sys/netinet6/in6_pcb.c

   1 /*      $FreeBSD: src/sys/netinet6/in6_pcb.c,v 1.10.2.9 2003/01/24 05:11:35 sam Exp $   */
   2 /*      $DragonFly: src/sys/netinet6/in6_pcb.c,v 1.28 2006/01/31 19:05:42 dillon Exp $  */
   3 /*      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $        */
   4
   5 /*
   6  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Neither the name of the project nor the names of its contributors
  18  *    may be used to endorse or promote products derived from this software
  19  *    without specific prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  */
  34
  35 /*
  36  * Copyright (c) 1982, 1986, 1991, 1993
  37  *      The Regents of the University of California.  All rights reserved.
  38  *
  39  * Redistribution and use in source and binary forms, with or without
  40  * modification, are permitted provided that the following conditions
  41  * are met:
  42  * 1. Redistributions of source code must retain the above copyright
  43  *    notice, this list of conditions and the following disclaimer.
  44  * 2. Redistributions in binary form must reproduce the above copyright
  45  *    notice, this list of conditions and the following disclaimer in the
  46  *    documentation and/or other materials provided with the distribution.
  47  * 3. All advertising materials mentioning features or use of this software
  48  *    must display the following acknowledgement:
  49  *      This product includes software developed by the University of
  50  *      California, Berkeley and its contributors.
  51  * 4. Neither the name of the University nor the names of its contributors
  52  *    may be used to endorse or promote products derived from this software
  53  *    without specific prior written permission.
  54  *
  55  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  56  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  57  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  58  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  59  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  60  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  61  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  62  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  63  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  64  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  65  * SUCH DAMAGE.
  66  *
  67  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  68  */
  69
  70 #include "opt_inet.h"
  71 #include "opt_inet6.h"
  72 #include "opt_ipsec.h"
  73
  74 #include <sys/param.h>
  75 #include <sys/systm.h>
  76 #include <sys/malloc.h>
  77 #include <sys/mbuf.h>
  78 #include <sys/domain.h>
  79 #include <sys/protosw.h>
  80 #include <sys/socket.h>
  81 #include <sys/socketvar.h>
  82 #include <sys/sockio.h>
  83 #include <sys/errno.h>
  84 #include <sys/time.h>
  85 #include <sys/proc.h>
  86 #include <sys/jail.h>
  87 #include <sys/thread2.h>
  88
  89 #include <vm/vm_zone.h>
  90
  91 #include <net/if.h>
  92 #include <net/if_types.h>
  93 #include <net/route.h>
  94
  95 #include <netinet/in.h>
  96 #include <netinet/in_var.h>
  97 #include <netinet/in_systm.h>
  98 #include <netinet/ip6.h>
  99 #include <netinet/ip_var.h>
 100 #include <netinet6/ip6_var.h>
 101 #include <netinet6/nd6.h>
 102 #include <netinet/in_pcb.h>
 103 #include <netinet6/in6_pcb.h>
 104
 105 #ifdef IPSEC
 106 #include <netinet6/ipsec.h>
 107 #ifdef INET6
 108 #include <netinet6/ipsec6.h>
 109 #endif
 110 #include <netinet6/ah.h>
 111 #ifdef INET6
 112 #include <netinet6/ah6.h>
 113 #endif
 114 #include <netproto/key/key.h>
 115 #endif /* IPSEC */
 116
 117 #ifdef FAST_IPSEC
 118 #include <netproto/ipsec/ipsec.h>
 119 #include <netproto/ipsec/ipsec6.h>
 120 #include <netproto/ipsec/key.h>
 121 #define IPSEC
 122 #endif /* FAST_IPSEC */
 123
 124 struct  in6_addr zeroin6_addr;
 125
 126 int
 127 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 128 {
 129         struct socket *so = inp->inp_socket;
 130         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 131         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 132         u_short lport = 0;
 133         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 134
 135         if (!in6_ifaddr) /* XXX broken! */
 136                 return (EADDRNOTAVAIL);
 137         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 138                 return(EINVAL);
 139         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 140                 wild = 1;
 141         if (nam) {
 142                 sin6 = (struct sockaddr_in6 *)nam;
 143                 if (nam->sa_len != sizeof(*sin6))
 144                         return(EINVAL);
 145                 /*
 146                  * family check.
 147                  */
 148                 if (nam->sa_family != AF_INET6)
 149                         return(EAFNOSUPPORT);
 150
 151                 /* KAME hack: embed scopeid */
 152                 if (in6_embedscope(&sin6->sin6_addr, sin6, inp, NULL) != 0)
 153                         return EINVAL;
 154                 /* this must be cleared for ifa_ifwithaddr() */
 155                 sin6->sin6_scope_id = 0;
 156
 157                 lport = sin6->sin6_port;
 158                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 159                         /*
 160                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 161                          * allow compepte duplication of binding if
 162                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 163                          * and a multicast address is bound on both
 164                          * new and duplicated sockets.
