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 #include "opt_mac.h"
  70
  71 #include <sys/param.h>
  72 #include <sys/systm.h>
  73 #include <sys/malloc.h>
  74 #include <sys/mbuf.h>
  75 #include <sys/domain.h>
  76 #include <sys/protosw.h>
  77 #include <sys/socket.h>
  78 #include <sys/socketvar.h>
  79 #include <sys/sockio.h>
  80 #include <sys/errno.h>
  81 #include <sys/time.h>
  82 #include <sys/priv.h>
  83 #include <sys/proc.h>
  84 #include <sys/jail.h>
  85
  86 #include <vm/uma.h>
  87
  88 #include <net/if.h>
  89 #include <net/if_types.h>
  90 #include <net/route.h>
  91
  92 #include <netinet/in.h>
  93 #include <netinet/in_var.h>
  94 #include <netinet/in_systm.h>
  95 #include <netinet/tcp_var.h>
  96 #include <netinet/ip6.h>
  97 #include <netinet/ip_var.h>
  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 (!in6_ifaddr) /* 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, 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 *ia = NULL;
 161
 162                         sin6->sin6_port = 0;            /* yech... */
 163                         if ((ia = ifa_ifwithaddr((struct sockaddr *)sin6)) == 0)
 164                                 return (EADDRNOTAVAIL);
 165
 166                         /*
 167                          * XXX: bind to an anycast address might accidentally
 168                          * cause sending a packet with anycast source address.
 169                          * We should allow to bind to a deprecated address, since
 170                          * the application dares to use it.
 171                          */
 172                         if (ia &&
 173                             ((struct in6_ifaddr *)ia)->ia6_flags &
 174                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 175                                 return (EADDRNOTAVAIL);
 176                         }
 177                 }
 178                 if (lport) {
 179                         struct inpcb *t;
 180
 181                         /* GROSS */
 182                         if (ntohs(lport) <= ipport_reservedhigh &&
 183                             ntohs(lport) >= ipport_reservedlow &&
 184                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
 185                             0))
 186                                 return (EACCES);
 187                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 188                             priv_check_cred(inp->inp_cred,
 189                             PRIV_NETINET_REUSEPORT, 0) != 0) {
 190                                 t = in6_pcblookup_local(pcbinfo,
 191                                     &sin6->sin6_addr, lport,
 192                                     INPLOOKUP_WILDCARD, cred);
 193                                 if (t &&
 194                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 195                                     (so->so_type != SOCK_STREAM ||
 196                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 197                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 198                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 199                                      (t->inp_socket->so_options & SO_REUSEPORT)
 200                                       == 0) && (inp->inp_cred->cr_uid !=
 201                                      t->inp_cred->cr_uid))
 202                                         return (EADDRINUSE);
 203                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 204                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 205                                         struct sockaddr_in sin;
 206
 207                                         in6_sin6_2_sin(&sin, sin6);
 208                                         t = in_pcblookup_local(pcbinfo,
 209                                             sin.sin_addr, lport,
 210                                             INPLOOKUP_WILDCARD, cred);
 211                                         if (t &&
 212                                             ((t->inp_flags &
 213                                               INP_TIMEWAIT) == 0) &&
 214                                             (so->so_type != SOCK_STREAM ||
 215                                              ntohl(t->inp_faddr.s_addr) ==
 216                                               INADDR_ANY) &&
 217                                             (inp->inp_cred->cr_uid !=
 218                                              t->inp_cred->cr_uid))
 219                                                 return (EADDRINUSE);
 220                                 }
 221                         }
 222                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 223                             lport, wild, cred);
 224                         if (t && (reuseport & ((t->inp_flags & INP_TIMEWAIT) ?
