sys/netinet6/in6_pcb.c

   1 /*-
   2  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the project nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  30  */
  31
  32 /*-
  33  * Copyright (c) 1982, 1986, 1991, 1993
  34  *      The Regents of the University of California.  All rights reserved.
  35  *
  36  * Redistribution and use in source and binary forms, with or without
  37  * modification, are permitted provided that the following conditions
  38  * are met:
  39  * 1. Redistributions of source code must retain the above copyright
  40  *    notice, this list of conditions and the following disclaimer.
  41  * 2. Redistributions in binary form must reproduce the above copyright
  42  *    notice, this list of conditions and the following disclaimer in the
  43  *    documentation and/or other materials provided with the distribution.
  44  * 4. Neither the name of the University nor the names of its contributors
  45  *    may be used to endorse or promote products derived from this software
  46  *    without specific prior written permission.
  47  *
  48  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  49  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  50  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  51  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  52  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  53  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  54  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  55  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  56  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  57  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  58  * SUCH DAMAGE.
  59  *
  60  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  61  */
  62
  63 #include <sys/cdefs.h>
  64 __FBSDID("$FreeBSD$");
  65
  66 #include "opt_inet.h"
  67 #include "opt_inet6.h"
  68 #include "opt_ipsec.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/malloc.h>
  73 #include <sys/mbuf.h>
  74 #include <sys/domain.h>
  75 #include <sys/protosw.h>
  76 #include <sys/socket.h>
  77 #include <sys/socketvar.h>
  78 #include <sys/sockio.h>
  79 #include <sys/errno.h>
  80 #include <sys/time.h>
  81 #include <sys/priv.h>
  82 #include <sys/proc.h>
  83 #include <sys/jail.h>
  84
  85 #include <vm/uma.h>
  86
  87 #include <net/if.h>
  88 #include <net/if_types.h>
  89 #include <net/route.h>
  90
  91 #include <netinet/in.h>
  92 #include <netinet/in_var.h>
  93 #include <netinet/in_systm.h>
  94 #include <netinet/tcp_var.h>
  95 #include <netinet/ip6.h>
  96 #include <netinet/ip_var.h>
  97
  98 #include <netinet6/ip6_var.h>
  99 #include <netinet6/nd6.h>
 100 #include <netinet/in_pcb.h>
 101 #include <netinet6/in6_pcb.h>
 102 #include <netinet6/scope6_var.h>
 103
 104 #include <security/mac/mac_framework.h>
 105
 106 struct  in6_addr zeroin6_addr;
 107
 108 int
 109 in6_pcbbind(register struct inpcb *inp, struct sockaddr *nam,
 110     struct ucred *cred)
 111 {
 112         struct socket *so = inp->inp_socket;
 113         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 114         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 115         u_short lport = 0;
 116         int error, wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 117
 118         INP_INFO_WLOCK_ASSERT(pcbinfo);
 119         INP_WLOCK_ASSERT(inp);
 120
 121         if (TAILQ_EMPTY(&V_in6_ifaddrhead))     /* XXX broken! */
 122                 return (EADDRNOTAVAIL);
 123         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 124                 return (EINVAL);
 125         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 126                 wild = INPLOOKUP_WILDCARD;
 127         if (nam == NULL) {
 128                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 129                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 130                         return (error);
 131         } else {
 132                 sin6 = (struct sockaddr_in6 *)nam;
 133                 if (nam->sa_len != sizeof(*sin6))
 134                         return (EINVAL);
 135                 /*
 136                  * family check.
 137                  */
 138                 if (nam->sa_family != AF_INET6)
 139                         return (EAFNOSUPPORT);
 140
 141                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 142                         return(error);
 143
 144                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 145                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 146                         return (error);
 147
 148                 lport = sin6->sin6_port;
 149                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 150                         /*
 151                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 152                          * allow compepte duplication of binding if
 153                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 154                          * and a multicast address is bound on both
 155                          * new and duplicated sockets.
