Home | History | Annotate | Line # | Download | only in net
if_tun.c revision 1.53
      1 /*	$NetBSD: if_tun.c,v 1.53 2002/09/06 13:18:43 gehenna Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 1988, Julian Onions <jpo (at) 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 1.53 2002/09/06 13:18:43 gehenna Exp $");
     19 
     20 #include "tun.h"
     21 
     22 #include "opt_inet.h"
     23 #include "opt_ns.h"
     24 
     25 #include <sys/param.h>
     26 #include <sys/proc.h>
     27 #include <sys/systm.h>
     28 #include <sys/mbuf.h>
     29 #include <sys/buf.h>
     30 #include <sys/protosw.h>
     31 #include <sys/socket.h>
     32 #include <sys/ioctl.h>
     33 #include <sys/errno.h>
     34 #include <sys/syslog.h>
     35 #include <sys/select.h>
     36 #include <sys/poll.h>
     37 #include <sys/file.h>
     38 #include <sys/signalvar.h>
     39 #include <sys/conf.h>
     40 
     41 #include <machine/cpu.h>
     42 
     43 #include <net/if.h>
     44 #include <net/if_ether.h>
     45 #include <net/netisr.h>
     46 #include <net/route.h>
     47 
     48 
     49 #ifdef INET
     50 #include <netinet/in.h>
     51 #include <netinet/in_systm.h>
     52 #include <netinet/in_var.h>
     53 #include <netinet/ip.h>
     54 #include <netinet/if_inarp.h>
     55 #endif
     56 
     57 #ifdef NS
     58 #include <netns/ns.h>
     59 #include <netns/ns_if.h>
     60 #endif
     61 
     62 #include "bpfilter.h"
     63 #if NBPFILTER > 0
     64 #include <sys/time.h>
     65 #include <net/bpf.h>
     66 #endif
     67 
     68 #include <net/if_tun.h>
     69 
     70 #define TUNDEBUG	if (tundebug) printf
     71 int	tundebug = 0;
     72 
     73 extern int ifqmaxlen;
     74 void	tunattach __P((int));
     75 LIST_HEAD(, tun_softc) tun_softc_list;
     76 static struct simplelock tun_softc_lock;
     77 
     78 int	tun_ioctl __P((struct ifnet *, u_long, caddr_t));
     79 int	tun_output __P((struct ifnet *, struct mbuf *, struct sockaddr *,
     80 		       struct rtentry *rt));
     81 int	tun_clone_create __P((struct if_clone *, int));
     82 void	tun_clone_destroy __P((struct ifnet *));
     83 
     84 struct if_clone tun_cloner =
     85     IF_CLONE_INITIALIZER("tun", tun_clone_create, tun_clone_destroy);
     86 
     87 static void tunattach0 __P((struct tun_softc *));
     88 static void tuninit __P((struct tun_softc *));
     89 #ifdef ALTQ
     90 static void tunstart __P((struct ifnet *));
     91 #endif
     92 static struct tun_softc *tun_find_unit __P((dev_t));
     93 
     94 dev_type_open(tunopen);
     95 dev_type_close(tunclose);
     96 dev_type_read(tunread);
     97 dev_type_write(tunwrite);
     98 dev_type_ioctl(tunioctl);
     99 dev_type_poll(tunpoll);
    100 
    101 const struct cdevsw tun_cdevsw = {
    102 	tunopen, tunclose, tunread, tunwrite, tunioctl,
    103 	nostop, notty, tunpoll, nommap,
    104 };
    105 
    106 void
    107 tunattach(unused)
    108 	int unused;
    109 {
    110 
    111 	simple_lock_init(&tun_softc_lock);
    112 	LIST_INIT(&tun_softc_list);
    113 	if_clone_attach(&tun_cloner);
    114 }
    115 
    116 int
    117 tun_clone_create(ifc, unit)
    118 	struct if_clone *ifc;
