Home | History | Annotate | Line # | Download | only in net
if_tun.c revision 1.164
      1 /*	$NetBSD: if_tun.c,v 1.164 2021/09/26 15:58:33 thorpej 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 /*
     18  * tun - tunnel software network interface.
     19  */
     20 
     21 #include <sys/cdefs.h>
     22 __KERNEL_RCSID(0, "$NetBSD: if_tun.c,v 1.164 2021/09/26 15:58:33 thorpej Exp $");
     23 
     24 #ifdef _KERNEL_OPT
     25 #include "opt_inet.h"
     26 #endif
     27 
     28 #include <sys/param.h>
     29 
     30 #include <sys/buf.h>
     31 #include <sys/conf.h>
     32 #include <sys/cpu.h>
     33 #include <sys/device.h>
     34 #include <sys/file.h>
     35 #include <sys/ioctl.h>
     36 #include <sys/kauth.h>
     37 #include <sys/kmem.h>
     38 #include <sys/lwp.h>
     39 #include <sys/mbuf.h>
     40 #include <sys/module.h>
     41 #include <sys/mutex.h>
     42 #include <sys/poll.h>
     43 #include <sys/select.h>
     44 #include <sys/signalvar.h>
     45 #include <sys/socket.h>
     46 
     47 #include <net/bpf.h>
     48 #include <net/if.h>
     49 #include <net/if_types.h>
     50 #include <net/route.h>
     51 
     52 #ifdef INET
     53 #include <netinet/in.h>
     54 #include <netinet/in_systm.h>
     55 #include <netinet/in_var.h>
     56 #include <netinet/ip.h>
     57 #include <netinet/if_inarp.h>
     58 #endif
     59 
     60 #include <net/if_tun.h>
     61 
     62 #include "ioconf.h"
     63 
     64 #define TUNDEBUG	if (tundebug) printf
     65 int	tundebug = 0;
     66 
     67 extern int ifqmaxlen;
     68 
     69 static LIST_HEAD(, tun_softc) tun_softc_list;
     70 static LIST_HEAD(, tun_softc) tunz_softc_list;
     71 static kmutex_t tun_softc_lock;
     72 
     73 static int	tun_ioctl(struct ifnet *, u_long, void *);
     74 static int	tun_output(struct ifnet *, struct mbuf *,
     75 			const struct sockaddr *, const struct rtentry *rt);
     76 static int	tun_clone_create(struct if_clone *, int);
     77 static int	tun_clone_destroy(struct ifnet *);
     78 
     79 static struct if_clone tun_cloner =
     80     IF_CLONE_INITIALIZER("tun", tun_clone_create, tun_clone_destroy);
     81 
     82 static void tunattach0(struct tun_softc *);
     83 static void tun_enable(struct tun_softc *, const struct ifaddr *);
     84 static void tun_i_softintr(void *);
     85 static void tun_o_softintr(void *);
     86 #ifdef ALTQ
     87 static void tunstart(struct ifnet *);
     88 #endif
     89 static struct tun_softc *tun_find_unit(dev_t);
     90 static struct tun_softc *tun_find_zunit(int);
     91 
     92 static dev_type_open(tunopen);
     93 static dev_type_close(tunclose);
     94 static dev_type_read(tunread);
     95 static dev_type_write(tunwrite);
     96 static dev_type_ioctl(tunioctl);
     97 static dev_type_poll(tunpoll);
     98 static dev_type_kqfilter(tunkqfilter);
     99 
    100 const struct cdevsw tun_cdevsw = {
    101 	.d_open = tunopen,
    102 	.d_close = tunclose,
    103 	.d_read = tunread,
    104 	.d_write = tunwrite,
    105 	.d_ioctl = tunioctl,
    106 	.d_stop = nostop,
    107 	.d_tty = notty,
    108 	.d_poll = tunpoll,
    109 	.d_mmap = nommap,
    110 	.d_kqfilter = tunkqfilter,
    111 	.d_discard = nodiscard,
    112 	.d_flag = D_OTHER | D_MPSAFE
    113 };
    114 
    115 #ifdef _MODULE
    116 devmajor_t tun_bmajor = -1, tun_cmajor = -1;
    117 #endif
    118 
    119 void
    120 tunattach(int unused)
    121 {
    122 
    123 	/*
    124 	 * Nothing to do here, initialization is handled by the
    125 	 * module initialization code in tuninit() below).
    126 	 */
    127 }
    128 
    129 static void
    130 tuninit(void)
    131 {
    132 
    133 	mutex_init(&tun_softc_lock, MUTEX_DEFAULT, IPL_NET);
    134 	LIST_INIT(&tun_softc_list);
    135 	LIST_INIT(&tunz_softc_list);
    136 	if_clone_attach(&tun_cloner);
    137 #ifdef _MODULE
    138 	devsw_attach("tun", NULL, &tun_bmajor, &tun_cdevsw, &tun_cmajor);
    139 #endif
    140 }
    141 
    142 static int
    143 tundetach(void)
    144 {
    145 #ifdef _MODULE
    146 	int error;
    147 #endif
    148 
    149 	if_clone_detach(&tun_cloner);
    150 #ifdef _MODULE
    151 	error = devsw_detach(NULL, &tun_cdevsw);
    152 	if (error != 0) {
    153 		if_clone_attach(&tun_cloner);
    154 		return error;
    155 	}
    156 #endif
    157 
    158 	if (!LIST_EMPTY(&tun_softc_list) || !LIST_EMPTY(&tunz_softc_list)) {
    159 #ifdef _MODULE
    160 		devsw_attach("tun", NULL, &tun_bmajor, &tun_cdevsw, &tun_cmajor);
    161 #endif
    162 		if_clone_attach(&tun_cloner);
    163 		return EBUSY;
    164 }
    165 
    166 	mutex_destroy(&tun_softc_lock);
    167 
    168 	return 0;
    169 }
    170 
    171 /*
    172  * Find driver instance from dev_t.