 165                          */
 166                         if (so->so_options & SO_REUSEADDR)
 167                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 168                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 169                         struct ifaddr *ia = NULL;
 170
 171                         sin6->sin6_port = 0;            /* yech... */
 172                         if ((ia = ifa_ifwithaddr((struct sockaddr *)sin6)) == 0)
 173                                 return(EADDRNOTAVAIL);
 174
 175                         /*
 176                          * XXX: bind to an anycast address might accidentally
 177                          * cause sending a packet with anycast source address.
 178                          * We should allow to bind to a deprecated address, since
 179                          * the application dare to use it.
 180                          */
 181                         if (ia &&
 182                             ((struct in6_ifaddr *)ia)->ia6_flags &
 183                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 184                                 return(EADDRNOTAVAIL);
 185                         }
 186                 }
 187                 if (lport) {
 188                         struct inpcb *t;
 189                         struct proc *p = td->td_proc; /* may be NULL */
 190
 191                         /* GROSS */
 192                         if (ntohs(lport) < IPV6PORT_RESERVED && p &&
 193                             suser_cred(p->p_ucred, PRISON_ROOT))
 194                                 return(EACCES);
 195                         if (so->so_cred->cr_uid != 0 &&
 196                             !IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 197                                 t = in6_pcblookup_local(pcbinfo,
 198                                     &sin6->sin6_addr, lport,
 199                                     INPLOOKUP_WILDCARD);
 200                                 if (t &&
 201                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 202                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 203                                      (t->inp_socket->so_options &
 204                                       SO_REUSEPORT) == 0) &&
 205                                     (so->so_cred->cr_uid !=
 206                                      t->inp_socket->so_cred->cr_uid))
 207                                         return (EADDRINUSE);
 208                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 209                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 210                                         struct sockaddr_in sin;
 211
 212                                         in6_sin6_2_sin(&sin, sin6);
 213                                         t = in_pcblookup_local(pcbinfo,
 214                                                 sin.sin_addr, lport,
 215                                                 INPLOOKUP_WILDCARD);
 216                                         if (t &&
 217                                             (so->so_cred->cr_uid !=
 218                                              t->inp_socket->so_cred->cr_uid) &&
 219                                             (ntohl(t->inp_laddr.s_addr) !=
 220                                              INADDR_ANY ||
 221                                              INP_SOCKAF(so) ==
 222                                              INP_SOCKAF(t->inp_socket)))
 223                                                 return (EADDRINUSE);
 224                                 }
 225                         }
 226                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 227                                                 lport, wild);
 228                         if (t && (reuseport & t->inp_socket->so_options) == 0)
 229                                 return(EADDRINUSE);
 230                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 231                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 232                                 struct sockaddr_in sin;
 233
 234                                 in6_sin6_2_sin(&sin, sin6);
 235                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 236                                                        lport, wild);
 237                                 if (t &&
 238                                     (reuseport & t->inp_socket->so_options)
 239                                     == 0 &&
 240                                     (ntohl(t->inp_laddr.s_addr)
 241                                      != INADDR_ANY ||
 242                                      INP_SOCKAF(so) ==
 243                                      INP_SOCKAF(t->inp_socket)))
 244                                         return (EADDRINUSE);
 245                         }
 246                 }
 247                 inp->in6p_laddr = sin6->sin6_addr;
 248         }
 249         if (lport == 0) {
 250                 int e;
 251                 if ((e = in6_pcbsetport(&inp->in6p_laddr, inp, td)) != 0)
 252                         return(e);
 253         }
 254         else {
 255                 inp->inp_lport = lport;
 256                 if (in_pcbinsporthash(inp) != 0) {
 257                         inp->in6p_laddr = in6addr_any;
 258                         inp->inp_lport = 0;
 259                         return (EAGAIN);
 260                 }
 261         }
 262         return(0);
 263 }
 264
 265 /*
 266  *   Transform old in6_pcbconnect() into an inner subroutine for new
 267  *   in6_pcbconnect(): Do some validity-checking on the remote
 268  *   address (in mbuf 'nam') and then determine local host address
 269  *   (i.e., which interface) to use to access that remote host.
 270  *
 271  *   This preserves definition of in6_pcbconnect(), while supporting a
 272  *   slightly different version for T/TCP.  (This is more than
 273  *   a bit of a kludge, but cleaning up the internal interfaces would
 274  *   have forced minor changes in every protocol).
