, M_DONTWAIT, MT_DATA);
905 	if (m == NULL) {
906 		error = ENOBUFS;
907 		goto out0;
908 	}
909 	mlen = MHLEN;
910 
911 	top = NULL;
912 	mp = &top;
913 	while (error == 0 && uio->uio_resid > 0) {
914 		m->m_len = min(mlen, uio->uio_resid);
915 		error = uiomove(mtod(m, void *), m->m_len, uio);
916 		*mp = m;
917 		mp = &m->m_next;
918 		if (error == 0 && uio->uio_resid > 0) {
919 			MGET(m, M_DONTWAIT, MT_DATA);
920 			if (m == NULL) {
921 				error = ENOBUFS;
922 				break;
923 			}
924 			mlen = MLEN;
925 		}
926 	}
927 	if (error) {
928 		if (top != NULL)
929 			m_freem (top);
930 		ifp->if_ierrors++;
931 		goto out0;
932 	}
933 
934 	top->m_pkthdr.len = tlen;
935 	top->m_pkthdr.rcvif = ifp;
936 
937 	bpf_mtap_af(ifp, dst.sa_family, top);
938 
939 	s = splnet();
940 	simple_lock(&tp->tun_lock);
941 	if ((tp->tun_flags & TUN_INITED) == 0) {
942 		/* Interface was destroyed */
943 		error = ENXIO;
944 		goto out;
945 	}
946 	if (IF_QFULL(ifq)) {
947 		IF_DROP(ifq);
948 		ifp->if_collisions++;
949 		m_freem(top);
950 		error = ENOBUFS;
951 		goto out;
952 	}
953 
954 	IF_ENQUEUE(ifq, top);
955 	ifp->if_ipackets++;
956 	ifp->if_ibytes += tlen;
957 	schednetisr(isr);
958 out:
959 	simple_unlock(&tp->tun_lock);
960 out_nolock:
961 	splx(s);
962 out0:
963 	return (error);
964 }
965 
966 #ifdef ALTQ
967 /*
968  * Start packet transmission on the interface.
969  * when the interface queue is rate-limited by ALTQ or TBR,
970  * if_start is needed to drain packets from the queue in order
971  * to notify readers when outgoing packets become ready.
972  *
973  * Should be called at splnet.
974  */
975 static void
976 tunstart(struct ifnet *ifp)
977 {
978 	struct tun_softc *tp = ifp->if_softc;
979 
980 	if (!ALTQ_IS_ENABLED(&ifp->if_snd) && !TBR_IS_ENABLED(&ifp->if_snd))
981 		return;
982 
983 	simple_lock(&tp->tun_lock);
984 	if (!IF_IS_EMPTY(&ifp->if_snd)) {
985 		if (tp->tun_flags & TUN_RWAIT) {
986 			tp->tun_flags &= ~TUN_RWAIT;
987 			wakeup((void *)tp);
988 		}
989 		if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
990 			softint_schedule(tp->tun_osih);
991 
992 		selnotify(&tp->tun_rsel, 0, 0);
993 	}
994 	simple_unlock(&tp->tun_lock);
995 }
996 #endif /* ALTQ */
997 /*
998  * tunpoll - the poll interface, this is only useful on reads
999  * really. The write detect always returns true, write never blocks
1000  * anyway, it either accepts the packet or drops it.
1001  */
1002 int
1003 tunpoll(dev_t dev, int events, struct lwp *l)
1004 {
1005 	struct tun_softc *tp;
1006 	struct ifnet	*ifp;
1007 	int		s, revents = 0;
1008 
1009 	s = splnet();
1010 	tp = tun_find_unit(dev);
1011 
1012 	/* interface was "destroyed" already */
1013 	if (tp == NULL)
1014 		goto out_nolock;
1015 
1016 	ifp = &tp->tun_if;
1017 
1018 	TUNDEBUG("%s: tunpoll\n", ifp->if_xname);
1019 
1020 	if (events & (POLLIN | POLLRDNORM)) {
1021 		if (!IFQ_IS_EMPTY(&ifp->if_snd)) {
1022 			TUNDEBUG("%s: tunpoll q=%d\n", ifp->if_xname,
1023 			    ifp->if_snd.ifq_len);
1024 			revents |= events & (POLLIN | POLLRDNORM);
1025 		} else {
1026 			TUNDEBUG("%s: tunpoll waiting\n", ifp->if_xname);
1027 			selrecord(l, &tp->tun_rsel);
1028 		}
1029 	}
1030 
1031 	if (events & (POLLOUT | POLLWRNORM))
1032 		revents |= events & (POLLOUT | POLLWRNORM);
1033 
1034 	simple_unlock(&tp->tun_lock);
1035 out_nolock:
1036 	splx(s);
1037 	return (revents);
1038 }
1039 
1040 static void
1041 filt_tunrdetach(struct knote *kn)
1042 {
1043 	struct tun_softc *tp = kn->kn_hook;
1044 	int s;
1045 
1046 	s = splnet();
1047 	SLIST_REMOVE(&tp->tun_rsel.sel_klist, kn, knote, kn_selnext);
1048 	splx(s);
1049 }
1050 
1051 static int
1052 filt_tunread(struct knote *kn, long hint)
1053 {
1054 	struct tun_softc *tp = kn->kn_hook;
1055 	struct ifnet *ifp = &tp->tun_if;
1056 	struct mbuf *m;
1057 	int s;
1058 
1059 	s = splnet();
1060 	IF_POLL(&ifp->if_snd, m);
1061 	if (m == NULL) {
1062 		splx(s);
1063 		return (0);
1064 	}
1065 
1066 	for (kn->kn_data = 0; m != NULL; m = m->m_next)
1067 		kn->kn_data += m->m_len;
1068 
1069 	splx(s);
1070 	return (1);
1071 }
1072 
1073 static const struct filterops tunread_filtops =
1074 	{ 1, NULL, filt_tunrdetach, filt_tunread };
1075 
1076 static const struct filterops tun_seltrue_filtops =
1077 	{ 1, NULL, filt_tunrdetach, filt_seltrue };
1078 
1079 int
1080 tunkqfilter(dev_t dev, struct knote *kn)
1081 {
1082 	struct tun_softc *tp;
1083 	struct klist *klist;
1084 	int rv = 0, s;
1085 
1086 	s = splnet();
1087 	tp = tun_find_unit(dev);
1088 	if (tp == NULL)
1089 		goto out_nolock;
1090 
1091 	switch (kn->kn_filter) {
1092 	case EVFILT_READ:
1093 		klist = &tp->tun_rsel.sel_klist;
1094 		kn->kn_fop = &tunread_filtops;
1095 		break;
1096 
1097 	case EVFILT_WRITE:
1098 		klist = &tp->tun_rsel.sel_klist;
1099 		kn->kn_fop = &tun_seltrue_filtops;
1100 		break;
1101 
1102 	default:
1103 		rv = EINVAL;
1104 		goto out;
1105 	}
1106 
1107 	kn->kn_hook = tp;
1108 
1109 	SLIST_INSERT_HEAD(klist, kn, kn_selnext);
1110 
1111 out:
1112 	simple_unlock(&tp->tun_lock);
1113 out_nolock:
1114 	splx(s);
1115 	return (rv);
1116 }
1117