    173  * Returns with tp locked (if found).
    174  */
    175 static struct tun_softc *
    176 tun_find_unit(dev_t dev)
    177 {
    178 	struct tun_softc *tp;
    179 	int unit = minor(dev);
    180 
    181 	mutex_enter(&tun_softc_lock);
    182 	LIST_FOREACH(tp, &tun_softc_list, tun_list)
    183 		if (unit == tp->tun_unit)
    184 			break;
    185 	if (tp)
    186 		mutex_enter(&tp->tun_lock);
    187 	mutex_exit(&tun_softc_lock);
    188 
    189 	return tp;
    190 }
    191 
    192 /*
    193  * Find zombie driver instance by unit number.
    194  * Remove tp from list and return it unlocked (if found).
    195  */
    196 static struct tun_softc *
    197 tun_find_zunit(int unit)
    198 {
    199 	struct tun_softc *tp;
    200 
    201 	mutex_enter(&tun_softc_lock);
    202 	LIST_FOREACH(tp, &tunz_softc_list, tun_list)
    203 		if (unit == tp->tun_unit)
    204 			break;
    205 	if (tp)
    206 		LIST_REMOVE(tp, tun_list);
    207 	mutex_exit(&tun_softc_lock);
    208 	KASSERTMSG(!tp || (tp->tun_flags & (TUN_INITED|TUN_OPEN)) == TUN_OPEN,
    209 	    "tun%d: inconsistent flags: %x", unit, tp->tun_flags);
    210 
    211 	return tp;
    212 }
    213 
    214 static int
    215 tun_clone_create(struct if_clone *ifc, int unit)
    216 {
    217 	struct tun_softc *tp;
    218 
    219 	if ((tp = tun_find_zunit(unit)) == NULL) {
    220 		tp = kmem_zalloc(sizeof(*tp), KM_SLEEP);
    221 
    222 		tp->tun_unit = unit;
    223 		mutex_init(&tp->tun_lock, MUTEX_DEFAULT, IPL_NET);
    224 		cv_init(&tp->tun_cv, "tunread");
    225 		selinit(&tp->tun_rsel);
    226 		selinit(&tp->tun_wsel);
    227 	} else {
    228 		/* Revive tunnel instance; clear ifp part */
    229 		(void)memset(&tp->tun_if, 0, sizeof(struct ifnet));
    230 	}
    231 
    232 	if_initname(&tp->tun_if, ifc->ifc_name, unit);
    233 	tunattach0(tp);
    234 	tp->tun_flags |= TUN_INITED;
    235 	tp->tun_osih = softint_establish(SOFTINT_CLOCK, tun_o_softintr, tp);
    236 	tp->tun_isih = softint_establish(SOFTINT_CLOCK, tun_i_softintr, tp);
    237 
    238 	mutex_enter(&tun_softc_lock);
    239 	LIST_INSERT_HEAD(&tun_softc_list, tp, tun_list);
    240 	mutex_exit(&tun_softc_lock);
    241 
    242 	return 0;
    243 }
    244 
    245 static void
    246 tunattach0(struct tun_softc *tp)
    247 {
    248 	struct ifnet *ifp;
    249 
    250 	ifp = &tp->tun_if;
    251 	ifp->if_softc = tp;
    252 	ifp->if_mtu = TUNMTU;
    253 	ifp->if_ioctl = tun_ioctl;
    254 	ifp->if_output = tun_output;
    255 #ifdef ALTQ
    256 	ifp->if_start = tunstart;
    257 #endif
    258 	ifp->if_flags = IFF_POINTOPOINT;
    259 	ifp->if_type = IFT_TUNNEL;
    260 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
    261 	ifp->if_dlt = DLT_NULL;
    262 	IFQ_SET_READY(&ifp->if_snd);
    263 	if_attach(ifp);
    264 	ifp->if_link_state = LINK_STATE_DOWN;
    265 	if_alloc_sadl(ifp);
    266 	bpf_attach(ifp, DLT_NULL, sizeof(uint32_t));
    267 }
    268 
    269 static int
    270 tun_clone_destroy(struct ifnet *ifp)
    271 {
    272 	struct tun_softc *tp = (void *)ifp;
    273 	bool zombie = false;
    274 
    275 	IF_PURGE(&ifp->if_snd);
    276 	ifp->if_flags &= ~IFF_RUNNING;
    277 
    278 	mutex_enter(&tun_softc_lock);
    279 	mutex_enter(&tp->tun_lock);
    280 	LIST_REMOVE(tp, tun_list);
    281 	if (tp->tun_flags & TUN_OPEN) {
    282 		/* Hang on to storage until last close. */
    283 		tp->tun_flags &= ~TUN_INITED;
    284 		LIST_INSERT_HEAD(&tunz_softc_list, tp, tun_list);
    285 		zombie = true;
    286 	}
    287 	mutex_exit(&tun_softc_lock);
    288 
    289 	if (tp->tun_flags & TUN_RWAIT) {
    290 		tp->tun_flags &= ~TUN_RWAIT;
    291 		cv_broadcast(&tp->tun_cv);
    292 	}
    293 	selnotify(&tp->tun_rsel, 0, NOTE_SUBMIT);
    294 
    295 	mutex_exit(&tp->tun_lock);
