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                         /* Undo an address bind that may have occurred. */
 263                         inp->in6p_laddr = in6addr_any;
 264                         return (error);
 265                 }
 266         } else {
 267                 inp->inp_lport = lport;
 268                 if (in_pcbinshash(inp) != 0) {
 269                         inp->in6p_laddr = in6addr_any;
 270                         inp->inp_lport = 0;
 271                         return (EAGAIN);
 272                 }
 273         }
 274         return (0);
 275 }
 276
 277 /*
 278  *   Transform old in6_pcbconnect() into an inner subroutine for new
 279  *   in6_pcbconnect(): Do some validity-checking on the remote
 280  *   address (in mbuf 'nam') and then determine local host address
 281  *   (i.e., which interface) to use to access that remote host.
 282  *
 283  *   This preserves definition of in6_pcbconnect(), while supporting a
 284  *   slightly different version for T/TCP.  (This is more than
 285  *   a bit of a kludge, but cleaning up the internal interfaces would
 286  *   have forced minor changes in every protocol).
 287  */
 288 int
 289 in6_pcbladdr(register struct inpcb *inp, struct sockaddr *nam,
 290     struct in6_addr *plocal_addr6)
 291 {
 292         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 293         int error = 0;
 294         struct ifnet *ifp = NULL;
 295         int scope_ambiguous = 0;
 296         struct in6_addr in6a;
 297
 298         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 299         INP_WLOCK_ASSERT(inp);
 300
 301         if (nam->sa_len != sizeof (*sin6))
 302                 return (EINVAL);
 303         if (sin6->sin6_family != AF_INET6)
 304                 return (EAFNOSUPPORT);
 305         if (sin6->sin6_port == 0)
 306                 return (EADDRNOTAVAIL);
 307
 308         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 309                 scope_ambiguous = 1;
 310         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 311                 return(error);
 312
 313         if (!TAILQ_EMPTY(&V_in6_ifaddrhead)) {
 314                 /*
 315                  * If the destination address is UNSPECIFIED addr,
 316                  * use the loopback addr, e.g ::1.
 317                  */
 318                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 319                         sin6->sin6_addr = in6addr_loopback;
 320         }
 321         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 322                 return (error);
 323
 324         error = in6_selectsrc(sin6, inp->in6p_outputopts,
 325             inp, NULL, inp->inp_cred, &ifp, &in6a);
 326         if (error)
 327                 return (error);
 328
 329         if (ifp && scope_ambiguous &&
 330             (error = in6_setscope(&sin6->sin6_addr, ifp, NULL)) != 0) {
 331                 return(error);
 332         }
 333
 334         /*
 335          * Do not update this earlier, in case we return with an error.
 336          *
 337          * XXX: this in6_selectsrc result might replace the bound local
 338          * address with the address specified by setsockopt(IPV6_PKTINFO).
 339          * Is it the intended behavior?
 340          */
 341         *plocal_addr6 = in6a;
 342
 343         /*
 344          * Don't do pcblookup call here; return interface in
 345          * plocal_addr6
 346          * and exit to caller, that will do the lookup.
 347          */
 348
 349         return (0);
 350 }
 351
 352 /*
 353  * Outer subroutine:
 354  * Connect from a socket to a specified address.
 355  * Both address and port must be specified in argument sin.
 356  * If don't have a local address for this socket yet,
 357  * then pick one.
 358  */
 359 int
 360 in6_pcbconnect(register struct inpcb *inp, struct sockaddr *nam,
 361     struct ucred *cred)
 362 {
 363         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 364         struct in6_addr addr6;
 365         int error;
 366
 367         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 368         INP_WLOCK_ASSERT(inp);
 369
 370         /*
 371          * Call inner routine, to assign local interface address.