 225                             intotw(t)->tw_so_options :
 226                             t->inp_socket->so_options)) == 0)
 227                                 return (EADDRINUSE);
 228                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 229                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 230                                 struct sockaddr_in sin;
 231
 232                                 in6_sin6_2_sin(&sin, sin6);
 233                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 234                                     lport, wild, cred);
 235                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 236                                         if ((reuseport &
 237                                             intotw(t)->tw_so_options) == 0 &&
 238                                             (ntohl(t->inp_laddr.s_addr) !=
 239                                              INADDR_ANY || ((inp->inp_vflag &
 240                                              INP_IPV6PROTO) ==
 241                                              (t->inp_vflag & INP_IPV6PROTO))))
 242                                                 return (EADDRINUSE);
 243                                 }
 244                                 else if (t &&
 245                                     (reuseport & t->inp_socket->so_options)
 246                                     == 0 && (ntohl(t->inp_laddr.s_addr) !=
 247                                     INADDR_ANY || INP_SOCKAF(so) ==
 248                                      INP_SOCKAF(t->inp_socket)))
 249                                         return (EADDRINUSE);
 250                         }
 251                 }
 252                 inp->in6p_laddr = sin6->sin6_addr;
 253         }
 254         if (lport == 0) {
 255                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0)
 256                         return (error);
 257         } else {
 258                 inp->inp_lport = lport;
 259                 if (in_pcbinshash(inp) != 0) {
 260                         inp->in6p_laddr = in6addr_any;
 261                         inp->inp_lport = 0;
 262                         return (EAGAIN);
 263                 }
 264         }
 265         return (0);
 266 }
 267
 268 /*
 269  *   Transform old in6_pcbconnect() into an inner subroutine for new
 270  *   in6_pcbconnect(): Do some validity-checking on the remote
 271  *   address (in mbuf 'nam') and then determine local host address
 272  *   (i.e., which interface) to use to access that remote host.
 273  *
 274  *   This preserves definition of in6_pcbconnect(), while supporting a
 275  *   slightly different version for T/TCP.  (This is more than
 276  *   a bit of a kludge, but cleaning up the internal interfaces would
 277  *   have forced minor changes in every protocol).
 278  */
 279 int
 280 in6_pcbladdr(register struct inpcb *inp, struct sockaddr *nam,
 281     struct in6_addr **plocal_addr6)
 282 {
 283         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 284         int error = 0;
 285         struct ifnet *ifp = NULL;
 286         int scope_ambiguous = 0;
 287
 288         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 289         INP_WLOCK_ASSERT(inp);
 290
 291         if (nam->sa_len != sizeof (*sin6))
 292                 return (EINVAL);
 293         if (sin6->sin6_family != AF_INET6)
 294                 return (EAFNOSUPPORT);
 295         if (sin6->sin6_port == 0)
 296                 return (EADDRNOTAVAIL);
 297
 298         if (sin6->sin6_scope_id == 0 && !ip6_use_defzone)
 299                 scope_ambiguous = 1;
 300         if ((error = sa6_embedscope(sin6, ip6_use_defzone)) != 0)
 301                 return(error);
 302
 303         if (in6_ifaddr) {
 304                 /*
 305                  * If the destination address is UNSPECIFIED addr,
 306                  * use the loopback addr, e.g ::1.
 307                  */
 308                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 309                         sin6->sin6_addr = in6addr_loopback;
 310         }
 311         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 312                 return (error);
 313
 314         /*
 315          * XXX: in6_selectsrc might replace the bound local address
 316          * with the address specified by setsockopt(IPV6_PKTINFO).
 317          * Is it the intended behavior?
 318          */
 319         *plocal_addr6 = in6_selectsrc(sin6, inp->in6p_outputopts,
 320                                       inp, NULL,
 321                                       inp->inp_cred,
 322                                       &ifp, &error);
 323         if (ifp && scope_ambiguous &&
 324             (error = in6_setscope(&sin6->sin6_addr, ifp, NULL)) != 0) {
 325                 return(error);
 326         }
 327
 328         if (*plocal_addr6 == NULL) {
 329                 if (error == 0)
 330                         error = EADDRNOTAVAIL;
 331                 return (error);
 332         }
 333         /*
 334          * Don't do pcblookup call here; return interface in
 335          * plocal_addr6
 336          * and exit to caller, that will do the lookup.
 337          */
 338
 339         return (0);
 340 }
 341
 342 /*
 343  * Outer subroutine:
 344  * Connect from a socket to a specified address.
 345  * Both address and port must be specified in argument sin.
 346  * If don't have a local address for this socket yet,
 347  * then pick one.
 348  */
 349 int
 350 in6_pcbconnect(register struct inpcb *inp, struct sockaddr *nam,
 351     struct ucred *cred)
 352 {
 353         struct in6_addr *addr6;
 354         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 355         int error;
 356
 357         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 358         INP_WLOCK_ASSERT(inp);
 359
 360         /*
 361          * Call inner routine, to assign local interface address.
