sys/net/tun/if_tun.c

   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 its 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/poll mode of
  14  * operation though.
  15  *
  16  * $FreeBSD: src/sys/net/if_tun.c,v 1.74.2.8 2002/02/13 00:43:11 dillon Exp $
  17  */
  18
  19 #include "use_tun.h"
  20 #include "opt_inet.h"
  21 #include "opt_inet6.h"
  22
  23 #include <sys/param.h>
  24 #include <sys/proc.h>
  25 #include <sys/priv.h>
  26 #include <sys/systm.h>
  27 #include <sys/mbuf.h>
  28 #include <sys/socket.h>
  29 #include <sys/conf.h>
  30 #include <sys/device.h>
  31 #include <sys/filio.h>
  32 #include <sys/sockio.h>
  33 #include <sys/thread2.h>
  34 #include <sys/ttycom.h>
  35 #include <sys/signalvar.h>
  36 #include <sys/filedesc.h>
  37 #include <sys/kernel.h>
  38 #include <sys/sysctl.h>
  39 #include <sys/uio.h>
  40 #include <sys/vnode.h>
  41 #include <sys/malloc.h>
  42
  43 #include <sys/mplock2.h>
  44
  45 #include <net/if.h>
  46 #include <net/if_types.h>
  47 #include <net/ifq_var.h>
  48 #include <net/netisr.h>
  49 #include <net/route.h>
  50 #include <sys/devfs.h>
  51
  52 #ifdef INET
  53 #include <netinet/in.h>
  54 #endif
  55
  56 #include <net/bpf.h>
  57
  58 #include "if_tunvar.h"
  59 #include "if_tun.h"
  60
  61 static MALLOC_DEFINE(M_TUN, "tun", "Tunnel Interface");
  62
  63 static void tunattach (void *);
  64 PSEUDO_SET(tunattach, if_tun);
  65
  66 static void tuncreate (cdev_t dev);
  67
  68 #define TUNDEBUG        if (tundebug) if_printf
  69 static int tundebug = 0;
  70 SYSCTL_INT(_debug, OID_AUTO, if_tun_debug, CTLFLAG_RW, &tundebug, 0,
  71     "Enable debug output");
  72
  73 static int tunoutput (struct ifnet *, struct mbuf *, struct sockaddr *,
  74             struct rtentry *rt);
  75 static int tunifioctl (struct ifnet *, u_long, caddr_t, struct ucred *);
  76 static int tuninit (struct ifnet *);
  77 static void tunstart(struct ifnet *, struct ifaltq_subque *);
  78 static void tun_filter_detach(struct knote *);
  79 static int tun_filter_read(struct knote *, long);
  80 static int tun_filter_write(struct knote *, long);
  81
  82 static  d_open_t        tunopen;
  83 static  d_close_t       tunclose;
  84 static  d_read_t        tunread;
  85 static  d_write_t       tunwrite;
  86 static  d_ioctl_t       tunioctl;
  87 static  d_kqfilter_t    tunkqfilter;
  88
  89 static d_clone_t tunclone;
  90 DEVFS_DEFINE_CLONE_BITMAP(tun);
  91
  92 #if NTUN <= 1
  93 #define TUN_PREALLOCATED_UNITS  4
  94 #else
  95 #define TUN_PREALLOCATED_UNITS  NTUN
  96 #endif
  97
  98 static struct dev_ops tun_ops = {
  99         { "tun", 0, 0 },
 100         .d_open =       tunopen,
 101         .d_close =      tunclose,
 102         .d_read =       tunread,
 103         .d_write =      tunwrite,
 104         .d_ioctl =      tunioctl,
 105         .d_kqfilter =   tunkqfilter
 106 };
 107
 108 static void
 109 tunattach(void *dummy)
 110 {
 111         int i;
 112         make_autoclone_dev(&tun_ops, &DEVFS_CLONE_BITMAP(tun),
 113                 tunclone, UID_UUCP, GID_DIALER, 0600, "tun");
 114         for (i = 0; i < TUN_PREALLOCATED_UNITS; i++) {
 115                 make_dev(&tun_ops, i, UID_UUCP, GID_DIALER, 0600, "tun%d", i);
 116                 devfs_clone_bitmap_set(&DEVFS_CLONE_BITMAP(tun), i);
 117         }
 118         /* Doesn't need uninit because unloading is not possible, see PSEUDO_SET */
 119 }
 120
 121 static int
 122 tunclone(struct dev_clone_args *ap)
 123 {
 124         int unit;
 125
 126         unit = devfs_clone_bitmap_get(&DEVFS_CLONE_BITMAP(tun), 0);
 127         ap->a_dev = make_only_dev(&tun_ops, unit, UID_UUCP, GID_DIALER, 0600,
 128                                                                 "tun%d", unit);
 129
 130         return 0;
 131 }
 132
 133 static void
 134 tuncreate(cdev_t dev)
 135 {
 136         struct tun_softc *sc;
 137         struct ifnet *ifp;
 138
 139 #if 0
