xref: /netbsd-src/sys/net/if_tun.c (revision cac8e449158efc7261bebc8657cbb0125a2cfdde)
1 /*	$NetBSD: if_tun.c,v 1.107 2008/06/15 16:37:21 christos 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 1.107 2008/06/15 16:37:21 christos 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 	if ((error = kauth_authorize_generic(l->l_cred, KAUTH_GENERIC_ISSUSER,
287 	    NULL)) != 0)
288 		return (error);
289 
290 	s = splnet();
291 	tp = tun_find_unit(dev);
292 
293 	if (tp == NULL) {
294 		(void)tun_clone_create(&tun_cloner, minor(dev));
295 		tp = tun_find_unit(dev);
296 		if (tp == NULL) {
297 			error = ENXIO;
298 			goto out_nolock;
299 		}
300 	}
301 
302 	if (tp->tun_flags & TUN_OPEN) {
303 		error = EBUSY;
304 		goto out;
305 	}
306 
307 	ifp = &tp->tun_if;
308 	tp->tun_flags |= TUN_OPEN;
309 	TUNDEBUG("%s: open\n", ifp->if_xname);
310 out:
311 	simple_unlock(&tp->tun_lock);
312 out_nolock:
313 	splx(s);
314 	return (error);
315 }
316 
317 /*
318  * tunclose - close the device - mark i/f down & delete
319  * routing info
320  */
321 int
322 tunclose(dev_t dev, int flag, int mode,
323     struct lwp *l)
324 {
325 	int	s;
326 	struct tun_softc *tp;
327 	struct ifnet	*ifp;
328 
329 	s = splnet();
330 	if ((tp = tun_find_zunit(minor(dev))) != NULL) {
331 		/* interface was "destroyed" before the close */
332 		seldestroy(&tp->tun_rsel);
333 		seldestroy(&tp->tun_wsel);
334 		softint_disestablish(tp->tun_osih);
335 		softint_disestablish(tp->tun_isih);
336 		free(tp, M_DEVBUF);
337 		goto out_nolock;
338 	}
339 
340 	if ((tp = tun_find_unit(dev)) == NULL)
341 		goto out_nolock;
342 
343 	ifp = &tp->tun_if;
344 
345 	tp->tun_flags &= ~TUN_OPEN;
346 
347 	/*
348 	 * junk all pending output
349 	 */
350 	IFQ_PURGE(&ifp->if_snd);
351 
352 	if (ifp->if_flags & IFF_UP) {
353 		if_down(ifp);
354 		if (ifp->if_flags & IFF_RUNNING) {
355 			/* find internet addresses and delete routes */
356 			struct ifaddr *ifa;
357 			IFADDR_FOREACH(ifa, ifp) {
358 #if defined(INET) || defined(INET6)
359 				if (ifa->ifa_addr->sa_family == AF_INET ||
360 				    ifa->ifa_addr->sa_family == AF_INET6) {
361 					rtinit(ifa, (int)RTM_DELETE,
362 					       tp->tun_flags & TUN_DSTADDR
363 							? RTF_HOST
364 							: 0);
365 				}
366 #endif
367 			}
368 		}
369 	}
370 	tp->tun_pgid = 0;
371 	selnotify(&tp->tun_rsel, 0, 0);
372 
373 	TUNDEBUG ("%s: closed\n", ifp->if_xname);
374 	simple_unlock(&tp->tun_lock);
375 out_nolock:
376 	splx(s);
377 	return (0);
378 }
379 
380 /*
381  * Call at splnet() with tp locked.