 362          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 363          */
 364         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 365                 return (error);
 366
 367         if (in6_pcblookup_hash(inp->inp_pcbinfo, &sin6->sin6_addr,
 368                                sin6->sin6_port,
 369                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 370                               ? addr6 : &inp->in6p_laddr,
 371                               inp->inp_lport, 0, NULL) != NULL) {
 372                 return (EADDRINUSE);
 373         }
 374         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 375                 if (inp->inp_lport == 0) {
 376                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 377                         if (error)
 378                                 return (error);
 379                 }
 380                 inp->in6p_laddr = *addr6;
 381         }
 382         inp->in6p_faddr = sin6->sin6_addr;
 383         inp->inp_fport = sin6->sin6_port;
 384         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 385         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 386         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 387                 inp->inp_flow |=
 388                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 389
 390         in_pcbrehash(inp);
 391
 392         return (0);
 393 }
 394
 395 void
 396 in6_pcbdisconnect(struct inpcb *inp)
 397 {
 398
 399         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 400         INP_WLOCK_ASSERT(inp);
 401
 402         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 403         inp->inp_fport = 0;
 404         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 405         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 406         in_pcbrehash(inp);
 407 }
 408
 409 struct sockaddr *
 410 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 411 {
 412         struct sockaddr_in6 *sin6;
 413
 414         MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6, M_SONAME, M_WAITOK);
 415         bzero(sin6, sizeof *sin6);
 416         sin6->sin6_family = AF_INET6;
 417         sin6->sin6_len = sizeof(*sin6);
 418         sin6->sin6_port = port;
 419         sin6->sin6_addr = *addr_p;
 420         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 421
 422         return (struct sockaddr *)sin6;
 423 }
 424
 425 struct sockaddr *
 426 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 427 {
 428         struct sockaddr_in sin;
 429         struct sockaddr_in6 *sin6_p;
 430
 431         bzero(&sin, sizeof sin);
 432         sin.sin_family = AF_INET;
 433         sin.sin_len = sizeof(sin);
 434         sin.sin_port = port;
 435         sin.sin_addr = *addr_p;
 436
 437         MALLOC(sin6_p, struct sockaddr_in6 *, sizeof *sin6_p, M_SONAME,
 438                 M_WAITOK);
 439         in6_sin_2_v4mapsin6(&sin, sin6_p);
 440
 441         return (struct sockaddr *)sin6_p;
 442 }
 443
 444 int
 445 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 446 {
 447         register struct inpcb *inp;
 448         struct in6_addr addr;
 449         in_port_t port;
 450
 451         inp = sotoinpcb(so);
 452         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 453
 454         INP_RLOCK(inp);
 455         port = inp->inp_lport;
 456         addr = inp->in6p_laddr;
 457         INP_RUNLOCK(inp);
 458
 459         *nam = in6_sockaddr(port, &addr);
 460         return 0;
 461 }
 462
 463 int
 464 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 465 {
 466         struct inpcb *inp;
 467         struct in6_addr addr;
 468         in_port_t port;
 469
 470         inp = sotoinpcb(so);
 471         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 472
 473         INP_RLOCK(inp);
 474         port = inp->inp_fport;
 475         addr = inp->in6p_faddr;
 476         INP_RUNLOCK(inp);
 477
 478         *nam = in6_sockaddr(port, &addr);
 479         return 0;
 480 }
 481
 482 int
 483 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 484 {
 485         struct  inpcb *inp;
 486         int     error;
 487
 488         inp = sotoinpcb(so);
 489         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 490
 491         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 492                 error = in_getsockaddr(so, nam);
 493                 if (error == 0)
 494                         in6_sin_2_v4mapsin6_in_sock(nam);
 495         } else {
 496                 /* scope issues will be handled in in6_getsockaddr(). */
 497                 error = in6_getsockaddr(so, nam);
 498         }
 499
 500         return error;
 501 }
 502
 503 int
 504 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 505 {
 506         struct  inpcb *inp;
 507         int     error;
 508
 509         inp = sotoinpcb(so);
 510         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 511
 512         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 513                 error = in_getpeeraddr(so, nam);
 514                 if (error == 0)
 515                         in6_sin_2_v4mapsin6_in_sock(nam);
 516         } else
 517         /* scope issues will be handled in in6_getpeeraddr(). */
 518         error = in6_getpeeraddr(so, nam);
 519
 520         return error;
 521 }
 522
 523 /*
 524  * Pass some notification to all connections of a protocol
 525  * associated with address dst.  The local address and/or port numbers
 526  * may be specified to limit the search.  The "usual action" will be
 527  * taken, depending on the ctlinput cmd.  The caller must filter any
 528  * cmds that are uninteresting (e.g., no error in the map).
