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  * $DragonFly: src/sys/net/tun/if_tun.c,v 1.37 2008/06/05 18:06:32 swildner Exp $
  18  */
  19
  20 #include "use_tun.h"
  21 #include "opt_atalk.h"
  22 #include "opt_inet.h"
  23 #include "opt_inet6.h"
  24 #include "opt_ipx.h"
  25
  26 #include <sys/param.h>
  27 #include <sys/proc.h>
  28 #include <sys/priv.h>
  29 #include <sys/systm.h>
  30 #include <sys/mbuf.h>
  31 #include <sys/socket.h>
  32 #include <sys/conf.h>
  33 #include <sys/device.h>
  34 #include <sys/filio.h>
  35 #include <sys/sockio.h>
  36 #include <sys/thread2.h>
  37 #include <sys/ttycom.h>
  38 #include <sys/signalvar.h>
  39 #include <sys/filedesc.h>
  40 #include <sys/kernel.h>
  41 #include <sys/sysctl.h>
  42 #include <sys/uio.h>
  43 #include <sys/vnode.h>
  44 #include <sys/malloc.h>
  45
  46 #include <sys/mplock2.h>
  47
  48 #include <net/if.h>
  49 #include <net/if_types.h>
  50 #include <net/ifq_var.h>
  51 #include <net/netisr.h>
  52 #include <net/route.h>
  53 #include <sys/devfs.h>
  54
  55 #ifdef INET
  56 #include <netinet/in.h>
  57 #endif
  58
  59 #include <net/bpf.h>
  60
  61 #include "if_tunvar.h"
  62 #include "if_tun.h"
  63
  64 static MALLOC_DEFINE(M_TUN, "tun", "Tunnel Interface");
  65
  66 static void tunattach (void *);
  67 PSEUDO_SET(tunattach, if_tun);
  68
  69 static void tuncreate (cdev_t dev);
  70
  71 #define TUNDEBUG        if (tundebug) if_printf
  72 static int tundebug = 0;
  73 SYSCTL_INT(_debug, OID_AUTO, if_tun_debug, CTLFLAG_RW, &tundebug, 0, "");
  74
  75 static int tunoutput (struct ifnet *, struct mbuf *, struct sockaddr *,
  76             struct rtentry *rt);
  77 static int tunifioctl (struct ifnet *, u_long, caddr_t, struct ucred *);
  78 static int tuninit (struct ifnet *);
  79 static void tunstart(struct ifnet *);
  80 static void tun_filter_detach(struct knote *);
  81 static int tun_filter_read(struct knote *, long);
  82 static int tun_filter_write(struct knote *, long);
  83
  84 static  d_open_t        tunopen;
  85 static  d_close_t       tunclose;
  86 static  d_read_t        tunread;
  87 static  d_write_t       tunwrite;
  88 static  d_ioctl_t       tunioctl;
  89 static  d_kqfilter_t    tunkqfilter;
  90
  91 static d_clone_t tunclone;
  92 DEVFS_DECLARE_CLONE_BITMAP(tun);
  93
  94 #if NTUN <= 1
  95 #define TUN_PREALLOCATED_UNITS  4
  96 #else
  97 #define TUN_PREALLOCATED_UNITS  NTUN
  98 #endif
  99
 100 #define CDEV_MAJOR 52
 101 static struct dev_ops tun_ops = {
 102         { "tun", CDEV_MAJOR, 0 },
 103         .d_open =       tunopen,
 104         .d_close =      tunclose,
 105         .d_read =       tunread,
 106         .d_write =      tunwrite,
 107         .d_ioctl =      tunioctl,
 108         .d_kqfilter =   tunkqfilter
 109 };
 110
 111 static void
 112 tunattach(void *dummy)
 113 {
 114         int i;
 115         make_autoclone_dev(&tun_ops, &DEVFS_CLONE_BITMAP(tun),
 116                 tunclone, UID_UUCP, GID_DIALER, 0600, "tun");
 117         for (i = 0; i < TUN_PREALLOCATED_UNITS; i++) {
 118                 make_dev(&tun_ops, i, UID_UUCP, GID_DIALER, 0600, "tun%d", i);
 119                 devfs_clone_bitmap_set(&DEVFS_CLONE_BITMAP(tun), i);
 120         }
 121         /* Doesn't need uninit because unloading is not possible, see PSEUDO_SET */
 122 }
 123
 124 static int
 125 tunclone(struct dev_clone_args *ap)
 126 {
 127         int unit;
 128
 129         unit = devfs_clone_bitmap_get(&DEVFS_CLONE_BITMAP(tun), 0);
 130         ap->a_dev = make_only_dev(&tun_ops, unit, UID_UUCP, GID_DIALER, 0600,
 131                                                                 "tun%d", unit);
 132
 133         return 0;
 134 }
 135
 136 static void
 137 tuncreate(cdev_t dev)
 138 {
 139         struct tun_softc *sc;
 140         struct ifnet *ifp;
 141
 142 #if 0
 143         dev = make_dev(&tun_ops, minor(dev),