382  */
383 static void
384 tuninit(struct tun_softc *tp)
385 {
386 	struct ifnet	*ifp = &tp->tun_if;
387 	struct ifaddr	*ifa;
388 
389 	TUNDEBUG("%s: tuninit\n", ifp->if_xname);
390 
391 	ifp->if_flags |= IFF_UP | IFF_RUNNING;
392 
393 	tp->tun_flags &= ~(TUN_IASET|TUN_DSTADDR);
394 	IFADDR_FOREACH(ifa, ifp) {
395 #ifdef INET
396 		if (ifa->ifa_addr->sa_family == AF_INET) {
397 			struct sockaddr_in *sin;
398 
399 			sin = satosin(ifa->ifa_addr);
400 			if (sin && sin->sin_addr.s_addr)
401 				tp->tun_flags |= TUN_IASET;
402 
403 			if (ifp->if_flags & IFF_POINTOPOINT) {
404 				sin = satosin(ifa->ifa_dstaddr);
405 				if (sin && sin->sin_addr.s_addr)
406 					tp->tun_flags |= TUN_DSTADDR;
407 			}
408 		}
409 #endif
410 #ifdef INET6
411 		if (ifa->ifa_addr->sa_family == AF_INET6) {
412 			struct sockaddr_in6 *sin;
413 
414 			sin = (struct sockaddr_in6 *)ifa->ifa_addr;
415 			if (!IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
416 				tp->tun_flags |= TUN_IASET;
417 
418 			if (ifp->if_flags & IFF_POINTOPOINT) {
419 				sin = (struct sockaddr_in6 *)ifa->ifa_dstaddr;
420 				if (sin &&
421 				    !IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
422 					tp->tun_flags |= TUN_DSTADDR;
423 			} else
424 				tp->tun_flags &= ~TUN_DSTADDR;
425 		}
426 #endif /* INET6 */
427 	}
428 
429 	return;
430 }
431 
432 /*
433  * Process an ioctl request.
434  */
435 static int
436 tun_ioctl(struct ifnet *ifp, u_long cmd, void *data)
437 {
438 	int		error = 0, s;
439 	struct tun_softc *tp = (struct tun_softc *)(ifp->if_softc);
440 	struct ifreq *ifr = data;
441 
442 	s = splnet();
443 	simple_lock(&tp->tun_lock);
444 
445 	switch (cmd) {
446 	case SIOCSIFADDR:
447 		tuninit(tp);
448 		TUNDEBUG("%s: address set\n", ifp->if_xname);
449 		break;
450 	case SIOCSIFDSTADDR:
451 		tuninit(tp);
452 		TUNDEBUG("%s: destination address set\n", ifp->if_xname);
453 		break;
454 	case SIOCSIFBRDADDR:
455 		TUNDEBUG("%s: broadcast address set\n", ifp->if_xname);
456 		break;
457 	case SIOCSIFMTU:
458 		if (ifr->ifr_mtu > TUNMTU || ifr->ifr_mtu < 576) {
459 			error = EINVAL;
460 			break;
461 		}
462 		TUNDEBUG("%s: interface mtu set\n", ifp->if_xname);
463 		if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET)
464 			error = 0;
465 		break;
466 	case SIOCADDMULTI:
467 	case SIOCDELMULTI:
468 		if (ifr == NULL) {
469 	        	error = EAFNOSUPPORT;           /* XXX */
470 			break;
471 		}
472 		switch (ifreq_getaddr(cmd, ifr)->sa_family) {
473 #ifdef INET
474 		case AF_INET:
475 			break;
476 #endif
477 #ifdef INET6
478 		case AF_INET6:
479 			break;
480 #endif
481 		default:
482 			error = EAFNOSUPPORT;
483 			break;
484 		}
485 		break;
486 	case SIOCSIFFLAGS:
487 		break;
488 	default:
489 		error = EINVAL;
490 	}
491 
492 	simple_unlock(&tp->tun_lock);
493 	splx(s);
494 	return (error);
495 }
496 
497 /*
498  * tun_output - queue packets from higher level ready to put out.
499  */
500 static int
501 tun_output(struct ifnet *ifp, struct mbuf *m0, const struct sockaddr *dst,
502     struct rtentry *rt)
503 {
504 	struct tun_softc *tp = ifp->if_softc;
505 	int		s;
506 	int		error;
507 #if defined(INET) || defined(INET6)
508 	int		mlen;
509 	uint32_t	*af;
510 #endif
511 	ALTQ_DECL(struct altq_pktattr pktattr;)
512 
513 	s = splnet();
514 	simple_lock(&tp->tun_lock);
515 	TUNDEBUG ("%s: tun_output\n", ifp->if_xname);
516 
517 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
518 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname,
519 			  tp->tun_flags);
520 		m_freem (m0);
521 		error = EHOSTDOWN;
522 		goto out;
523 	}
524 
525 	/*
526 	 * if the queueing discipline needs packet classification,
527 	 * do it before prepending link headers.