    119 	int unit;
    120 {
    121 	struct tun_softc *sc;
    122 
    123 	sc = malloc(sizeof(struct tun_softc), M_DEVBUF, M_WAITOK);
    124 	(void)memset(sc, 0, sizeof(struct tun_softc));
    125 
    126 	(void)snprintf(sc->tun_if.if_xname, sizeof(sc->tun_if.if_xname),
    127 	    "%s%d", ifc->ifc_name, unit);
    128 	sc->tun_unit = unit;
    129 	simple_lock_init(&sc->tun_lock);
    130 
    131 	tunattach0(sc);
    132 
    133 	simple_lock(&tun_softc_lock);
    134 	LIST_INSERT_HEAD(&tun_softc_list, sc, tun_list);
    135 	simple_unlock(&tun_softc_lock);
    136 
    137 	return (0);
    138 }
    139 
    140 void
    141 tunattach0(sc)
    142 	struct tun_softc *sc;
    143 {
    144 	struct ifnet *ifp = (void *)sc;
    145 
    146 	sc->tun_flags = TUN_INITED;
    147 
    148 	ifp = &sc->tun_if;
    149 	ifp->if_softc = sc;
    150 	ifp->if_mtu = TUNMTU;
    151 	ifp->if_ioctl = tun_ioctl;
    152 	ifp->if_output = tun_output;
    153 #ifdef ALTQ
    154 	ifp->if_start = tunstart;
    155 #endif
    156 	ifp->if_flags = IFF_POINTOPOINT;
    157 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
    158 	ifp->if_collisions = 0;
    159 	ifp->if_ierrors = 0;
    160 	ifp->if_oerrors = 0;
    161 	ifp->if_ipackets = 0;
    162 	ifp->if_opackets = 0;
    163 	ifp->if_dlt = DLT_NULL;
    164 	IFQ_SET_READY(&ifp->if_snd);
    165 	if_attach(ifp);
    166 	if_alloc_sadl(ifp);
    167 #if NBPFILTER > 0
    168 	bpfattach(ifp, DLT_NULL, sizeof(u_int32_t));
    169 #endif
    170 }
    171 
    172 void
    173 tun_clone_destroy(ifp)
    174 	struct ifnet *ifp;
    175 {
    176 	struct tun_softc *tp = (void *)ifp;
    177 	struct proc *p;
    178 
    179 	simple_lock(&tun_softc_lock);
    180 	simple_lock(&tp->tun_lock);
    181 	LIST_REMOVE(tp, tun_list);
    182 	simple_unlock(&tp->tun_lock);
    183 	simple_unlock(&tun_softc_lock);
    184 
    185 	if (tp->tun_flags & TUN_RWAIT) {
    186 		tp->tun_flags &= ~TUN_RWAIT;
    187 		wakeup((caddr_t)tp);
    188 	}
    189 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgrp) {
    190 		if (tp->tun_pgrp > 0)
    191 			gsignal(tp->tun_pgrp, SIGIO);
    192 		else if ((p = pfind(-tp->tun_pgrp)) != NULL)
    193 			psignal(p, SIGIO);
    194 	}
    195 	selwakeup(&tp->tun_rsel);
    196 
    197 #if NBPFILTER > 0
    198 	bpfdetach(ifp);
    199 #endif
    200 	if_detach(ifp);
    201 
    202 	free(tp, M_DEVBUF);
    203 }
    204 
    205 static struct tun_softc *
    206 tun_find_unit(dev)
    207 	dev_t dev;
    208 {
    209 	struct tun_softc *tp;
    210 	int unit = minor(dev);
    211 
    212 	simple_lock(&tun_softc_lock);
    213 	LIST_FOREACH(tp, &tun_softc_list, tun_list)
    214 		if (unit == tp->tun_unit)
    215 			break;
    216 	if (tp)
    217 		simple_lock(&tp->tun_lock);
    218 	simple_unlock(&tun_softc_lock);
    219 
    220 	return (tp);
    221 }
    222 
    223 /*
    224  * tunnel open - must be superuser & the device must be
    225  * configured in
    226  */
    227 int
    228 tunopen(dev, flag, mode, p)
    229 	dev_t	dev;
    230 	int	flag, mode;
    231 	struct proc *p;
    232 {
    233 	struct ifnet	*ifp;
    234 	struct tun_softc *tp;