 144             UID_UUCP, GID_DIALER, 0600, "tun%d", lminor(dev));
 145 #endif
 146
 147         MALLOC(sc, struct tun_softc *, sizeof(*sc), M_TUN, M_WAITOK | M_ZERO);
 148         sc->tun_flags = TUN_INITED;
 149
 150         ifp = &sc->tun_if;
 151         if_initname(ifp, "tun", lminor(dev));
 152         ifp->if_mtu = TUNMTU;
 153         ifp->if_ioctl = tunifioctl;
 154         ifp->if_output = tunoutput;
 155         ifp->if_start = tunstart;
 156         ifp->if_flags = IFF_POINTOPOINT | IFF_MULTICAST;
 157         ifp->if_type = IFT_PPP;
 158         ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
 159         ifq_set_ready(&ifp->if_snd);
 160         ifp->if_softc = sc;
 161         if_attach(ifp, NULL);
 162         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 163         dev->si_drv1 = sc;
 164 }
 165
 166 /*
 167  * tunnel open - must be superuser & the device must be
 168  * configured in
 169  */
 170 static  int
 171 tunopen(struct dev_open_args *ap)
 172 {
 173         cdev_t dev = ap->a_head.a_dev;
 174         struct ifnet    *ifp;
 175         struct tun_softc *tp;
 176         int     error;
 177
 178         if ((error = priv_check_cred(ap->a_cred, PRIV_ROOT, 0)) != 0)
 179                 return (error);
 180
 181         tp = dev->si_drv1;
 182         if (!tp) {
 183                 tuncreate(dev);
 184                 tp = dev->si_drv1;
 185         }
 186         if (tp->tun_flags & TUN_OPEN)
 187                 return EBUSY;
 188         tp->tun_pid = curproc->p_pid;
 189         ifp = &tp->tun_if;
 190         tp->tun_flags |= TUN_OPEN;
 191         TUNDEBUG(ifp, "open\n");
 192         return (0);
 193 }
 194
 195 /*
 196  * tunclose - close the device - mark i/f down & delete
 197  * routing info
 198  */
 199 static  int
 200 tunclose(struct dev_close_args *ap)
 201 {
 202         cdev_t dev = ap->a_head.a_dev;
 203         struct tun_softc *tp;
 204         struct ifnet    *ifp;
 205
 206         tp = dev->si_drv1;
 207         ifp = &tp->tun_if;
 208
 209         tp->tun_flags &= ~TUN_OPEN;
 210         tp->tun_pid = 0;
 211
 212         /* Junk all pending output. */
 213         ifq_purge(&ifp->if_snd);
 214
 215         if (ifp->if_flags & IFF_UP)
 216                 if_down(ifp);
 217         ifp->if_flags &= ~IFF_RUNNING;
 218         if_purgeaddrs_nolink(ifp);
 219
 220         funsetown(tp->tun_sigio);
 221         KNOTE(&tp->tun_rkq.ki_note, 0);
 222
 223         TUNDEBUG(ifp, "closed\n");
 224 #if 0
 225         if (dev->si_uminor >= TUN_PREALLOCATED_UNITS) {
 226                 devfs_clone_bitmap_put(&DEVFS_CLONE_BITMAP(tun), dev->si_uminor);
 227         }
 228 #endif
 229         return (0);
 230 }
 231
 232 static int
 233 tuninit(struct ifnet *ifp)
 234 {
 235         struct tun_softc *tp = ifp->if_softc;
 236         struct ifaddr_container *ifac;
 237         int error = 0;
 238
 239         TUNDEBUG(ifp, "tuninit\n");
 240
 241         ifp->if_flags |= IFF_UP | IFF_RUNNING;
 242         getmicrotime(&ifp->if_lastchange);
 243
 244         TAILQ_FOREACH(ifac, &ifp->if_addrheads[mycpuid], ifa_link) {
 245                 struct ifaddr *ifa = ifac->ifa;
 246
 247                 if (ifa->ifa_addr == NULL) {
 248                         error = EFAULT;
 249                         /* XXX: Should maybe return straight off? */
 250                 } else {
 251 #ifdef INET
 252                         if (ifa->ifa_addr->sa_family == AF_INET) {
 253                             struct sockaddr_in *si;
 254
 255                             si = (struct sockaddr_in *)ifa->ifa_addr;
 256                             if (si->sin_addr.s_addr)
 257                                     tp->tun_flags |= TUN_IASET;
 258                         }
 259 #endif
 260                 }
 261         }
 262         return (error);
 263 }
 264
 265 /*
 266  * Process an ioctl request.