528 	 */
529 	IFQ_CLASSIFY(&ifp->if_snd, m0, dst->sa_family, &pktattr);
530 
531 #if NBPFILTER > 0
532 	if (ifp->if_bpf)
533 		bpf_mtap_af(ifp->if_bpf, dst->sa_family, m0);
534 #endif
535 
536 	switch(dst->sa_family) {
537 #ifdef INET6
538 	case AF_INET6:
539 #endif
540 #ifdef INET
541 	case AF_INET:
542 #endif
543 #if defined(INET) || defined(INET6)
544 		if (tp->tun_flags & TUN_PREPADDR) {
545 			/* Simple link-layer header */
546 			M_PREPEND(m0, dst->sa_len, M_DONTWAIT);
547 			if (m0 == NULL) {
548 				IF_DROP(&ifp->if_snd);
549 				error = ENOBUFS;
550 				goto out;
551 			}
552 			bcopy(dst, mtod(m0, char *), dst->sa_len);
553 		}
554 
555 		if (tp->tun_flags & TUN_IFHEAD) {
556 			/* Prepend the address family */
557 			M_PREPEND(m0, sizeof(*af), M_DONTWAIT);
558 			if (m0 == NULL) {
559 				IF_DROP(&ifp->if_snd);
560 				error = ENOBUFS;
561 				goto out;
562 			}
563 			af = mtod(m0,uint32_t *);
564 			*af = htonl(dst->sa_family);
565 		} else {
566 #ifdef INET
567 			if (dst->sa_family != AF_INET)
568 #endif
569 			{
570 				m_freem(m0);
571 				error = EAFNOSUPPORT;
572 				goto out;
573 			}
574 		}
575 		/* FALLTHROUGH */
576 	case AF_UNSPEC:
577 		IFQ_ENQUEUE(&ifp->if_snd, m0, &pktattr, error);
578 		if (error) {
579 			ifp->if_collisions++;
580 			error = EAFNOSUPPORT;
581 			goto out;
582 		}
583 		mlen = m0->m_pkthdr.len;
584 		ifp->if_opackets++;
585 		ifp->if_obytes += mlen;
586 		break;
587 #endif
588 	default:
589 		m_freem(m0);
590 		error = EAFNOSUPPORT;
591 		goto out;
592 	}
593 
594 	if (tp->tun_flags & TUN_RWAIT) {
595 		tp->tun_flags &= ~TUN_RWAIT;
596 		wakeup((void *)tp);
597 	}
598 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
599 		softint_schedule(tp->tun_isih);
600 
601 	selnotify(&tp->tun_rsel, 0, 0);
602 out:
603 	simple_unlock(&tp->tun_lock);
604 	splx(s);
605 	return (0);
606 }
607 
608 static void
609 tun_i_softintr(void *cookie)
610 {
611 	struct tun_softc *tp = cookie;
612 
613 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
614 		fownsignal(tp->tun_pgid, SIGIO, POLL_IN, POLLIN|POLLRDNORM,
615 		    NULL);
616 }
617 
618 static void
619 tun_o_softintr(void *cookie)
620 {
621 	struct tun_softc *tp = cookie;
622 
623 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
624 		fownsignal(tp->tun_pgid, SIGIO, POLL_OUT, POLLOUT|POLLWRNORM,
625 		    NULL);
626 }
627 
628 /*
629  * the cdevsw interface is now pretty minimal.