 529  * Call the protocol specific routine (if any) to report
 530  * any errors for each matching socket.
 531  */
 532 void
 533 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 534     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 535     int cmd, void *cmdarg,
 536     struct inpcb *(*notify)(struct inpcb *, int))
 537 {
 538         struct inpcb *inp, *inp_temp;
 539         struct sockaddr_in6 sa6_src, *sa6_dst;
 540         u_short fport = fport_arg, lport = lport_arg;
 541         u_int32_t flowinfo;
 542         int errno;
 543
 544         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 545                 return;
 546
 547         sa6_dst = (struct sockaddr_in6 *)dst;
 548         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 549                 return;
 550
 551         /*
 552          * note that src can be NULL when we get notify by local fragmentation.
 553          */
 554         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 555         flowinfo = sa6_src.sin6_flowinfo;
 556
 557         /*
 558          * Redirects go to all references to the destination,
 559          * and use in6_rtchange to invalidate the route cache.
 560          * Dead host indications: also use in6_rtchange to invalidate
 561          * the cache, and deliver the error to all the sockets.
 562          * Otherwise, if we have knowledge of the local port and address,
 563          * deliver only to that socket.
 564          */
 565         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 566                 fport = 0;
 567                 lport = 0;
 568                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 569
 570                 if (cmd != PRC_HOSTDEAD)
 571                         notify = in6_rtchange;
 572         }
 573         errno = inet6ctlerrmap[cmd];
 574         INP_INFO_WLOCK(pcbinfo);
 575         LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 576                 INP_WLOCK(inp);
 577                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 578                         INP_WUNLOCK(inp);
 579                         continue;
 580                 }
 581
 582                 /*
 583                  * If the error designates a new path MTU for a destination
 584                  * and the application (associated with this socket) wanted to
 585                  * know the value, notify. Note that we notify for all
 586                  * disconnected sockets if the corresponding application
 587                  * wanted. This is because some UDP applications keep sending
 588                  * sockets disconnected.
 589                  * XXX: should we avoid to notify the value to TCP sockets?
 590                  */
 591                 if (cmd == PRC_MSGSIZE && (inp->inp_flags & IN6P_MTU) != 0 &&
 592                     (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 593                      IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr))) {
 594                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 595                                         (u_int32_t *)cmdarg);
 596                 }
 597
 598                 /*
 599                  * Detect if we should notify the error. If no source and
 600                  * destination ports are specifed, but non-zero flowinfo and
 601                  * local address match, notify the error. This is the case
 602                  * when the error is delivered with an encrypted buffer
 603                  * by ESP. Otherwise, just compare addresses and ports
 604                  * as usual.
 605                  */
 606                 if (lport == 0 && fport == 0 && flowinfo &&
 607                     inp->inp_socket != NULL &&
 608                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 609                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 610                         goto do_notify;
 611                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 612                                              &sa6_dst->sin6_addr) ||
 613                          inp->inp_socket == 0 ||
 614                          (lport && inp->inp_lport != lport) ||
 615                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 616                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 617                                               &sa6_src.sin6_addr)) ||
 618                          (fport && inp->inp_fport != fport)) {
 619                         INP_WUNLOCK(inp);
 620                         continue;
 621                 }
 622
 623           do_notify:
 624                 if (notify) {
 625                         if ((*notify)(inp, errno))
 626                                 INP_WUNLOCK(inp);
 627                 } else
 628                         INP_WUNLOCK(inp);
 629         }
 630         INP_INFO_WUNLOCK(pcbinfo);
 631 }
 632
 633 /*
 634  * Lookup a PCB based on the local address and port.
 635  */
 636 struct inpcb *
 637 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 638     u_short lport, int wild_okay, struct ucred *cred)
 639 {
 640         register struct inpcb *inp;
 641         int matchwild = 3, wildcard;
 642
 643         INP_INFO_WLOCK_ASSERT(pcbinfo);
 644
 645         if (!wild_okay) {
 646                 struct inpcbhead *head;
 647                 /*
 648                  * Look for an unconnected (wildcard foreign addr) PCB that
 649                  * matches the local address and port we're looking for.
