if_tun.c revision 26566 
1/*	$NetBSD: if_tun.c,v 1.14 1994/06/29 06:36:25 cgd 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 it's wicked way with. This driver has it's
12 * roots in a similar driver written by Phil Cockcroft (formerly) at
13 * UCL. This driver is based much more on read/write/select mode of
14 * operation though.
15 */
16
17#include "tun.h"
18#if NTUN > 0
19
20#include <sys/param.h>
21#include <sys/proc.h>
22#include <sys/systm.h>
23#include <sys/mbuf.h>
24#include <sys/buf.h>
25#include <sys/protosw.h>
26#include <sys/socket.h>
27#include <sys/filio.h>
28#include <sys/sockio.h>
29#include <sys/ttycom.h>
30#include <sys/errno.h>
31#include <sys/syslog.h>
32#include <sys/select.h>
33#include <sys/fcntl.h>
34#include <sys/signalvar.h>
35#include <sys/kernel.h>
36#include <sys/sysctl.h>
37#ifdef DEVFS
38#include <sys/devfsext.h>
39#endif /*DEVFS*/
40#include <sys/conf.h>
41
42#include <net/if.h>
43#include <net/netisr.h>
44#include <net/route.h>
45
46#ifdef INET
47#include <netinet/in.h>
48#include <netinet/in_systm.h>
49#include <netinet/in_var.h>
50#include <netinet/ip.h>
51#include <netinet/if_ether.h>
52#endif
53
54#ifdef NS
55#include <netns/ns.h>
56#include <netns/ns_if.h>
57#endif
58
59#include "bpfilter.h"
60#if NBPFILTER > 0
61#include <sys/time.h>
62#include <net/bpf.h>
63#endif
64
65#include <net/if_tun.h>
66
67static void tunattach __P((void *));
68PSEUDO_SET(tunattach, if_tun);
69
70#define TUNDEBUG	if (tundebug) printf
71static int tundebug = 0;
72SYSCTL_INT(_debug, OID_AUTO, if_tun_debug, CTLFLAG_RW, &tundebug, 0, "");
73
74static struct tun_softc tunctl[NTUN];
75
76static int tunoutput __P((struct ifnet *, struct mbuf *, struct sockaddr *,
77	    struct rtentry *rt));
78static int tunifioctl __P((struct ifnet *, int, caddr_t));
79static int tuninit __P((int));
80
81static	d_open_t	tunopen;
82static	d_close_t	tunclose;
83static	d_read_t	tunread;
84static	d_write_t	tunwrite;
85static	d_ioctl_t	tunioctl;
86static	d_select_t	tunselect;
87
88#define CDEV_MAJOR 52
89static struct cdevsw tun_cdevsw = {
90	tunopen,	tunclose,	tunread,	tunwrite,
91	tunioctl,	nullstop,	noreset,	nodevtotty,
92	tunselect,	nommap,		nostrategy,	"tun",	NULL,	-1
93};
94
95
96static tun_devsw_installed = 0;
97#ifdef	DEVFS
98static	void	*tun_devfs_token[NTUN];
99#endif
100
101static void
102tunattach(dummy)
103	void *dummy;
104{
105	register int i;
106	struct ifnet *ifp;
107	dev_t dev;
108
109	if ( tun_devsw_installed )
110		return;
111	dev = makedev(CDEV_MAJOR, 0);
112	cdevsw_add(&dev, &tun_cdevsw, NULL);
113	tun_devsw_installed = 1;
114	for ( i = 0; i < NTUN; i++ ) {
115#ifdef DEVFS
116		tun_devfs_token[i] = devfs_add_devswf(&tun_cdevsw, i, DV_CHR,
117						      UID_UUCP, GID_DIALER,
118						      0600, "tun%d", i);
119#endif
120		tunctl[i].tun_flags = TUN_INITED;
121
122		ifp = &tunctl[i].tun_if;
123		ifp->if_unit = i;
124		ifp->if_name = "tun";
125		ifp->if_mtu = TUNMTU;
126		ifp->if_ioctl = tunifioctl;