    296 
    297 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
    298 		fownsignal(tp->tun_pgid, SIGIO, POLL_HUP, 0, NULL);
    299 
    300 	bpf_detach(ifp);
    301 	if_detach(ifp);
    302 
    303 	if (!zombie) {
    304 		seldestroy(&tp->tun_rsel);
    305 		seldestroy(&tp->tun_wsel);
    306 		softint_disestablish(tp->tun_osih);
    307 		softint_disestablish(tp->tun_isih);
    308 		mutex_destroy(&tp->tun_lock);
    309 		cv_destroy(&tp->tun_cv);
    310 		kmem_free(tp, sizeof(*tp));
    311 	}
    312 
    313 	return 0;
    314 }
    315 
    316 /*
    317  * tunnel open - must be superuser & the device must be
    318  * configured in
    319  */
    320 static int
    321 tunopen(dev_t dev, int flag, int mode, struct lwp *l)
    322 {
    323 	struct ifnet	*ifp;
    324 	struct tun_softc *tp;
    325 	int	error;
    326 
    327 	error = kauth_authorize_network(l->l_cred, KAUTH_NETWORK_INTERFACE_TUN,
    328 	    KAUTH_REQ_NETWORK_INTERFACE_TUN_ADD, NULL, NULL, NULL);
    329 	if (error)
    330 		return error;
    331 
    332 	tp = tun_find_unit(dev);
    333 
    334 	if (tp == NULL) {
    335 		(void)tun_clone_create(&tun_cloner, minor(dev));
    336 		tp = tun_find_unit(dev);
    337 		if (tp == NULL) {
    338 			return ENXIO;
    339 		}
    340 	}
    341 
    342 	if (tp->tun_flags & TUN_OPEN) {
    343 		mutex_exit(&tp->tun_lock);
    344 		return EBUSY;
    345 	}
    346 
    347 	ifp = &tp->tun_if;
    348 	tp->tun_flags |= TUN_OPEN;
    349 	TUNDEBUG("%s: open\n", ifp->if_xname);
    350 	if_link_state_change(ifp, LINK_STATE_UP);
    351 
    352 	mutex_exit(&tp->tun_lock);
    353 
    354 	return error;
    355 }
    356 
    357 /*
    358  * tunclose - close the device - mark i/f down & delete
    359  * routing info
    360  */
    361 int
    362 tunclose(dev_t dev, int flag, int mode,
    363     struct lwp *l)
    364 {
    365 	struct tun_softc *tp;
    366 	struct ifnet	*ifp;
    367 
    368 	if ((tp = tun_find_zunit(minor(dev))) != NULL) {
    369 		/* interface was "destroyed" before the close */
    370 		seldestroy(&tp->tun_rsel);
    371 		seldestroy(&tp->tun_wsel);
    372 		softint_disestablish(tp->tun_osih);
    373 		softint_disestablish(tp->tun_isih);
    374 		mutex_destroy(&tp->tun_lock);
    375 		kmem_free(tp, sizeof(*tp));
    376 		return 0;
    377 	}
    378 
    379 	if ((tp = tun_find_unit(dev)) == NULL)
    380 		goto out_nolock;
    381 
    382 	ifp = &tp->tun_if;
    383 
    384 	tp->tun_flags &= ~TUN_OPEN;
    385 
    386 	tp->tun_pgid = 0;
    387 	selnotify(&tp->tun_rsel, 0, NOTE_SUBMIT);
    388 
    389 	TUNDEBUG ("%s: closed\n", ifp->if_xname);
    390 	mutex_exit(&tp->tun_lock);
    391 
    392 	/*
    393 	 * junk all pending output
    394 	 */
    395 	IFQ_PURGE(&ifp->if_snd);
    396 
    397 	if (ifp->if_flags & IFF_UP) {
    398 		if_down(ifp);
    399 		if (ifp->if_flags & IFF_RUNNING) {
    400 			/* find internet addresses and delete routes */
    401 			struct ifaddr *ifa;
    402 			IFADDR_READER_FOREACH(ifa, ifp) {
    403 #if defined(INET) || defined(INET6)
    404 				if (ifa->ifa_addr->sa_family == AF_INET ||
    405 				    ifa->ifa_addr->sa_family == AF_INET6) {
    406 					rtinit(ifa, (int)RTM_DELETE,
    407 					       tp->tun_flags & TUN_DSTADDR
    408 							? RTF_HOST
    409 							: 0);
    410 				}
    411 #endif
    412 			}
    413 		}
    414 	}
    415 
    416 	if_link_state_change(ifp, LINK_STATE_DOWN);
    417 
    418 out_nolock:
    419 	return 0;
    420 }
    421 
    422 static void
    423 tun_enable(struct tun_softc *tp, const struct ifaddr *ifa)
    424 {
    425 	struct ifnet	*ifp = &tp->tun_if;
    426 
    427 	TUNDEBUG("%s: %s\n", __func__, ifp->if_xname);
    428 
    429 	mutex_enter(&tp->tun_lock);
    430 	tp->tun_flags &= ~(TUN_IASET|TUN_DSTADDR);