630  */
631 int
632 tunioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
633 {
634 	struct tun_softc *tp;
635 	int s, error = 0;
636 
637 	s = splnet();
638 	tp = tun_find_unit(dev);
639 
640 	/* interface was "destroyed" already */
641 	if (tp == NULL) {
642 		error = ENXIO;
643 		goto out_nolock;
644 	}
645 
646 	switch (cmd) {
647 	case TUNSDEBUG:
648 		tundebug = *(int *)data;
649 		break;
650 
651 	case TUNGDEBUG:
652 		*(int *)data = tundebug;
653 		break;
654 
655 	case TUNSIFMODE:
656 		switch (*(int *)data & (IFF_POINTOPOINT|IFF_BROADCAST)) {
657 		case IFF_POINTOPOINT:
658 		case IFF_BROADCAST:
659 			if (tp->tun_if.if_flags & IFF_UP) {
660 				error = EBUSY;
661 				goto out;
662 			}
663 			tp->tun_if.if_flags &=
664 				~(IFF_BROADCAST|IFF_POINTOPOINT|IFF_MULTICAST);
665 			tp->tun_if.if_flags |= *(int *)data;
666 			break;
667 		default:
668 			error = EINVAL;
669 			goto out;
670 		}
671 		break;
672 
673 	case TUNSLMODE:
674 		if (*(int *)data) {
675 			tp->tun_flags |= TUN_PREPADDR;
676 			tp->tun_flags &= ~TUN_IFHEAD;
677 		} else
678 			tp->tun_flags &= ~TUN_PREPADDR;
679 		break;
680 
681 	case TUNSIFHEAD:
682 		if (*(int *)data) {
683 			tp->tun_flags |= TUN_IFHEAD;
684 			tp->tun_flags &= ~TUN_PREPADDR;
685 		} else
686 			tp->tun_flags &= ~TUN_IFHEAD;
687 		break;
688 
689 	case TUNGIFHEAD:
690 		*(int *)data = (tp->tun_flags & TUN_IFHEAD);
691 		break;
692 
693 	case FIONBIO:
694 		if (*(int *)data)
695 			tp->tun_flags |= TUN_NBIO;
696 		else
697 			tp->tun_flags &= ~TUN_NBIO;
698 		break;
699 
700 	case FIOASYNC:
701 		if (*(int *)data)
702 			tp->tun_flags |= TUN_ASYNC;
703 		else
704 			tp->tun_flags &= ~TUN_ASYNC;
705 		break;
706 
707 	case FIONREAD:
708 		if (tp->tun_if.if_snd.ifq_head)
709 			*(int *)data = tp->tun_if.if_snd.ifq_head->m_pkthdr.len;
710 		else
711 			*(int *)data = 0;
712 		break;
713 
714 	case TIOCSPGRP:
715 	case FIOSETOWN:
716 		error = fsetown(&tp->tun_pgid, cmd, data);
717 		break;
718 
719 	case TIOCGPGRP:
720 	case FIOGETOWN:
721 		error = fgetown(tp->tun_pgid, cmd, data);
722 		break;
723 
724 	default:
725 		error = ENOTTY;
726 	}
727 
728 out:
729 	simple_unlock(&tp->tun_lock);
730 out_nolock:
731 	splx(s);
732 	return (error);
733 }
734 
735 /*
736  * The cdevsw read interface - reads a packet at a time, or at
737  * least as much of a packet as can be read.
738  */
739 int
740 tunread(dev_t dev, struct uio *uio, int ioflag)
741 {
742 	struct tun_softc *tp;
743 	struct ifnet	*ifp;
744 	struct mbuf	*m, *m0;
745 	int		error = 0, len, s, index;
746 
747 	s = splnet();
748 	tp = tun_find_unit(dev);
749 
750 	/* interface was "destroyed" already */
751 	if (tp == NULL) {
752 		error = ENXIO;
753 		goto out_nolock;
754 	}
755 
756 	index = tp->tun_if.if_index;
757 	ifp = &tp->tun_if;
758 
759 	TUNDEBUG ("%s: read\n", ifp->if_xname);
760 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
761 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname, tp->tun_flags);
762 		error = EHOSTDOWN;
763 		goto out;
764 	}
765 
766 	tp->tun_flags &= ~TUN_RWAIT;
767 
768 	do {
769 		IFQ_DEQUEUE(&ifp->if_snd, m0);
770 		if (m0 == 0) {
771 			if (tp->tun_flags & TUN_NBIO) {
772 				error = EWOULDBLOCK;
773 				goto out;
774 			}
775 			tp->tun_flags |= TUN_RWAIT;
776 			if (ltsleep((void *)tp, PZERO|PCATCH|PNORELOCK,
777 					"tunread", 0, &tp->tun_lock) != 0) {
778 				error = EINTR;
779 				goto out_nolock;
780 			} else {
781 				/*
782 				 * Maybe the interface was destroyed while
783 				 * we were sleeping, so let's ensure that
784 				 * we're looking at the same (valid) tun
785 				 * interface before looping.
