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 #include "opt_ipx.h"
  23
  24 #include <sys/param.h>
  25 #include <sys/proc.h>
  26 #include <sys/priv.h>
  27 #include <sys/systm.h>
  28 #include <sys/mbuf.h>
  29 #include <sys/socket.h>
  30 #include <sys/conf.h>
  31 #include <sys/device.h>
  32 #include <sys/filio.h>
  33 #include <sys/sockio.h>
  34 #include <sys/thread2.h>
  35 #include <sys/ttycom.h>
  36 #include <sys/signalvar.h>
  37 #include <sys/filedesc.h>
  38 #include <sys/kernel.h>
  39 #include <sys/sysctl.h>
  40 #include <sys/uio.h>
  41 #include <sys/vnode.h>
  42 #include <sys/malloc.h>
  43
  44 #include <sys/mplock2.h>
  45
  46 #include <net/if.h>
  47 #include <net/if_types.h>
  48 #include <net/ifq_var.h>
  49 #include <net/netisr.h>
  50 #include <net/route.h>
  51 #include <sys/devfs.h>
  52
  53 #ifdef INET
  54 #include <netinet/in.h>
  55 #endif
  56
  57 #include <net/bpf.h>
  58
  59 #include "if_tunvar.h"
  60 #include "if_tun.h"
  61
  62 static MALLOC_DEFINE(M_TUN, "tun", "Tunnel Interface");
  63
  64 static void tunattach (void *);
  65 PSEUDO_SET(tunattach, if_tun);
  66
  67 static void tuncreate (cdev_t dev);
  68
  69 #define TUNDEBUG        if (tundebug) if_printf
  70 static int tundebug = 0;
  71 SYSCTL_INT(_debug, OID_AUTO, if_tun_debug, CTLFLAG_RW, &tundebug, 0,
  72     "Enable debug output");
  73
  74 static int tunoutput (struct ifnet *, struct mbuf *, struct sockaddr *,
  75             struct rtentry *rt);
  76 static int tunifioctl (struct ifnet *, u_long, caddr_t, struct ucred *);
  77 static int tuninit (struct ifnet *);
  78 static void tunstart(struct ifnet *, struct ifaltq_subque *);
  79 static void tun_filter_detach(struct knote *);
  80 static int tun_filter_read(struct knote *, long);
  81 static int tun_filter_write(struct knote *, long);
  82
  83 static  d_open_t        tunopen;
  84 static  d_close_t       tunclose;
  85 static  d_read_t        tunread;
  86 static  d_write_t       tunwrite;
  87 static  d_ioctl_t       tunioctl;
  88 static  d_kqfilter_t    tunkqfilter;
  89
  90 static d_clone_t tunclone;
  91 DEVFS_DECLARE_CLONE_BITMAP(tun);
  92
  93 #if NTUN <= 1
  94 #define TUN_PREALLOCATED_UNITS  4
  95 #else
  96 #define TUN_PREALLOCATED_UNITS  NTUN
  97 #endif
  98
  99 static struct dev_ops tun_ops = {
 100         { "tun", 0, 0 },
 101         .d_open =       tunopen,
 102         .d_close =      tunclose,
 103         .d_read =       tunread,
 104         .d_write =      tunwrite,
 105         .d_ioctl =      tunioctl,
 106         .d_kqfilter =   tunkqfilter
 107 };
 108
 109 static void
 110 tunattach(void *dummy)
 111 {
 112         int i;
 113         make_autoclone_dev(&tun_ops, &DEVFS_CLONE_BITMAP(tun),
 114                 tunclone, UID_UUCP, GID_DIALER, 0600, "tun");
 115         for (i = 0; i < TUN_PREALLOCATED_UNITS; i++) {
 116                 make_dev(&tun_ops, i, UID_UUCP, GID_DIALER, 0600, "tun%d", i);
 117                 devfs_clone_bitmap_set(&DEVFS_CLONE_BITMAP(tun), i);
 118         }
 119         /* Doesn't need uninit because unloading is not possible, see PSEUDO_SET */
 120 }
 121
 122 static int
 123 tunclone(struct dev_clone_args *ap)
 124 {
 125         int unit;
 126
 127         unit = devfs_clone_bitmap_get(&DEVFS_CLONE_BITMAP(tun), 0);
 128         ap->a_dev = make_only_dev(&tun_ops, unit, UID_UUCP, GID_DIALER, 0600,