127		ifp->if_output = tunoutput;
128		ifp->if_flags = IFF_POINTOPOINT | IFF_MULTICAST;
129		ifp->if_snd.ifq_maxlen = ifqmaxlen;
130		ifp->if_collisions = 0;
131		ifp->if_ierrors = 0;
132		ifp->if_oerrors = 0;
133		ifp->if_ipackets = 0;
134		ifp->if_opackets = 0;
135		if_attach(ifp);
136#if NBPFILTER > 0
137		bpfattach(ifp, DLT_NULL, sizeof(u_int));
138#endif
139	}
140}
141
142/*
143 * tunnel open - must be superuser & the device must be
144 * configured in
145 */
146static	int
147tunopen(dev, flag, mode, p)
148	dev_t	dev;
149	int	flag, mode;
150	struct proc *p;
151{
152	struct ifnet	*ifp;
153	struct tun_softc *tp;
154	register int	unit, error;
155
156	error = suser(p->p_ucred, &p->p_acflag);
157	if (error)
158		return (error);
159
160	if ((unit = minor(dev)) >= NTUN)
161		return (ENXIO);
162	tp = &tunctl[unit];
163	if (tp->tun_flags & TUN_OPEN)
164		return EBUSY;
165	ifp = &tp->tun_if;
166	tp->tun_flags |= TUN_OPEN;
167	TUNDEBUG("%s%d: open\n", ifp->if_name, ifp->if_unit);
168	return (0);
169}
170
171/*
172 * tunclose - close the device - mark i/f down & delete
173 * routing info
174 */
175static	int
176tunclose(dev_t dev, int foo, int bar, struct proc *p)
177{
178	register int	unit = minor(dev), s;
179	struct tun_softc *tp = &tunctl[unit];
180	struct ifnet	*ifp = &tp->tun_if;
181	struct mbuf	*m;
182
183	tp->tun_flags &= ~TUN_OPEN;
184
185	/*
186	 * junk all pending output
187	 */
188	do {
189		s = splimp();
190		IF_DEQUEUE(&ifp->if_snd, m);
191		splx(s);
192		if (m)
193			m_freem(m);
194	} while (m);
195
196	if (ifp->if_flags & IFF_UP) {
197		s = splimp();
198		if_down(ifp);
199		if (ifp->if_flags & IFF_RUNNING) {
200		    /* find internet addresses and delete routes */
201		    register struct ifaddr *ifa;
202		    for (ifa = ifp->if_addrhead.tqh_first; ifa;
203			 ifa = ifa->ifa_link.tqe_next) {
204			if (ifa->ifa_addr->sa_family == AF_INET) {
205			    rtinit(ifa, (int)RTM_DELETE,
206				   tp->tun_flags & TUN_DSTADDR ? RTF_HOST : 0);
207			}
208		    }
209		}
210		splx(s);
211	}
212	tp->tun_pgrp = 0;
213	selwakeup(&tp->tun_rsel);
214
215	TUNDEBUG ("%s%d: closed\n", ifp->if_name, ifp->if_unit);
216	return (0);
217}
218
219static int
220tuninit(unit)
221	int	unit;
222{
223	struct tun_softc *tp = &tunctl[unit];
224	struct ifnet	*ifp = &tp->tun_if;
225	register struct ifaddr *ifa;
226
227	TUNDEBUG("%s%d: tuninit\n", ifp->if_name, ifp->if_unit);
228
229	ifp->if_flags |= IFF_UP | IFF_RUNNING;
230	microtime(&ifp->if_lastchange);
231
232	for (ifa = ifp->if_addrhead.tqh_first; ifa;
233	     ifa = ifa->ifa_link.tqe_next)
234		if (ifa->ifa_addr->sa_family == AF_INET) {
235		    struct sockaddr_in *si;
236
237		    si = (struct sockaddr_in *)ifa->ifa_addr;
238		    if (si && si->sin_addr.s_addr)
239			    tp->tun_flags |= TUN_IASET;
240
241		    si = (struct sockaddr_in *)ifa->ifa_dstaddr;
242		    if (si && si->sin_addr.s_addr)
243			    tp->tun_flags |= TUN_DSTADDR;
244		}
245
246	return 0;
247}
248
249/*
250 * Process an ioctl request.