 372          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 373          */
 374         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 375                 return (error);
 376
 377         if (in6_pcblookup_hash(inp->inp_pcbinfo, &sin6->sin6_addr,
 378                                sin6->sin6_port,
 379                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 380                               ? &addr6 : &inp->in6p_laddr,
 381                               inp->inp_lport, 0, NULL) != NULL) {
 382                 return (EADDRINUSE);
 383         }
 384         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 385                 if (inp->inp_lport == 0) {
 386                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 387                         if (error)
 388                                 return (error);
 389                 }
 390                 inp->in6p_laddr = addr6;
 391         }
 392         inp->in6p_faddr = sin6->sin6_addr;
 393         inp->inp_fport = sin6->sin6_port;
 394         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 395         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 396         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 397                 inp->inp_flow |=
 398                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 399
 400         in_pcbrehash(inp);
 401
 402         return (0);
 403 }
 404
 405 void
 406 in6_pcbdisconnect(struct inpcb *inp)
 407 {
 408
 409         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 410         INP_WLOCK_ASSERT(inp);
 411
 412         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 413         inp->inp_fport = 0;
 414         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 415         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 416         in_pcbrehash(inp);
 417 }
 418
 419 struct sockaddr *
 420 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 421 {
 422         struct sockaddr_in6 *sin6;
 423
 424         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 425         bzero(sin6, sizeof *sin6);
 426         sin6->sin6_family = AF_INET6;
 427         sin6->sin6_len = sizeof(*sin6);
 428         sin6->sin6_port = port;
 429         sin6->sin6_addr = *addr_p;
 430         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 431
 432         return (struct sockaddr *)sin6;
 433 }
 434
 435 struct sockaddr *
 436 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 437 {
 438         struct sockaddr_in sin;
 439         struct sockaddr_in6 *sin6_p;
 440
 441         bzero(&sin, sizeof sin);
 442         sin.sin_family = AF_INET;
 443         sin.sin_len = sizeof(sin);
 444         sin.sin_port = port;
 445         sin.sin_addr = *addr_p;
 446
 447         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 448                 M_WAITOK);
 449         in6_sin_2_v4mapsin6(&sin, sin6_p);
 450
 451         return (struct sockaddr *)sin6_p;
 452 }
 453
 454 int
 455 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 456 {
 457         register struct inpcb *inp;
 458         struct in6_addr addr;
 459         in_port_t port;
 460
 461         inp = sotoinpcb(so);
 462         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 463
 464         INP_RLOCK(inp);
 465         port = inp->inp_lport;
 466         addr = inp->in6p_laddr;
 467         INP_RUNLOCK(inp);
 468
 469         *nam = in6_sockaddr(port, &addr);
 470         return 0;
 471 }
 472
 473 int
 474 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 475 {
 476         struct inpcb *inp;
 477         struct in6_addr addr;
 478         in_port_t port;
 479
 480         inp = sotoinpcb(so);
 481         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 482
 483         INP_RLOCK(inp);
 484         port = inp->inp_fport;
 485         addr = inp->in6p_faddr;
 486         INP_RUNLOCK(inp);
 487
 488         *nam = in6_sockaddr(port, &addr);
 489         return 0;
 490 }
 491
 492 int
 493 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 494 {
 495         struct  inpcb *inp;
 496         int     error;
 497
 498         inp = sotoinpcb(so);
 499         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 500
 501         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 502                 error = in_getsockaddr(so, nam);
 503                 if (error == 0)
 504                         in6_sin_2_v4mapsin6_in_sock(nam);
 505         } else {
 506                 /* scope issues will be handled in in6_getsockaddr(). */
 507                 error = in6_getsockaddr(so, nam);
 508         }
 509
 510         return error;
 511 }
 512
 513 int
 514 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 515 {
 516         struct  inpcb *inp;
 517         int     error;
 518
 519         inp = sotoinpcb(so);
 520         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 521
 522         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 523                 error = in_getpeeraddr(so, nam);
 524                 if (error == 0)
 525                         in6_sin_2_v4mapsin6_in_sock(nam);
 526         } else
 527         /* scope issues will be handled in in6_getpeeraddr(). */
 528         error = in6_getpeeraddr(so, nam);
 529
 530         return error;
 531 }
 532
 533 /*
 534  * Pass some notification to all connections of a protocol
 535  * associated with address dst.  The local address and/or port numbers
 536  * may be specified to limit the search.  The "usual action" will be
 537  * taken, depending on the ctlinput cmd.  The caller must filter any
 538  * cmds that are uninteresting (e.g., no error in the map).
 539  * Call the protocol specific routine (if any) to report
 540  * any errors for each matching socket.