    235 	int	error;
    236 
    237 	if ((error = suser(p->p_ucred, &p->p_acflag)) != 0)
    238 		return (error);
    239 
    240 	if (NTUN < 1)
    241 		return (ENXIO);
    242 
    243 	tp = tun_find_unit(dev);
    244 
    245 	if (!tp) {
    246 		(void)tun_clone_create(&tun_cloner, minor(dev));
    247 		tp = tun_find_unit(dev);
    248 	}
    249 
    250 	if (!tp)
    251 		return (ENXIO);
    252 
    253 	if (tp->tun_flags & TUN_OPEN) {
    254 		simple_unlock(&tp->tun_lock);
    255 		return (EBUSY);
    256 	}
    257 
    258 	ifp = &tp->tun_if;
    259 	tp->tun_flags |= TUN_OPEN;
    260 	TUNDEBUG("%s: open\n", ifp->if_xname);
    261 	simple_unlock(&tp->tun_lock);
    262 	return (0);
    263 }
    264 
    265 /*
    266  * tunclose - close the device - mark i/f down & delete
    267  * routing info
    268  */
    269 int
    270 tunclose(dev, flag, mode, p)
    271 	dev_t	dev;
    272 	int	flag;
    273 	int	mode;
    274 	struct proc *p;
    275 {
    276 	int	s;
    277 	struct tun_softc *tp;
    278 	struct ifnet	*ifp;
    279 
    280 	tp = tun_find_unit(dev);
    281 
    282 	/* interface was "destroyed" before the close */
    283 	if (tp == NULL)
    284 		return (0);
    285 
    286 	ifp = &tp->tun_if;
    287 
    288 	tp->tun_flags &= ~TUN_OPEN;
    289 
    290 	/*
    291 	 * junk all pending output
    292 	 */
    293 	s = splnet();
    294 	IFQ_PURGE(&ifp->if_snd);
    295 	splx(s);
    296 
    297 	if (ifp->if_flags & IFF_UP) {
    298 		s = splnet();
    299 		if_down(ifp);
    300 		if (ifp->if_flags & IFF_RUNNING) {
    301 			/* find internet addresses and delete routes */
    302 			struct ifaddr *ifa;
    303 			TAILQ_FOREACH(ifa, &ifp->if_addrlist, ifa_list) {
    304 #ifdef INET
    305 				if (ifa->ifa_addr->sa_family == AF_INET) {
    306 					rtinit(ifa, (int)RTM_DELETE,
    307 					       tp->tun_flags & TUN_DSTADDR
    308 							? RTF_HOST
    309 							: 0);
    310 				}
    311 #endif
    312 			}
    313 		}
    314 		splx(s);
    315 	}
    316 	tp->tun_pgrp = 0;
    317 	selwakeup(&tp->tun_rsel);
    318 
    319 	TUNDEBUG ("%s: closed\n", ifp->if_xname);
    320 	simple_unlock(&tp->tun_lock);
    321 	return (0);
    322 }
    323 
    324 static void
    325 tuninit(tp)
    326 	struct tun_softc *tp;
    327 {
    328 	struct ifnet	*ifp = &tp->tun_if;
    329 	struct ifaddr	*ifa;
    330 
    331 	TUNDEBUG("%s: tuninit\n", ifp->if_xname);
    332 
    333 	ifp->if_flags |= IFF_UP | IFF_RUNNING;
    334 
    335 	tp->tun_flags &= ~(TUN_IASET|TUN_DSTADDR);
    336 	TAILQ_FOREACH(ifa, &ifp->if_addrlist, ifa_list) {
    337 #ifdef INET
    338 		if (ifa->ifa_addr->sa_family == AF_INET) {
    339 			struct sockaddr_in *sin;
    340 
    341 			sin = satosin(ifa->ifa_addr);
    342 			if (sin && sin->sin_addr.s_addr)
    343 				tp->tun_flags |= TUN_IASET;
    344 
    345 			if (ifp->if_flags & IFF_POINTOPOINT) {
    346 				sin = satosin(ifa->ifa_dstaddr);
    347 				if (sin && sin->sin_addr.s_addr)
    348 					tp->tun_flags |= TUN_DSTADDR;
    349 			}
    350 		}
    351 #endif
    352 	}
    353 
    354 	return;
    355 }
    356 
    357 /*
    358  * Process an ioctl request.