251 */
252int
253tunifioctl(ifp, cmd, data)
254	struct ifnet *ifp;
255	int	cmd;
256	caddr_t	data;
257{
258	register struct ifreq *ifr = (struct ifreq *)data;
259	int		error = 0, s;
260
261	s = splimp();
262	switch(cmd) {
263	case SIOCSIFADDR:
264		tuninit(ifp->if_unit);
265		TUNDEBUG("%s%d: address set\n",
266			 ifp->if_name, ifp->if_unit);
267		break;
268	case SIOCSIFDSTADDR:
269		tuninit(ifp->if_unit);
270		TUNDEBUG("%s%d: destination address set\n",
271			 ifp->if_name, ifp->if_unit);
272		break;
273	case SIOCSIFMTU:
274		ifp->if_mtu = ifr->ifr_mtu;
275		TUNDEBUG("%s%d: mtu set\n",
276			 ifp->if_name, ifp->if_unit);
277		break;
278	case SIOCADDMULTI:
279	case SIOCDELMULTI:
280		break;
281
282
283	default:
284		error = EINVAL;
285	}
286	splx(s);
287	return (error);
288}
289
290/*
291 * tunoutput - queue packets from higher level ready to put out.
292 */
293int
294tunoutput(ifp, m0, dst, rt)
295	struct ifnet   *ifp;
296	struct mbuf    *m0;
297	struct sockaddr *dst;
298	struct rtentry *rt;
299{
300	struct tun_softc *tp = &tunctl[ifp->if_unit];
301	struct proc	*p;
302	int		s;
303
304	TUNDEBUG ("%s%d: tunoutput\n", ifp->if_name, ifp->if_unit);
305
306	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
307		TUNDEBUG ("%s%d: not ready 0%o\n", ifp->if_name,
308			  ifp->if_unit, tp->tun_flags);
309		m_freem (m0);
310		return EHOSTDOWN;
311	}
312
313#if NBPFILTER > 0
314	/* BPF write needs to be handled specially */
315	if (dst->sa_family == AF_UNSPEC) {
316		dst->sa_family = *(mtod(m0, int *));
317		m0->m_len -= sizeof(int);
318		m0->m_pkthdr.len -= sizeof(int);
319		m0->m_data += sizeof(int);
320	}
321
322	if (ifp->if_bpf) {
323		/*
324		 * We need to prepend the address family as
325		 * a four byte field.  Cons up a dummy header
326		 * to pacify bpf.  This is safe because bpf
327		 * will only read from the mbuf (i.e., it won't
328		 * try to free it or keep a pointer to it).
329		 */
330		struct mbuf m;
331		u_int af = dst->sa_family;
332
333		m.m_next = m0;
334		m.m_len = 4;
335		m.m_data = (char *)&af;
336
337		bpf_mtap(ifp, &m);
338	}
339#endif
340
341	switch(dst->sa_family) {
342#ifdef INET
343	case AF_INET:
344		s = splimp();
345		if (IF_QFULL(&ifp->if_snd)) {
346			IF_DROP(&ifp->if_snd);
347			m_freem(m0);
348			splx(s);
349			ifp->if_collisions++;
350			return (ENOBUFS);
351		}
352		ifp->if_obytes += m0->m_pkthdr.len;
353		IF_ENQUEUE(&ifp->if_snd, m0);
354		splx(s);
355		ifp->if_opackets++;
356		break;
357#endif
358	default:
359		m_freem(m0);
360		return EAFNOSUPPORT;
361	}
362
363	if (tp->tun_flags & TUN_RWAIT) {
364		tp->tun_flags &= ~TUN_RWAIT;
365		wakeup((caddr_t)tp);
366	}
367	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgrp) {
368		if (tp->tun_pgrp > 0)
369			gsignal(tp->tun_pgrp, SIGIO);
370		else if ((p = pfind(-tp->tun_pgrp)) != 0)
371			psignal(p, SIGIO);
372	}
373	selwakeup(&tp->tun_rsel);
374	return 0;
375}
376
377/*
378 * the cdevsw interface is now pretty minimal.