 129                                                                 "tun%d", unit);
 130
 131         return 0;
 132 }
 133
 134 static void
 135 tuncreate(cdev_t dev)
 136 {
 137         struct tun_softc *sc;
 138         struct ifnet *ifp;
 139
 140 #if 0
 141         dev = make_dev(&tun_ops, minor(dev),
 142             UID_UUCP, GID_DIALER, 0600, "tun%d", lminor(dev));
 143 #endif
 144
 145         sc = kmalloc(sizeof(*sc), M_TUN, M_WAITOK | M_ZERO);
 146         sc->tun_flags = TUN_INITED;
 147
 148         ifp = &sc->tun_if;
 149         if_initname(ifp, "tun", lminor(dev));
 150         ifp->if_mtu = TUNMTU;
 151         ifp->if_ioctl = tunifioctl;
 152         ifp->if_output = tunoutput;
 153         ifp->if_start = tunstart;
 154         ifp->if_flags = IFF_POINTOPOINT | IFF_MULTICAST;
 155         ifp->if_type = IFT_PPP;
 156         ifq_set_maxlen(&ifp->if_snd, ifqmaxlen);
 157         ifq_set_ready(&ifp->if_snd);
 158         ifp->if_softc = sc;
 159         if_attach(ifp, NULL);
 160         bpfattach(ifp, DLT_NULL, sizeof(u_int));
 161         dev->si_drv1 = sc;
 162 }
 163
 164 /*
 165  * tunnel open - must be superuser & the device must be
 166  * configured in
 167  */
 168 static  int
 169 tunopen(struct dev_open_args *ap)
 170 {
 171         cdev_t dev = ap->a_head.a_dev;
 172         struct ifnet    *ifp;
 173         struct tun_softc *tp;
 174         int     error;
 175
 176         if ((error = priv_check_cred(ap->a_cred, PRIV_ROOT, 0)) != 0)
 177                 return (error);
 178
 179         tp = dev->si_drv1;
 180         if (!tp) {
 181                 tuncreate(dev);
 182                 tp = dev->si_drv1;
 183         }
 184         if (tp->tun_flags & TUN_OPEN)
 185                 return EBUSY;
 186         tp->tun_pid = curproc->p_pid;
 187         ifp = &tp->tun_if;
 188         tp->tun_flags |= TUN_OPEN;
 189         TUNDEBUG(ifp, "open\n");
 190         return (0);
 191 }
 192
 193 /*
 194  * tunclose - close the device - mark i/f down & delete
 195  * routing info
 196  */
 197 static  int
 198 tunclose(struct dev_close_args *ap)
 199 {
 200         cdev_t dev = ap->a_head.a_dev;
 201         struct tun_softc *tp;
 202         struct ifnet    *ifp;
 203
 204         tp = dev->si_drv1;
 205         ifp = &tp->tun_if;
 206
 207         tp->tun_flags &= ~TUN_OPEN;
 208         tp->tun_pid = 0;
 209
 210         /* Junk all pending output. */
 211         ifq_purge_all(&ifp->if_snd);
 212
 213         if (ifp->if_flags & IFF_UP)
 214                 if_down(ifp);
 215         ifp->if_flags &= ~IFF_RUNNING;
 216         if_purgeaddrs_nolink(ifp);
 217
 218         funsetown(&tp->tun_sigio);
 219         KNOTE(&tp->tun_rkq.ki_note, 0);
 220
 221         TUNDEBUG(ifp, "closed\n");
 222 #if 0
 223         if (dev->si_uminor >= TUN_PREALLOCATED_UNITS) {
 224                 devfs_clone_bitmap_put(&DEVFS_CLONE_BITMAP(tun), dev->si_uminor);
 225         }
 226 #endif
 227         return (0);
 228 }
 229
 230 static int
 231 tuninit(struct ifnet *ifp)
 232 {
 233 #ifdef INET
 234         struct tun_softc *tp = ifp->if_softc;
 235 #endif
 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                 bpf_gettoken();
 344                 if (ifp->if_bpf) {
 345                         /*
 346                          * We need to prepend the address family as
 347                          * a four byte field.