 275  */
 276
 277 int
 278 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 279              struct in6_addr **plocal_addr6)
 280 {
 281         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 282         struct ifnet *ifp = NULL;
 283         int error = 0;
 284
 285         if (nam->sa_len != sizeof (*sin6))
 286                 return (EINVAL);
 287         if (sin6->sin6_family != AF_INET6)
 288                 return (EAFNOSUPPORT);
 289         if (sin6->sin6_port == 0)
 290                 return (EADDRNOTAVAIL);
 291
 292         /* KAME hack: embed scopeid */
 293         if (in6_embedscope(&sin6->sin6_addr, sin6, inp, &ifp) != 0)
 294                 return EINVAL;
 295
 296         if (in6_ifaddr) {
 297                 /*
 298                  * If the destination address is UNSPECIFIED addr,
 299                  * use the loopback addr, e.g ::1.
 300                  */
 301                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 302                         sin6->sin6_addr = in6addr_loopback;
 303         }
 304         {
 305                 /*
 306                  * XXX: in6_selectsrc might replace the bound local address
 307                  * with the address specified by setsockopt(IPV6_PKTINFO).
 308                  * Is it the intended behavior?
 309                  */
 310                 *plocal_addr6 = in6_selectsrc(sin6, inp->in6p_outputopts,
 311                                               inp->in6p_moptions,
 312                                               &inp->in6p_route,
 313                                               &inp->in6p_laddr, &error);
 314                 if (*plocal_addr6 == 0) {
 315                         if (error == 0)
 316                                 error = EADDRNOTAVAIL;
 317                         return(error);
 318                 }
 319                 /*
 320                  * Don't do pcblookup call here; return interface in
 321                  * plocal_addr6
 322                  * and exit to caller, that will do the lookup.
 323                  */
 324         }
 325
 326         if (inp->in6p_route.ro_rt)
 327                 ifp = inp->in6p_route.ro_rt->rt_ifp;
 328
 329         return(0);
 330 }
 331
 332 /*
 333  * Outer subroutine:
 334  * Connect from a socket to a specified address.
 335  * Both address and port must be specified in argument sin.
 336  * If don't have a local address for this socket yet,
 337  * then pick one.
 338  */
 339 int
 340 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 341 {
 342         struct in6_addr *addr6;
 343         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 344         int error;
 345
 346         /*
 347          * Call inner routine, to assign local interface address.
 348          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 349          */
 350         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 351                 return(error);
 352
 353         if (in6_pcblookup_hash(inp->inp_cpcbinfo, &sin6->sin6_addr,
 354                                sin6->sin6_port,
 355                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 356                               ? addr6 : &inp->in6p_laddr,
 357                               inp->inp_lport, 0, NULL) != NULL) {
 358                 return (EADDRINUSE);
 359         }
 360         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 361                 if (inp->inp_lport == 0) {
 362                         error = in6_pcbbind(inp, (struct sockaddr *)0, td);
 363                         if (error)
 364                                 return (error);
 365                 }
 366                 inp->in6p_laddr = *addr6;
 367         }
 368         inp->in6p_faddr = sin6->sin6_addr;
 369         inp->inp_fport = sin6->sin6_port;
 370         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 371         inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
 372         if (inp->in6p_flags & IN6P_AUTOFLOWLABEL)
 373                 inp->in6p_flowinfo |=
 374                     (htonl(ip6_flow_seq++) & IPV6_FLOWLABEL_MASK);
 375
 376         in_pcbinsconnhash(inp);
 377         return (0);
 378 }
 379
 380 #if 0
 381 /*
 382  * Return an IPv6 address, which is the most appropriate for given
 383  * destination and user specified options.
 384  * If necessary, this function lookups the routing table and return
 385  * an entry to the caller for later use.
 386  */
 387 struct in6_addr *
 388 in6_selectsrc(struct sockaddr_in6 *dstsock, struct ip6_pktopts *opts,
 389               struct ip6_moptions *mopts, struct route_in6 *ro,
 390               struct in6_addr *laddr, int *errorp)
 391 {
 392         struct in6_addr *dst;
 393         struct in6_ifaddr *ia6 = 0;
 394         struct in6_pktinfo *pi = NULL;
 395
 396         dst = &dstsock->sin6_addr;
 397         *errorp = 0;
 398
 399         /*
 400          * If the source address is explicitly specified by the caller,
 401          * use it.
 402          */
 403         if (opts && (pi = opts->ip6po_pktinfo) &&
 404             !IN6_IS_ADDR_UNSPECIFIED(&pi->ipi6_addr))
 405                 return(&pi->ipi6_addr);
 406
 407         /*
 408          * If the source address is not specified but the socket(if any)
 409          * is already bound, use the bound address.
 410          */
 411         if (laddr && !IN6_IS_ADDR_UNSPECIFIED(laddr))
 412                 return(laddr);
 413
 414         /*
 415          * If the caller doesn't specify the source address but
 416          * the outgoing interface, use an address associated with
 417          * the interface.