 541  */
 542 void
 543 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 544     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 545     int cmd, void *cmdarg,
 546     struct inpcb *(*notify)(struct inpcb *, int))
 547 {
 548         struct inpcb *inp, *inp_temp;
 549         struct sockaddr_in6 sa6_src, *sa6_dst;
 550         u_short fport = fport_arg, lport = lport_arg;
 551         u_int32_t flowinfo;
 552         int errno;
 553
 554         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 555                 return;
 556
 557         sa6_dst = (struct sockaddr_in6 *)dst;
 558         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 559                 return;
 560
 561         /*
 562          * note that src can be NULL when we get notify by local fragmentation.
 563          */
 564         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 565         flowinfo = sa6_src.sin6_flowinfo;
 566
 567         /*
 568          * Redirects go to all references to the destination,
 569          * and use in6_rtchange to invalidate the route cache.
 570          * Dead host indications: also use in6_rtchange to invalidate
 571          * the cache, and deliver the error to all the sockets.
 572          * Otherwise, if we have knowledge of the local port and address,
 573          * deliver only to that socket.
 574          */
 575         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 576                 fport = 0;
 577                 lport = 0;
 578                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 579
 580                 if (cmd != PRC_HOSTDEAD)
 581                         notify = in6_rtchange;
 582         }
 583         errno = inet6ctlerrmap[cmd];
 584         INP_INFO_WLOCK(pcbinfo);
 585         LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 586                 INP_WLOCK(inp);
 587                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 588                         INP_WUNLOCK(inp);
 589                         continue;
 590                 }
 591
 592                 /*
 593                  * If the error designates a new path MTU for a destination
 594                  * and the application (associated with this socket) wanted to
 595                  * know the value, notify. Note that we notify for all
 596                  * disconnected sockets if the corresponding application
 597                  * wanted. This is because some UDP applications keep sending
 598                  * sockets disconnected.
 599                  * XXX: should we avoid to notify the value to TCP sockets?
 600                  */
 601                 if (cmd == PRC_MSGSIZE && (inp->inp_flags & IN6P_MTU) != 0 &&
 602                     (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 603                      IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr))) {
 604                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 605                                         (u_int32_t *)cmdarg);
 606                 }
 607
 608                 /*
 609                  * Detect if we should notify the error. If no source and
 610                  * destination ports are specifed, but non-zero flowinfo and
 611                  * local address match, notify the error. This is the case
 612                  * when the error is delivered with an encrypted buffer
 613                  * by ESP. Otherwise, just compare addresses and ports
 614                  * as usual.
 615                  */
 616                 if (lport == 0 && fport == 0 && flowinfo &&
 617                     inp->inp_socket != NULL &&
 618                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 619                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 620                         goto do_notify;
 621                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 622                                              &sa6_dst->sin6_addr) ||
 623                          inp->inp_socket == 0 ||
 624                          (lport && inp->inp_lport != lport) ||
 625                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 626                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 627                                               &sa6_src.sin6_addr)) ||
 628                          (fport && inp->inp_fport != fport)) {
 629                         INP_WUNLOCK(inp);
 630                         continue;
 631                 }
 632
 633           do_notify:
 634                 if (notify) {
 635                         if ((*notify)(inp, errno))
 636                                 INP_WUNLOCK(inp);
 637                 } else
 638                         INP_WUNLOCK(inp);
 639         }
 640         INP_INFO_WUNLOCK(pcbinfo);
 641 }
 642
 643 /*
 644  * Lookup a PCB based on the local address and port.
 645  */
 646 struct inpcb *
 647 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 648     u_short lport, int wild_okay, struct ucred *cred)
 649 {
 650         register struct inpcb *inp;
 651         int matchwild = 3, wildcard;
 652
 653         INP_INFO_WLOCK_ASSERT(pcbinfo);
 654
 655         if (!wild_okay) {
 656                 struct inpcbhead *head;
 657                 /*
 658                  * Look for an unconnected (wildcard foreign addr) PCB that
 659                  * matches the local address and port we're looking for.
 660                  */
 661                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 662                     0, pcbinfo->ipi_hashmask)];
 663                 LIST_FOREACH(inp, head, inp_hash) {
 664                         /* XXX inp locking */
 665                         if ((inp->inp_vflag & INP_IPV6) == 0)
 666                                 continue;
 667                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 668                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 669                             inp->inp_lport == lport) {
 670                                 /* Found. */
 671                                 if (cred == NULL ||
 672                                     prison_equal_ip6(cred->cr_prison,
 673                                         inp->inp_cred->cr_prison))
 674                                         return (inp);
 675                         }
 676                 }
 677                 /*
 678                  * Not found.