 348                          */
 349                         uint32_t af = dst->sa_family;
 350
 351                         bpf_ptap(ifp->if_bpf, m0, &af, sizeof(af));
 352                 }
 353                 bpf_reltoken();
 354         }
 355
 356         /* prepend sockaddr? this may abort if the mbuf allocation fails */
 357         if (tp->tun_flags & TUN_LMODE) {
 358                 /* allocate space for sockaddr */
 359                 M_PREPEND(m0, dst->sa_len, MB_DONTWAIT);
 360
 361                 /* if allocation failed drop packet */
 362                 if (m0 == NULL){
 363                         IFNET_STAT_INC(ifp, oerrors, 1);
 364                         return (ENOBUFS);
 365                 } else {
 366                         bcopy(dst, m0->m_data, dst->sa_len);
 367                 }
 368         }
 369
 370         if (tp->tun_flags & TUN_IFHEAD) {
 371                 /* Prepend the address family */
 372                 M_PREPEND(m0, 4, MB_DONTWAIT);
 373
 374                 /* if allocation failed drop packet */
 375                 if (m0 == NULL){
 376                         IFNET_STAT_INC(ifp, oerrors, 1);
 377                         return ENOBUFS;
 378                 } else
 379                         *(u_int32_t *)m0->m_data = htonl(dst->sa_family);
 380         } else {
 381 #ifdef INET
 382                 if (dst->sa_family != AF_INET)
 383 #endif
 384                 {
 385                         m_freem(m0);
 386                         return EAFNOSUPPORT;
 387                 }
 388         }
 389
 390         error = ifq_handoff(ifp, m0, &pktattr);
 391         if (error) {
 392                 IFNET_STAT_INC(ifp, collisions, 1);
 393         } else {
 394                 IFNET_STAT_INC(ifp, opackets, 1);
 395                 if (tp->tun_flags & TUN_RWAIT) {
 396                         tp->tun_flags &= ~TUN_RWAIT;
 397                         wakeup((caddr_t)tp);
 398                 }
 399                 get_mplock();
 400                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 401                         pgsigio(tp->tun_sigio, SIGIO, 0);
 402                 rel_mplock();
 403                 ifnet_deserialize_all(ifp);
 404                 KNOTE(&tp->tun_rkq.ki_note, 0);
 405                 ifnet_serialize_all(ifp);
 406         }
 407         return (error);
 408 }
 409
 410 static int
 411 tunoutput(struct ifnet *ifp, struct mbuf *m0, struct sockaddr *dst,
 412           struct rtentry *rt)
 413 {
 414         int error;
 415
 416         ifnet_serialize_all(ifp);
 417         error = tunoutput_serialized(ifp, m0, dst, rt);
 418         ifnet_deserialize_all(ifp);
 419
 420         return error;
 421 }
 422
 423 /*
 424  * the ops interface is now pretty minimal.
 425  */
 426 static  int
 427 tunioctl(struct dev_ioctl_args *ap)
 428 {
 429         cdev_t dev = ap->a_head.a_dev;
 430         struct tun_softc *tp = dev->si_drv1;
 431         struct tuninfo *tunp;
 432
 433         switch (ap->a_cmd) {
 434         case TUNSIFINFO:
 435                 tunp = (struct tuninfo *)ap->a_data;
 436                 if (tunp->mtu < IF_MINMTU)
 437                         return (EINVAL);
 438                 tp->tun_if.if_mtu = tunp->mtu;
 439                 tp->tun_if.if_type = tunp->type;
 440                 tp->tun_if.if_baudrate = tunp->baudrate;
 441                 break;
 442         case TUNGIFINFO:
 443                 tunp = (struct tuninfo *)ap->a_data;
 444                 tunp->mtu = tp->tun_if.if_mtu;
 445                 tunp->type = tp->tun_if.if_type;
 446                 tunp->baudrate = tp->tun_if.if_baudrate;
 447                 break;
 448         case TUNSDEBUG:
 449                 tundebug = *(int *)ap->a_data;
 450                 break;
 451         case TUNGDEBUG:
 452                 *(int *)ap->a_data = tundebug;
 453                 break;
 454         case TUNSLMODE:
 455                 if (*(int *)ap->a_data) {
 456                         tp->tun_flags |= TUN_LMODE;
 457                         tp->tun_flags &= ~TUN_IFHEAD;
 458                 } else
 459                         tp->tun_flags &= ~TUN_LMODE;
 460                 break;
 461         case TUNSIFHEAD:
 462                 if (*(int *)ap->a_data) {