    431 
    432 	switch (ifa->ifa_addr->sa_family) {
    433 #ifdef INET
    434 	case AF_INET: {
    435 		struct sockaddr_in *sin;
    436 
    437 		sin = satosin(ifa->ifa_addr);
    438 		if (sin && sin->sin_addr.s_addr)
    439 			tp->tun_flags |= TUN_IASET;
    440 
    441 		if (ifp->if_flags & IFF_POINTOPOINT) {
    442 			sin = satosin(ifa->ifa_dstaddr);
    443 			if (sin && sin->sin_addr.s_addr)
    444 				tp->tun_flags |= TUN_DSTADDR;
    445 		}
    446 		break;
    447 	    }
    448 #endif
    449 #ifdef INET6
    450 	case AF_INET6: {
    451 		struct sockaddr_in6 *sin;
    452 
    453 		sin = satosin6(ifa->ifa_addr);
    454 		if (!IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
    455 			tp->tun_flags |= TUN_IASET;
    456 
    457 		if (ifp->if_flags & IFF_POINTOPOINT) {
    458 			sin = satosin6(ifa->ifa_dstaddr);
    459 			if (sin && !IN6_IS_ADDR_UNSPECIFIED(&sin->sin6_addr))
    460 				tp->tun_flags |= TUN_DSTADDR;
    461 		} else
    462 			tp->tun_flags &= ~TUN_DSTADDR;
    463 		break;
    464 	    }
    465 #endif /* INET6 */
    466 	default:
    467 		break;
    468 	}
    469 	ifp->if_flags |= IFF_UP | IFF_RUNNING;
    470 	mutex_exit(&tp->tun_lock);
    471 }
    472 
    473 /*
    474  * Process an ioctl request.
    475  */
    476 static int
    477 tun_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    478 {
    479 	struct tun_softc *tp = (struct tun_softc *)(ifp->if_softc);
    480 	struct ifreq *ifr = (struct ifreq *)data;
    481 	struct ifaddr *ifa = (struct ifaddr *)data;
    482 	int error = 0;
    483 
    484 	switch (cmd) {
    485 	case SIOCINITIFADDR:
    486 		tun_enable(tp, ifa);
    487 		ifa->ifa_rtrequest = p2p_rtrequest;
    488 		TUNDEBUG("%s: address set\n", ifp->if_xname);
    489 		break;
    490 	case SIOCSIFBRDADDR:
    491 		TUNDEBUG("%s: broadcast address set\n", ifp->if_xname);
    492 		break;
    493 	case SIOCSIFMTU:
    494 		if (ifr->ifr_mtu > TUNMTU || ifr->ifr_mtu < 576) {
    495 			error = EINVAL;
    496 			break;
    497 		}
    498 		TUNDEBUG("%s: interface mtu set\n", ifp->if_xname);
    499 		if ((error = ifioctl_common(ifp, cmd, data)) == ENETRESET)
    500 			error = 0;
    501 		break;
    502 	case SIOCADDMULTI:
    503 	case SIOCDELMULTI:
    504 		if (ifr == NULL) {
    505 	        	error = EAFNOSUPPORT;           /* XXX */
    506 			break;
    507 		}
    508 		switch (ifreq_getaddr(cmd, ifr)->sa_family) {
    509 #ifdef INET
    510 		case AF_INET:
    511 			break;
    512 #endif
    513 #ifdef INET6
    514 		case AF_INET6:
    515 			break;
    516 #endif
    517 		default:
    518 			error = EAFNOSUPPORT;
    519 			break;
    520 		}
    521 		break;
    522 	default:
    523 		error = ifioctl_common(ifp, cmd, data);
    524 	}
    525 
    526 	return error;
    527 }
    528 
    529 /*
    530  * tun_output - queue packets from higher level ready to put out.
    531  */
    532 static int
    533 tun_output(struct ifnet *ifp, struct mbuf *m0, const struct sockaddr *dst,
    534     const struct rtentry *rt)
    535 {
    536 	struct tun_softc *tp = ifp->if_softc;
    537 	int		error;
    538 #if defined(INET) || defined(INET6)
    539 	int		mlen;
    540 	uint32_t	*af;
    541 #endif
    542 
    543 	mutex_enter(&tp->tun_lock);
    544 	TUNDEBUG ("%s: tun_output\n", ifp->if_xname);
    545 
    546 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
    547 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname,
    548 			  tp->tun_flags);
    549 		error = EHOSTDOWN;
    550 		mutex_exit(&tp->tun_lock);
    551 		goto out;
    552 	}
    553 	// XXXrmind
    554 	mutex_exit(&tp->tun_lock);
    555 
    556 	/*
    557 	 * if the queueing discipline needs packet classification,
    558 	 * do it before prepending link headers.