 267  *
 268  * MPSAFE
 269  */
 270 int
 271 tunifioctl(struct ifnet *ifp, u_long cmd, caddr_t data, struct ucred *cr)
 272 {
 273         struct ifreq *ifr = (struct ifreq *)data;
 274         struct tun_softc *tp = ifp->if_softc;
 275         struct ifstat *ifs;
 276         int error = 0;
 277
 278         switch(cmd) {
 279         case SIOCGIFSTATUS:
 280                 ifs = (struct ifstat *)data;
 281                 if (tp->tun_pid)
 282                         ksprintf(ifs->ascii + strlen(ifs->ascii),
 283                             "\tOpened by PID %d\n", tp->tun_pid);
 284                 break;
 285         case SIOCSIFADDR:
 286                 error = tuninit(ifp);
 287                 TUNDEBUG(ifp, "address set, error=%d\n", error);
 288                 break;
 289         case SIOCSIFDSTADDR:
 290                 error = tuninit(ifp);
 291                 TUNDEBUG(ifp, "destination address set, error=%d\n", error);
 292                 break;
 293         case SIOCSIFMTU:
 294                 ifp->if_mtu = ifr->ifr_mtu;
 295                 TUNDEBUG(ifp, "mtu set\n");
 296                 break;
 297         case SIOCSIFFLAGS:
 298         case SIOCADDMULTI:
 299         case SIOCDELMULTI:
 300                 break;
 301         default:
 302                 error = EINVAL;
 303         }
 304         return (error);
 305 }
 306
 307 /*
 308  * tunoutput - queue packets from higher level ready to put out.
 309  *
 310  * MPSAFE
 311  */
 312 static int
 313 tunoutput_serialized(struct ifnet *ifp, struct mbuf *m0, struct sockaddr *dst,
 314                      struct rtentry *rt)
 315 {
 316         struct tun_softc *tp = ifp->if_softc;
 317         int error;
 318         struct altq_pktattr pktattr;
 319
 320         TUNDEBUG(ifp, "tunoutput\n");
 321
 322         if ((tp->tun_flags & TUN_READY) != TUN_READY) {
 323                 TUNDEBUG(ifp, "not ready 0%o\n", tp->tun_flags);
 324                 m_freem (m0);
 325                 return EHOSTDOWN;
 326         }
 327
 328         /*
 329          * if the queueing discipline needs packet classification,
 330          * do it before prepending link headers.
 331          */
 332         ifq_classify(&ifp->if_snd, m0, dst->sa_family, &pktattr);
 333
 334         /* BPF write needs to be handled specially */
 335         if (dst->sa_family == AF_UNSPEC) {
 336                 dst->sa_family = *(mtod(m0, int *));
 337                 m0->m_len -= sizeof(int);
 338                 m0->m_pkthdr.len -= sizeof(int);
 339                 m0->m_data += sizeof(int);
 340         }
 341
 342         if (ifp->if_bpf) {
 343                 /*
 344                  * We need to prepend the address family as
 345                  * a four byte field.