 418          */
 419         if (pi && pi->ipi6_ifindex) {
 420                 /* XXX boundary check is assumed to be already done. */
 421                 ia6 = in6_ifawithscope(ifindex2ifnet[pi->ipi6_ifindex],
 422                                        dst);
 423                 if (ia6 == 0) {
 424                         *errorp = EADDRNOTAVAIL;
 425                         return(0);
 426                 }
 427                 return(&satosin6(&ia6->ia_addr)->sin6_addr);
 428         }
 429
 430         /*
 431          * If the destination address is a link-local unicast address or
 432          * a multicast address, and if the outgoing interface is specified
 433          * by the sin6_scope_id filed, use an address associated with the
 434          * interface.
 435          * XXX: We're now trying to define more specific semantics of
 436          *      sin6_scope_id field, so this part will be rewritten in
 437          *      the near future.
 438          */
 439         if ((IN6_IS_ADDR_LINKLOCAL(dst) || IN6_IS_ADDR_MULTICAST(dst)) &&
 440             dstsock->sin6_scope_id) {
 441                 /*
 442                  * I'm not sure if boundary check for scope_id is done
 443                  * somewhere...
 444                  */
 445                 if (dstsock->sin6_scope_id < 0 ||
 446                     if_index < dstsock->sin6_scope_id) {
 447                         *errorp = ENXIO; /* XXX: better error? */
 448                         return(0);
 449                 }
 450                 ia6 = in6_ifawithscope(ifindex2ifnet[dstsock->sin6_scope_id],
 451                                        dst);
 452                 if (ia6 == 0) {
 453                         *errorp = EADDRNOTAVAIL;
 454                         return(0);
 455                 }
 456                 return(&satosin6(&ia6->ia_addr)->sin6_addr);
 457         }
 458
 459         /*
 460          * If the destination address is a multicast address and
 461          * the outgoing interface for the address is specified
 462          * by the caller, use an address associated with the interface.
 463          * There is a sanity check here; if the destination has node-local
 464          * scope, the outgoing interfacde should be a loopback address.
 465          * Even if the outgoing interface is not specified, we also
 466          * choose a loopback interface as the outgoing interface.
 467          */
 468         if (IN6_IS_ADDR_MULTICAST(dst)) {
 469                 struct ifnet *ifp = mopts ? mopts->im6o_multicast_ifp : NULL;
 470
 471                 if (ifp == NULL && IN6_IS_ADDR_MC_NODELOCAL(dst)) {
 472                         ifp = &loif[0];
 473                 }
 474
 475                 if (ifp) {
 476                         ia6 = in6_ifawithscope(ifp, dst);
 477                         if (ia6 == 0) {
 478                                 *errorp = EADDRNOTAVAIL;
 479                                 return(0);
 480                         }
 481                         return(&ia6->ia_addr.sin6_addr);
 482                 }
 483         }
 484
 485         /*
 486          * If the next hop address for the packet is specified
 487          * by caller, use an address associated with the route
 488          * to the next hop.
 489          */
 490         {
 491                 struct sockaddr_in6 *sin6_next;
 492                 struct rtentry *rt;
 493
 494                 if (opts && opts->ip6po_nexthop) {
 495                         sin6_next = satosin6(opts->ip6po_nexthop);
 496                         rt = nd6_lookup(&sin6_next->sin6_addr, 1, NULL);
 497                         if (rt) {
 498                                 ia6 = in6_ifawithscope(rt->rt_ifp, dst);
 499                                 if (ia6 == 0)
 500                                         ia6 = ifatoia6(rt->rt_ifa);
 501                         }
 502                         if (ia6 == 0) {
 503                                 *errorp = EADDRNOTAVAIL;
 504                                 return(0);
 505                         }
 506                         return(&satosin6(&ia6->ia_addr)->sin6_addr);
 507                 }
 508         }
 509
 510         /*
 511          * If route is known or can be allocated now,
 512          * our src addr is taken from the i/f, else punt.
 513          */
 514         if (ro) {
 515                 if (ro->ro_rt &&
 516                     !IN6_ARE_ADDR_EQUAL(&satosin6(&ro->ro_dst)->sin6_addr, dst)) {
 517                         RTFREE(ro->ro_rt);
 518                         ro->ro_rt = NULL;
 519                 }
 520                 if (ro->ro_rt == NULL || ro->ro_rt->rt_ifp == NULL) {
 521                         struct sockaddr_in6 *dst6;
 522
 523                         /* No route yet, so try to acquire one */
 524                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in6));
 525                         dst6 = &ro->ro_dst;
 526                         dst6->sin6_family = AF_INET6;
 527                         dst6->sin6_len = sizeof(struct sockaddr_in6);
 528                         dst6->sin6_addr = *dst;
 529                         if (IN6_IS_ADDR_MULTICAST(dst)) {
 530                                 ro->ro_rt =
 531                                   rtpurelookup((struct sockaddr *)&ro->ro_dst);
 532                         } else {
 533                                 rtalloc((struct route *)ro);
 534                         }
 535                 }
 536
 537                 /*
 538                  * in_pcbconnect() checks out IFF_LOOPBACK to skip using
 539                  * the address. But we don't know why it does so.
