sys/netinet6/in6_pcb.c

   1 /*      $FreeBSD$       */
   2 /*      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $        */
   3
   4 /*-
   5  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. Neither the name of the project nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  */
  33
  34 /*-
  35  * Copyright (c) 1982, 1986, 1991, 1993
  36  *      The Regents of the University of California.  All rights reserved.
  37  *
  38  * Redistribution and use in source and binary forms, with or without
  39  * modification, are permitted provided that the following conditions
  40  * are met:
  41  * 1. Redistributions of source code must retain the above copyright
  42  *    notice, this list of conditions and the following disclaimer.
  43  * 2. Redistributions in binary form must reproduce the above copyright
  44  *    notice, this list of conditions and the following disclaimer in the
  45  *    documentation and/or other materials provided with the distribution.
  46  * 4. Neither the name of the University nor the names of its contributors
  47  *    may be used to endorse or promote products derived from this software
  48  *    without specific prior written permission.
  49  *
  50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  60  * SUCH DAMAGE.
  61  *
  62  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  63  */
  64
  65 #include "opt_inet.h"
  66 #include "opt_inet6.h"
  67 #include "opt_ipsec.h"
  68
  69 #include <sys/param.h>
  70 #include <sys/systm.h>
  71 #include <sys/malloc.h>
  72 #include <sys/mbuf.h>
  73 #include <sys/domain.h>
  74 #include <sys/protosw.h>
  75 #include <sys/socket.h>
  76 #include <sys/socketvar.h>
  77 #include <sys/sockio.h>
  78 #include <sys/errno.h>
  79 #include <sys/time.h>
  80 #include <sys/proc.h>
  81 #include <sys/jail.h>
  82
  83 #include <vm/uma.h>
  84
  85 #include <net/if.h>
  86 #include <net/if_types.h>
  87 #include <net/route.h>
  88
  89 #include <netinet/in.h>
  90 #include <netinet/in_var.h>
  91 #include <netinet/in_systm.h>
  92 #include <netinet/tcp_var.h>
  93 #include <netinet/ip6.h>
  94 #include <netinet/ip_var.h>
  95 #include <netinet6/ip6_var.h>
  96 #include <netinet6/nd6.h>
  97 #include <netinet/in_pcb.h>
  98 #include <netinet6/in6_pcb.h>
  99
 100 #ifdef IPSEC
 101 #include <netinet6/ipsec.h>
 102 #ifdef INET6
 103 #include <netinet6/ipsec6.h>
 104 #endif
 105 #include <netinet6/ah.h>
 106 #ifdef INET6
 107 #include <netinet6/ah6.h>
 108 #endif
 109 #include <netkey/key.h>
 110 #endif /* IPSEC */
 111
 112 #ifdef FAST_IPSEC
 113 #include <netipsec/ipsec.h>
 114 #include <netipsec/ipsec6.h>
 115 #include <netipsec/key.h>
 116 #endif /* FAST_IPSEC */
 117
 118 struct  in6_addr zeroin6_addr;
 119
 120 int
 121 in6_pcbbind(inp, nam, cred)
 122         register struct inpcb *inp;
 123         struct sockaddr *nam;
 124         struct ucred *cred;
 125 {
 126         struct socket *so = inp->inp_socket;
 127         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 128         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 129         u_short lport = 0;
 130         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 131
 132         INP_INFO_WLOCK_ASSERT(pcbinfo);
 133         INP_LOCK_ASSERT(inp);
 134
 135         if (!in6_ifaddr) /* XXX broken! */
 136                 return (EADDRNOTAVAIL);
 137         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 138                 return (EINVAL);
 139         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 140                 wild = 1;
 141         if (nam) {
 142                 sin6 = (struct sockaddr_in6 *)nam;
 143                 if (nam->sa_len != sizeof(*sin6))
 144                         return (EINVAL);
 145                 /*
 146                  * family check.
 147                  */
 148                 if (nam->sa_family != AF_INET6)
 149                         return (EAFNOSUPPORT);
 150
 151                 /* KAME hack: embed scopeid */
 152                 if (in6_embedscope(&sin6->sin6_addr, sin6, inp, NULL) != 0)
 153                         return EINVAL;
 154                 /* this must be cleared for ifa_ifwithaddr() */
 155                 sin6->sin6_scope_id = 0;
 156
 157                 lport = sin6->sin6_port;
 158                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 159                         /*
 160                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 161                          * allow compepte duplication of binding if
 162                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 163                          * and a multicast address is bound on both
 164                          * new and duplicated sockets.