 156                          */
 157                         if (so->so_options & SO_REUSEADDR)
 158                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 159                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 160                         struct ifaddr *ifa;
 161
 162                         sin6->sin6_port = 0;            /* yech... */
 163                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 164                             NULL &&
 165                             (inp->inp_flags & INP_BINDANY) == 0) {
 166                                 return (EADDRNOTAVAIL);
 167                         }
 168
 169                         /*
 170                          * XXX: bind to an anycast address might accidentally
 171                          * cause sending a packet with anycast source address.
 172                          * We should allow to bind to a deprecated address, since
 173                          * the application dares to use it.
 174                          */
 175                         if (ifa != NULL &&
 176                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 177                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 178                                 ifa_free(ifa);
 179                                 return (EADDRNOTAVAIL);
 180                         }
 181                         if (ifa != NULL)
 182                                 ifa_free(ifa);
 183                 }
 184                 if (lport) {
 185                         struct inpcb *t;
 186
 187                         /* GROSS */
 188                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 189                             ntohs(lport) >= V_ipport_reservedlow &&
 190                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
 191                             0))
 192                                 return (EACCES);
 193                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 194                             priv_check_cred(inp->inp_cred,
 195                             PRIV_NETINET_REUSEPORT, 0) != 0) {
 196                                 t = in6_pcblookup_local(pcbinfo,
 197                                     &sin6->sin6_addr, lport,
 198                                     INPLOOKUP_WILDCARD, cred);
 199                                 if (t &&
 200                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 201                                     (so->so_type != SOCK_STREAM ||
 202                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 203                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 204                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 205                                      (t->inp_socket->so_options & SO_REUSEPORT)
 206                                       == 0) && (inp->inp_cred->cr_uid !=
 207                                      t->inp_cred->cr_uid))
 208                                         return (EADDRINUSE);
 209                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 210                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 211                                         struct sockaddr_in sin;
 212
 213                                         in6_sin6_2_sin(&sin, sin6);
 214                                         t = in_pcblookup_local(pcbinfo,
 215                                             sin.sin_addr, lport,
 216                                             INPLOOKUP_WILDCARD, cred);
 217                                         if (t &&
 218                                             ((t->inp_flags &
 219                                               INP_TIMEWAIT) == 0) &&
 220                                             (so->so_type != SOCK_STREAM ||
 221                                              ntohl(t->inp_faddr.s_addr) ==
 222                                               INADDR_ANY) &&
 223                                             (inp->inp_cred->cr_uid !=
 224                                              t->inp_cred->cr_uid))
 225                                                 return (EADDRINUSE);
 226                                 }
 227                         }
 228                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 229                             lport, wild, cred);
 230                         if (t && (reuseport & ((t->inp_flags & INP_TIMEWAIT) ?
 231                             intotw(t)->tw_so_options :
 232                             t->inp_socket->so_options)) == 0)
 233                                 return (EADDRINUSE);
 234                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 235                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 236                                 struct sockaddr_in sin;
 237
 238                                 in6_sin6_2_sin(&sin, sin6);
 239                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 240                                     lport, wild, cred);
 241                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 242                                         if ((reuseport &
 243                                             intotw(t)->tw_so_options) == 0 &&
 244                                             (ntohl(t->inp_laddr.s_addr) !=
 245                                              INADDR_ANY || ((inp->inp_vflag &
 246                                              INP_IPV6PROTO) ==
 247                                              (t->inp_vflag & INP_IPV6PROTO))))
 248                                                 return (EADDRINUSE);
 249                                 }
 250                                 else if (t &&
 251                                     (reuseport & t->inp_socket->so_options)
 252                                     == 0 && (ntohl(t->inp_laddr.s_addr) !=
 253                                     INADDR_ANY || INP_SOCKAF(so) ==
 254                                      INP_SOCKAF(t->inp_socket)))
 255                                         return (EADDRINUSE);
 256                         }
 257                 }
 258                 inp->in6p_laddr = sin6->sin6_addr;
 259         }
 260         if (lport == 0) {
 261                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0)
 262                         return (error);
 263         } else {
 264                 inp->inp_lport = lport;
 265                 if (in_pcbinshash(inp) != 0) {
 266                         inp->in6p_laddr = in6addr_any;
 267                         inp->inp_lport = 0;
 268                         return (EAGAIN);
 269                 }
 270         }
 271         return (0);
 272 }
 273
 274 /*
 275  *   Transform old in6_pcbconnect() into an inner subroutine for new
 276  *   in6_pcbconnect(): Do some validity-checking on the remote
 277  *   address (in mbuf 'nam') and then determine local host address
 278  *   (i.e., which interface) to use to access that remote host.