    359  */
    360 int
    361 tun_ioctl(ifp, cmd, data)
    362 	struct ifnet *ifp;
    363 	u_long cmd;
    364 	caddr_t	data;
    365 {
    366 	int		error = 0, s;
    367 	struct tun_softc *tp = (struct tun_softc *)(ifp->if_softc);
    368 
    369 	simple_lock(&tp->tun_lock);
    370 
    371 	s = splnet();
    372 	switch(cmd) {
    373 	case SIOCSIFADDR:
    374 		tuninit((struct tun_softc *)(ifp->if_softc));
    375 		TUNDEBUG("%s: address set\n", ifp->if_xname);
    376 		break;
    377 	case SIOCSIFDSTADDR:
    378 		tuninit((struct tun_softc *)(ifp->if_softc));
    379 		TUNDEBUG("%s: destination address set\n", ifp->if_xname);
    380 		break;
    381 	case SIOCSIFBRDADDR:
    382 		TUNDEBUG("%s: broadcast address set\n", ifp->if_xname);
    383 		break;
    384 	case SIOCSIFMTU: {
    385 		struct ifreq *ifr = (struct ifreq *) data;
    386 		if (ifr->ifr_mtu > TUNMTU || ifr->ifr_mtu < 576) {
    387 		    error = EINVAL;
    388 		    break;
    389 		}
    390 		TUNDEBUG("%s: interface mtu set\n", ifp->if_xname);
    391 		ifp->if_mtu = ifr->ifr_mtu;
    392 		break;
    393 	}
    394 	case SIOCADDMULTI:
    395 	case SIOCDELMULTI: {
    396 		struct ifreq *ifr = (struct ifreq *) data;
    397 		if (ifr == 0) {
    398 	        	error = EAFNOSUPPORT;           /* XXX */
    399 			break;
    400 		}
    401 		switch (ifr->ifr_addr.sa_family) {
    402 
    403 #ifdef INET
    404 		case AF_INET:
    405 			break;
    406 #endif
    407 
    408 		default:
    409 			error = EAFNOSUPPORT;
    410 			break;
    411 		}
    412 		break;
    413 	}
    414 	case SIOCSIFFLAGS:
    415 		break;
    416 	default:
    417 		error = EINVAL;
    418 	}
    419 	splx(s);
    420 	simple_unlock(&tp->tun_lock);
    421 	return (error);
    422 }
    423 
    424 /*
    425  * tun_output - queue packets from higher level ready to put out.
    426  */
    427 int
    428 tun_output(ifp, m0, dst, rt)
    429 	struct ifnet   *ifp;
    430 	struct mbuf    *m0;
    431 	struct sockaddr *dst;
    432 	struct rtentry *rt;
    433 {
    434 	struct tun_softc *tp = ifp->if_softc;
    435 	struct proc	*p;
    436 #ifdef INET
    437 	int		s;
    438 	int		error;
    439 #endif
    440 	ALTQ_DECL(struct altq_pktattr pktattr;)
    441 
    442 	simple_lock(&tp->tun_lock);
    443 	TUNDEBUG ("%s: tun_output\n", ifp->if_xname);
    444 
    445 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
    446 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname,
    447 			  tp->tun_flags);
    448 		m_freem (m0);
    449 		simple_unlock(&tp->tun_lock);
    450 		return (EHOSTDOWN);
    451 	}
    452 
    453 	/*
    454 	 * if the queueing discipline needs packet classification,
    455 	 * do it before prepending link headers.
    456 	 */
    457 	IFQ_CLASSIFY(&ifp->if_snd, m0, dst->sa_family, &pktattr);
    458 
    459 #if NBPFILTER > 0
    460 	if (ifp->if_bpf) {
    461 		/*
    462 		 * We need to prepend the address family as
    463 		 * a four byte field.  Cons up a dummy header
    464 		 * to pacify bpf.  This is safe because bpf
    465 		 * will only read from the mbuf (i.e., it won't
    466 		 * try to free it or keep a pointer to it).