 540                  * It is necessary to ensure the scope even for lo0
 541                  * so doesn't check out IFF_LOOPBACK.
 542                  */
 543
 544                 if (ro->ro_rt) {
 545                         ia6 = in6_ifawithscope(ro->ro_rt->rt_ifa->ifa_ifp, dst);
 546                         if (ia6 == 0) /* xxx scope error ?*/
 547                                 ia6 = ifatoia6(ro->ro_rt->rt_ifa);
 548                 }
 549                 if (ia6 == 0) {
 550                         *errorp = EHOSTUNREACH; /* no route */
 551                         return(0);
 552                 }
 553                 return(&satosin6(&ia6->ia_addr)->sin6_addr);
 554         }
 555
 556         *errorp = EADDRNOTAVAIL;
 557         return(0);
 558 }
 559
 560 /*
 561  * Default hop limit selection. The precedence is as follows:
 562  * 1. Hoplimit valued specified via ioctl.
 563  * 2. (If the outgoing interface is detected) the current
 564  *     hop limit of the interface specified by router advertisement.
 565  * 3. The system default hoplimit.
 566 */
 567 int
 568 in6_selecthlim(struct in6pcb *in6p, struct ifnet *ifp)
 569 {
 570         if (in6p && in6p->in6p_hops >= 0)
 571                 return(in6p->in6p_hops);
 572         else if (ifp)
 573                 return(ND_IFINFO(ifp)->chlim);
 574         else
 575                 return(ip6_defhlim);
 576 }
 577 #endif
 578
 579 void
 580 in6_pcbdisconnect(struct inpcb *inp)
 581 {
 582         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 583         inp->inp_fport = 0;
 584         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 585         inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
 586         in_pcbremconnhash(inp);
 587         if (inp->inp_socket->so_state & SS_NOFDREF)
 588                 in6_pcbdetach(inp);
 589 }
 590
 591 void
 592 in6_pcbdetach(struct inpcb *inp)
 593 {
 594         struct socket *so = inp->inp_socket;
 595         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 596
 597 #ifdef IPSEC
 598         if (inp->in6p_sp != NULL)
 599                 ipsec6_delete_pcbpolicy(inp);
 600 #endif /* IPSEC */
 601         inp->inp_gencnt = ++ipi->ipi_gencnt;
 602         in_pcbremlists(inp);
 603         so->so_pcb = NULL;
 604         sofree(so);
 605
 606         if (inp->in6p_options)
 607                 m_freem(inp->in6p_options);
 608         ip6_freepcbopts(inp->in6p_outputopts);
 609         ip6_freemoptions(inp->in6p_moptions);
 610         if (inp->in6p_route.ro_rt)
 611                 rtfree(inp->in6p_route.ro_rt);
 612         /* Check and free IPv4 related resources in case of mapped addr */
 613         if (inp->inp_options)
 614                 m_free(inp->inp_options);
 615         ip_freemoptions(inp->inp_moptions);
 616
 617         inp->inp_vflag = 0;
 618         zfree(ipi->ipi_zone, inp);
 619 }
 620
 621 /*
 622  * The calling convention of in6_setsockaddr() and in6_setpeeraddr() was
 623  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 624  * in struct pr_usrreqs, so that protocols can just reference then directly
 625  * without the need for a wrapper function.  The socket must have a valid
 626  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 627  * except through a kernel programming error, so it is acceptable to panic
 628  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 629  * because there actually /is/ a programming error somewhere... XXX)
 630  */
 631 int
 632 in6_setsockaddr(struct socket *so, struct sockaddr **nam)
 633 {
 634         struct inpcb *inp;
 635         struct sockaddr_in6 *sin6;
 636
 637         /*
 638          * Do the malloc first in case it blocks.