786 				 */
787 				tp = tun_find_unit(dev);
788 				if (tp == NULL) {
789 					error = ENXIO;
790 					goto out_nolock;
791 				}
792 				if (tp->tun_if.if_index != index) {
793 					error = ENXIO;
794 					goto out;
795 				}
796 			}
797 		}
798 	} while (m0 == 0);
799 
800 	simple_unlock(&tp->tun_lock);
801 	splx(s);
802 
803 	/* Copy the mbuf chain */
804 	while (m0 && uio->uio_resid > 0 && error == 0) {
805 		len = min(uio->uio_resid, m0->m_len);
806 		if (len != 0)
807 			error = uiomove(mtod(m0, void *), len, uio);
808 		MFREE(m0, m);
809 		m0 = m;
810 	}
811 
812 	if (m0) {
813 		TUNDEBUG("Dropping mbuf\n");
814 		m_freem(m0);
815 	}
816 	if (error)
817 		ifp->if_ierrors++;
818 
819 	return (error);
820 
821 out:
822 	simple_unlock(&tp->tun_lock);
823 out_nolock:
824 	splx(s);
825 	return (error);
826 }
827 
828 /*
829  * the cdevsw write interface - an atomic write is a packet - or else!
830  */
831 int
832 tunwrite(dev_t dev, struct uio *uio, int ioflag)
833 {
834 	struct tun_softc *tp;
835 	struct ifnet	*ifp;
836 	struct mbuf	*top, **mp, *m;
837 	struct ifqueue	*ifq;
838 	struct sockaddr	dst;
839 	int		isr, error = 0, s, tlen, mlen;
840 	uint32_t	family;
841 
842 	s = splnet();
843 	tp = tun_find_unit(dev);
844 
845 	/* interface was "destroyed" already */
846 	if (tp == NULL) {
847 		error = ENXIO;
848 		goto out_nolock;
849 	}
850 
851 	/* Unlock until we've got the data */
852 	simple_unlock(&tp->tun_lock);
853 	splx(s);
854 
855 	ifp = &tp->tun_if;
856 
857 	TUNDEBUG("%s: tunwrite\n", ifp->if_xname);
858 
859 	if (tp->tun_flags & TUN_PREPADDR) {
860 		if (uio->uio_resid < sizeof(dst)) {
861 			error = EIO;
862 			goto out0;
863 		}
864 		error = uiomove((void *)&dst, sizeof(dst), uio);
865 		if (dst.sa_len > sizeof(dst)) {
866 			/* Duh.. */
867 			char discard;
868 			int n = dst.sa_len - sizeof(dst);
869 			while (n--)
870 				if ((error = uiomove(&discard, 1, uio)) != 0) {
871 					goto out0;
872 				}
873 		}
874 	} else if (tp->tun_flags & TUN_IFHEAD) {
875 		if (uio->uio_resid < sizeof(family)){
876 			error = EIO;
877 			goto out0;
878 		}
879 		error = uiomove((void *)&family, sizeof(family), uio);
880 		dst.sa_family = ntohl(family);
881 	} else {
882 #ifdef INET
883 		dst.sa_family = AF_INET;
884 #endif
885 	}
886 
887 	if (uio->uio_resid > TUNMTU) {
888 		TUNDEBUG("%s: len=%lu!\n", ifp->if_xname,
889 		    (unsigned long)uio->uio_resid);
890 		error = EIO;
891 		goto out0;
892 	}
893 
894 	switch (dst.sa_family) {
895 #ifdef INET
896 	case AF_INET:
897 		ifq = &ipintrq;
898 		isr = NETISR_IP;
899 		break;
900 #endif
901 #ifdef INET6
902 	case AF_INET6:
903 		ifq = &ip6intrq;
904 		isr = NETISR_IPV6;