 279  *
 280  *   This preserves definition of in6_pcbconnect(), while supporting a
 281  *   slightly different version for T/TCP.  (This is more than
 282  *   a bit of a kludge, but cleaning up the internal interfaces would
 283  *   have forced minor changes in every protocol).
 284  */
 285 int
 286 in6_pcbladdr(register struct inpcb *inp, struct sockaddr *nam,
 287     struct in6_addr *plocal_addr6)
 288 {
 289         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 290         int error = 0;
 291         struct ifnet *ifp = NULL;
 292         int scope_ambiguous = 0;
 293         struct in6_addr in6a;
 294
 295         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 296         INP_WLOCK_ASSERT(inp);
 297
 298         if (nam->sa_len != sizeof (*sin6))
 299                 return (EINVAL);
 300         if (sin6->sin6_family != AF_INET6)
 301                 return (EAFNOSUPPORT);
 302         if (sin6->sin6_port == 0)
 303                 return (EADDRNOTAVAIL);
 304
 305         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 306                 scope_ambiguous = 1;
 307         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 308                 return(error);
 309
 310         if (!TAILQ_EMPTY(&V_in6_ifaddrhead)) {
 311                 /*
 312                  * If the destination address is UNSPECIFIED addr,
 313                  * use the loopback addr, e.g ::1.
 314                  */
 315                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 316                         sin6->sin6_addr = in6addr_loopback;
 317         }
 318         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 319                 return (error);
 320
 321         error = in6_selectsrc(sin6, inp->in6p_outputopts,
 322             inp, NULL, inp->inp_cred, &ifp, &in6a);
 323         if (error)
 324                 return (error);
 325
 326         if (ifp && scope_ambiguous &&
 327             (error = in6_setscope(&sin6->sin6_addr, ifp, NULL)) != 0) {
 328                 return(error);
 329         }
 330
 331         /*
 332          * Do not update this earlier, in case we return with an error.
 333          *
 334          * XXX: this in6_selectsrc result might replace the bound local
 335          * address with the address specified by setsockopt(IPV6_PKTINFO).
 336          * Is it the intended behavior?
 337          */
 338         *plocal_addr6 = in6a;
 339
 340         /*
 341          * Don't do pcblookup call here; return interface in
 342          * plocal_addr6
 343          * and exit to caller, that will do the lookup.
 344          */
 345
 346         return (0);
 347 }
 348
 349 /*
 350  * Outer subroutine:
 351  * Connect from a socket to a specified address.
 352  * Both address and port must be specified in argument sin.
 353  * If don't have a local address for this socket yet,
 354  * then pick one.
 355  */
 356 int
 357 in6_pcbconnect(register struct inpcb *inp, struct sockaddr *nam,
 358     struct ucred *cred)
 359 {
 360         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 361         struct in6_addr addr6;
 362         int error;
 363
 364         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 365         INP_WLOCK_ASSERT(inp);
 366
 367         /*
 368          * Call inner routine, to assign local interface address.