 346                  */
 347                 uint32_t af = dst->sa_family;
 348
 349                 bpf_ptap(ifp->if_bpf, m0, &af, sizeof(af));
 350         }
 351
 352         /* prepend sockaddr? this may abort if the mbuf allocation fails */
 353         if (tp->tun_flags & TUN_LMODE) {
 354                 /* allocate space for sockaddr */
 355                 M_PREPEND(m0, dst->sa_len, MB_DONTWAIT);
 356
 357                 /* if allocation failed drop packet */
 358                 if (m0 == NULL){
 359                         IF_DROP(&ifp->if_snd);
 360                         ifp->if_oerrors++;
 361                         return (ENOBUFS);
 362                 } else {
 363                         bcopy(dst, m0->m_data, dst->sa_len);
 364                 }
 365         }
 366
 367         if (tp->tun_flags & TUN_IFHEAD) {
 368                 /* Prepend the address family */
 369                 M_PREPEND(m0, 4, MB_DONTWAIT);
 370
 371                 /* if allocation failed drop packet */
 372                 if (m0 == NULL){
 373                         IF_DROP(&ifp->if_snd);
 374                         ifp->if_oerrors++;
 375                         return ENOBUFS;
 376                 } else
 377                         *(u_int32_t *)m0->m_data = htonl(dst->sa_family);
 378         } else {
 379 #ifdef INET
 380                 if (dst->sa_family != AF_INET)
 381 #endif
 382                 {
 383                         m_freem(m0);
 384                         return EAFNOSUPPORT;
 385                 }
 386         }
 387
 388         error = ifq_handoff(ifp, m0, &pktattr);
 389         if (error) {
 390                 ifp->if_collisions++;
 391         } else {
 392                 ifp->if_opackets++;
 393                 if (tp->tun_flags & TUN_RWAIT) {
 394                         tp->tun_flags &= ~TUN_RWAIT;
 395                         wakeup((caddr_t)tp);
 396                 }
 397                 get_mplock();
 398                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 399                         pgsigio(tp->tun_sigio, SIGIO, 0);
 400                 rel_mplock();
 401                 ifnet_deserialize_all(ifp);
 402                 KNOTE(&tp->tun_rkq.ki_note, 0);
 403                 ifnet_serialize_all(ifp);
 404         }
 405         return (error);
 406 }
 407
 408 static int
 409 tunoutput(struct ifnet *ifp, struct mbuf *m0, struct sockaddr *dst,
 410           struct rtentry *rt)
 411 {
 412         int error;
 413
 414         ifnet_serialize_all(ifp);
 415         error = tunoutput_serialized(ifp, m0, dst, rt);
 416         ifnet_deserialize_all(ifp);
 417
 418         return error;
 419 }
 420
 421 /*
 422  * the ops interface is now pretty minimal.
 423  */
 424 static  int
 425 tunioctl(struct dev_ioctl_args *ap)
 426 {
 427         cdev_t dev = ap->a_head.a_dev;
 428         struct tun_softc *tp = dev->si_drv1;
 429         struct tuninfo *tunp;
 430
 431         switch (ap->a_cmd) {
 432         case TUNSIFINFO:
 433                 tunp = (struct tuninfo *)ap->a_data;
 434                 if (tunp->mtu < IF_MINMTU)
 435                         return (EINVAL);
 436                 tp->tun_if.if_mtu = tunp->mtu;
 437                 tp->tun_if.if_type = tunp->type;
 438                 tp->tun_if.if_baudrate = tunp->baudrate;
 439                 break;
 440         case TUNGIFINFO:
 441                 tunp = (struct tuninfo *)ap->a_data;
 442                 tunp->mtu = tp->tun_if.if_mtu;
 443                 tunp->type = tp->tun_if.if_type;
 444                 tunp->baudrate = tp->tun_if.if_baudrate;
 445                 break;
 446         case TUNSDEBUG:
 447                 tundebug = *(int *)ap->a_data;
 448                 break;
 449         case TUNGDEBUG:
 450                 *(int *)ap->a_data = tundebug;
 451                 break;
 452         case TUNSLMODE:
 453                 if (*(int *)ap->a_data) {
 454                         tp->tun_flags |= TUN_LMODE;
 455                         tp->tun_flags &= ~TUN_IFHEAD;
 456                 } else
 457                         tp->tun_flags &= ~TUN_LMODE;
 458                 break;
 459         case TUNSIFHEAD:
 460                 if (*(int *)ap->a_data) {
 461                         tp->tun_flags |= TUN_IFHEAD;
 462                         tp->tun_flags &= ~TUN_LMODE;
 463                 } else
