if_tun.c revision 12708 
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/ioctl.h>
28#include <sys/errno.h>
29#include <sys/syslog.h>
30#include <sys/select.h>
31#include <sys/file.h>
32#include <sys/signalvar.h>
33#include <sys/kernel.h>
34#include <sys/sysctl.h>
35#ifdef DEVFS
36#include <sys/devfsext.h>
37#endif /*DEVFS*/
38#include <sys/conf.h>
39
40#include <machine/cpu.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_RD, &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#ifdef DEVFS
109	char	name[32];
110#endif
111
112	if( tun_devsw_installed ) return;
113	dev = makedev(CDEV_MAJOR, 0);
114	cdevsw_add(&dev,&tun_cdevsw, NULL);
115	tun_devsw_installed = 1;
116	for ( i = 0; i < NTUN; i++ ) {
117#ifdef DEVFS
118		sprintf(name, "tun%d", i );
119		tun_devfs_token[i] = devfs_add_devsw(
120			"/", name, &tun_cdevsw , i,
121			DV_CHR, 0, 0, 0600);
122#endif
123		tunctl[i].tun_flags = TUN_INITED;
124
125		ifp = &tunctl[i].tun_if;
126		ifp->if_unit = i;
127		ifp->if_name = "tun";
128		ifp->if_mtu = TUNMTU;
129		ifp->if_ioctl = tunifioctl;
130		ifp->if_output = tunoutput;
131		ifp->if_flags = IFF_POINTOPOINT | IFF_MULTICAST;
132		ifp->if_snd.ifq_maxlen = ifqmaxlen;
133		ifp->if_collisions = 0;
134		ifp->if_ierrors = 0;
135		ifp->if_oerrors = 0;
136		ifp->if_ipackets = 0;
137		ifp->if_opackets = 0;
138		if_attach(ifp);
139#if NBPFILTER > 0
140		bpfattach(&tunctl[i].tun_bpf, ifp, DLT_NULL, sizeof(u_int));
141#endif
142	}
143}
144
145/*
146 * tunnel open - must be superuser & the device must be
147 * configured in
148 */
149static	int
150tunopen(dev, flag, mode, p)
151	dev_t	dev;
152	int	flag, mode;
153	struct proc *p;
154{
155	struct ifnet	*ifp;
156	struct tun_softc *tp;
157	register int	unit, error;
158
159	error = suser(p->p_ucred, &p->p_acflag);
160	if (error)
161		return (error);
162
163	if ((unit = minor(dev)) >= NTUN)
164		return (ENXIO);
165	tp = &tunctl[unit];
166	if (tp->tun_flags & TUN_OPEN)
167		return ENXIO;
168	ifp = &tp->tun_if;
169	tp->tun_flags |= TUN_OPEN;
170	TUNDEBUG("%s%d: open\n", ifp->if_name, ifp->if_unit);
171	return (0);
172}
173
174/*
175 * tunclose - close the device - mark i/f down & delete
176 * routing info
177 */
178static	int
179tunclose(dev_t dev, int foo, int bar, struct proc *p)
180{
181	register int	unit = minor(dev), s;
182	struct tun_softc *tp = &tunctl[unit];
183	struct ifnet	*ifp = &tp->tun_if;
184	struct mbuf	*m;
185
186	tp->tun_flags &= ~TUN_OPEN;
187
188	/*
189	 * junk all pending output
190	 */
191	do {
192		s = splimp();
193		IF_DEQUEUE(&ifp->if_snd, m);
194		splx(s);
195		if (m)
196			m_freem(m);
197	} while (m);
198
199	if (ifp->if_flags & IFF_UP) {
200		s = splimp();
201		if_down(ifp);
202		if (ifp->if_flags & IFF_RUNNING) {
203		    /* find internet addresses and delete routes */
204		    register struct ifaddr *ifa;
205		    for (ifa = ifp->if_addrlist; ifa; ifa = ifa->ifa_next) {
206			if (ifa->ifa_addr->sa_family == AF_INET) {
207			    rtinit(ifa, (int)RTM_DELETE,
208				   tp->tun_flags & TUN_DSTADDR ? RTF_HOST : 0);
209			}
210		    }
211		}