 369          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 370          */
 371         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 372                 return (error);
 373
 374         if (in6_pcblookup_hash(inp->inp_pcbinfo, &sin6->sin6_addr,
 375                                sin6->sin6_port,
 376                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 377                               ? &addr6 : &inp->in6p_laddr,
 378                               inp->inp_lport, 0, NULL) != NULL) {
 379                 return (EADDRINUSE);
 380         }
 381         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 382                 if (inp->inp_lport == 0) {
 383                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 384                         if (error)
 385                                 return (error);
 386                 }
 387                 inp->in6p_laddr = addr6;
 388         }
 389         inp->in6p_faddr = sin6->sin6_addr;
 390         inp->inp_fport = sin6->sin6_port;
 391         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 392         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 393         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 394                 inp->inp_flow |=
 395                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 396
 397         in_pcbrehash(inp);
 398
 399         return (0);
 400 }
 401
 402 void
 403 in6_pcbdisconnect(struct inpcb *inp)
 404 {
 405
 406         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 407         INP_WLOCK_ASSERT(inp);
 408
 409         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 410         inp->inp_fport = 0;
 411         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 412         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 413         in_pcbrehash(inp);
 414 }
 415
 416 struct sockaddr *
 417 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 418 {
 419         struct sockaddr_in6 *sin6;
 420
 421         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 422         bzero(sin6, sizeof *sin6);
 423         sin6->sin6_family = AF_INET6;
 424         sin6->sin6_len = sizeof(*sin6);
 425         sin6->sin6_port = port;
 426         sin6->sin6_addr = *addr_p;
 427         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 428
 429         return (struct sockaddr *)sin6;
 430 }
 431
 432 struct sockaddr *
 433 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 434 {
 435         struct sockaddr_in sin;
 436         struct sockaddr_in6 *sin6_p;
 437
 438         bzero(&sin, sizeof sin);
 439         sin.sin_family = AF_INET;
 440         sin.sin_len = sizeof(sin);
 441         sin.sin_port = port;
 442         sin.sin_addr = *addr_p;
 443
 444         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 445                 M_WAITOK);
 446         in6_sin_2_v4mapsin6(&sin, sin6_p);
 447
 448         return (struct sockaddr *)sin6_p;
 449 }
 450
 451 int
 452 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 453 {
 454         register struct inpcb *inp;
 455         struct in6_addr addr;
 456         in_port_t port;
 457
 458         inp = sotoinpcb(so);
 459         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 460
 461         INP_RLOCK(inp);
 462         port = inp->inp_lport;
 463         addr = inp->in6p_laddr;
 464         INP_RUNLOCK(inp);
 465
 466         *nam = in6_sockaddr(port, &addr);
 467         return 0;
 468 }
 469
 470 int
 471 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 472 {
 473         struct inpcb *inp;
 474         struct in6_addr addr;
 475         in_port_t port;
 476
 477         inp = sotoinpcb(so);
 478         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 479
 480         INP_RLOCK(inp);
 481         port = inp->inp_fport;
 482         addr = inp->in6p_faddr;
 483         INP_RUNLOCK(inp);
 484
 485         *nam = in6_sockaddr(port, &addr);
 486         return 0;
 487 }
 488
 489 int
 490 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 491 {
 492         struct  inpcb *inp;
 493         int     error;
 494
 495         inp = sotoinpcb(so);
 496         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 497
 498         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 499                 error = in_getsockaddr(so, nam);
 500                 if (error == 0)
 501                         in6_sin_2_v4mapsin6_in_sock(nam);
 502         } else {
 503                 /* scope issues will be handled in in6_getsockaddr(). */
 504                 error = in6_getsockaddr(so, nam);
 505         }
 506
 507         return error;
 508 }
 509
 510 int
 511 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 512 {
 513         struct  inpcb *inp;
 514         int     error;
 515
 516         inp = sotoinpcb(so);
 517         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 518
 519         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 520                 error = in_getpeeraddr(so, nam);
 521                 if (error == 0)
 522                         in6_sin_2_v4mapsin6_in_sock(nam);
 523         } else
 524         /* scope issues will be handled in in6_getpeeraddr(). */
 525         error = in6_getpeeraddr(so, nam);
 526
 527         return error;
 528 }
 529
 530 /*
 531  * Pass some notification to all connections of a protocol
 532  * associated with address dst.  The local address and/or port numbers
 533  * may be specified to limit the search.  The "usual action" will be
 534  * taken, depending on the ctlinput cmd.  The caller must filter any
 535  * cmds that are uninteresting (e.g., no error in the map).
 536  * Call the protocol specific routine (if any) to report
 537  * any errors for each matching socket.
 538  */
 539 void
 540 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 541     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 542     int cmd, void *cmdarg,
 543     struct inpcb *(*notify)(struct inpcb *, int))
 544 {
 545         struct inpcb *inp, *inp_temp;
 546         struct sockaddr_in6 sa6_src, *sa6_dst;
 547         u_short fport = fport_arg, lport = lport_arg;
 548         u_int32_t flowinfo;
 549         int errno;
 550
 551         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 552                 return;
 553
 554         sa6_dst = (struct sockaddr_in6 *)dst;
 555         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 556                 return;
 557
 558         /*
 559          * note that src can be NULL when we get notify by local fragmentation.