 639          */
 640         MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6, M_SONAME, M_WAITOK);
 641         bzero(sin6, sizeof *sin6);
 642         sin6->sin6_family = AF_INET6;
 643         sin6->sin6_len = sizeof(*sin6);
 644
 645         crit_enter();
 646         inp = so->so_pcb;
 647         if (!inp) {
 648                 crit_exit();
 649                 free(sin6, M_SONAME);
 650                 return EINVAL;
 651         }
 652         sin6->sin6_port = inp->inp_lport;
 653         sin6->sin6_addr = inp->in6p_laddr;
 654         crit_exit();
 655         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 656                 sin6->sin6_scope_id = ntohs(sin6->sin6_addr.s6_addr16[1]);
 657         else
 658                 sin6->sin6_scope_id = 0;        /*XXX*/
 659         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 660                 sin6->sin6_addr.s6_addr16[1] = 0;
 661
 662         *nam = (struct sockaddr *)sin6;
 663         return 0;
 664 }
 665
 666 int
 667 in6_setpeeraddr(struct socket *so, struct sockaddr **nam)
 668 {
 669         struct inpcb *inp;
 670         struct sockaddr_in6 *sin6;
 671
 672         /*
 673          * Do the malloc first in case it blocks.
 674          */
 675         MALLOC(sin6, struct sockaddr_in6 *, sizeof(*sin6), M_SONAME, M_WAITOK);
 676         bzero((caddr_t)sin6, sizeof (*sin6));
 677         sin6->sin6_family = AF_INET6;
 678         sin6->sin6_len = sizeof(struct sockaddr_in6);
 679
 680         crit_enter();
 681         inp = so->so_pcb;
 682         if (!inp) {
 683                 crit_exit();
 684                 free(sin6, M_SONAME);
 685                 return EINVAL;
 686         }
 687         sin6->sin6_port = inp->inp_fport;
 688         sin6->sin6_addr = inp->in6p_faddr;
 689         crit_exit();
 690         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 691                 sin6->sin6_scope_id = ntohs(sin6->sin6_addr.s6_addr16[1]);
 692         else
 693                 sin6->sin6_scope_id = 0;        /*XXX*/
 694         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 695                 sin6->sin6_addr.s6_addr16[1] = 0;
 696
 697         *nam = (struct sockaddr *)sin6;
 698         return 0;
 699 }
 700
 701 int
 702 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 703 {
 704         struct  inpcb *inp = so->so_pcb;
 705         int     error;
 706
 707         if (inp == NULL)
 708                 return EINVAL;
 709         if (inp->inp_vflag & INP_IPV4) {
 710                 error = in_setsockaddr(so, nam);
 711                 if (error == 0)
 712                         in6_sin_2_v4mapsin6_in_sock(nam);
 713         } else
 714         /* scope issues will be handled in in6_setsockaddr(). */
 715         error = in6_setsockaddr(so, nam);
 716
 717         return error;
 718 }
 719
 720 int
 721 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 722 {
 723         struct  inpcb *inp = so->so_pcb;
 724         int     error;
 725
 726         if (inp == NULL)
 727                 return EINVAL;
 728         if (inp->inp_vflag & INP_IPV4) {
 729                 error = in_setpeeraddr(so, nam);
 730                 if (error == 0)
 731                         in6_sin_2_v4mapsin6_in_sock(nam);
 732         } else
 733         /* scope issues will be handled in in6_setpeeraddr(). */
 734         error = in6_setpeeraddr(so, nam);
 735
 736         return error;
 737 }
 738
 739 /*
 740  * Pass some notification to all connections of a protocol
 741  * associated with address dst.  The local address and/or port numbers
 742  * may be specified to limit the search.  The "usual action" will be
 743  * taken, depending on the ctlinput cmd.  The caller must filter any
 744  * cmds that are uninteresting (e.g., no error in the map).
 745  * Call the protocol specific routine (if any) to report
 746  * any errors for each matching socket.
 747  *
 748  * Must be called under crit_enter().
 749  */
 750 void
 751 in6_pcbnotify(struct inpcbhead *head, struct sockaddr *dst, in_port_t fport,
 752               const struct sockaddr *src, in_port_t lport, int cmd, int arg,
 753               void (*notify) (struct inpcb *, int))
 754 {
 755         struct inpcb *inp, *ninp;
 756         struct sockaddr_in6 sa6_src, *sa6_dst;
 757         u_int32_t flowinfo;
 758
 759         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 760                 return;
 761
 762         sa6_dst = (struct sockaddr_in6 *)dst;
 763         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 764                 return;
 765
 766         /*
 767          * note that src can be NULL when we get notify by local fragmentation.
 768          */
 769         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 770         flowinfo = sa6_src.sin6_flowinfo;
 771
 772         /*
 773          * Redirects go to all references to the destination,
 774          * and use in6_rtchange to invalidate the route cache.
 775          * Dead host indications: also use in6_rtchange to invalidate
 776          * the cache, and deliver the error to all the sockets.
 777          * Otherwise, if we have knowledge of the local port and address,
 778          * deliver only to that socket.