 464                         tp->tun_flags &= ~TUN_IFHEAD;
 465                 break;
 466         case TUNGIFHEAD:
 467                 *(int *)ap->a_data = (tp->tun_flags & TUN_IFHEAD) ? 1 : 0;
 468                 break;
 469         case TUNSIFMODE:
 470                 /* deny this if UP */
 471                 if (tp->tun_if.if_flags & IFF_UP)
 472                         return(EBUSY);
 473
 474                 switch (*(int *)ap->a_data & ~IFF_MULTICAST) {
 475                 case IFF_POINTOPOINT:
 476                 case IFF_BROADCAST:
 477                         tp->tun_if.if_flags &= ~(IFF_BROADCAST|IFF_POINTOPOINT);
 478                         tp->tun_if.if_flags |= *(int *)ap->a_data;
 479                         break;
 480                 default:
 481                         return(EINVAL);
 482                 }
 483                 break;
 484         case TUNSIFPID:
 485                 tp->tun_pid = curproc->p_pid;
 486                 break;
 487         case FIOASYNC:
 488                 if (*(int *)ap->a_data)
 489                         tp->tun_flags |= TUN_ASYNC;
 490                 else
 491                         tp->tun_flags &= ~TUN_ASYNC;
 492                 break;
 493         case FIONREAD:
 494                 if (!ifq_is_empty(&tp->tun_if.if_snd)) {
 495                         struct mbuf *mb;
 496
 497                         mb = ifq_poll(&tp->tun_if.if_snd);
 498                         for( *(int *)ap->a_data = 0; mb != 0; mb = mb->m_next)
 499                                 *(int *)ap->a_data += mb->m_len;
 500                 } else {
 501                         *(int *)ap->a_data = 0;
 502                 }
 503                 break;
 504         case FIOSETOWN:
 505                 return (fsetown(*(int *)ap->a_data, &tp->tun_sigio));
 506
 507         case FIOGETOWN:
 508                 *(int *)ap->a_data = fgetown(tp->tun_sigio);
 509                 return (0);
 510
 511         /* This is deprecated, FIOSETOWN should be used instead. */
 512         case TIOCSPGRP:
 513                 return (fsetown(-(*(int *)ap->a_data), &tp->tun_sigio));
 514
 515         /* This is deprecated, FIOGETOWN should be used instead. */
 516         case TIOCGPGRP:
 517                 *(int *)ap->a_data = -fgetown(tp->tun_sigio);
 518                 return (0);
 519
 520         default:
 521                 return (ENOTTY);
 522         }
 523         return (0);
 524 }
 525
 526 /*
 527  * The ops read interface - reads a packet at a time, or at
 528  * least as much of a packet as can be read.
 529  */
 530 static  int
 531 tunread(struct dev_read_args *ap)
 532 {
 533         cdev_t dev = ap->a_head.a_dev;
 534         struct uio *uio = ap->a_uio;
 535         struct tun_softc *tp = dev->si_drv1;
 536         struct ifnet    *ifp = &tp->tun_if;
 537         struct mbuf     *m0;
 538         int             error=0, len;
 539
 540         TUNDEBUG(ifp, "read\n");
 541         if ((tp->tun_flags & TUN_READY) != TUN_READY) {
 542                 TUNDEBUG(ifp, "not ready 0%o\n", tp->tun_flags);
 543                 return EHOSTDOWN;
 544         }
 545
 546         tp->tun_flags &= ~TUN_RWAIT;
 547
 548         ifnet_serialize_all(ifp);
 549
 550         while ((m0 = ifq_dequeue(&ifp->if_snd, NULL)) == NULL) {
 551                 if (ap->a_ioflag & IO_NDELAY) {
 552                         ifnet_deserialize_all(ifp);
 553                         return EWOULDBLOCK;
 554                 }
 555                 tp->tun_flags |= TUN_RWAIT;
 556                 ifnet_deserialize_all(ifp);
 557                 if ((error = tsleep(tp, PCATCH, "tunread", 0)) != 0)
 558                         return error;
 559                 ifnet_serialize_all(ifp);
 560         }
 561
 562         ifnet_deserialize_all(ifp);
 563
 564         while (m0 && uio->uio_resid > 0 && error == 0) {
 565                 len = (int)szmin(uio->uio_resid, m0->m_len);
 566                 if (len != 0)
 567                         error = uiomove(mtod(m0, caddr_t), (size_t)len, uio);
 568                 m0 = m_free(m0);
 569         }
 570
 571         if (m0) {
 572                 TUNDEBUG(ifp, "Dropping mbuf\n");
 573                 m_freem(m0);
 574         }
 575         return error;
 576 }
 577
 578 /*
 579  * the ops write interface - an atomic write is a packet - or else!