 560          */
 561         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 562         flowinfo = sa6_src.sin6_flowinfo;
 563
 564         /*
 565          * Redirects go to all references to the destination,
 566          * and use in6_rtchange to invalidate the route cache.
 567          * Dead host indications: also use in6_rtchange to invalidate
 568          * the cache, and deliver the error to all the sockets.
 569          * Otherwise, if we have knowledge of the local port and address,
 570          * deliver only to that socket.
 571          */
 572         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 573                 fport = 0;
 574                 lport = 0;
 575                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 576
 577                 if (cmd != PRC_HOSTDEAD)
 578                         notify = in6_rtchange;
 579         }
 580         errno = inet6ctlerrmap[cmd];
 581         INP_INFO_WLOCK(pcbinfo);
 582         LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 583                 INP_WLOCK(inp);
 584                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 585                         INP_WUNLOCK(inp);
 586                         continue;
 587                 }
 588
 589                 /*
 590                  * If the error designates a new path MTU for a destination
 591                  * and the application (associated with this socket) wanted to
 592                  * know the value, notify. Note that we notify for all
 593                  * disconnected sockets if the corresponding application
 594                  * wanted. This is because some UDP applications keep sending
 595                  * sockets disconnected.
 596                  * XXX: should we avoid to notify the value to TCP sockets?
 597                  */
 598                 if (cmd == PRC_MSGSIZE && (inp->inp_flags & IN6P_MTU) != 0 &&
 599                     (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 600                      IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr))) {
 601                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 602                                         (u_int32_t *)cmdarg);
 603                 }
 604
 605                 /*
 606                  * Detect if we should notify the error. If no source and
 607                  * destination ports are specifed, but non-zero flowinfo and
 608                  * local address match, notify the error. This is the case
 609                  * when the error is delivered with an encrypted buffer
 610                  * by ESP. Otherwise, just compare addresses and ports
 611                  * as usual.
 612                  */
 613                 if (lport == 0 && fport == 0 && flowinfo &&
 614                     inp->inp_socket != NULL &&
 615                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 616                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 617                         goto do_notify;
 618                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 619                                              &sa6_dst->sin6_addr) ||
 620                          inp->inp_socket == 0 ||
 621                          (lport && inp->inp_lport != lport) ||
 622                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 623                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 624                                               &sa6_src.sin6_addr)) ||
 625                          (fport && inp->inp_fport != fport)) {
 626                         INP_WUNLOCK(inp);
 627                         continue;
 628                 }
 629
 630           do_notify:
 631                 if (notify) {
 632                         if ((*notify)(inp, errno))
 633                                 INP_WUNLOCK(inp);
 634                 } else
 635                         INP_WUNLOCK(inp);
 636         }
 637         INP_INFO_WUNLOCK(pcbinfo);
 638 }
 639
 640 /*
 641  * Lookup a PCB based on the local address and port.
 642  */
 643 struct inpcb *
 644 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 645     u_short lport, int wild_okay, struct ucred *cred)
 646 {
 647         register struct inpcb *inp;
 648         int matchwild = 3, wildcard;
 649
 650         INP_INFO_WLOCK_ASSERT(pcbinfo);
 651
 652         if (!wild_okay) {
 653                 struct inpcbhead *head;
 654                 /*
 655                  * Look for an unconnected (wildcard foreign addr) PCB that
 656                  * matches the local address and port we're looking for.
 657                  */
 658                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 659                     0, pcbinfo->ipi_hashmask)];
 660                 LIST_FOREACH(inp, head, inp_hash) {
 661                         /* XXX inp locking */
 662                         if ((inp->inp_vflag & INP_IPV6) == 0)
 663                                 continue;
 664                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 665                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 666                             inp->inp_lport == lport) {
 667                                 /* Found. */
 668                                 if (cred == NULL ||
 669                                     prison_equal_ip6(cred->cr_prison,
 670                                         inp->inp_cred->cr_prison))
 671                                         return (inp);
 672                         }
 673                 }
 674                 /*
 675                  * Not found.