 140         dev = make_dev(&tun_ops, minor(dev),
 141             UID_UUCP, GID_DIALER, 0600, "tun%d", lminor(dev));
 142 #endif
 143
 144         sc = kmalloc(sizeof(*sc), M_TUN, M_WAITOK | M_ZERO);
 145         sc->tun_flags = TUN_INITED;
 146
 147         ifp = &sc->tun_if;
 148         if_initname(ifp, "tun", lminor(dev));
 149         ifp->if_mtu = TUNMTU;
 150         ifp->if_ioctl = tunifioctl;
 151         ifp->if_output = tunoutput;
 152         ifp->if_start = tunstart;
 153         ifp->if_flags = IFF_POINTOPOINT | IFF_MULTICAST;
 154         ifp->if_type = IFT_PPP;
 155         ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
 156         ifq_set_ready(&ifp->if_snd);
 157         ifp->if_softc = sc;
 158         if_attach(ifp, NULL);
 159         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 160         dev->si_drv1 = sc;
 161 }
 162
 163 /*
 164  * tunnel open - must be superuser & the device must be
 165  * configured in
 166  */
 167 static  int
 168 tunopen(struct dev_open_args *ap)
 169 {
 170         cdev_t dev = ap->a_head.a_dev;
 171         struct ifnet    *ifp;
 172         struct tun_softc *tp;
 173         int     error;
 174
 175         if ((error = priv_check_cred(ap->a_cred, PRIV_ROOT, 0)) != 0)
 176                 return (error);
 177
 178         tp = dev->si_drv1;
 179         if (!tp) {
 180                 tuncreate(dev);
 181                 tp = dev->si_drv1;
 182         }
 183         if (tp->tun_flags & TUN_OPEN)
 184                 return EBUSY;
 185         tp->tun_pid = curproc->p_pid;
 186         ifp = &tp->tun_if;
 187         tp->tun_flags |= TUN_OPEN;
 188         TUNDEBUG(ifp, "open\n");
 189         return (0);
 190 }
 191
 192 /*
 193  * tunclose - close the device - mark i/f down & delete
 194  * routing info
 195  */
 196 static  int
 197 tunclose(struct dev_close_args *ap)
 198 {
 199         cdev_t dev = ap->a_head.a_dev;
 200         struct tun_softc *tp;
 201         struct ifnet    *ifp;
 202
 203         tp = dev->si_drv1;
 204         ifp = &tp->tun_if;
 205
 206         tp->tun_flags &= ~TUN_OPEN;
 207         tp->tun_pid = 0;
 208
 209         /* Junk all pending output. */
 210         ifq_purge_all(&ifp->if_snd);
 211
 212         if (ifp->if_flags & IFF_UP)
 213                 if_down(ifp);
 214         ifp->if_flags &= ~IFF_RUNNING;
 215         if_purgeaddrs_nolink(ifp);
 216
 217         funsetown(&tp->tun_sigio);
 218         KNOTE(&tp->tun_rkq.ki_note, 0);
 219
 220         TUNDEBUG(ifp, "closed\n");
 221 #if 0
 222         if (dev->si_uminor >= TUN_PREALLOCATED_UNITS) {
 223                 devfs_clone_bitmap_put(&DEVFS_CLONE_BITMAP(tun), dev->si_uminor);
 224         }
 225 #endif
 226         return (0);
 227 }
 228
 229 static int
 230 tuninit(struct ifnet *ifp)
 231 {
 232 #ifdef INET
 233         struct tun_softc *tp = ifp->if_softc;
 234 #endif
 235         struct ifaddr_container *ifac;
 236         int error = 0;
 237
 238         TUNDEBUG(ifp, "tuninit\n");
 239
 240         ifp->if_flags |= IFF_UP | IFF_RUNNING;
 241         getmicrotime(&ifp->if_lastchange);
 242
 243         TAILQ_FOREACH(ifac, &ifp->if_addrheads[mycpuid], ifa_link) {
 244                 struct ifaddr *ifa = ifac->ifa;
 245
 246                 if (ifa->ifa_addr == NULL) {
 247                         error = EFAULT;
 248                         /* XXX: Should maybe return straight off? */
 249                 } else {
 250 #ifdef INET
 251                         if (ifa->ifa_addr->sa_family == AF_INET) {
 252                             struct sockaddr_in *si;
 253
 254                             si = (struct sockaddr_in *)ifa->ifa_addr;
 255                             if (si->sin_addr.s_addr)
 256                                     tp->tun_flags |= TUN_IASET;
 257                         }
 258 #endif
 259                 }
 260         }
 261         return (error);
 262 }
 263
 264 /*
 265  * Process an ioctl request.