 679                  */
 680                 return (NULL);
 681         } else {
 682                 struct inpcbporthead *porthash;
 683                 struct inpcbport *phd;
 684                 struct inpcb *match = NULL;
 685                 /*
 686                  * Best fit PCB lookup.
 687                  *
 688                  * First see if this local port is in use by looking on the
 689                  * port hash list.
 690                  */
 691                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 692                     pcbinfo->ipi_porthashmask)];
 693                 LIST_FOREACH(phd, porthash, phd_hash) {
 694                         if (phd->phd_port == lport)
 695                                 break;
 696                 }
 697                 if (phd != NULL) {
 698                         /*
 699                          * Port is in use by one or more PCBs. Look for best
 700                          * fit.
 701                          */
 702                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 703                                 wildcard = 0;
 704                                 if (cred != NULL &&
 705                                     !prison_equal_ip6(cred->cr_prison,
 706                                         inp->inp_cred->cr_prison))
 707                                         continue;
 708                                 /* XXX inp locking */
 709                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 710                                         continue;
 711                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 712                                         wildcard++;
 713                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 714                                         &inp->in6p_laddr)) {
 715                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 716                                                 wildcard++;
 717                                         else if (!IN6_ARE_ADDR_EQUAL(
 718                                             &inp->in6p_laddr, laddr))
 719                                                 continue;
 720                                 } else {
 721                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 722                                                 wildcard++;
 723                                 }
 724                                 if (wildcard < matchwild) {
 725                                         match = inp;
 726                                         matchwild = wildcard;
 727                                         if (matchwild == 0)
 728                                                 break;
 729                                 }
 730                         }
 731                 }
 732                 return (match);
 733         }
 734 }
 735
 736 void
 737 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 738 {
 739         struct inpcb *in6p;
 740         struct ip6_moptions *im6o;
 741         int i, gap;
 742
 743         INP_INFO_RLOCK(pcbinfo);
 744         LIST_FOREACH(in6p, pcbinfo->ipi_listhead, inp_list) {
 745                 INP_WLOCK(in6p);
 746                 im6o = in6p->in6p_moptions;
 747                 if ((in6p->inp_vflag & INP_IPV6) && im6o != NULL) {
 748                         /*
 749                          * Unselect the outgoing ifp for multicast if it
 750                          * is being detached.
 751                          */
 752                         if (im6o->im6o_multicast_ifp == ifp)
 753                                 im6o->im6o_multicast_ifp = NULL;
 754                         /*
 755                          * Drop multicast group membership if we joined
 756                          * through the interface being detached.
 757                          */
 758                         gap = 0;
 759                         for (i = 0; i < im6o->im6o_num_memberships; i++) {
 760                                 if (im6o->im6o_membership[i]->in6m_ifp ==
 761                                     ifp) {
 762                                         in6_mc_leave(im6o->im6o_membership[i],
 763                                             NULL);
 764                                         gap++;
 765                                 } else if (gap != 0) {
 766                                         im6o->im6o_membership[i - gap] =
 767                                             im6o->im6o_membership[i];
 768                                 }
 769                         }
 770                         im6o->im6o_num_memberships -= gap;
 771                 }
 772                 INP_WUNLOCK(in6p);
 773         }
 774         INP_INFO_RUNLOCK(pcbinfo);
 775 }
 776
 777 /*
 778  * Check for alternatives when higher level complains
 779  * about service problems.  For now, invalidate cached
 780  * routing information.  If the route was created dynamically
 781  * (by a redirect), time to try a default gateway again.
 782  */
 783 void
 784 in6_losing(struct inpcb *in6p)
 785 {
 786
 787         /*
 788          * We don't store route pointers in the routing table anymore
 789          */
 790         return;
 791 }
 792
 793 /*
 794  * After a routing change, flush old routing
 795  * and allocate a (hopefully) better one.
 796  */
 797 struct inpcb *
 798 in6_rtchange(struct inpcb *inp, int errno)
 799 {
 800         /*
 801          * We don't store route pointers in the routing table anymore
 802          */
 803         return inp;
 804 }
 805
 806 /*
 807  * Lookup PCB in hash list.