 779          */
 780         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 781                 fport = 0;
 782                 lport = 0;
 783                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 784
 785                 if (cmd != PRC_HOSTDEAD)
 786                         notify = in6_rtchange;
 787         }
 788         if (cmd != PRC_MSGSIZE)
 789                 arg = inet6ctlerrmap[cmd];
 790         crit_enter();
 791         for (inp = LIST_FIRST(head); inp != NULL; inp = ninp) {
 792                 ninp = LIST_NEXT(inp, inp_list);
 793
 794                 if (inp->inp_flags & INP_PLACEMARKER)
 795                         continue;
 796
 797                 if ((inp->inp_vflag & INP_IPV6) == 0)
 798                         continue;
 799
 800                 /*
 801                  * Detect if we should notify the error. If no source and
 802                  * destination ports are specifed, but non-zero flowinfo and
 803                  * local address match, notify the error. This is the case
 804                  * when the error is delivered with an encrypted buffer
 805                  * by ESP. Otherwise, just compare addresses and ports
 806                  * as usual.
 807                  */
 808                 if (lport == 0 && fport == 0 && flowinfo &&
 809                     inp->inp_socket != NULL &&
 810                     flowinfo == (inp->in6p_flowinfo & IPV6_FLOWLABEL_MASK) &&
 811                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 812                         goto do_notify;
 813                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 814                                              &sa6_dst->sin6_addr) ||
 815                          inp->inp_socket == 0 ||
 816                          (lport && inp->inp_lport != lport) ||
 817                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 818                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 819                                               &sa6_src.sin6_addr)) ||
 820                          (fport && inp->inp_fport != fport))
 821                         continue;
 822
 823 do_notify:
 824                 if (notify)
 825                         (*notify)(inp, arg);
 826         }
 827         crit_exit();
 828 }
 829
 830 /*
 831  * Lookup a PCB based on the local address and port.
 832  */
 833 struct inpcb *
 834 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 835                     u_int lport_arg, int wild_okay)
 836 {
 837         struct inpcb *inp;
 838         int matchwild = 3, wildcard;
 839         u_short lport = lport_arg;
 840         struct inpcbporthead *porthash;
 841         struct inpcbport *phd;
 842         struct inpcb *match = NULL;
 843
 844         /*
 845          * Best fit PCB lookup.
 846          *
 847          * First see if this local port is in use by looking on the
 848          * port hash list.
 849          */
 850         porthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(lport,
 851             pcbinfo->porthashmask)];
 852         LIST_FOREACH(phd, porthash, phd_hash) {
 853                 if (phd->phd_port == lport)
 854                         break;
 855         }
 856         if (phd != NULL) {
 857                 /*
 858                  * Port is in use by one or more PCBs. Look for best
 859                  * fit.
 860                  */
 861                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 862                         wildcard = 0;
 863                         if ((inp->inp_vflag & INP_IPV6) == 0)
 864                                 continue;
 865                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 866                                 wildcard++;
 867                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 868                                 if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 869                                         wildcard++;
 870                                 else if (!IN6_ARE_ADDR_EQUAL(
 871                                         &inp->in6p_laddr, laddr))
 872                                         continue;
 873                         } else {
 874                                 if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 875                                         wildcard++;
 876                         }
 877                         if (wildcard && !wild_okay)
 878                                 continue;
 879                         if (wildcard < matchwild) {
 880                                 match = inp;
 881                                 if (wildcard == 0)
 882                                         break;
 883                                 else
 884                                         matchwild = wildcard;
 885                         }
 886                 }
 887         }
 888         return (match);
 889 }
 890
 891 void
 892 in6_pcbpurgeif0(struct in6pcb *head, struct ifnet *ifp)
 893 {
 894         struct in6pcb *in6p;
 895         struct ip6_moptions *im6o;
 896         struct in6_multi_mship *imm, *nimm;
 897
 898         for (in6p = head; in6p != NULL; in6p = LIST_NEXT(in6p, inp_list)) {
 899                 if (in6p->in6p_flags & INP_PLACEMARKER)
 900                         continue;
 901                 im6o = in6p->in6p_moptions;
 902                 if ((in6p->inp_vflag & INP_IPV6) &&
 903                     im6o) {
 904                         /*
 905                          * Unselect the outgoing interface if it is being
 906                          * detached.
 907                          */
 908                         if (im6o->im6o_multicast_ifp == ifp)
 909                                 im6o->im6o_multicast_ifp = NULL;
 910
 911                         /*
 912                          * Drop multicast group membership if we joined
 913                          * through the interface being detached.
 914                          * XXX controversial - is it really legal for kernel
 915                          * to force this?
 916                          */
 917                         for (imm = im6o->im6o_memberships.lh_first;
 918                              imm != NULL; imm = nimm) {
 919                                 nimm = imm->i6mm_chain.le_next;
 920                                 if (imm->i6mm_maddr->in6m_ifp == ifp) {
 921                                         LIST_REMOVE(imm, i6mm_chain);
 922                                         in6_delmulti(imm->i6mm_maddr);
 923                                         free(imm, M_IPMADDR);
 924                                 }
 925                         }
 926                 }
 927         }
 928 }
 929
 930 /*
 931  * Check for alternatives when higher level complains
 932  * about service problems.  For now, invalidate cached
 933  * routing information.  If the route was created dynamically
 934  * (by a redirect), time to try a default gateway again.