905 		break;
906 #endif
907 	default:
908 		error = EAFNOSUPPORT;
909 		goto out0;
910 	}
911 
912 	tlen = uio->uio_resid;
913 
914 	/* get a header mbuf */
915 	MGETHDR(m, M_DONTWAIT, MT_DATA);
916 	if (m == NULL) {
917 		error = ENOBUFS;
918 		goto out0;
919 	}
920 	mlen = MHLEN;
921 
922 	top = NULL;
923 	mp = &top;
924 	while (error == 0 && uio->uio_resid > 0) {
925 		m->m_len = min(mlen, uio->uio_resid);
926 		error = uiomove(mtod(m, void *), m->m_len, uio);
927 		*mp = m;
928 		mp = &m->m_next;
929 		if (error == 0 && uio->uio_resid > 0) {
930 			MGET(m, M_DONTWAIT, MT_DATA);
931 			if (m == NULL) {
932 				error = ENOBUFS;
933 				break;
934 			}
935 			mlen = MLEN;
936 		}
937 	}
938 	if (error) {
939 		if (top != NULL)
940 			m_freem (top);
941 		ifp->if_ierrors++;
942 		goto out0;
943 	}
944 
945 	top->m_pkthdr.len = tlen;
946 	top->m_pkthdr.rcvif = ifp;
947 
948 #if NBPFILTER > 0
949 	if (ifp->if_bpf)
950 		bpf_mtap_af(ifp->if_bpf, dst.sa_family, top);
951 #endif
952 
953 	s = splnet();
954 	simple_lock(&tp->tun_lock);
955 	if ((tp->tun_flags & TUN_INITED) == 0) {
956 		/* Interface was destroyed */
957 		error = ENXIO;
958 		goto out;
959 	}
960 	if (IF_QFULL(ifq)) {
961 		IF_DROP(ifq);
962 		ifp->if_collisions++;
963 		m_freem(top);
964 		error = ENOBUFS;
965 		goto out;
966 	}
967 
968 	IF_ENQUEUE(ifq, top);
969 	ifp->if_ipackets++;
970 	ifp->if_ibytes += tlen;
971 	schednetisr(isr);
972 out:
973 	simple_unlock(&tp->tun_lock);
974 out_nolock:
975 	splx(s);
976 out0:
977 	return (error);
978 }
979 
980 #ifdef ALTQ
981 /*
982  * Start packet transmission on the interface.
983  * when the interface queue is rate-limited by ALTQ or TBR,
984  * if_start is needed to drain packets from the queue in order
985  * to notify readers when outgoing packets become ready.
986  *
987  * Should be called at splnet.
988  */
989 static void
990 tunstart(struct ifnet *ifp)
991 {
992 	struct tun_softc *tp = ifp->if_softc;
993 
994 	if (!ALTQ_IS_ENABLED(&ifp->if_snd) && !TBR_IS_ENABLED(&ifp->if_snd))
995 		return;
996 
997 	simple_lock(&tp->tun_lock);
998 	if (!IF_IS_EMPTY(&ifp->if_snd)) {
999 		if (tp->tun_flags & TUN_RWAIT) {
1000 			tp->tun_flags &= ~TUN_RWAIT;
1001 			wakeup((void *)tp);
1002 		}
1003 		if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
1004 			softint_schedule(tp->tun_osih);
1005 
1006 		selnotify(&tp->tun_rsel, 0, 0);
1007 	}
1008 	simple_unlock(&tp->tun_lock);
1009 }
1010 #endif /* ALTQ */
1011 /*
1012  * tunpoll - the poll interface, this is only useful on reads
1013  * really. The write detect always returns true, write never blocks
1014  * anyway, it either accepts the packet or drops it.