 580  */
 581 static  int
 582 tunwrite(struct dev_write_args *ap)
 583 {
 584         cdev_t dev = ap->a_head.a_dev;
 585         struct uio *uio = ap->a_uio;
 586         struct tun_softc *tp = dev->si_drv1;
 587         struct ifnet    *ifp = &tp->tun_if;
 588         struct mbuf     *top, **mp, *m;
 589         int             error=0;
 590         size_t          tlen, mlen;
 591         uint32_t        family;
 592         int             isr;
 593
 594         TUNDEBUG(ifp, "tunwrite\n");
 595
 596         if (uio->uio_resid == 0)
 597                 return 0;
 598
 599         if (uio->uio_resid > TUNMRU) {
 600                 TUNDEBUG(ifp, "len=%zd!\n", uio->uio_resid);
 601                 return EIO;
 602         }
 603         tlen = uio->uio_resid;
 604
 605         /* get a header mbuf */
 606         MGETHDR(m, MB_DONTWAIT, MT_DATA);
 607         if (m == NULL)
 608                 return ENOBUFS;
 609         mlen = MHLEN;
 610
 611         top = 0;
 612         mp = &top;
 613         while (error == 0 && uio->uio_resid > 0) {
 614                 m->m_len = (int)szmin(mlen, uio->uio_resid);
 615                 error = uiomove(mtod (m, caddr_t), (size_t)m->m_len, uio);
 616                 *mp = m;
 617                 mp = &m->m_next;
 618                 if (uio->uio_resid > 0) {
 619                         MGET (m, MB_DONTWAIT, MT_DATA);
 620                         if (m == 0) {
 621                                 error = ENOBUFS;
 622                                 break;
 623                         }
 624                         mlen = MLEN;
 625                 }
 626         }
 627         if (error) {
 628                 if (top)
 629                         m_freem (top);
 630                 ifp->if_ierrors++;
 631                 return error;
 632         }
 633
 634         top->m_pkthdr.len = (int)tlen;
 635         top->m_pkthdr.rcvif = ifp;
 636
 637         if (ifp->if_bpf) {
 638                 if (tp->tun_flags & TUN_IFHEAD) {
 639                         /*
 640                          * Conveniently, we already have a 4-byte address
 641                          * family prepended to our packet !
 642                          * Inconveniently, it's in the wrong byte order !
 643                          */
 644                         if ((top = m_pullup(top, sizeof(family))) == NULL)
 645                                 return ENOBUFS;
 646                         *mtod(top, u_int32_t *) =
 647                             ntohl(*mtod(top, u_int32_t *));
 648                         bpf_mtap(ifp->if_bpf, top);
 649                         *mtod(top, u_int32_t *) =
 650                             htonl(*mtod(top, u_int32_t *));
 651                 } else {
 652                         /*
 653                          * We need to prepend the address family as
 654                          * a four byte field.