379 */
380static	int
381tunioctl(dev, cmd, data, flag, p)
382	dev_t		dev;
383	int		cmd;
384	caddr_t		data;
385	int		flag;
386	struct proc	*p;
387{
388	int		unit = minor(dev), s;
389	struct tun_softc *tp = &tunctl[unit];
390 	struct tuninfo *tunp;
391
392	switch (cmd) {
393 	case TUNSIFINFO:
394 	        tunp = (struct tuninfo *)data;
395 		tp->tun_if.if_mtu = tunp->mtu;
396 		tp->tun_if.if_type = tunp->type;
397 		tp->tun_if.if_baudrate = tunp->baudrate;
398 		break;
399 	case TUNGIFINFO:
400 		tunp = (struct tuninfo *)data;
401 		tunp->mtu = tp->tun_if.if_mtu;
402 		tunp->type = tp->tun_if.if_type;
403 		tunp->baudrate = tp->tun_if.if_baudrate;
404 		break;
405	case TUNSDEBUG:
406		tundebug = *(int *)data;
407		break;
408	case TUNGDEBUG:
409		*(int *)data = tundebug;
410		break;
411	case FIONBIO:
412		if (*(int *)data)
413			tp->tun_flags |= TUN_NBIO;
414		else
415			tp->tun_flags &= ~TUN_NBIO;
416		break;
417	case FIOASYNC:
418		if (*(int *)data)
419			tp->tun_flags |= TUN_ASYNC;
420		else
421			tp->tun_flags &= ~TUN_ASYNC;
422		break;
423	case FIONREAD:
424		s = splimp();
425		if (tp->tun_if.if_snd.ifq_head) {
426			struct mbuf *mb = tp->tun_if.if_snd.ifq_head;
427			for( *(int *)data = 0; mb != 0; mb = mb->m_next)
428				*(int *)data += mb->m_len;
429		} else
430			*(int *)data = 0;
431		splx(s);
432		break;
433	case TIOCSPGRP:
434		tp->tun_pgrp = *(int *)data;
435		break;
436	case TIOCGPGRP:
437		*(int *)data = tp->tun_pgrp;
438		break;
439	default:
440		return (ENOTTY);
441	}
442	return (0);
443}
444
445/*
446 * The cdevsw read interface - reads a packet at a time, or at
447 * least as much of a packet as can be read.
448 */
449static	int
450tunread(dev_t dev, struct uio *uio, int flag)
451{
452	int		unit = minor(dev);
453	struct tun_softc *tp = &tunctl[unit];
454	struct ifnet	*ifp = &tp->tun_if;
455	struct mbuf	*m, *m0;
456	int		error=0, len, s;
457
458	TUNDEBUG ("%s%d: read\n", ifp->if_name, ifp->if_unit);
459	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
460		TUNDEBUG ("%s%d: not ready 0%o\n", ifp->if_name,
461			  ifp->if_unit, tp->tun_flags);
462		return EHOSTDOWN;
463	}
464
465	tp->tun_flags &= ~TUN_RWAIT;
466
467	s = splimp();
468	do {
469		IF_DEQUEUE(&ifp->if_snd, m0);
470		if (m0 == 0) {
471			if (tp->tun_flags & TUN_NBIO) {
472				splx(s);
473				return EWOULDBLOCK;
474			}
475			tp->tun_flags |= TUN_RWAIT;
476			if( error = tsleep((caddr_t)tp, PCATCH | (PZERO + 1),
477					"tunread", 0)) {
478				splx(s);
479				return error;
480			}
481		}
482	} while (m0 == 0);
483	splx(s);
484
485	while (m0 && uio->uio_resid > 0 && error == 0) {
486		len = min(uio->uio_resid, m0->m_len);
487		if (len == 0)
488			break;
489		error = uiomove(mtod(m0, caddr_t), len, uio);
490		MFREE(m0, m);
491		m0 = m;
492	}
493
494	if (m0) {
495		TUNDEBUG("Dropping mbuf\n");
496		m_freem(m0);
497	}
498	return error;
499}
500
501/*
502 * the cdevsw write interface - an atomic write is a packet - or else!