 165                          */
 166                         if (so->so_options & SO_REUSEADDR)
 167                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 168                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 169                         struct ifaddr *ia = NULL;
 170
 171                         sin6->sin6_port = 0;            /* yech... */
 172                         if ((ia = ifa_ifwithaddr((struct sockaddr *)sin6)) == 0)
 173                                 return (EADDRNOTAVAIL);
 174
 175                         /*
 176                          * XXX: bind to an anycast address might accidentally
 177                          * cause sending a packet with anycast source address.
 178                          * We should allow to bind to a deprecated address, since
 179                          * the application dares to use it.
 180                          */
 181                         if (ia &&
 182                             ((struct in6_ifaddr *)ia)->ia6_flags &
 183                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 184                                 return (EADDRNOTAVAIL);
 185                         }
 186                 }
 187                 if (lport) {
 188                         struct inpcb *t;
 189
 190                         /* GROSS */
 191                         if (ntohs(lport) < IPV6PORT_RESERVED &&
 192                             suser_cred(cred, SUSER_ALLOWJAIL))
 193                                 return (EACCES);
 194                         if (so->so_cred->cr_uid != 0 &&
 195                             !IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 196                                 t = in6_pcblookup_local(pcbinfo,
 197                                     &sin6->sin6_addr, lport,
 198                                     INPLOOKUP_WILDCARD);
 199                                 if (t &&
 200                                     ((t->inp_vflag & 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) && (so->so_cred->cr_uid !=
 207                                      t->inp_socket->so_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);
 217                                         if (t &&
 218                                             ((t->inp_vflag &
 219                                               INP_TIMEWAIT) == 0) &&
 220                                             (so->so_type != SOCK_STREAM ||
 221                                              ntohl(t->inp_faddr.s_addr) ==
 222                                               INADDR_ANY) &&
 223                                             (so->so_cred->cr_uid !=
 224                                              t->inp_socket->so_cred->cr_uid))
 225                                                 return (EADDRINUSE);
 226                                 }
 227                         }
 228                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 229                                                 lport, wild);
 230                         if (t && (reuseport & ((t->inp_vflag & 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);
 241                                 if (t && t->inp_vflag & 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                 int e;
 262                 if ((e = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0)
 263                         return (e);
 264         }
 265         else {
 266                 inp->inp_lport = lport;
 267                 if (in_pcbinshash(inp) != 0) {
 268                         inp->in6p_laddr = in6addr_any;
 269                         inp->inp_lport = 0;
 270                         return (EAGAIN);
 271                 }
 272         }
 273         return (0);
 274 }
 275
 276 /*
 277  *   Transform old in6_pcbconnect() into an inner subroutine for new
 278  *   in6_pcbconnect(): Do some validity-checking on the remote
 279  *   address (in mbuf 'nam') and then determine local host address
 280  *   (i.e., which interface) to use to access that remote host.
 281  *
 282  *   This preserves definition of in6_pcbconnect(), while supporting a
 283  *   slightly different version for T/TCP.  (This is more than
 284  *   a bit of a kludge, but cleaning up the internal interfaces would
 285  *   have forced minor changes in every protocol).
 286  */
 287
 288 int
 289 in6_pcbladdr(inp, nam, plocal_addr6)
 290         register struct inpcb *inp;
 291         struct sockaddr *nam;
 292         struct in6_addr **plocal_addr6;
 293 {
 294         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 295         struct ifnet *ifp = NULL;
 296         int error = 0;
 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         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 306         INP_LOCK_ASSERT(inp);
 307
 308         /* KAME hack: embed scopeid */
 309         if (in6_embedscope(&sin6->sin6_addr, sin6, inp, &ifp) != 0)
 310                 return EINVAL;
 311
 312         if (in6_ifaddr) {
 313                 /*
 314                  * If the destination address is UNSPECIFIED addr,
 315                  * use the loopback addr, e.g ::1.