 266  *
 267  * MPSAFE
 268  */
 269 static int
 270 tunifioctl(struct ifnet *ifp, u_long cmd, caddr_t data, struct ucred *cr)
 271 {
 272         struct ifreq *ifr = (struct ifreq *)data;
 273         struct tun_softc *tp = ifp->if_softc;
 274         struct ifstat *ifs;
 275         int error = 0;
 276
 277         switch(cmd) {
 278         case SIOCGIFSTATUS:
 279                 ifs = (struct ifstat *)data;
 280                 if (tp->tun_pid)
 281                         ksprintf(ifs->ascii + strlen(ifs->ascii),
 282                             "\tOpened by PID %d\n", tp->tun_pid);
 283                 break;
 284         case SIOCSIFADDR:
 285                 error = tuninit(ifp);
 286                 TUNDEBUG(ifp, "address set, error=%d\n", error);
 287                 break;
 288         case SIOCSIFDSTADDR:
 289                 error = tuninit(ifp);
 290                 TUNDEBUG(ifp, "destination address set, error=%d\n", error);
 291                 break;
 292         case SIOCSIFMTU:
 293                 ifp->if_mtu = ifr->ifr_mtu;
 294                 TUNDEBUG(ifp, "mtu set\n");
 295                 break;
 296         case SIOCSIFFLAGS:
 297         case SIOCADDMULTI:
 298         case SIOCDELMULTI:
 299                 break;
 300         default:
 301                 error = EINVAL;
 302         }
 303         return (error);
 304 }
 305
 306 /*
 307  * tunoutput - queue packets from higher level ready to put out.
 308  *
 309  * MPSAFE
 310  */
 311 static int
 312 tunoutput_serialized(struct ifnet *ifp, struct mbuf *m0, struct sockaddr *dst,
 313                      struct rtentry *rt)
 314 {
 315         struct tun_softc *tp = ifp->if_softc;
 316         int error;
 317         struct altq_pktattr pktattr;
 318
 319         TUNDEBUG(ifp, "tunoutput\n");
 320
 321         if ((tp->tun_flags & TUN_READY) != TUN_READY) {
 322                 TUNDEBUG(ifp, "not ready 0%o\n", tp->tun_flags);
 323                 m_freem (m0);
 324                 return EHOSTDOWN;
 325         }
 326
 327         /*
 328          * if the queueing discipline needs packet classification,
 329          * do it before prepending link headers.
 330          */
 331         ifq_classify(&ifp->if_snd, m0, dst->sa_family, &pktattr);
 332
 333         /* BPF write needs to be handled specially */
 334         if (dst->sa_family == AF_UNSPEC) {
 335                 dst->sa_family = *(mtod(m0, int *));
 336                 m0->m_len -= sizeof(int);
 337                 m0->m_pkthdr.len -= sizeof(int);
 338                 m0->m_data += sizeof(int);
 339         }
 340
 341         if (ifp->if_bpf) {
 342                 bpf_gettoken();
 343                 if (ifp->if_bpf) {
 344                         /*
 345                          * We need to prepend the address family as
 346                          * a four byte field.