 650                  */
 651                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 652                     0, pcbinfo->ipi_hashmask)];
 653                 LIST_FOREACH(inp, head, inp_hash) {
 654                         /* XXX inp locking */
 655                         if ((inp->inp_vflag & INP_IPV6) == 0)
 656                                 continue;
 657                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 658                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 659                             inp->inp_lport == lport) {
 660                                 /* Found. */
 661                                 if (cred == NULL ||
 662                                     inp->inp_cred->cr_prison == cred->cr_prison)
 663                                         return (inp);
 664                         }
 665                 }
 666                 /*
 667                  * Not found.
 668                  */
 669                 return (NULL);
 670         } else {
 671                 struct inpcbporthead *porthash;
 672                 struct inpcbport *phd;
 673                 struct inpcb *match = NULL;
 674                 /*
 675                  * Best fit PCB lookup.
 676                  *
 677                  * First see if this local port is in use by looking on the
 678                  * port hash list.
 679                  */
 680                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 681                     pcbinfo->ipi_porthashmask)];
 682                 LIST_FOREACH(phd, porthash, phd_hash) {
 683                         if (phd->phd_port == lport)
 684                                 break;
 685                 }
 686                 if (phd != NULL) {
 687                         /*
 688                          * Port is in use by one or more PCBs. Look for best
 689                          * fit.
 690                          */
 691                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 692                                 wildcard = 0;
 693                                 if (cred != NULL &&
 694                                     inp->inp_cred->cr_prison != cred->cr_prison)
 695                                         continue;
 696                                 /* XXX inp locking */
 697                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 698                                         continue;
 699                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 700                                         wildcard++;
 701                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 702                                         &inp->in6p_laddr)) {
 703                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 704                                                 wildcard++;
 705                                         else if (!IN6_ARE_ADDR_EQUAL(
 706                                             &inp->in6p_laddr, laddr))
 707                                                 continue;
 708                                 } else {
 709                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 710                                                 wildcard++;
 711                                 }
 712                                 if (wildcard < matchwild) {
 713                                         match = inp;
 714                                         matchwild = wildcard;
 715                                         if (matchwild == 0)
 716                                                 break;
 717                                 }
 718                         }
 719                 }
 720                 return (match);
 721         }
 722 }
 723
 724 void
 725 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 726 {
 727         struct inpcb *in6p;
 728         struct ip6_moptions *im6o;
 729         struct in6_multi_mship *imm, *nimm;
 730
 731         INP_INFO_RLOCK(pcbinfo);
 732         LIST_FOREACH(in6p, pcbinfo->ipi_listhead, inp_list) {
 733                 INP_WLOCK(in6p);
 734                 im6o = in6p->in6p_moptions;
 735                 if ((in6p->inp_vflag & INP_IPV6) &&
 736                     im6o) {
 737                         /*
 738                          * Unselect the outgoing interface if it is being
 739                          * detached.
 740                          */
 741                         if (im6o->im6o_multicast_ifp == ifp)
 742                                 im6o->im6o_multicast_ifp = NULL;
 743
 744                         /*
 745                          * Drop multicast group membership if we joined
 746                          * through the interface being detached.
 747                          * XXX controversial - is it really legal for kernel
 748                          * to force this?
 749                          */
 750                         for (imm = im6o->im6o_memberships.lh_first;
 751                              imm != NULL; imm = nimm) {
 752                                 nimm = imm->i6mm_chain.le_next;
 753                                 if (imm->i6mm_maddr->in6m_ifp == ifp) {
 754                                         LIST_REMOVE(imm, i6mm_chain);
 755                                         in6_delmulti(imm->i6mm_maddr);
 756                                         free(imm, M_IP6MADDR);
 757                                 }
 758                         }
 759                 }
 760                 INP_WUNLOCK(in6p);
 761         }
 762         INP_INFO_RUNLOCK(pcbinfo);
 763 }
 764
 765 /*
 766  * Check for alternatives when higher level complains
 767  * about service problems.  For now, invalidate cached
 768  * routing information.  If the route was created dynamically
 769  * (by a redirect), time to try a default gateway again.
 770  */
 771 void
 772 in6_losing(struct inpcb *in6p)
 773 {
 774
 775         /*
 776          * We don't store route pointers in the routing table anymore
 777          */
 778         return;
 779 }
 780
 781 /*
 782  * After a routing change, flush old routing
 783  * and allocate a (hopefully) better one.