 655                          */
 656                         static const uint32_t af = AF_INET;
 657
 658                         bpf_ptap(ifp->if_bpf, top, &af, sizeof(af));
 659                 }
 660         }
 661
 662         if (tp->tun_flags & TUN_IFHEAD) {
 663                 if (top->m_len < sizeof(family) &&
 664                     (top = m_pullup(top, sizeof(family))) == NULL)
 665                                 return ENOBUFS;
 666                 family = ntohl(*mtod(top, u_int32_t *));
 667                 m_adj(top, sizeof(family));
 668         } else
 669                 family = AF_INET;
 670
 671         ifp->if_ibytes += top->m_pkthdr.len;
 672         ifp->if_ipackets++;
 673
 674         switch (family) {
 675 #ifdef INET
 676         case AF_INET:
 677                 isr = NETISR_IP;
 678                 break;
 679 #endif
 680 #ifdef INET6
 681         case AF_INET6:
 682                 isr = NETISR_IPV6;
 683                 break;
 684 #endif
 685 #ifdef IPX
 686         case AF_IPX:
 687                 isr = NETISR_IPX;
 688                 break;
 689 #endif
 690 #ifdef NETATALK
 691         case AF_APPLETALK:
 692                 isr = NETISR_ATALK2;
 693                 break;
 694 #endif
 695         default:
 696                 m_freem(m);
 697                 return (EAFNOSUPPORT);
 698         }
 699
 700         netisr_dispatch(isr, top);
 701         return (0);
 702 }
 703
 704 static struct filterops tun_read_filtops =
 705         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_read };
 706 static struct filterops tun_write_filtops =
 707         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_write };
 708
 709 static int
 710 tunkqfilter(struct dev_kqfilter_args *ap)
 711 {
 712         cdev_t dev = ap->a_head.a_dev;
 713         struct tun_softc *tp = dev->si_drv1;
 714         struct knote *kn = ap->a_kn;
 715         struct klist *klist;
 716
 717         ap->a_result = 0;
 718         ifnet_serialize_all(&tp->tun_if);
 719
 720         switch (kn->kn_filter) {
 721         case EVFILT_READ:
 722                 kn->kn_fop = &tun_read_filtops;
 723                 kn->kn_hook = (caddr_t)tp;
 724                 break;
 725         case EVFILT_WRITE:
 726                 kn->kn_fop = &tun_write_filtops;
 727                 kn->kn_hook = (caddr_t)tp;
 728                 break;
 729         default:
 730                 ifnet_deserialize_all(&tp->tun_if);
 731                 ap->a_result = EOPNOTSUPP;
 732                 return (0);
 733         }
 734
 735         klist = &tp->tun_rkq.ki_note;
 736         knote_insert(klist, kn);
 737         ifnet_deserialize_all(&tp->tun_if);
 738
 739         return (0);
 740 }
 741
 742 static void
 743 tun_filter_detach(struct knote *kn)
 744 {
 745         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 746         struct klist *klist = &tp->tun_rkq.ki_note;
 747
 748         knote_remove(klist, kn);
 749 }
 750
 751 static int
 752 tun_filter_write(struct knote *kn, long hint)
 753 {
 754         /* Always ready for a write */
 755         return (1);
 756 }
 757
 758 static int
 759 tun_filter_read(struct knote *kn, long hint)
 760 {
 761         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 762         int ready = 0;
 763
 764         ifnet_serialize_all(&tp->tun_if);
 765         if (!ifq_is_empty(&tp->tun_if.if_snd))
 766                 ready = 1;
 767         ifnet_deserialize_all(&tp->tun_if);
 768
 769         return (ready);
 770 }
 771
 772 /*
 773  * Start packet transmission on the interface.
 774  * when the interface queue is rate-limited by ALTQ,
 775  * if_start is needed to drain packets from the queue in order
 776  * to notify readers when outgoing packets become ready.
 777  */
 778 static void
 779 tunstart(struct ifnet *ifp)
 780 {
 781         struct tun_softc *tp = ifp->if_softc;
 782         struct mbuf *m;
 783
 784         if (!ifq_is_enabled(&ifp->if_snd))
 785                 return;
 786
 787         m = ifq_poll(&ifp->if_snd);
 788         if (m != NULL) {
 789                 if (tp->tun_flags & TUN_RWAIT) {
 790                         tp->tun_flags &= ~TUN_RWAIT;
 791                         wakeup((caddr_t)tp);
 792                 }
 793                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 794                         pgsigio(tp->tun_sigio, SIGIO, 0);
 795                 ifnet_deserialize_tx(ifp);
 796                 KNOTE(&tp->tun_rkq.ki_note, 0);
 797                 ifnet_serialize_tx(ifp);
 798         }
 799 }