 463                         tp->tun_flags |= TUN_IFHEAD;
 464                         tp->tun_flags &= ~TUN_LMODE;
 465                 } else
 466                         tp->tun_flags &= ~TUN_IFHEAD;
 467                 break;
 468         case TUNGIFHEAD:
 469                 *(int *)ap->a_data = (tp->tun_flags & TUN_IFHEAD) ? 1 : 0;
 470                 break;
 471         case TUNSIFMODE:
 472                 /* deny this if UP */
 473                 if (tp->tun_if.if_flags & IFF_UP)
 474                         return(EBUSY);
 475
 476                 switch (*(int *)ap->a_data & ~IFF_MULTICAST) {
 477                 case IFF_POINTOPOINT:
 478                 case IFF_BROADCAST:
 479                         tp->tun_if.if_flags &= ~(IFF_BROADCAST|IFF_POINTOPOINT);
 480                         tp->tun_if.if_flags |= *(int *)ap->a_data;
 481                         break;
 482                 default:
 483                         return(EINVAL);
 484                 }
 485                 break;
 486         case TUNSIFPID:
 487                 tp->tun_pid = curproc->p_pid;
 488                 break;
 489         case FIOASYNC:
 490                 if (*(int *)ap->a_data)
 491                         tp->tun_flags |= TUN_ASYNC;
 492                 else
 493                         tp->tun_flags &= ~TUN_ASYNC;
 494                 break;
 495         case FIONREAD:
 496                 *(int *)ap->a_data = ifsq_poll_pktlen(
 497                     ifq_get_subq_default(&tp->tun_if.if_snd));
 498                 break;
 499         case FIOSETOWN:
 500                 return (fsetown(*(int *)ap->a_data, &tp->tun_sigio));
 501
 502         case FIOGETOWN:
 503                 *(int *)ap->a_data = fgetown(&tp->tun_sigio);
 504                 return (0);
 505
 506         /* This is deprecated, FIOSETOWN should be used instead. */
 507         case TIOCSPGRP:
 508                 return (fsetown(-(*(int *)ap->a_data), &tp->tun_sigio));
 509
 510         /* This is deprecated, FIOGETOWN should be used instead. */
 511         case TIOCGPGRP:
 512                 *(int *)ap->a_data = -fgetown(&tp->tun_sigio);
 513                 return (0);
 514
 515         default:
 516                 return (ENOTTY);
 517         }
 518         return (0);
 519 }
 520
 521 /*
 522  * The ops read interface - reads a packet at a time, or at
 523  * least as much of a packet as can be read.
 524  */
 525 static  int
 526 tunread(struct dev_read_args *ap)
 527 {
 528         cdev_t dev = ap->a_head.a_dev;
 529         struct uio *uio = ap->a_uio;
 530         struct tun_softc *tp = dev->si_drv1;
 531         struct ifnet    *ifp = &tp->tun_if;
 532         struct ifaltq_subque *ifsq = ifq_get_subq_default(&ifp->if_snd);
 533         struct mbuf     *m0;
 534         int             error=0, len;
 535
 536         TUNDEBUG(ifp, "read\n");
 537         if ((tp->tun_flags & TUN_READY) != TUN_READY) {
 538                 TUNDEBUG(ifp, "not ready 0%o\n", tp->tun_flags);
 539                 return EHOSTDOWN;
 540         }
 541
 542         tp->tun_flags &= ~TUN_RWAIT;
 543
 544         ifnet_serialize_all(ifp);
 545
 546         while ((m0 = ifsq_dequeue(ifsq)) == NULL) {
 547                 if (ap->a_ioflag & IO_NDELAY) {
 548                         ifnet_deserialize_all(ifp);
 549                         return EWOULDBLOCK;
 550                 }
 551                 tp->tun_flags |= TUN_RWAIT;
 552                 ifnet_deserialize_all(ifp);
 553                 if ((error = tsleep(tp, PCATCH, "tunread", 0)) != 0)
 554                         return error;
 555                 ifnet_serialize_all(ifp);
 556         }
 557
 558         ifnet_deserialize_all(ifp);
 559
 560         while (m0 && uio->uio_resid > 0 && error == 0) {
 561                 len = (int)szmin(uio->uio_resid, m0->m_len);
 562                 if (len != 0)
 563                         error = uiomove(mtod(m0, caddr_t), (size_t)len, uio);
 564                 m0 = m_free(m0);
 565         }
 566
 567         if (m0) {
 568                 TUNDEBUG(ifp, "Dropping mbuf\n");
 569                 m_freem(m0);
 570         }
 571         return error;
 572 }
 573
 574 /*
 575  * the ops write interface - an atomic write is a packet - or else!