212		splx(s);
213	}
214	tp->tun_pgrp = 0;
215	selwakeup(&tp->tun_rsel);
216
217	TUNDEBUG ("%s%d: closed\n", ifp->if_name, ifp->if_unit);
218	return (0);
219}
220
221static int
222tuninit(unit)
223	int	unit;
224{
225	struct tun_softc *tp = &tunctl[unit];
226	struct ifnet	*ifp = &tp->tun_if;
227	register struct ifaddr *ifa;
228
229	TUNDEBUG("%s%d: tuninit\n", ifp->if_name, ifp->if_unit);
230
231	ifp->if_flags |= IFF_UP | IFF_RUNNING;
232
233	for (ifa = ifp->if_addrlist; ifa; ifa = ifa->ifa_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 SIOCADDMULTI:
274	case SIOCDELMULTI:
275		if (ifr == 0) {
276			error = EAFNOSUPPORT;		/* XXX */
277			break;
278		}
279		switch (ifr->ifr_addr.sa_family) {
280
281#ifdef INET
282		case AF_INET:
283			break;
284#endif
285
286		default:
287			error = EAFNOSUPPORT;
288			break;
289		}
290		break;
291
292
293	default:
294		error = EINVAL;
295	}
296	splx(s);
297	return (error);
298}
299
300/*
301 * tunoutput - queue packets from higher level ready to put out.
302 */
303int
304tunoutput(ifp, m0, dst, rt)
305	struct ifnet   *ifp;
306	struct mbuf    *m0;
307	struct sockaddr *dst;
308	struct rtentry *rt;
309{
310	struct tun_softc *tp = &tunctl[ifp->if_unit];
311	struct proc	*p;
312	int		s;
313
314	TUNDEBUG ("%s%d: tunoutput\n", ifp->if_name, ifp->if_unit);
315
316	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
317		TUNDEBUG ("%s%d: not ready 0%o\n", ifp->if_name,
318			  ifp->if_unit, tp->tun_flags);
319		m_freem (m0);
320		return EHOSTDOWN;
321	}
322
323#if NBPFILTER > 0
324	/* BPF write needs to be handled specially */
325	if (dst->sa_family == AF_UNSPEC) {
326		dst->sa_family = *(mtod(m0, int *));
327		m0->m_len -= sizeof(int);
328		m0->m_pkthdr.len -= sizeof(int);
329		m0->m_data += sizeof(int);
330	}
331
332	if (tp->tun_bpf) {
333		/*
334		 * We need to prepend the address family as
335		 * a four byte field.  Cons up a dummy header
336		 * to pacify bpf.  This is safe because bpf
337		 * will only read from the mbuf (i.e., it won't
338		 * try to free it or keep a pointer to it).
339		 */
340		struct mbuf m;
341		u_int af = dst->sa_family;
342
343		m.m_next = m0;
344		m.m_len = 4;
345		m.m_data = (char *)&af;
346
347		bpf_mtap(tp->tun_bpf, &m);
348	}
349#endif
350
351	switch(dst->sa_family) {
352#ifdef INET
353	case AF_INET:
354		s = splimp();
355		if (IF_QFULL(&ifp->if_snd)) {
356			IF_DROP(&ifp->if_snd);
357			m_freem(m0);
358			splx(s);
359			ifp->if_collisions++;
360			return (ENOBUFS);
361		}
362		IF_ENQUEUE(&ifp->if_snd, m0);
363		splx(s);
364		ifp->if_opackets++;
365		break;
366#endif
367	default:
368		m_freem(m0);
369		return EAFNOSUPPORT;
370	}
371
372	if (tp->tun_flags & TUN_RWAIT) {
373		tp->tun_flags &= ~TUN_RWAIT;
374		wakeup((caddr_t)tp);
375	}
376	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgrp) {
377		if (tp->tun_pgrp > 0)
378			gsignal(tp->tun_pgrp, SIGIO);
379		else if ((p = pfind(-tp->tun_pgrp)) != 0)
380			psignal(p, SIGIO);
381	}
382	selwakeup(&tp->tun_rsel);
383	return 0;
384}
385
386/*
387 * the cdevsw interface is now pretty minimal.