 347                          */
 348                         uint32_t af = dst->sa_family;
 349
 350                         bpf_ptap(ifp->if_bpf, m0, &af, sizeof(af));
 351                 }
 352                 bpf_reltoken();
 353         }
 354
 355         /* prepend sockaddr? this may abort if the mbuf allocation fails */
 356         if (tp->tun_flags & TUN_LMODE) {
 357                 /* allocate space for sockaddr */
 358                 M_PREPEND(m0, dst->sa_len, M_NOWAIT);
 359
 360                 /* if allocation failed drop packet */
 361                 if (m0 == NULL){
 362                         IFNET_STAT_INC(ifp, oerrors, 1);
 363                         return (ENOBUFS);
 364                 } else {
 365                         bcopy(dst, m0->m_data, dst->sa_len);
 366                 }
 367         }
 368
 369         if (tp->tun_flags & TUN_IFHEAD) {
 370                 /* Prepend the address family */
 371                 M_PREPEND(m0, 4, M_NOWAIT);
 372
 373                 /* if allocation failed drop packet */
 374                 if (m0 == NULL){
 375                         IFNET_STAT_INC(ifp, oerrors, 1);
 376                         return ENOBUFS;
 377                 } else
 378                         *(u_int32_t *)m0->m_data = htonl(dst->sa_family);
 379         } else {
 380 #ifdef INET
 381                 if (dst->sa_family != AF_INET)
 382 #endif
 383                 {
 384                         m_freem(m0);
 385                         return EAFNOSUPPORT;
 386                 }
 387         }
 388
 389         error = ifq_handoff(ifp, m0, &pktattr);
 390         if (error) {
 391                 IFNET_STAT_INC(ifp, collisions, 1);
 392         } else {
 393                 IFNET_STAT_INC(ifp, opackets, 1);
 394                 if (tp->tun_flags & TUN_RWAIT) {
 395                         tp->tun_flags &= ~TUN_RWAIT;
 396                         wakeup((caddr_t)tp);
 397                 }
 398                 get_mplock();
 399                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 400                         pgsigio(tp->tun_sigio, SIGIO, 0);
 401                 rel_mplock();
 402                 ifnet_deserialize_all(ifp);
 403                 KNOTE(&tp->tun_rkq.ki_note, 0);
 404                 ifnet_serialize_all(ifp);
 405         }
 406         return (error);
 407 }
 408
 409 static int
 410 tunoutput(struct ifnet *ifp, struct mbuf *m0, struct sockaddr *dst,
 411           struct rtentry *rt)
 412 {
 413         int error;
 414
 415         ifnet_serialize_all(ifp);
 416         error = tunoutput_serialized(ifp, m0, dst, rt);
 417         ifnet_deserialize_all(ifp);
 418
 419         return error;
 420 }
 421
 422 /*
 423  * the ops interface is now pretty minimal.
 424  */
 425 static  int
 426 tunioctl(struct dev_ioctl_args *ap)
 427 {
 428         cdev_t dev = ap->a_head.a_dev;
 429         struct tun_softc *tp = dev->si_drv1;
 430         struct tuninfo *tunp;
 431
 432         switch (ap->a_cmd) {
 433         case TUNSIFINFO:
 434                 tunp = (struct tuninfo *)ap->a_data;
 435                 if (tunp->mtu < IF_MINMTU)
 436                         return (EINVAL);
 437                 tp->tun_if.if_mtu = tunp->mtu;
 438                 tp->tun_if.if_type = tunp->type;
 439                 tp->tun_if.if_baudrate = tunp->baudrate;
 440                 break;
 441         case TUNGIFINFO:
 442                 tunp = (struct tuninfo *)ap->a_data;
 443                 tunp->mtu = tp->tun_if.if_mtu;
 444                 tunp->type = tp->tun_if.if_type;
 445                 tunp->baudrate = tp->tun_if.if_baudrate;
 446                 break;
 447         case TUNSDEBUG:
 448                 tundebug = *(int *)ap->a_data;
 449                 break;
 450         case TUNGDEBUG:
 451                 *(int *)ap->a_data = tundebug;
 452                 break;
 453         case TUNSLMODE:
 454                 if (*(int *)ap->a_data) {
 455                         tp->tun_flags |= TUN_LMODE;
 456                         tp->tun_flags &= ~TUN_IFHEAD;
 457                 } else
 458                         tp->tun_flags &= ~TUN_LMODE;
 459                 break;
 460         case TUNSIFHEAD:
 461                 if (*(int *)ap->a_data) {