 576  */
 577 static  int
 578 tunwrite(struct dev_write_args *ap)
 579 {
 580         cdev_t dev = ap->a_head.a_dev;
 581         struct uio *uio = ap->a_uio;
 582         struct tun_softc *tp = dev->si_drv1;
 583         struct ifnet    *ifp = &tp->tun_if;
 584         struct mbuf     *top, **mp, *m;
 585         int             error=0;
 586         size_t          tlen, mlen;
 587         uint32_t        family;
 588         int             isr;
 589
 590         TUNDEBUG(ifp, "tunwrite\n");
 591
 592         if (uio->uio_resid == 0)
 593                 return 0;
 594
 595         if (uio->uio_resid > TUNMRU) {
 596                 TUNDEBUG(ifp, "len=%zd!\n", uio->uio_resid);
 597                 return EIO;
 598         }
 599         tlen = uio->uio_resid;
 600
 601         /* get a header mbuf */
 602         MGETHDR(m, MB_WAIT, MT_DATA);
 603         if (m == NULL)
 604                 return ENOBUFS;
 605         mlen = MHLEN;
 606
 607         top = NULL;
 608         mp = &top;
 609         while (error == 0 && uio->uio_resid > 0) {
 610                 m->m_len = (int)szmin(mlen, uio->uio_resid);
 611                 error = uiomove(mtod (m, caddr_t), (size_t)m->m_len, uio);
 612                 *mp = m;
 613                 mp = &m->m_next;
 614                 if (uio->uio_resid > 0) {
 615                         MGET (m, MB_WAIT, MT_DATA);
 616                         if (m == NULL) {
 617                                 error = ENOBUFS;
 618                                 break;
 619                         }
 620                         mlen = MLEN;
 621                 }
 622         }
 623         if (error) {
 624                 if (top)
 625                         m_freem (top);
 626                 IFNET_STAT_INC(ifp, ierrors, 1);
 627                 return error;
 628         }
 629
 630         top->m_pkthdr.len = (int)tlen;
 631         top->m_pkthdr.rcvif = ifp;
 632
 633         if (ifp->if_bpf) {
 634                 bpf_gettoken();
 635
 636                 if (ifp->if_bpf) {
 637                         if (tp->tun_flags & TUN_IFHEAD) {
 638                                 /*
 639                                  * Conveniently, we already have a 4-byte
 640                                  * address family prepended to our packet !
 641                                  * Inconveniently, it's in the wrong byte
 642                                  * order !
 643                                  */
 644                                 if ((top = m_pullup(top, sizeof(family)))
 645                                     == NULL) {
 646                                         bpf_reltoken();
 647                                         return ENOBUFS;
 648                                 }
 649                                 *mtod(top, u_int32_t *) =
 650                                     ntohl(*mtod(top, u_int32_t *));
 651                                 bpf_mtap(ifp->if_bpf, top);
 652                                 *mtod(top, u_int32_t *) =
 653                                     htonl(*mtod(top, u_int32_t *));
 654                         } else {
 655                                 /*
 656                                  * We need to prepend the address family as
 657                                  * a four byte field.