388 */
389static	int
390tunioctl(dev, cmd, data, flag, p)
391	dev_t		dev;
392	int		cmd;
393	caddr_t		data;
394	int		flag;
395	struct proc	*p;
396{
397	int		unit = minor(dev), s;
398	struct tun_softc *tp = &tunctl[unit];
399 	struct tuninfo *tunp;
400
401	switch (cmd) {
402 	case TUNSIFINFO:
403 	        tunp = (struct tuninfo *)data;
404 		tp->tun_if.if_mtu = tunp->mtu;
405 		tp->tun_if.if_type = tunp->type;
406 		tp->tun_if.if_baudrate = tunp->baudrate;
407 		break;
408 	case TUNGIFINFO:
409 		tunp = (struct tuninfo *)data;
410 		tunp->mtu = tp->tun_if.if_mtu;
411 		tunp->type = tp->tun_if.if_type;
412 		tunp->baudrate = tp->tun_if.if_baudrate;
413 		break;
414	case TUNSDEBUG:
415		tundebug = *(int *)data;
416		break;
417	case TUNGDEBUG:
418		*(int *)data = tundebug;
419		break;
420	case FIONBIO:
421		if (*(int *)data)
422			tp->tun_flags |= TUN_NBIO;
423		else
424			tp->tun_flags &= ~TUN_NBIO;
425		break;
426	case FIOASYNC:
427		if (*(int *)data)
428			tp->tun_flags |= TUN_ASYNC;
429		else
430			tp->tun_flags &= ~TUN_ASYNC;
431		break;
432	case FIONREAD:
433		s = splimp();
434		if (tp->tun_if.if_snd.ifq_head)
435			*(int *)data = tp->tun_if.if_snd.ifq_head->m_len;
436		else
437			*(int *)data = 0;
438		splx(s);
439		break;
440	case TIOCSPGRP:
441		tp->tun_pgrp = *(int *)data;
442		break;
443	case TIOCGPGRP:
444		*(int *)data = tp->tun_pgrp;
445		break;
446	default:
447		return (ENOTTY);
448	}
449	return (0);
450}
451
452/*
453 * The cdevsw read interface - reads a packet at a time, or at
454 * least as much of a packet as can be read.
455 */
456static	int
457tunread(dev_t dev, struct uio *uio, int flag)
458{
459	int		unit = minor(dev);
460	struct tun_softc *tp = &tunctl[unit];
461	struct ifnet	*ifp = &tp->tun_if;
462	struct mbuf	*m, *m0;
463	int		error=0, len, s;
464
465	TUNDEBUG ("%s%d: read\n", ifp->if_name, ifp->if_unit);
466	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
467		TUNDEBUG ("%s%d: not ready 0%o\n", ifp->if_name,
468			  ifp->if_unit, tp->tun_flags);
469		return EHOSTDOWN;
470	}
471
472	tp->tun_flags &= ~TUN_RWAIT;
473
474	s = splimp();
475	do {
476		IF_DEQUEUE(&ifp->if_snd, m0);
477		if (m0 == 0) {
478			if (tp->tun_flags & TUN_NBIO) {
479				splx(s);
480				return EWOULDBLOCK;
481			}
482			tp->tun_flags |= TUN_RWAIT;
483			tsleep((caddr_t)tp, PZERO + 1, "tunread", 0);
484		}
485	} while (m0 == 0);
486	splx(s);
487
488	while (m0 && uio->uio_resid > 0 && error == 0) {
489		len = min(uio->uio_resid, m0->m_len);
490		if (len == 0)
491			break;
492		error = uiomove(mtod(m0, caddr_t), len, uio);
493		MFREE(m0, m);
494		m0 = m;
495	}
496
497	if (m0) {
498		TUNDEBUG("Dropping mbuf\n");
499		m_freem(m0);
500	}
501	return error;
502}
503
504/*
505 * the cdevsw write interface - an atomic write is a packet - or else!