    559 	 */
    560 	IFQ_CLASSIFY(&ifp->if_snd, m0, dst->sa_family);
    561 
    562 	bpf_mtap_af(ifp, dst->sa_family, m0, BPF_D_OUT);
    563 
    564 	if ((error = pfil_run_hooks(ifp->if_pfil, &m0, ifp, PFIL_OUT)) != 0)
    565 		goto out;
    566 	if (m0 == NULL)
    567 		goto out;
    568 
    569 	switch(dst->sa_family) {
    570 #ifdef INET6
    571 	case AF_INET6:
    572 #endif
    573 #ifdef INET
    574 	case AF_INET:
    575 #endif
    576 #if defined(INET) || defined(INET6)
    577 		if (tp->tun_flags & TUN_PREPADDR) {
    578 			/* Simple link-layer header */
    579 			M_PREPEND(m0, dst->sa_len, M_DONTWAIT);
    580 			if (m0 == NULL) {
    581 				IF_DROP(&ifp->if_snd);
    582 				error = ENOBUFS;
    583 				goto out;
    584 			}
    585 			memcpy(mtod(m0, char *), dst, dst->sa_len);
    586 		}
    587 
    588 		if (tp->tun_flags & TUN_IFHEAD) {
    589 			/* Prepend the address family */
    590 			M_PREPEND(m0, sizeof(*af), M_DONTWAIT);
    591 			if (m0 == NULL) {
    592 				IF_DROP(&ifp->if_snd);
    593 				error = ENOBUFS;
    594 				goto out;
    595 			}
    596 			af = mtod(m0,uint32_t *);
    597 			*af = htonl(dst->sa_family);
    598 		} else {
    599 #ifdef INET
    600 			if (dst->sa_family != AF_INET)
    601 #endif
    602 			{
    603 				error = EAFNOSUPPORT;
    604 				goto out;
    605 			}
    606 		}
    607 		/* FALLTHROUGH */
    608 	case AF_UNSPEC:
    609 		mlen = m0->m_pkthdr.len;
    610 		IFQ_ENQUEUE(&ifp->if_snd, m0, error);
    611 		if (error) {
    612 			if_statinc(ifp, if_collisions);
    613 			error = EAFNOSUPPORT;
    614 			m0 = NULL;
    615 			goto out;
    616 		}
    617 		if_statadd2(ifp, if_opackets, 1, if_obytes, mlen);
    618 		break;
    619 #endif
    620 	default:
    621 		error = EAFNOSUPPORT;
    622 		goto out;
    623 	}
    624 
    625 	mutex_enter(&tp->tun_lock);
    626 	if (tp->tun_flags & TUN_RWAIT) {
    627 		tp->tun_flags &= ~TUN_RWAIT;
    628 		cv_broadcast(&tp->tun_cv);
    629 	}
    630 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
    631 		softint_schedule(tp->tun_isih);
    632 
    633 	selnotify(&tp->tun_rsel, 0, NOTE_SUBMIT);
    634 
    635 	mutex_exit(&tp->tun_lock);
    636 out:
    637 	if (error && m0)
    638 		m_freem(m0);
    639 
    640 	return error;
    641 }
    642 
    643 static void
    644 tun_i_softintr(void *cookie)
    645 {
    646 	struct tun_softc *tp = cookie;
    647 
    648 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
    649 		fownsignal(tp->tun_pgid, SIGIO, POLL_IN, POLLIN|POLLRDNORM,
    650 		    NULL);
    651 }
    652 
    653 static void
    654 tun_o_softintr(void *cookie)
    655 {
    656 	struct tun_softc *tp = cookie;
    657 
    658 	if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
    659 		fownsignal(tp->tun_pgid, SIGIO, POLL_OUT, POLLOUT|POLLWRNORM,
    660 		    NULL);
    661 }
    662 
    663 /*
    664  * the cdevsw interface is now pretty minimal.