503 */
504static	int
505tunwrite(dev_t dev, struct uio *uio, int flag)
506{
507	int		unit = minor (dev);
508	struct ifnet	*ifp = &tunctl[unit].tun_if;
509	struct mbuf	*top, **mp, *m;
510	int		error=0, s, tlen, mlen;
511
512	TUNDEBUG("%s%d: tunwrite\n", ifp->if_name, ifp->if_unit);
513
514	if (uio->uio_resid < 0 || uio->uio_resid > TUNMRU) {
515		TUNDEBUG("%s%d: len=%d!\n", ifp->if_name, ifp->if_unit,
516		    uio->uio_resid);
517		return EIO;
518	}
519	tlen = uio->uio_resid;
520
521	/* get a header mbuf */
522	MGETHDR(m, M_DONTWAIT, MT_DATA);
523	if (m == NULL)
524		return ENOBUFS;
525	mlen = MHLEN;
526
527	top = 0;
528	mp = &top;
529	while (error == 0 && uio->uio_resid > 0) {
530		m->m_len = min(mlen, uio->uio_resid);
531		error = uiomove(mtod (m, caddr_t), m->m_len, uio);
532		*mp = m;
533		mp = &m->m_next;
534		if (uio->uio_resid > 0) {
535			MGET (m, M_DONTWAIT, MT_DATA);
536			if (m == 0) {
537				error = ENOBUFS;
538				break;
539			}
540			mlen = MLEN;
541		}
542	}
543	if (error) {
544		if (top)
545			m_freem (top);
546		return error;
547	}
548
549	top->m_pkthdr.len = tlen;
550	top->m_pkthdr.rcvif = ifp;
551
552#if NBPFILTER > 0
553	if (ifp->if_bpf) {
554		/*
555		 * We need to prepend the address family as
556		 * a four byte field.  Cons up a dummy header
557		 * to pacify bpf.  This is safe because bpf
558		 * will only read from the mbuf (i.e., it won't
559		 * try to free it or keep a pointer to it).
560		 */
561		struct mbuf m;
562		u_int af = AF_INET;
563
564		m.m_next = top;
565		m.m_len = 4;
566		m.m_data = (char *)&af;
567
568		bpf_mtap(ifp, &m);
569	}
570#endif
571
572	s = splimp();
573	if (IF_QFULL (&ipintrq)) {
574		IF_DROP(&ipintrq);
575		splx(s);
576		ifp->if_collisions++;
577		m_freem(top);
578		return ENOBUFS;
579	}
580	IF_ENQUEUE(&ipintrq, top);
581	splx(s);
582	ifp->if_ibytes += tlen;
583	ifp->if_ipackets++;
584	schednetisr(NETISR_IP);
585	return error;
586}
587
588/*
589 * tunselect - the select interface, this is only useful on reads
590 * really. The write detect always returns true, write never blocks
591 * anyway, it either accepts the packet or drops it.
592 */
593static	int
594tunselect(dev_t dev, int rw, struct proc *p)
595{
596	int		unit = minor(dev), s;
597	struct tun_softc *tp = &tunctl[unit];
598	struct ifnet	*ifp = &tp->tun_if;
599
600	s = splimp();
601	TUNDEBUG("%s%d: tunselect\n", ifp->if_name, ifp->if_unit);
602
603	switch (rw) {
604	case FREAD:
605		if (ifp->if_snd.ifq_len > 0) {
606			splx(s);
607			TUNDEBUG("%s%d: tunselect q=%d\n", ifp->if_name,
608			    ifp->if_unit, ifp->if_snd.ifq_len);
609			return 1;
610		}
611		selrecord(p, &tp->tun_rsel);
612		break;
613	case FWRITE:
614		splx(s);
615		return 1;
616	}
617	splx(s);
618	TUNDEBUG("%s%d: tunselect waiting\n", ifp->if_name, ifp->if_unit);
619	return 0;
620}
621
622
623#endif  /* NTUN */
624