506 */
507static	int
508tunwrite(dev_t dev, struct uio *uio, int flag)
509{
510	int		unit = minor (dev);
511	struct ifnet	*ifp = &tunctl[unit].tun_if;
512	struct mbuf	*top, **mp, *m;
513	int		error=0, s, tlen, mlen;
514
515	TUNDEBUG("%s%d: tunwrite\n", ifp->if_name, ifp->if_unit);
516
517	if (uio->uio_resid < 0 || uio->uio_resid > TUNMTU) {
518		TUNDEBUG("%s%d: len=%d!\n", ifp->if_name, ifp->if_unit,
519		    uio->uio_resid);
520		return EIO;
521	}
522	tlen = uio->uio_resid;
523
524	/* get a header mbuf */
525	MGETHDR(m, M_DONTWAIT, MT_DATA);
526	if (m == NULL)
527		return ENOBUFS;
528	mlen = MHLEN;
529
530	top = 0;
531	mp = &top;
532	while (error == 0 && uio->uio_resid > 0) {
533		m->m_len = min(mlen, uio->uio_resid);
534		error = uiomove(mtod (m, caddr_t), m->m_len, uio);
535		*mp = m;
536		mp = &m->m_next;
537		if (uio->uio_resid > 0) {
538			MGET (m, M_DONTWAIT, MT_DATA);
539			if (m == 0) {
540				error = ENOBUFS;
541				break;
542			}
543			mlen = MLEN;
544		}
545	}
546	if (error) {
547		if (top)
548			m_freem (top);
549		return error;
550	}
551
552	top->m_pkthdr.len = tlen;
553	top->m_pkthdr.rcvif = ifp;
554
555#if NBPFILTER > 0
556	if (tunctl[unit].tun_bpf) {
557		/*
558		 * We need to prepend the address family as
559		 * a four byte field.  Cons up a dummy header
560		 * to pacify bpf.  This is safe because bpf
561		 * will only read from the mbuf (i.e., it won't
562		 * try to free it or keep a pointer to it).
563		 */
564		struct mbuf m;
565		u_int af = AF_INET;
566
567		m.m_next = top;
568		m.m_len = 4;
569		m.m_data = (char *)&af;
570
571		bpf_mtap(tunctl[unit].tun_bpf, &m);
572	}
573#endif
574
575	s = splimp();
576	if (IF_QFULL (&ipintrq)) {
577		IF_DROP(&ipintrq);
578		splx(s);
579		ifp->if_collisions++;
580		m_freem(top);
581		return ENOBUFS;
582	}
583	IF_ENQUEUE(&ipintrq, top);
584	splx(s);
585	ifp->if_ipackets++;
586	schednetisr(NETISR_IP);
587	return error;
588}
589
590/*
591 * tunselect - the select interface, this is only useful on reads
592 * really. The write detect always returns true, write never blocks
593 * anyway, it either accepts the packet or drops it.
594 */
595static	int
596tunselect(dev_t dev, int rw, struct proc *p)
597{
598	int		unit = minor(dev), s;
599	struct tun_softc *tp = &tunctl[unit];
600	struct ifnet	*ifp = &tp->tun_if;
601
602	s = splimp();
603	TUNDEBUG("%s%d: tunselect\n", ifp->if_name, ifp->if_unit);
604
605	switch (rw) {
606	case FREAD:
607		if (ifp->if_snd.ifq_len > 0) {
608			splx(s);
609			TUNDEBUG("%s%d: tunselect q=%d\n", ifp->if_name,
610			    ifp->if_unit, ifp->if_snd.ifq_len);
611			return 1;
612		}
613		selrecord(p, &tp->tun_rsel);
614		break;
615	case FWRITE:
616		splx(s);
617		return 1;
618	}
619	splx(s);
620	TUNDEBUG("%s%d: tunselect waiting\n", ifp->if_name, ifp->if_unit);
621	return 0;
622}
623
624
625#endif  /* NTUN */
626