 784  */
 785 struct inpcb *
 786 in6_rtchange(struct inpcb *inp, int errno)
 787 {
 788         /*
 789          * We don't store route pointers in the routing table anymore
 790          */
 791         return inp;
 792 }
 793
 794 /*
 795  * Lookup PCB in hash list.
 796  */
 797 struct inpcb *
 798 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
 799     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg, int wildcard,
 800     struct ifnet *ifp)
 801 {
 802         struct inpcbhead *head;
 803         struct inpcb *inp, *tmpinp;
 804         u_short fport = fport_arg, lport = lport_arg;
 805         int faith;
 806
 807         INP_INFO_LOCK_ASSERT(pcbinfo);
 808
 809         if (faithprefix_p != NULL)
 810                 faith = (*faithprefix_p)(laddr);
 811         else
 812                 faith = 0;
 813
 814         /*
 815          * First look for an exact match.
 816          */
 817         tmpinp = NULL;
 818         head = &pcbinfo->ipi_hashbase[
 819             INP_PCBHASH(faddr->s6_addr32[3] /* XXX */, lport, fport,
 820             pcbinfo->ipi_hashmask)];
 821         LIST_FOREACH(inp, head, inp_hash) {
 822                 /* XXX inp locking */
 823                 if ((inp->inp_vflag & INP_IPV6) == 0)
 824                         continue;
 825                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 826                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 827                     inp->inp_fport == fport &&
 828                     inp->inp_lport == lport) {
 829                         /*
 830                          * XXX We should be able to directly return
 831                          * the inp here, without any checks.
 832                          * Well unless both bound with SO_REUSEPORT?
 833                          */
 834                         if (jailed(inp->inp_cred))
 835                                 return (inp);
 836                         if (tmpinp == NULL)
 837                                 tmpinp = inp;
 838                 }
 839         }
 840         if (tmpinp != NULL)
 841                 return (tmpinp);
 842
 843         /*
 844          * Then look for a wildcard match, if requested.
 845          */
 846         if (wildcard == INPLOOKUP_WILDCARD) {
 847                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 848                 struct inpcb *jail_wild = NULL;
 849                 int injail;
 850
 851                 /*
 852                  * Order of socket selection - we always prefer jails.
 853                  *      1. jailed, non-wild.
 854                  *      2. jailed, wild.
 855                  *      3. non-jailed, non-wild.
 856                  *      4. non-jailed, wild.
 857                  */
 858                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 859                     0, pcbinfo->ipi_hashmask)];
 860                 LIST_FOREACH(inp, head, inp_hash) {
 861                         /* XXX inp locking */
 862                         if ((inp->inp_vflag & INP_IPV6) == 0)
 863                                 continue;
 864
 865                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 866                             inp->inp_lport != lport) {
 867                                 continue;
 868                         }
 869
 870                         /* XXX inp locking */
 871                         if (faith && (inp->inp_flags & INP_FAITH) == 0)
 872                                 continue;
 873
 874                         injail = jailed(inp->inp_cred);
 875                         if (injail) {
 876                                 if (prison_check_ip6(inp->inp_cred,
 877                                     laddr) != 0)
 878                                         continue;
 879                         } else {
 880                                 if (local_exact != NULL)
 881                                         continue;
 882                         }
 883
 884                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
 885                                 if (injail)
 886                                         return (inp);
 887                                 else
 888                                         local_exact = inp;
 889                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 890                                 if (injail)
 891                                         jail_wild = inp;
 892                                 else
 893                                         local_wild = inp;
 894                         }
 895                 } /* LIST_FOREACH */
 896
 897                 if (jail_wild != NULL)
 898                         return (jail_wild);
 899                 if (local_exact != NULL)
 900                         return (local_exact);
 901                 if (local_wild != NULL)
 902                         return (local_wild);
 903         } /* if (wildcard == INPLOOKUP_WILDCARD) */
 904
 905         /*
 906          * Not found.
 907          */
 908         return (NULL);
 909 }
 910
 911 void
 912 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m)
 913 {
 914         struct ip6_hdr *ip;
 915
 916         ip = mtod(m, struct ip6_hdr *);
 917         bzero(sin6, sizeof(*sin6));
 918         sin6->sin6_len = sizeof(*sin6);
 919         sin6->sin6_family = AF_INET6;
 920         sin6->sin6_addr = ip->ip6_src;
 921
 922         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
 923
 924         return;
 925 }