 462                         tp->tun_flags |= TUN_IFHEAD;
 463                         tp->tun_flags &= ~TUN_LMODE;
 464                 } else
 465                         tp->tun_flags &= ~TUN_IFHEAD;
 466                 break;
 467         case TUNGIFHEAD:
 468                 *(int *)ap->a_data = (tp->tun_flags & TUN_IFHEAD) ? 1 : 0;
 469                 break;
 470         case TUNSIFMODE:
 471                 /* deny this if UP */
 472                 if (tp->tun_if.if_flags & IFF_UP)
 473                         return(EBUSY);
 474
 475                 switch (*(int *)ap->a_data & ~IFF_MULTICAST) {
 476                 case IFF_POINTOPOINT:
 477                 case IFF_BROADCAST:
 478                         tp->tun_if.if_flags &= ~(IFF_BROADCAST|IFF_POINTOPOINT);
 479                         tp->tun_if.if_flags |= *(int *)ap->a_data;
 480                         break;
 481                 default:
 482                         return(EINVAL);
 483                 }
 484                 break;
 485         case TUNSIFPID:
 486                 tp->tun_pid = curproc->p_pid;
 487                 break;
 488         case FIOASYNC:
 489                 if (*(int *)ap->a_data)
 490                         tp->tun_flags |= TUN_ASYNC;
 491                 else
 492                         tp->tun_flags &= ~TUN_ASYNC;
 493                 break;
 494         case FIONREAD:
 495                 *(int *)ap->a_data = ifsq_poll_pktlen(
 496                     ifq_get_subq_default(&tp->tun_if.if_snd));
 497                 break;
 498         case FIOSETOWN:
 499                 return (fsetown(*(int *)ap->a_data, &tp->tun_sigio));
 500
 501         case FIOGETOWN:
 502                 *(int *)ap->a_data = fgetown(&tp->tun_sigio);
 503                 return (0);
 504
 505         /* This is deprecated, FIOSETOWN should be used instead. */
 506         case TIOCSPGRP:
 507                 return (fsetown(-(*(int *)ap->a_data), &tp->tun_sigio));
 508
 509         /* This is deprecated, FIOGETOWN should be used instead. */
 510         case TIOCGPGRP:
 511                 *(int *)ap->a_data = -fgetown(&tp->tun_sigio);
 512                 return (0);
 513
 514         default:
 515                 return (ENOTTY);
 516         }
 517         return (0);
 518 }
 519
 520 /*
 521  * The ops read interface - reads a packet at a time, or at
 522  * least as much of a packet as can be read.
 523  */
 524 static  int
 525 tunread(struct dev_read_args *ap)
 526 {
 527         cdev_t dev = ap->a_head.a_dev;
 528         struct uio *uio = ap->a_uio;
 529         struct tun_softc *tp = dev->si_drv1;
 530         struct ifnet    *ifp = &tp->tun_if;
 531         struct ifaltq_subque *ifsq = ifq_get_subq_default(&ifp->if_snd);
 532         struct mbuf     *m0;
 533         int             error=0, len;
 534
 535         TUNDEBUG(ifp, "read\n");
 536         if ((tp->tun_flags & TUN_READY) != TUN_READY) {
 537                 TUNDEBUG(ifp, "not ready 0%o\n", tp->tun_flags);
 538                 return EHOSTDOWN;
 539         }
 540
 541         tp->tun_flags &= ~TUN_RWAIT;
 542
 543         ifnet_serialize_all(ifp);
 544
 545         while ((m0 = ifsq_dequeue(ifsq)) == NULL) {
 546                 if (ap->a_ioflag & IO_NDELAY) {
 547                         ifnet_deserialize_all(ifp);
 548                         return EWOULDBLOCK;
 549                 }
 550                 tp->tun_flags |= TUN_RWAIT;
 551                 ifnet_deserialize_all(ifp);
 552                 if ((error = tsleep(tp, PCATCH, "tunread", 0)) != 0)
 553                         return error;
 554                 ifnet_serialize_all(ifp);
 555         }
 556
 557         ifnet_deserialize_all(ifp);
 558
 559         while (m0 && uio->uio_resid > 0 && error == 0) {
 560                 len = (int)szmin(uio->uio_resid, m0->m_len);
 561                 if (len != 0)
 562                         error = uiomove(mtod(m0, caddr_t), (size_t)len, uio);
 563                 m0 = m_free(m0);
 564         }
 565
 566         if (m0) {
 567                 TUNDEBUG(ifp, "Dropping mbuf\n");
 568                 m_freem(m0);
 569         }
 570         return error;
 571 }
 572
 573 /*
 574  * the ops write interface - an atomic write is a packet - or else!