    665  */
    666 int
    667 tunioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
    668 {
    669 	struct tun_softc *tp;
    670 	int error = 0;
    671 
    672 	tp = tun_find_unit(dev);
    673 
    674 	/* interface was "destroyed" already */
    675 	if (tp == NULL) {
    676 		return ENXIO;
    677 	}
    678 
    679 	switch (cmd) {
    680 	case TUNSDEBUG:
    681 		tundebug = *(int *)data;
    682 		break;
    683 
    684 	case TUNGDEBUG:
    685 		*(int *)data = tundebug;
    686 		break;
    687 
    688 	case TUNSIFMODE:
    689 		switch (*(int *)data & (IFF_POINTOPOINT|IFF_BROADCAST)) {
    690 		case IFF_POINTOPOINT:
    691 		case IFF_BROADCAST:
    692 			if (tp->tun_if.if_flags & IFF_UP) {
    693 				error = EBUSY;
    694 				goto out;
    695 			}
    696 			tp->tun_if.if_flags &=
    697 				~(IFF_BROADCAST|IFF_POINTOPOINT|IFF_MULTICAST);
    698 			tp->tun_if.if_flags |= *(int *)data;
    699 			break;
    700 		default:
    701 			error = EINVAL;
    702 			goto out;
    703 		}
    704 		break;
    705 
    706 	case TUNSLMODE:
    707 		if (*(int *)data) {
    708 			tp->tun_flags |= TUN_PREPADDR;
    709 			tp->tun_flags &= ~TUN_IFHEAD;
    710 		} else
    711 			tp->tun_flags &= ~TUN_PREPADDR;
    712 		break;
    713 
    714 	case TUNSIFHEAD:
    715 		if (*(int *)data) {
    716 			tp->tun_flags |= TUN_IFHEAD;
    717 			tp->tun_flags &= ~TUN_PREPADDR;
    718 		} else
    719 			tp->tun_flags &= ~TUN_IFHEAD;
    720 		break;
    721 
    722 	case TUNGIFHEAD:
    723 		*(int *)data = (tp->tun_flags & TUN_IFHEAD);
    724 		break;
    725 
    726 	case FIONBIO:
    727 		if (*(int *)data)
    728 			tp->tun_flags |= TUN_NBIO;
    729 		else
    730 			tp->tun_flags &= ~TUN_NBIO;
    731 		break;
    732 
    733 	case FIOASYNC:
    734 		if (*(int *)data)
    735 			tp->tun_flags |= TUN_ASYNC;
    736 		else
    737 			tp->tun_flags &= ~TUN_ASYNC;
    738 		break;
    739 
    740 	case FIONREAD:
    741 		if (tp->tun_if.if_snd.ifq_head)
    742 			*(int *)data = tp->tun_if.if_snd.ifq_head->m_pkthdr.len;
    743 		else
    744 			*(int *)data = 0;
    745 		break;
    746 
    747 	case TIOCSPGRP:
    748 	case FIOSETOWN:
    749 		error = fsetown(&tp->tun_pgid, cmd, data);
    750 		break;
    751 
    752 	case TIOCGPGRP:
    753 	case FIOGETOWN:
    754 		error = fgetown(tp->tun_pgid, cmd, data);
    755 		break;
    756 
    757 	default:
    758 		error = ENOTTY;
    759 	}
    760 
    761 out:
    762 	mutex_exit(&tp->tun_lock);
    763 
    764 	return error;
    765 }
    766 
    767 /*
    768  * The cdevsw read interface - reads a packet at a time, or at
    769  * least as much of a packet as can be read.
    770  */
    771 int
    772 tunread(dev_t dev, struct uio *uio, int ioflag)
    773 {
    774 	struct tun_softc *tp;
    775 	struct ifnet	*ifp;
    776 	struct mbuf	*m, *m0;
    777 	int		error = 0, len;
    778 
    779 	tp = tun_find_unit(dev);
    780 
    781 	/* interface was "destroyed" already */
    782 	if (tp == NULL) {
    783 		return ENXIO;
    784 	}
    785 
    786 	ifp = &tp->tun_if;
    787 
    788 	TUNDEBUG ("%s: read\n", ifp->if_xname);
    789 	if ((tp->tun_flags & TUN_READY) != TUN_READY) {
    790 		TUNDEBUG ("%s: not ready 0%o\n", ifp->if_xname, tp->tun_flags);
    791 		error = EHOSTDOWN;
    792 		goto out;
    793 	}
    794 
    795 	tp->tun_flags &= ~TUN_RWAIT;
    796 
    797 	do {
    798 		IFQ_DEQUEUE(&ifp->if_snd, m0);
    799 		if (m0 == 0) {
    800 			if (tp->tun_flags & TUN_NBIO) {
    801 				error = EWOULDBLOCK;
    802 				goto out;
    803 			}
    804 			tp->tun_flags |= TUN_RWAIT;
    805 			if (cv_wait_sig(&tp->tun_cv, &tp->tun_lock)) {
    806 				error = EINTR;
    807 				goto out;
    808 			}
    809 		}
    810 	} while (m0 == 0);
    811 
    812 	mutex_exit(&tp->tun_lock);
    813 
    814 	/* Copy the mbuf chain */
    815 	while (m0 && uio->uio_resid > 0 && error == 0) {
    816 		len = uimin(uio->uio_resid, m0->m_len);
    817 		if (len != 0)
    818 			error = uiomove(mtod(m0, void *), len, uio);
    819 		m0 = m = m_free(m0);
    820 	}
    821 
    822 	if (m0) {
    823 		TUNDEBUG("Dropping mbuf\n");
    824 		m_freem(m0);
    825 	}
    826 	if (error)
    827 		if_statinc(ifp, if_ierrors);
    828 
    829 	return error;
    830 
    831 out:
    832 	mutex_exit(&tp->tun_lock);
    833 
    834 	return error;
    835 }
    836 
    837 /*
    838  * the cdevsw write interface - an atomic write is a packet - or else!