 658                                  */
 659                                 static const uint32_t af = AF_INET;
 660
 661                                 bpf_ptap(ifp->if_bpf, top, &af, sizeof(af));
 662                         }
 663                 }
 664
 665                 bpf_reltoken();
 666         }
 667
 668         if (tp->tun_flags & TUN_IFHEAD) {
 669                 if (top->m_len < sizeof(family) &&
 670                     (top = m_pullup(top, sizeof(family))) == NULL)
 671                                 return ENOBUFS;
 672                 family = ntohl(*mtod(top, u_int32_t *));
 673                 m_adj(top, sizeof(family));
 674         } else
 675                 family = AF_INET;
 676
 677         IFNET_STAT_INC(ifp, ibytes, top->m_pkthdr.len);
 678         IFNET_STAT_INC(ifp, ipackets, 1);
 679
 680         switch (family) {
 681 #ifdef INET
 682         case AF_INET:
 683                 isr = NETISR_IP;
 684                 break;
 685 #endif
 686 #ifdef INET6
 687         case AF_INET6:
 688                 isr = NETISR_IPV6;
 689                 break;
 690 #endif
 691 #ifdef IPX
 692         case AF_IPX:
 693                 isr = NETISR_IPX;
 694                 break;
 695 #endif
 696         default:
 697                 m_freem(m);
 698                 return (EAFNOSUPPORT);
 699         }
 700
 701         netisr_queue(isr, top);
 702         return (0);
 703 }
 704
 705 static struct filterops tun_read_filtops =
 706         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_read };
 707 static struct filterops tun_write_filtops =
 708         { FILTEROP_ISFD, NULL, tun_filter_detach, tun_filter_write };
 709
 710 static int
 711 tunkqfilter(struct dev_kqfilter_args *ap)
 712 {
 713         cdev_t dev = ap->a_head.a_dev;
 714         struct tun_softc *tp = dev->si_drv1;
 715         struct knote *kn = ap->a_kn;
 716         struct klist *klist;
 717
 718         ap->a_result = 0;
 719         ifnet_serialize_all(&tp->tun_if);
 720
 721         switch (kn->kn_filter) {
 722         case EVFILT_READ:
 723                 kn->kn_fop = &tun_read_filtops;
 724                 kn->kn_hook = (caddr_t)tp;
 725                 break;
 726         case EVFILT_WRITE:
 727                 kn->kn_fop = &tun_write_filtops;
 728                 kn->kn_hook = (caddr_t)tp;
 729                 break;
 730         default:
 731                 ifnet_deserialize_all(&tp->tun_if);
 732                 ap->a_result = EOPNOTSUPP;
 733                 return (0);
 734         }
 735
 736         klist = &tp->tun_rkq.ki_note;
 737         knote_insert(klist, kn);
 738         ifnet_deserialize_all(&tp->tun_if);
 739
 740         return (0);
 741 }
 742
 743 static void
 744 tun_filter_detach(struct knote *kn)
 745 {
 746         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 747         struct klist *klist = &tp->tun_rkq.ki_note;
 748
 749         knote_remove(klist, kn);
 750 }
 751
 752 static int
 753 tun_filter_write(struct knote *kn, long hint)
 754 {
 755         /* Always ready for a write */
 756         return (1);
 757 }
 758
 759 static int
 760 tun_filter_read(struct knote *kn, long hint)
 761 {
 762         struct tun_softc *tp = (struct tun_softc *)kn->kn_hook;
 763         int ready = 0;
 764
 765         ifnet_serialize_all(&tp->tun_if);
 766         if (!ifsq_is_empty(ifq_get_subq_default(&tp->tun_if.if_snd)))
 767                 ready = 1;
 768         ifnet_deserialize_all(&tp->tun_if);
 769
 770         return (ready);
 771 }
 772
 773 /*
 774  * Start packet transmission on the interface.
 775  * when the interface queue is rate-limited by ALTQ,
 776  * if_start is needed to drain packets from the queue in order
 777  * to notify readers when outgoing packets become ready.
 778  */
 779 static void
 780 tunstart(struct ifnet *ifp, struct ifaltq_subque *ifsq)
 781 {
 782         struct tun_softc *tp = ifp->if_softc;
 783         struct mbuf *m;
 784
 785         ASSERT_ALTQ_SQ_DEFAULT(ifp, ifsq);
 786
 787         if (!ifq_is_enabled(&ifp->if_snd))
 788                 return;
 789
 790         m = ifsq_poll(ifsq);
 791         if (m != NULL) {
 792                 if (tp->tun_flags & TUN_RWAIT) {
 793                         tp->tun_flags &= ~TUN_RWAIT;
 794                         wakeup((caddr_t)tp);
 795                 }
 796                 if (tp->tun_flags & TUN_ASYNC && tp->tun_sigio)
 797                         pgsigio(tp->tun_sigio, SIGIO, 0);
 798                 ifsq_deserialize_hw(ifsq);
 799                 KNOTE(&tp->tun_rkq.ki_note, 0);
 800                 ifsq_serialize_hw(ifsq);
 801         }
 802 }