 575  */
 576 static  int
 577 tunwrite(struct dev_write_args *ap)
 578 {
 579         cdev_t dev = ap->a_head.a_dev;
 580         struct uio *uio = ap->a_uio;
 581         struct tun_softc *tp = dev->si_drv1;
 582         struct ifnet    *ifp = &tp->tun_if;
 583         struct mbuf     *top, **mp, *m;
 584         int             error=0;
 585         size_t          tlen, mlen;
 586         uint32_t        family;
 587         int             isr;
 588
 589         TUNDEBUG(ifp, "tunwrite\n");
 590
 591         if (uio->uio_resid == 0)
 592                 return 0;
 593
 594         if (uio->uio_resid > TUNMRU) {
 595                 TUNDEBUG(ifp, "len=%zd!\n", uio->uio_resid);
 596                 return EIO;
 597         }
 598         tlen = uio->uio_resid;
 599
 600         /* get a header mbuf */
 601         MGETHDR(m, M_WAITOK, MT_DATA);
 602         if (m == NULL)
 603                 return ENOBUFS;
 604         mlen = MHLEN;
 605
 606         top = NULL;
 607         mp = &top;
 608         while (error == 0 && uio->uio_resid > 0) {
 609                 m->m_len = (int)szmin(mlen, uio->uio_resid);
 610                 error = uiomove(mtod (m, caddr_t), (size_t)m->m_len, uio);
 611                 *mp = m;
 612                 mp = &m->m_next;
 613                 if (uio->uio_resid > 0) {
 614                         MGET (m, M_WAITOK, MT_DATA);
 615                         if (m == NULL) {
 616                                 error = ENOBUFS;
 617                                 break;
 618                         }
 619                         mlen = MLEN;
 620                 }
 621         }
 622         if (error) {
 623                 if (top)
 624                         m_freem (top);
 625                 IFNET_STAT_INC(ifp, ierrors, 1);
 626                 return error;
 627         }
 628
 629         top->m_pkthdr.len = (int)tlen;
 630         top->m_pkthdr.rcvif = ifp;
 631
 632         if (ifp->if_bpf) {
 633                 bpf_gettoken();
 634
 635                 if (ifp->if_bpf) {
 636                         if (tp->tun_flags & TUN_IFHEAD) {
 637                                 /*
 638                                  * Conveniently, we already have a 4-byte
 639                                  * address family prepended to our packet !
 640                                  * Inconveniently, it's in the wrong byte
 641                                  * order !
 642                                  */
 643                                 if ((top = m_pullup(top, sizeof(family)))
 644                                     == NULL) {
 645                                         bpf_reltoken();
 646                                         return ENOBUFS;
 647                                 }
 648                                 *mtod(top, u_int32_t *) =
 649                                     ntohl(*mtod(top, u_int32_t *));
 650                                 bpf_mtap(ifp->if_bpf, top);
 651                                 *mtod(top, u_int32_t *) =
 652                                     htonl(*mtod(top, u_int32_t *));
 653                         } else {
 654                                 /*
 655                                  * We need to prepend the address family as
 656                                  * a four byte field.