    839  */
    840 int
    841 tunwrite(dev_t dev, struct uio *uio, int ioflag)
    842 {
    843 	struct tun_softc *tp;
    844 	struct ifnet	*ifp;
    845 	struct mbuf	*top, **mp, *m;
    846 	pktqueue_t	*pktq;
    847 	struct sockaddr	dst;
    848 	int		error = 0, tlen, mlen;
    849 	uint32_t	family;
    850 
    851 	tp = tun_find_unit(dev);
    852 	if (tp == NULL) {
    853 		/* Interface was "destroyed" already. */
    854 		return ENXIO;
    855 	}
    856 
    857 	/* Unlock until we've got the data */
    858 	mutex_exit(&tp->tun_lock);
    859 
    860 	ifp = &tp->tun_if;
    861 
    862 	TUNDEBUG("%s: tunwrite\n", ifp->if_xname);
    863 
    864 	if (tp->tun_flags & TUN_PREPADDR) {
    865 		if (uio->uio_resid < sizeof(dst)) {
    866 			error = EIO;
    867 			goto out0;
    868 		}
    869 		error = uiomove((void *)&dst, sizeof(dst), uio);
    870 		if (dst.sa_len > sizeof(dst)) {
    871 			/* Duh.. */
    872 			int n = dst.sa_len - sizeof(dst);
    873 			while (n--) {
    874 				char discard;
    875 				error = uiomove(&discard, 1, uio);
    876 				if (error) {
    877 					goto out0;
    878 				}
    879 			}
    880 		}
    881 	} else if (tp->tun_flags & TUN_IFHEAD) {
    882 		if (uio->uio_resid < sizeof(family)){
    883 			error = EIO;
    884 			goto out0;
    885 		}
    886 		error = uiomove((void *)&family, sizeof(family), uio);
    887 		dst.sa_family = ntohl(family);
    888 	} else {
    889 #ifdef INET
    890 		dst.sa_family = AF_INET;
    891 #endif
    892 	}
    893 
    894 	if (uio->uio_resid == 0 || uio->uio_resid > TUNMTU) {
    895 		TUNDEBUG("%s: len=%lu!\n", ifp->if_xname,
    896 		    (unsigned long)uio->uio_resid);
    897 		error = EIO;
    898 		goto out0;
    899 	}
    900 
    901 	switch (dst.sa_family) {
    902 #ifdef INET
    903 	case AF_INET:
    904 		pktq = ip_pktq;
    905 		break;
    906 #endif
    907 #ifdef INET6
    908 	case AF_INET6:
    909 		pktq = ip6_pktq;
    910 		break;
    911 #endif
    912 	default:
    913 		error = EAFNOSUPPORT;
    914 		goto out0;
    915 	}
    916 
    917 	tlen = uio->uio_resid;
    918 
    919 	/* get a header mbuf */
    920 	MGETHDR(m, M_DONTWAIT, MT_DATA);
    921 	if (m == NULL) {
    922 		return ENOBUFS;
    923 	}
    924 	mlen = MHLEN;
    925 
    926 	top = NULL;
    927 	mp = &top;
    928 	while (error == 0 && uio->uio_resid > 0) {
    929 		m->m_len = uimin(mlen, uio->uio_resid);
    930 		error = uiomove(mtod(m, void *), m->m_len, uio);
    931 		*mp = m;
    932 		mp = &m->m_next;
    933 		if (error == 0 && uio->uio_resid > 0) {
    934 			MGET(m, M_DONTWAIT, MT_DATA);
    935 			if (m == NULL) {
    936 				error = ENOBUFS;
    937 				break;
    938 			}
    939 			mlen = MLEN;
    940 		}
    941 	}
    942 	if (error) {
    943 		if (top != NULL)
    944 			m_freem(top);
    945 		if_statinc(ifp, if_ierrors);
    946 		goto out0;
    947 	}
    948 
    949 	top->m_pkthdr.len = tlen;
    950 	m_set_rcvif(top, ifp);
    951 
    952 	bpf_mtap_af(ifp, dst.sa_family, top, BPF_D_IN);
    953 
    954 	if ((error = pfil_run_hooks(ifp->if_pfil, &top, ifp, PFIL_IN)) != 0)
    955 		goto out0;
    956 	if (top == NULL)
    957 		goto out0;
    958 
    959 	mutex_enter(&tp->tun_lock);
    960 	if ((tp->tun_flags & TUN_INITED) == 0) {
    961 		/* Interface was destroyed */
    962 		error = ENXIO;
    963 		goto out;
    964 	}
    965 	if (__predict_false(!pktq_enqueue(pktq, top, 0))) {
    966 		if_statinc(ifp, if_collisions);
    967 		mutex_exit(&tp->tun_lock);
    968 		error = ENOBUFS;
    969 		m_freem(top);
    970 		goto out0;
    971 	}
    972 	if_statadd2(ifp, if_ipackets, 1, if_ibytes, tlen);
    973 out:
    974 	mutex_exit(&tp->tun_lock);
    975 out0:
    976 	return error;
    977 }
    978 
    979 #ifdef ALTQ
    980 /*
    981  * Start packet transmission on the interface.
    982  * when the interface queue is rate-limited by ALTQ or TBR,
    983  * if_start is needed to drain packets from the queue in order
    984  * to notify readers when outgoing packets become ready.