 808  */
 809 struct inpcb *
 810 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
 811     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg, int wildcard,
 812     struct ifnet *ifp)
 813 {
 814         struct inpcbhead *head;
 815         struct inpcb *inp, *tmpinp;
 816         u_short fport = fport_arg, lport = lport_arg;
 817         int faith;
 818
 819         INP_INFO_LOCK_ASSERT(pcbinfo);
 820
 821         if (faithprefix_p != NULL)
 822                 faith = (*faithprefix_p)(laddr);
 823         else
 824                 faith = 0;
 825
 826         /*
 827          * First look for an exact match.
 828          */
 829         tmpinp = NULL;
 830         head = &pcbinfo->ipi_hashbase[
 831             INP_PCBHASH(faddr->s6_addr32[3] /* XXX */, lport, fport,
 832             pcbinfo->ipi_hashmask)];
 833         LIST_FOREACH(inp, head, inp_hash) {
 834                 /* XXX inp locking */
 835                 if ((inp->inp_vflag & INP_IPV6) == 0)
 836                         continue;
 837                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 838                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 839                     inp->inp_fport == fport &&
 840                     inp->inp_lport == lport) {
 841                         /*
 842                          * XXX We should be able to directly return
 843                          * the inp here, without any checks.
 844                          * Well unless both bound with SO_REUSEPORT?
 845                          */
 846                         if (prison_flag(inp->inp_cred, PR_IP6))
 847                                 return (inp);
 848                         if (tmpinp == NULL)
 849                                 tmpinp = inp;
 850                 }
 851         }
 852         if (tmpinp != NULL)
 853                 return (tmpinp);
 854
 855         /*
 856          * Then look for a wildcard match, if requested.
 857          */
 858         if (wildcard == INPLOOKUP_WILDCARD) {
 859                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 860                 struct inpcb *jail_wild = NULL;
 861                 int injail;
 862
 863                 /*
 864                  * Order of socket selection - we always prefer jails.
 865                  *      1. jailed, non-wild.
 866                  *      2. jailed, wild.
 867                  *      3. non-jailed, non-wild.
 868                  *      4. non-jailed, wild.
 869                  */
 870                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(INADDR_ANY, lport,
 871                     0, pcbinfo->ipi_hashmask)];
 872                 LIST_FOREACH(inp, head, inp_hash) {
 873                         /* XXX inp locking */
 874                         if ((inp->inp_vflag & INP_IPV6) == 0)
 875                                 continue;
 876
 877                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 878                             inp->inp_lport != lport) {
 879                                 continue;
 880                         }
 881
 882                         /* XXX inp locking */
 883                         if (faith && (inp->inp_flags & INP_FAITH) == 0)
 884                                 continue;
 885
 886                         injail = prison_flag(inp->inp_cred, PR_IP6);
 887                         if (injail) {
 888                                 if (prison_check_ip6(inp->inp_cred,
 889                                     laddr) != 0)
 890                                         continue;
 891                         } else {
 892                                 if (local_exact != NULL)
 893                                         continue;
 894                         }
 895
 896                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
 897                                 if (injail)
 898                                         return (inp);
 899                                 else
 900                                         local_exact = inp;
 901                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 902                                 if (injail)
 903                                         jail_wild = inp;
 904                                 else
 905                                         local_wild = inp;
 906                         }
 907                 } /* LIST_FOREACH */
 908
 909                 if (jail_wild != NULL)
 910                         return (jail_wild);
 911                 if (local_exact != NULL)
 912                         return (local_exact);
 913                 if (local_wild != NULL)
 914                         return (local_wild);
 915         } /* if (wildcard == INPLOOKUP_WILDCARD) */
 916
 917         /*
 918          * Not found.
 919          */
 920         return (NULL);
 921 }
 922
 923 void
 924 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m)
 925 {
 926         struct ip6_hdr *ip;
 927
 928         ip = mtod(m, struct ip6_hdr *);
 929         bzero(sin6, sizeof(*sin6));
 930         sin6->sin6_len = sizeof(*sin6);
 931         sin6->sin6_family = AF_INET6;
 932         sin6->sin6_addr = ip->ip6_src;
 933
 934         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
 935
 936         return;
 937 }