 316                  */
 317                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 318                         sin6->sin6_addr = in6addr_loopback;
 319         }
 320         {
 321                 /*
 322                  * XXX: in6_selectsrc might replace the bound local address
 323                  * with the address specified by setsockopt(IPV6_PKTINFO).
 324                  * Is it the intended behavior?
 325                  */
 326                 *plocal_addr6 = in6_selectsrc(sin6, inp->in6p_outputopts,
 327                                               inp->in6p_moptions, NULL,
 328                                               &inp->in6p_laddr, &error);
 329                 if (*plocal_addr6 == 0) {
 330                         if (error == 0)
 331                                 error = EADDRNOTAVAIL;
 332                         return (error);
 333                 }
 334                 /*
 335                  * Don't do pcblookup call here; return interface in
 336                  * plocal_addr6
 337                  * and exit to caller, that will do the lookup.
 338                  */
 339         }
 340         return (0);
 341 }
 342
 343 /*
 344  * Outer subroutine:
 345  * Connect from a socket to a specified address.
 346  * Both address and port must be specified in argument sin.
 347  * If don't have a local address for this socket yet,
 348  * then pick one.
 349  */
 350 int
 351 in6_pcbconnect(inp, nam, cred)
 352         register struct inpcb *inp;
 353         struct sockaddr *nam;
 354         struct ucred *cred;
 355 {
 356         struct in6_addr *addr6;
 357         register struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 358         int error;
 359
 360         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 361         INP_LOCK_ASSERT(inp);
 362
 363         /*
 364          * Call inner routine, to assign local interface address.
 365          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 366          */
 367         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 368                 return (error);
 369
 370         if (in6_pcblookup_hash(inp->inp_pcbinfo, &sin6->sin6_addr,
 371                                sin6->sin6_port,
 372                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 373                               ? addr6 : &inp->in6p_laddr,
 374                               inp->inp_lport, 0, NULL) != NULL) {
 375                 return (EADDRINUSE);
 376         }
 377         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 378                 if (inp->inp_lport == 0) {
 379                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 380                         if (error)
 381                                 return (error);
 382                 }
 383                 inp->in6p_laddr = *addr6;
 384         }
 385         inp->in6p_faddr = sin6->sin6_addr;
 386         inp->inp_fport = sin6->sin6_port;
 387         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 388         inp->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
 389         if (inp->in6p_flags & IN6P_AUTOFLOWLABEL)
 390                 inp->in6p_flowinfo |=
 391                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 392
 393         in_pcbrehash(inp);
 394 #ifdef IPSEC
 395         if (inp->inp_socket->so_type == SOCK_STREAM)
 396                 ipsec_pcbconn(inp->inp_sp);
 397 #endif
 398         return (0);
 399 }
 400
 401 void
 402 in6_pcbdisconnect(inp)
 403         struct inpcb *inp;
 404 {
 405
 406         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 407         INP_LOCK_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->in6p_flowinfo &= ~IPV6_FLOWLABEL_MASK;
 413         in_pcbrehash(inp);
 414 #ifdef IPSEC
 415         ipsec_pcbdisconn(inp->inp_sp);
 416 #endif
 417         if (inp->inp_socket->so_state & SS_NOFDREF)
 418                 in6_pcbdetach(inp);
 419 }
 420
 421 void
 422 in6_pcbdetach(inp)
 423         struct inpcb *inp;
 424 {
 425         struct socket *so = inp->inp_socket;
 426         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 427
 428         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 429         INP_LOCK_ASSERT(inp);
 430
 431 #if defined(IPSEC) || defined(FAST_IPSEC)
 432         if (inp->in6p_sp != NULL)
 433                 ipsec6_delete_pcbpolicy(inp);
 434 #endif /* IPSEC */
 435         inp->inp_gencnt = ++ipi->ipi_gencnt;
 436         in_pcbremlists(inp);
 437
 438         if (so) {
 439                 ACCEPT_LOCK();
 440                 SOCK_LOCK(so);