    467 		 */
    468 		struct mbuf m;
    469 		u_int32_t af = dst->sa_family;
    470 
    471 		m.m_next = m0;
    472 		m.m_len = sizeof(af);
    473 		m.m_data = (char *)&af;
    474 
    475 		bpf_mtap(ifp->if_bpf, &m);
    476 	}
    477 #endif
    478 
    479 	switch(dst->sa_family) {
    480 #ifdef INET
    481 	case AF_INET:
    482 		if (tp->tun_flags & TUN_PREPADDR) {
    483 			/* Simple link-layer header */
    484 			M_PREPEND(m0, dst->sa_len, M_DONTWAIT);
    485 			if (m0 == NULL) {
    486 				IF_DROP(&ifp->if_snd);
    487 				simple_unlock(&tp->tun_lock);
    488 				return (ENOBUFS);
    489 			}
    490 			bcopy(dst, mtod(m0, char *), dst->sa_len);
    491 		}
    492 		/* FALLTHROUGH */
    493 	case AF_UNSPEC:
    494 		s = splnet();
    495 		IFQ_ENQUEUE(&ifp->if_snd, m0, &pktattr, error);
    496 		if (error) {
    497 			splx(s);
    498 			ifp->if_collisions++;
    499 			return (error);
    500 		}
    501 		splx(s);
    502 		ifp->if_opackets++;
    503 		break;
    504 #endif
    505 	default:
    506 		m_freem(m0);
    507 		simple_unlock(&tp->tun_lock);
    508 		return (EAFNOSUPPORT);
    509 	}
    510 
    511 	if (tp->tun_flags & TUN_RWAIT) {
    512 		tp->tun_flags &= ~TUN_RWAIT;
    513 		wakeup((caddr_t)tp);
    514 	}
    515 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgrp) {
    516 		if (tp->tun_pgrp > 0)
    517 			gsignal(tp->tun_pgrp, SIGIO);
    518 		else if ((p = pfind(-tp->tun_pgrp)) != NULL)
    519 			psignal(p, SIGIO);
    520 	}
    521 	selwakeup(&tp->tun_rsel);
    522 	simple_unlock(&tp->tun_lock);
    523 	return (0);
    524 }
    525 
    526 /*
    527  * the cdevsw interface is now pretty minimal.
    528  */
    529 int
    530 tunioctl(dev, cmd, data, flag, p)
    531 	dev_t		dev;
    532 	u_long		cmd;
    533 	caddr_t		data;
    534 	int		flag;
    535 	struct proc	*p;
    536 {
    537 	int		s;
    538 	struct tun_softc *tp;
    539 
    540 	tp = tun_find_unit(dev);
    541 
    542 	/* interface was "destroyed" already */
    543 	if (tp == NULL)
    544 		return (ENXIO);
    545 
    546 	switch (cmd) {
    547 	case TUNSDEBUG:
    548 		tundebug = *(int *)data;
    549 		break;
    550 
    551 	case TUNGDEBUG:
    552 		*(int *)data = tundebug;
    553 		break;
    554 
    555 	case TUNSIFMODE:
    556 		switch (*(int *)data & (IFF_POINTOPOINT|IFF_BROADCAST)) {
    557 		case IFF_POINTOPOINT:
    558 		case IFF_BROADCAST:
    559 			s = splnet();
    560 			if (tp->tun_if.if_flags & IFF_UP) {
    561 				splx(s);
    562 				simple_unlock(&tp->tun_lock);
    563 				return (EBUSY);
    564 			}
    565 			tp->tun_if.if_flags &=
    566 				~(IFF_BROADCAST|IFF_POINTOPOINT|IFF_MULTICAST);
    567 			tp->tun_if.if_flags |= *(int *)data;
    568 			splx(s);
    569 			break;
    570 		default:
    571 			simple_unlock(&tp->tun_lock);
    572 			return (EINVAL);
    573 			break;
    574 		}
    575 		break;
    576 
    577 	case TUNSLMODE:
    578 		if (*(int *)data)
    579 			tp->tun_flags |= TUN_PREPADDR;
    580 		else
    581 			tp->tun_flags &= ~TUN_PREPADDR;
    582 		break;
    583 
    584 	case FIONBIO:
    585 		if (*(int *)data)
    586 			tp->tun_flags |= TUN_NBIO;
    587 		else
    588 			tp->tun_flags &= ~TUN_NBIO;
    589 		break;
    590 
    591 	case FIOASYNC:
    592 		if (*(int *)data)
    593 			tp->tun_flags |= TUN_ASYNC;
    594 		else
    595 			tp->tun_flags &= ~TUN_ASYNC;
    596 		break;
    597 
    598 	case FIONREAD:
    599 		s = splnet();
    600 		if (tp->tun_if.if_snd.ifq_head)
    601 			*(int *)data = tp->tun_if.if_snd.ifq_head->m_pkthdr.len;
    602 		else
    603 			*(int *)data = 0;
    604 		splx(s);
    605 		break;
    606 
    607 	case TIOCSPGRP:
    608 		tp->tun_pgrp = *(int *)data;
    609 		break;
    610 
    611 	case TIOCGPGRP:
    612 		*(int *)data = tp->tun_pgrp;
    613 		break;
    614 
    615 	default:
    616 		simple_unlock(&tp->tun_lock);
    617 		return (ENOTTY);
    618 	}
    619 	simple_unlock(&tp->tun_lock);
    620 	return (0);
    621 }
    622 
    623 /*
    624  * The cdevsw read interface - reads a packet at a time, or at
    625  * least as much of a packet as can be read.
    626  */
    627 int
    628 tunread(dev, uio, ioflag)
    629 	dev_t		dev;
    630 	struct uio	*uio;
    631 	int		ioflag;
    632 {
    633 	struct tun_softc *tp;
    634 	struct ifnet	*ifp;
    635 	struct mbuf	*m, *m0;
    636 	int		error=0, len, s, index;
    637 
    638 	tp = tun_find_unit(dev);
    639 
    640 	/* interface was "destroyed" already */
    641 	if (tp == NULL)
    642 		return (ENXIO);
    643 
    644 	index = tp->tun_if.if_index;
    645 	ifp = &tp->tun_if;
    646 
    647 	TUNDEBUG ("%s: read\n", ifp->if_xname);
    648 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
    649 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname, tp->tun_flags);
    650 		simple_unlock(&tp->tun_lock);
    651 		return EHOSTDOWN;
    652 	}
    653 
    654 	tp->tun_flags &= ~TUN_RWAIT;
    655 
    656 	s = splnet();
    657 	do {
    658 		IFQ_DEQUEUE(&ifp->if_snd, m0);
    659 		if (m0 == 0) {
    660 			if (tp->tun_flags & TUN_NBIO) {
    661 				splx(s);
    662 				simple_unlock(&tp->tun_lock);
    663 				return (EWOULDBLOCK);
    664 			}
    665 			tp->tun_flags |= TUN_RWAIT;
    666 			simple_unlock(&tp->tun_lock);
    667 			if (tsleep((caddr_t)tp, PZERO|PCATCH, "tunread", 0)) {
    668 				splx(s);
    669 				return (EINTR);
    670 			} else {
    671 				/*
    672 				 * Maybe the interface was destroyed while
    673 				 * we were sleeping, so let's ensure that
    674 				 * we're looking at the same (valid) tun
    675 				 * interface before looping.
    676 				 */
    677 				tp = tun_find_unit(dev);
    678 				if (tp == NULL ||
    679 				    tp->tun_if.if_index != index) {
    680 					splx(s);
    681 					if (tp)
    682 						simple_unlock(&tp->tun_lock);
    683 					return (ENXIO);
    684 				}
    685 			}
    686 		}
    687 	} while (m0 == 0);
    688 	splx(s);
    689 
    690 	while (m0 && uio->uio_resid > 0 && error == 0) {
    691 		len = min(uio->uio_resid, m0->m_len);
    692 		if (len != 0)
    693 			error = uiomove(mtod(m0, caddr_t), len, uio);
    694 		MFREE(m0, m);
    695 		m0 = m;
    696 	}
    697 
    698 	if (m0) {
    699 		TUNDEBUG("Dropping mbuf\n");
    700 		m_freem(m0);
    701 	}
    702 	if (error)
    703 		ifp->if_ierrors++;
    704 	simple_unlock(&tp->tun_lock);
    705 	return (error);
    706 }
    707 
    708 /*
    709  * the cdevsw write interface - an atomic write is a packet - or else!