 676                  */
 677                 return (NULL);
 678         } else {
 679                 struct inpcbporthead *porthash;
 680                 struct inpcbport *phd;
 681                 struct inpcb *match = NULL;
 682                 /*
 683                  * Best fit PCB lookup.
 684                  *
 685                  * First see if this local port is in use by looking on the
 686                  * port hash list.
 687                  */
 688                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 689                     pcbinfo->ipi_porthashmask)];
 690                 LIST_FOREACH(phd, porthash, phd_hash) {
 691                         if (phd->phd_port == lport)
 692                                 break;
 693                 }
 694                 if (phd != NULL) {
 695                         /*
 696                          * Port is in use by one or more PCBs. Look for best
 697                          * fit.
 698                          */
 699                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 700                                 wildcard = 0;
 701                                 if (cred != NULL &&
 702                                     !prison_equal_ip6(cred->cr_prison,
 703                                         inp->inp_cred->cr_prison))
 704                                         continue;
 705                                 /* XXX inp locking */
 706                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 707                                         continue;
 708                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 709                                         wildcard++;
 710                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 711                                         &inp->in6p_laddr)) {
 712                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 713                                                 wildcard++;
 714                                         else if (!IN6_ARE_ADDR_EQUAL(
 715                                             &inp->in6p_laddr, laddr))
 716                                                 continue;
 717                                 } else {
 718                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 719                                                 wildcard++;
 720                                 }
 721                                 if (wildcard < matchwild) {
 722                                         match = inp;
 723                                         matchwild = wildcard;
 724                                         if (matchwild == 0)
 725                                                 break;
 726                                 }
 727                         }
 728                 }
 729                 return (match);
 730         }
 731 }
 732
 733 void
 734 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 735 {
 736         struct inpcb *in6p;
 737         struct ip6_moptions *im6o;
 738         int i, gap;
 739
 740         INP_INFO_RLOCK(pcbinfo);
 741         LIST_FOREACH(in6p, pcbinfo->ipi_listhead, inp_list) {
 742                 INP_WLOCK(in6p);
 743                 im6o = in6p->in6p_moptions;
 744                 if ((in6p->inp_vflag & INP_IPV6) && im6o != NULL) {
 745                         /*
 746                          * Unselect the outgoing ifp for multicast if it
 747                          * is being detached.
 748                          */
 749                         if (im6o->im6o_multicast_ifp == ifp)
 750                                 im6o->im6o_multicast_ifp = NULL;
 751                         /*
 752                          * Drop multicast group membership if we joined
 753                          * through the interface being detached.
 754                          */
 755                         gap = 0;
 756                         for (i = 0; i < im6o->im6o_num_memberships; i++) {
 757                                 if (im6o->im6o_membership[i]->in6m_ifp ==
 758                                     ifp) {
 759                                         in6_mc_leave(im6o->im6o_membership[i],
 760                                             NULL);
 761                                         gap++;
 762                                 } else if (gap != 0) {
 763                                         im6o->im6o_membership[i - gap] =
 764                                             im6o->im6o_membership[i];
 765                                 }
 766                         }
 767                         im6o->im6o_num_memberships -= gap;
 768                 }
 769                 INP_WUNLOCK(in6p);
 770         }
 771         INP_INFO_RUNLOCK(pcbinfo);
 772 }
 773
 774 /*
 775  * Check for alternatives when higher level complains
 776  * about service problems.  For now, invalidate cached
 777  * routing information.  If the route was created dynamically
 778  * (by a redirect), time to try a default gateway again.
 779  */
 780 void
 781 in6_losing(struct inpcb *in6p)
 782 {
 783
 784         /*
 785          * We don't store route pointers in the routing table anymore
 786          */
 787         return;
 788 }
 789
 790 /*
 791  * After a routing change, flush old routing
 792  * and allocate a (hopefully) better one.
 793  */
 794 struct inpcb *
 795 in6_rtchange(struct inpcb *inp, int errno)
 796 {
 797         /*
 798          * We don't store route pointers in the routing table anymore
 799          */
 800         return inp;
 801 }
 802
 803 /*
 804  * Lookup PCB in hash list.