 935  */
 936 void
 937 in6_losing(struct inpcb *in6p)
 938 {
 939         struct rtentry *rt;
 940         struct rt_addrinfo info;
 941
 942         if ((rt = in6p->in6p_route.ro_rt) != NULL) {
 943                 bzero((caddr_t)&info, sizeof(info));
 944                 info.rti_flags = rt->rt_flags;
 945                 info.rti_info[RTAX_DST] = rt_key(rt);
 946                 info.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
 947                 info.rti_info[RTAX_NETMASK] = rt_mask(rt);
 948                 rt_missmsg(RTM_LOSING, &info, rt->rt_flags, 0);
 949                 if (rt->rt_flags & RTF_DYNAMIC)
 950                         rtrequest1_global(RTM_DELETE, &info, NULL, NULL);
 951                 in6p->in6p_route.ro_rt = NULL;
 952                 rtfree(rt);
 953                 /*
 954                  * A new route can be allocated
 955                  * the next time output is attempted.
 956                  */
 957         }
 958 }
 959
 960 /*
 961  * After a routing change, flush old routing
 962  * and allocate a (hopefully) better one.
 963  */
 964 void
 965 in6_rtchange(struct inpcb *inp, int errno)
 966 {
 967         if (inp->in6p_route.ro_rt) {
 968                 rtfree(inp->in6p_route.ro_rt);
 969                 inp->in6p_route.ro_rt = 0;
 970                 /*
 971                  * A new route can be allocated the next time
 972                  * output is attempted.
 973                  */
 974         }
 975 }
 976
 977 /*
 978  * Lookup PCB in hash list.
 979  */
 980 struct inpcb *
 981 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
 982                    u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
 983                    int wildcard, struct ifnet *ifp)
 984 {
 985         struct inpcbhead *head;
 986         struct inpcb *inp;
 987         u_short fport = fport_arg, lport = lport_arg;
 988         int faith;
 989
 990         if (faithprefix_p != NULL)
 991                 faith = (*faithprefix_p)(laddr);
 992         else
 993                 faith = 0;
 994
 995         /*
 996          * First look for an exact match.
 997          */
 998         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr->s6_addr32[3] /* XXX */,
 999                                               fport,
1000                                               laddr->s6_addr32[3], /* XXX JH */
1001                                               lport,
1002                                               pcbinfo->hashmask)];
1003         LIST_FOREACH(inp, head, inp_hash) {
1004                 if ((inp->inp_vflag & INP_IPV6) == 0)
1005                         continue;
1006                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1007                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1008                     inp->inp_fport == fport &&
1009                     inp->inp_lport == lport) {
1010                         /*
1011                          * Found.
1012                          */
1013                         return (inp);
1014                 }
1015         }
1016         if (wildcard) {
1017                 struct inpcontainerhead *chead;
1018                 struct inpcontainer *ic;
1019                 struct inpcb *local_wild = NULL;
1020
1021                 chead = &pcbinfo->wildcardhashbase[INP_PCBWILDCARDHASH(lport,
1022                     pcbinfo->wildcardhashmask)];
1023                 LIST_FOREACH(ic, chead, ic_list) {
1024                         inp = ic->ic_inp;
1025
1026                         if (!(inp->inp_vflag & INP_IPV6))
1027                                 continue;
1028                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
1029                             inp->inp_lport == lport) {
1030                                 if (faith && (inp->inp_flags & INP_FAITH) == 0)
1031                                         continue;
1032                                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
1033                                                        laddr))
1034                                         return (inp);
1035                                 else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
1036                                         local_wild = inp;
1037                         }
1038                 }
1039                 return (local_wild);
1040         }
1041
1042         /*
1043          * Not found.
1044          */
1045         return (NULL);
1046 }
1047
1048 void
1049 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m)
1050 {
1051         struct ip6_hdr *ip;
1052
1053         ip = mtod(m, struct ip6_hdr *);
1054         bzero(sin6, sizeof(*sin6));
1055         sin6->sin6_len = sizeof(*sin6);
1056         sin6->sin6_family = AF_INET6;
1057         sin6->sin6_addr = ip->ip6_src;
1058         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
1059                 sin6->sin6_addr.s6_addr16[1] = 0;
1060         sin6->sin6_scope_id =
1061                 (m->m_pkthdr.rcvif && IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
1062                 ? m->m_pkthdr.rcvif->if_index : 0;
1063
1064         return;
1065 }