    710  */
    711 int
    712 tunwrite(dev, uio, ioflag)
    713 	dev_t		dev;
    714 	struct uio	*uio;
    715 	int		ioflag;
    716 {
    717 	struct tun_softc *tp;
    718 	struct ifnet	*ifp;
    719 	struct mbuf	*top, **mp, *m;
    720 	struct ifqueue	*ifq;
    721 	struct sockaddr	dst;
    722 	int		isr, error=0, s, tlen, mlen;
    723 
    724 	tp = tun_find_unit(dev);
    725 
    726 	/* interface was "destroyed" already */
    727 	if (tp == NULL)
    728 		return (ENXIO);
    729 
    730 	ifp = &tp->tun_if;
    731 
    732 	TUNDEBUG("%s: tunwrite\n", ifp->if_xname);
    733 
    734 	if (tp->tun_flags & TUN_PREPADDR) {
    735 		if (uio->uio_resid < sizeof(dst)) {
    736 			simple_unlock(&tp->tun_lock);
    737 			return (EIO);
    738 		}
    739 		error = uiomove((caddr_t)&dst, sizeof(dst), uio);
    740 		if (dst.sa_len > sizeof(dst)) {
    741 			/* Duh.. */
    742 			char discard;
    743 			int n = dst.sa_len - sizeof(dst);
    744 			while (n--)
    745 				if ((error = uiomove(&discard, 1, uio)) != 0) {
    746 					simple_unlock(&tp->tun_lock);
    747 					return (error);
    748 				}
    749 		}
    750 	} else {
    751 #ifdef INET
    752 		dst.sa_family = AF_INET;
    753 #endif
    754 	}
    755 
    756 	if (uio->uio_resid < 0 || uio->uio_resid > TUNMTU) {
    757 		TUNDEBUG("%s: len=%lu!\n", ifp->if_xname,
    758 		    (unsigned long)uio->uio_resid);
    759 		simple_unlock(&tp->tun_lock);
    760 		return (EIO);
    761 	}
    762 
    763 	switch (dst.sa_family) {
    764 #ifdef INET
    765 	case AF_INET:
    766 		ifq = &ipintrq;
    767 		isr = NETISR_IP;
    768 		break;
    769 #endif
    770 	default:
    771 		simple_unlock(&tp->tun_lock);
    772 		return (EAFNOSUPPORT);
    773 	}
    774 
    775 	tlen = uio->uio_resid;
    776 
    777 	/* get a header mbuf */
    778 	MGETHDR(m, M_DONTWAIT, MT_DATA);
    779 	if (m == NULL) {
    780 		simple_unlock(&tp->tun_lock);
    781 		return (ENOBUFS);
    782 	}
    783 	mlen = MHLEN;
    784 
    785 	top = 0;
    786 	mp = &top;
    787 	while (error == 0 && uio->uio_resid > 0) {
    788 		m->m_len = min(mlen, uio->uio_resid);
    789 		error = uiomove(mtod (m, caddr_t), m->m_len, uio);
    790 		*mp = m;
    791 		mp = &m->m_next;
    792 		if (uio->uio_resid > 0) {
    793 			MGET (m, M_DONTWAIT, MT_DATA);
    794 			if (m == 0) {
    795 				error = ENOBUFS;
    796 				break;
    797 			}
    798 			mlen = MLEN;
    799 		}
    800 	}
    801 	if (error) {
    802 		if (top)
    803 			m_freem (top);
    804 		ifp->if_ierrors++;
    805 		simple_unlock(&tp->tun_lock);
    806 		return (error);
    807 	}
    808 
    809 	top->m_pkthdr.len = tlen;
    810 	top->m_pkthdr.rcvif = ifp;
    811 
    812 #if NBPFILTER > 0
    813 	if (ifp->if_bpf) {
    814 		/*
    815 		 * We need to prepend the address family as
    816 		 * a four byte field.  Cons up a dummy header
    817 		 * to pacify bpf.  This is safe because bpf
    818 		 * will only read from the mbuf (i.e., it won't
    819 		 * try to free it or keep a pointer to it).