 805  */
 806 struct inpcb *
 807 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
 808     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg, int wildcard,
 809     struct ifnet *ifp)
 810 {
 811         struct inpcbhead *head;
 812         struct inpcb *inp, *tmpinp;
 813         u_short fport = fport_arg, lport = lport_arg;
 814         int faith;
 815
 816         INP_INFO_LOCK_ASSERT(pcbinfo);
 817
 818         if (faithprefix_p != NULL)
 819                 faith = (*faithprefix_p)(laddr);
 820         else
 821                 faith = 0;
 822
 823         /*
 824          * First look for an exact match.
 825          */
 826         tmpinp = NULL;
 827         head = &pcbinfo->ipi_hashbase[
 828             INP_PCBHASH(faddr->s6_addr32[3] /* XXX */, lport, fport,
 829             pcbinfo->ipi_hashmask)];
 830         LIST_FOREACH(inp, head, inp_hash) {
 831                 /* XXX inp locking */
 832                 if ((inp->inp_vflag & INP_IPV6) == 0)
 833                         continue;
 834                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 835                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 836                     inp->inp_fport == fport &&
 837                     inp->inp_lport == lport) {
 838                         /*
 839                          * XXX We should be able to directly return
 840                          * the inp here, without any checks.
 841                          * Well unless both bound with SO_REUSEPORT?
 842                          */
 843                         if (prison_flag(inp->inp_cred, PR_IP6))
 844                                 return (inp);
 845                         if (tmpinp == NULL)
 846                                 tmpinp = inp;
 847                 }
 848         }
 849         if (tmpinp != NULL)
 850                 return (tmpinp);
 851
 852         /*
 853          * Then look for a wildcard match, if requested.
 854          */
 855         if (wildcard == INPLOOKUP_WILDCARD) {
 856                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 857                 struct inpcb *jail_wild = NULL;
 858                 int injail;
 859
 860                 /*
 861                  * Order of socket selection - we always prefer jails.
 862                  *      1. jailed, non-wild.
 863                  *      2. jailed, wild.
 864                  *      3. non-jailed, non-wild.
 865                  *      4. non-jailed, wild.
 866                  */
 867                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 868                     0, pcbinfo->ipi_hashmask)];
 869                 LIST_FOREACH(inp, head, inp_hash) {
 870                         /* XXX inp locking */
 871                         if ((inp->inp_vflag & INP_IPV6) == 0)
 872                                 continue;
 873
 874                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 875                             inp->inp_lport != lport) {
 876                                 continue;
 877                         }
 878
 879                         /* XXX inp locking */
 880                         if (faith && (inp->inp_flags & INP_FAITH) == 0)
 881                                 continue;
 882
 883                         injail = prison_flag(inp->inp_cred, PR_IP6);
 884                         if (injail) {
 885                                 if (prison_check_ip6(inp->inp_cred,
 886                                     laddr) != 0)
 887                                         continue;
 888                         } else {
 889                                 if (local_exact != NULL)
 890                                         continue;
 891                         }
 892
 893                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
 894                                 if (injail)
 895                                         return (inp);
 896                                 else
 897                                         local_exact = inp;
 898                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 899                                 if (injail)
 900                                         jail_wild = inp;
 901                                 else
 902                                         local_wild = inp;
 903                         }
 904                 } /* LIST_FOREACH */
 905
 906                 if (jail_wild != NULL)
 907                         return (jail_wild);
 908                 if (local_exact != NULL)
 909                         return (local_exact);
 910                 if (local_wild != NULL)
 911                         return (local_wild);
 912         } /* if (wildcard == INPLOOKUP_WILDCARD) */
 913
 914         /*
 915          * Not found.
 916          */
 917         return (NULL);
 918 }
 919
 920 void
 921 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m)
 922 {
 923         struct ip6_hdr *ip;
 924
 925         ip = mtod(m, struct ip6_hdr *);
 926         bzero(sin6, sizeof(*sin6));
 927         sin6->sin6_len = sizeof(*sin6);
 928         sin6->sin6_family = AF_INET6;
 929         sin6->sin6_addr = ip->ip6_src;
 930
 931         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
 932
 933         return;
 934 }