1015  */
1016 int
1017 tunpoll(dev_t dev, int events, struct lwp *l)
1018 {
1019 	struct tun_softc *tp;
1020 	struct ifnet	*ifp;
1021 	int		s, revents = 0;
1022 
1023 	s = splnet();
1024 	tp = tun_find_unit(dev);
1025 
1026 	/* interface was "destroyed" already */
1027 	if (tp == NULL)
1028 		goto out_nolock;
1029 
1030 	ifp = &tp->tun_if;
1031 
1032 	TUNDEBUG("%s: tunpoll\n", ifp->if_xname);
1033 
1034 	if (events & (POLLIN | POLLRDNORM)) {
1035 		if (!IFQ_IS_EMPTY(&ifp->if_snd)) {
1036 			TUNDEBUG("%s: tunpoll q=%d\n", ifp->if_xname,
1037 			    ifp->if_snd.ifq_len);
1038 			revents |= events & (POLLIN | POLLRDNORM);
1039 		} else {
1040 			TUNDEBUG("%s: tunpoll waiting\n", ifp->if_xname);
1041 			selrecord(l, &tp->tun_rsel);
1042 		}
1043 	}
1044 
1045 	if (events & (POLLOUT | POLLWRNORM))
1046 		revents |= events & (POLLOUT | POLLWRNORM);
1047 
1048 	simple_unlock(&tp->tun_lock);
1049 out_nolock:
1050 	splx(s);
1051 	return (revents);
1052 }
1053 
1054 static void
1055 filt_tunrdetach(struct knote *kn)
1056 {
1057 	struct tun_softc *tp = kn->kn_hook;
1058 	int s;
1059 
1060 	s = splnet();
1061 	SLIST_REMOVE(&tp->tun_rsel.sel_klist, kn, knote, kn_selnext);
1062 	splx(s);
1063 }
1064 
1065 static int
1066 filt_tunread(struct knote *kn, long hint)
1067 {
1068 	struct tun_softc *tp = kn->kn_hook;
1069 	struct ifnet *ifp = &tp->tun_if;
1070 	struct mbuf *m;
1071 	int s;
1072 
1073 	s = splnet();
1074 	IF_POLL(&ifp->if_snd, m);
1075 	if (m == NULL) {
1076 		splx(s);
1077 		return (0);
1078 	}
1079 
1080 	for (kn->kn_data = 0; m != NULL; m = m->m_next)
1081 		kn->kn_data += m->m_len;
1082 
1083 	splx(s);
1084 	return (1);
1085 }
1086 
1087 static const struct filterops tunread_filtops =
1088 	{ 1, NULL, filt_tunrdetach, filt_tunread };
1089 
1090 static const struct filterops tun_seltrue_filtops =
1091 	{ 1, NULL, filt_tunrdetach, filt_seltrue };
1092 
1093 int
1094 tunkqfilter(dev_t dev, struct knote *kn)
1095 {
1096 	struct tun_softc *tp;
1097 	struct klist *klist;
1098 	int rv = 0, s;
1099 
1100 	s = splnet();
1101 	tp = tun_find_unit(dev);
1102 	if (tp == NULL)
1103 		goto out_nolock;
1104 
1105 	switch (kn->kn_filter) {
1106 	case EVFILT_READ:
1107 		klist = &tp->tun_rsel.sel_klist;
1108 		kn->kn_fop = &tunread_filtops;
1109 		break;
1110 
1111 	case EVFILT_WRITE:
1112 		klist = &tp->tun_rsel.sel_klist;
1113 		kn->kn_fop = &tun_seltrue_filtops;
1114 		break;
1115 
1116 	default:
1117 		rv = EINVAL;
1118 		goto out;
1119 	}
1120 
1121 	kn->kn_hook = tp;
1122 
1123 	SLIST_INSERT_HEAD(klist, kn, kn_selnext);
1124 
1125 out:
1126 	simple_unlock(&tp->tun_lock);
1127 out_nolock:
1128 	splx(s);
1129 	return (rv);
1130 }
1131