    985  */
    986 static void
    987 tunstart(struct ifnet *ifp)
    988 {
    989 	struct tun_softc *tp = ifp->if_softc;
    990 
    991 	if (!ALTQ_IS_ENABLED(&ifp->if_snd) && !TBR_IS_ENABLED(&ifp->if_snd))
    992 		return;
    993 
    994 	mutex_enter(&tp->tun_lock);
    995 	if (!IF_IS_EMPTY(&ifp->if_snd)) {
    996 		if (tp->tun_flags & TUN_RWAIT) {
    997 			tp->tun_flags &= ~TUN_RWAIT;
    998 			cv_broadcast(&tp->tun_cv);
    999 		}
   1000 		if (tp->tun_flags & TUN_ASYNC && tp->tun_pgid)
   1001 			softint_schedule(tp->tun_osih);
   1002 
   1003 		selnotify(&tp->tun_rsel, 0, NOTE_SUBMIT);
   1004 	}
   1005 	mutex_exit(&tp->tun_lock);
   1006 }
   1007 #endif /* ALTQ */
   1008 /*
   1009  * tunpoll - the poll interface, this is only useful on reads
   1010  * really. The write detect always returns true, write never blocks
   1011  * anyway, it either accepts the packet or drops it.
   1012  */
   1013 int
   1014 tunpoll(dev_t dev, int events, struct lwp *l)
   1015 {
   1016 	struct tun_softc *tp;
   1017 	struct ifnet	*ifp;
   1018 	int revents = 0;
   1019 
   1020 	tp = tun_find_unit(dev);
   1021 	if (tp == NULL) {
   1022 		/* Interface was "destroyed" already. */
   1023 		return 0;
   1024 	}
   1025 	ifp = &tp->tun_if;
   1026 
   1027 	TUNDEBUG("%s: tunpoll\n", ifp->if_xname);
   1028 
   1029 	if (events & (POLLIN | POLLRDNORM)) {
   1030 		if (!IFQ_IS_EMPTY(&ifp->if_snd)) {
   1031 			TUNDEBUG("%s: tunpoll q=%d\n", ifp->if_xname,
   1032 			    ifp->if_snd.ifq_len);
   1033 			revents |= events & (POLLIN | POLLRDNORM);
   1034 		} else {
   1035 			TUNDEBUG("%s: tunpoll waiting\n", ifp->if_xname);
   1036 			selrecord(l, &tp->tun_rsel);
   1037 		}
   1038 	}
   1039 
   1040 	if (events & (POLLOUT | POLLWRNORM))
   1041 		revents |= events & (POLLOUT | POLLWRNORM);
   1042 
   1043 	mutex_exit(&tp->tun_lock);
   1044 
   1045 	return revents;
   1046 }
   1047 
   1048 static void
   1049 filt_tunrdetach(struct knote *kn)
   1050 {
   1051 	struct tun_softc *tp = kn->kn_hook;
   1052 
   1053 	mutex_enter(&tp->tun_lock);
   1054 	selremove_knote(&tp->tun_rsel, kn);
   1055 	mutex_exit(&tp->tun_lock);
   1056 }
   1057 
   1058 static int
   1059 filt_tunread(struct knote *kn, long hint)
   1060 {
   1061 	struct tun_softc *tp = kn->kn_hook;
   1062 	struct ifnet *ifp = &tp->tun_if;
   1063 	struct mbuf *m;
   1064 	int ready;
   1065 
   1066 	if (hint & NOTE_SUBMIT)
   1067 		KASSERT(mutex_owned(&tp->tun_lock));
   1068 	else
   1069 		mutex_enter(&tp->tun_lock);
   1070 
   1071 	IF_POLL(&ifp->if_snd, m);
   1072 	ready = (m != NULL);
   1073 	for (kn->kn_data = 0; m != NULL; m = m->m_next)
   1074 		kn->kn_data += m->m_len;
   1075 
   1076 	if (hint & NOTE_SUBMIT)
   1077 		KASSERT(mutex_owned(&tp->tun_lock));
   1078 	else
   1079 		mutex_exit(&tp->tun_lock);
   1080 
   1081 	return ready;
   1082 }
   1083 
   1084 static const struct filterops tunread_filtops = {
   1085 	.f_flags = FILTEROP_ISFD,
   1086 	.f_attach = NULL,
   1087 	.f_detach = filt_tunrdetach,
   1088 	.f_event = filt_tunread,
   1089 };
   1090 
   1091 int
   1092 tunkqfilter(dev_t dev, struct knote *kn)
   1093 {
   1094 	struct tun_softc *tp;
   1095 	int rv = 0;
   1096 
   1097 	tp = tun_find_unit(dev);
   1098 	if (tp == NULL)
   1099 		goto out_nolock;
   1100 
   1101 	switch (kn->kn_filter) {
   1102 	case EVFILT_READ:
   1103 		kn->kn_fop = &tunread_filtops;
   1104 		kn->kn_hook = tp;
   1105 		selrecord_knote(&tp->tun_rsel, kn);
   1106 		break;
   1107 
   1108 	case EVFILT_WRITE:
   1109 		kn->kn_fop = &seltrue_filtops;
   1110 		break;
   1111 
   1112 	default:
   1113 		rv = EINVAL;
   1114 		goto out;
   1115 	}
   1116 
   1117 out:
   1118 	mutex_exit(&tp->tun_lock);
   1119 out_nolock:
   1120 	return rv;
   1121 }
   1122 
   1123 /*
   1124  * Module infrastructure
   1125  */
   1126 #include "if_module.h"
   1127 
   1128 IF_MODULE(MODULE_CLASS_DRIVER, tun, NULL)
   1129