    820 		 */
    821 		struct mbuf m;
    822 		u_int32_t af = AF_INET;
    823 
    824 		m.m_next = top;
    825 		m.m_len = sizeof(af);
    826 		m.m_data = (char *)&af;
    827 
    828 		bpf_mtap(ifp->if_bpf, &m);
    829 	}
    830 #endif
    831 
    832 	s = splnet();
    833 	if (IF_QFULL(ifq)) {
    834 		IF_DROP(ifq);
    835 		splx(s);
    836 		ifp->if_collisions++;
    837 		m_freem(top);
    838 		simple_unlock(&tp->tun_lock);
    839 		return (ENOBUFS);
    840 	}
    841 	IF_ENQUEUE(ifq, top);
    842 	splx(s);
    843 	ifp->if_ipackets++;
    844 	schednetisr(isr);
    845 	simple_unlock(&tp->tun_lock);
    846 	return (error);
    847 }
    848 
    849 #ifdef ALTQ
    850 /*
    851  * Start packet transmission on the interface.
    852  * when the interface queue is rate-limited by ALTQ or TBR,
    853  * if_start is needed to drain packets from the queue in order
    854  * to notify readers when outgoing packets become ready.
    855  */
    856 static void
    857 tunstart(ifp)
    858 	struct ifnet *ifp;
    859 {
    860 	struct tun_softc *tp = ifp->if_softc;
    861 	struct mbuf *m;
    862 	struct proc	*p;
    863 
    864 	if (!ALTQ_IS_ENABLED(&ifp->if_snd) && !TBR_IS_ENABLED(&ifp->if_snd))
    865 		return;
    866 
    867 	IFQ_POLL(&ifp->if_snd, m);
    868 	if (m != NULL) {
    869 		if (tp->tun_flags & TUN_RWAIT) {
    870 			tp->tun_flags &= ~TUN_RWAIT;
    871 			wakeup((caddr_t)tp);
    872 		}
    873 		if (tp->tun_flags & TUN_ASYNC && tp->tun_pgrp) {
    874 			if (tp->tun_pgrp > 0)
    875 				gsignal(tp->tun_pgrp, SIGIO);
    876 			else if ((p = pfind(-tp->tun_pgrp)) != NULL)
    877 				psignal(p, SIGIO);
    878 		}
    879 		selwakeup(&tp->tun_rsel);
    880 	}
    881 }
    882 #endif /* ALTQ */
    883 /*
    884  * tunpoll - the poll interface, this is only useful on reads
    885  * really. The write detect always returns true, write never blocks
    886  * anyway, it either accepts the packet or drops it.
    887  */
    888 int
    889 tunpoll(dev, events, p)
    890 	dev_t		dev;
    891 	int		events;
    892 	struct proc	*p;
    893 {
    894 	struct tun_softc *tp;
    895 	struct ifnet	*ifp;
    896 	int		s, revents = 0;
    897 
    898 	tp = tun_find_unit(dev);
    899 
    900 	/* interface was "destroyed" already */
    901 	if (tp == NULL)
    902 		return (0);
    903 
    904 	ifp = &tp->tun_if;
    905 
    906 	s = splnet();
    907 	TUNDEBUG("%s: tunpoll\n", ifp->if_xname);
    908 
    909 	if (events & (POLLIN | POLLRDNORM)) {
    910 		if (IFQ_IS_EMPTY(&ifp->if_snd) == 0) {
    911 			TUNDEBUG("%s: tunpoll q=%d\n", ifp->if_xname,
    912 			    ifp->if_snd.ifq_len);
    913 			revents |= events & (POLLIN | POLLRDNORM);
    914 		} else {
    915 			TUNDEBUG("%s: tunpoll waiting\n", ifp->if_xname);
    916 			selrecord(p, &tp->tun_rsel);
    917 		}
    918 	}
    919 
    920 	if (events & (POLLOUT | POLLWRNORM))
    921 		revents |= events & (POLLOUT | POLLWRNORM);
    922 
    923 	splx(s);
    924 	simple_unlock(&tp->tun_lock);
    925 	return (revents);
    926 }
    927