 657                                  */
 658                                 static const uint32_t af = AF_INET;
 659
 660                                 bpf_ptap(ifp->if_bpf, top, &af, sizeof(af));
 661                         }
 662                 }
 663
 664                 bpf_reltoken();
 665         }
 666
 667         if (tp->tun_flags & TUN_IFHEAD) {
 668                 if (top->m_len < sizeof(family) &&
 669                     (top = m_pullup(top, sizeof(family))) == NULL)
 670                                 return ENOBUFS;
 671                 family = ntohl(*mtod(top, u_int32_t *));
 672                 m_adj(top, sizeof(family));
 673         } else
 674                 family = AF_INET;
 675
 676         IFNET_STAT_INC(ifp, ibytes, top->m_pkthdr.len);
 677         IFNET_STAT_INC(ifp, ipackets, 1);
 678
 679         switch (family) {
 680 #ifdef INET
 681         case AF_INET:
 682                 isr = NETISR_IP;
 683                 break;
 684 #endif
 685 #ifdef INET6
 686         case AF_INET6:
 687                 isr = NETISR_IPV6;
 688                 break;
 689 #endif
 690         default:
 691                 m_freem(m);
 692                 return (EAFNOSUPPORT);
 693         }
 694
 695         netisr_queue(isr, top);
 696         return (0);
 697 }
 698
 699 static struct filterops tun_read_filtops =
 700         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_read };
 701 static struct filterops tun_write_filtops =
 702         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_write };
 703
 704 static int
 705 tunkqfilter(struct dev_kqfilter_args *ap)
 706 {
 707         cdev_t dev = ap->a_head.a_dev;
 708         struct tun_softc *tp = dev->si_drv1;
 709         struct knote *kn = ap->a_kn;
 710         struct klist *klist;
 711
 712         ap->a_result = 0;
 713         ifnet_serialize_all(&tp->tun_if);
 714
 715         switch (kn->kn_filter) {
 716         case EVFILT_READ:
 717                 kn->kn_fop = &tun_read_filtops;
 718                 kn->kn_hook = (caddr_t)tp;
 719                 break;
 720         case EVFILT_WRITE:
 721                 kn->kn_fop = &tun_write_filtops;
 722                 kn->kn_hook = (caddr_t)tp;
 723                 break;
 724         default:
 725                 ifnet_deserialize_all(&tp->tun_if);
 726                 ap->a_result = EOPNOTSUPP;
 727                 return (0);
 728         }
 729
 730         klist = &tp->tun_rkq.ki_note;
 731         knote_insert(klist, kn);
 732         ifnet_deserialize_all(&tp->tun_if);
 733
 734         return (0);
 735 }
 736
 737 static void
 738 tun_filter_detach(struct knote *kn)
 739 {
 740         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 741         struct klist *klist = &tp->tun_rkq.ki_note;
 742
 743         knote_remove(klist, kn);
 744 }
 745
 746 static int
 747 tun_filter_write(struct knote *kn, long hint)
 748 {
 749         /* Always ready for a write */
 750         return (1);
 751 }
 752
 753 static int
 754 tun_filter_read(struct knote *kn, long hint)
 755 {
 756         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 757         int ready = 0;
 758
 759         ifnet_serialize_all(&tp->tun_if);
 760         if (!ifsq_is_empty(ifq_get_subq_default(&tp->tun_if.if_snd)))
 761                 ready = 1;
 762         ifnet_deserialize_all(&tp->tun_if);
 763
 764         return (ready);
 765 }
 766
 767 /*
 768  * Start packet transmission on the interface.
 769  * when the interface queue is rate-limited by ALTQ,
 770  * if_start is needed to drain packets from the queue in order
 771  * to notify readers when outgoing packets become ready.
 772  */
 773 static void
 774 tunstart(struct ifnet *ifp, struct ifaltq_subque *ifsq)
 775 {
 776         struct tun_softc *tp = ifp->if_softc;
 777         struct mbuf *m;
 778
 779         ASSERT_ALTQ_SQ_DEFAULT(ifp, ifsq);
 780
 781         if (!ifq_is_enabled(&ifp->if_snd))
 782                 return;
 783
 784         m = ifsq_poll(ifsq);
 785         if (m != NULL) {
 786                 if (tp->tun_flags & TUN_RWAIT) {
 787                         tp->tun_flags &= ~TUN_RWAIT;
 788                         wakeup((caddr_t)tp);
 789                 }
 790                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 791                         pgsigio(tp->tun_sigio, SIGIO, 0);
 792                 ifsq_deserialize_hw(ifsq);
 793                 KNOTE(&tp->tun_rkq.ki_note, 0);
 794                 ifsq_serialize_hw(ifsq);
 795         }
 796 }