 441                 so->so_pcb = NULL;
 442                 sotryfree(so);
 443         }
 444
 445         ip6_freepcbopts(inp->in6p_outputopts);
 446         ip6_freemoptions(inp->in6p_moptions);
 447         /* Check and free IPv4 related resources in case of mapped addr */
 448         if (inp->inp_options)
 449                 (void)m_free(inp->inp_options);
 450         ip_freemoptions(inp->inp_moptions);
 451         inp->inp_vflag = 0;
 452         INP_LOCK_DESTROY(inp);
 453         uma_zfree(ipi->ipi_zone, inp);
 454 }
 455
 456 struct sockaddr *
 457 in6_sockaddr(port, addr_p)
 458         in_port_t port;
 459         struct in6_addr *addr_p;
 460 {
 461         struct sockaddr_in6 *sin6;
 462
 463         MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6, M_SONAME, M_WAITOK);
 464         bzero(sin6, sizeof *sin6);
 465         sin6->sin6_family = AF_INET6;
 466         sin6->sin6_len = sizeof(*sin6);
 467         sin6->sin6_port = port;
 468         sin6->sin6_addr = *addr_p;
 469         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 470                 sin6->sin6_scope_id = ntohs(sin6->sin6_addr.s6_addr16[1]);
 471         else
 472                 sin6->sin6_scope_id = 0;        /*XXX*/
 473         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 474                 sin6->sin6_addr.s6_addr16[1] = 0;
 475
 476         return (struct sockaddr *)sin6;
 477 }
 478
 479 struct sockaddr *
 480 in6_v4mapsin6_sockaddr(port, addr_p)
 481         in_port_t port;
 482         struct in_addr *addr_p;
 483 {
 484         struct sockaddr_in sin;
 485         struct sockaddr_in6 *sin6_p;
 486
 487         bzero(&sin, sizeof sin);
 488         sin.sin_family = AF_INET;
 489         sin.sin_len = sizeof(sin);
 490         sin.sin_port = port;
 491         sin.sin_addr = *addr_p;
 492
 493         MALLOC(sin6_p, struct sockaddr_in6 *, sizeof *sin6_p, M_SONAME,
 494                 M_WAITOK);
 495         in6_sin_2_v4mapsin6(&sin, sin6_p);
 496
 497         return (struct sockaddr *)sin6_p;
 498 }
 499
 500 /*
 501  * The calling convention of in6_setsockaddr() and in6_setpeeraddr() was
 502  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 503  * in struct pr_usrreqs, so that protocols can just reference then directly
 504  * without the need for a wrapper function.  The socket must have a valid
 505  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 506  * except through a kernel programming error, so it is acceptable to panic
 507  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 508  * because there actually /is/ a programming error somewhere... XXX)
 509  */
 510 int
 511 in6_setsockaddr(so, nam)
 512         struct socket *so;
 513         struct sockaddr **nam;
 514 {
 515         int s;
 516         register struct inpcb *inp;
 517         struct in6_addr addr;
 518         in_port_t port;
 519
 520         s = splnet();
 521         inp = sotoinpcb(so);
 522         if (!inp) {
 523                 splx(s);
 524                 return EINVAL;
 525         }
 526         port = inp->inp_lport;
 527         addr = inp->in6p_laddr;
 528         splx(s);
 529
 530         *nam = in6_sockaddr(port, &addr);
 531         return 0;
 532 }
 533
 534 int
 535 in6_setpeeraddr(so, nam)
 536         struct socket *so;
 537         struct sockaddr **nam;
 538 {
 539         int s;
 540         struct inpcb *inp;
 541         struct in6_addr addr;
 542         in_port_t port;
 543
 544         s = splnet();
 545         inp = sotoinpcb(so);
 546         if (!inp) {
 547                 splx(s);
 548                 return EINVAL;
 549         }
 550         port = inp->inp_fport;
 551         addr = inp->in6p_faddr;
 552         splx(s);
 553
 554         *nam = in6_sockaddr(port, &addr);
 555         return 0;
 556 }
 557
 558 int
 559 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 560 {
 561         struct  inpcb *inp = sotoinpcb(so);
 562         int     error;
 563
 564         if (inp == NULL)
 565                 return EINVAL;
 566         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 567                 error = in_setsockaddr(so, nam, &tcbinfo);
 568                 if (error == 0)
 569                         in6_sin_2_v4mapsin6_in_sock(nam);
 570         } else {
 571                 /* scope issues will be handled in in6_setsockaddr(). */
 572                 error = in6_setsockaddr(so, nam);
 573         }
 574
 575         return error;
 576 }
 577
 578 int
 579 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 580 {
 581         struct  inpcb *inp = sotoinpcb(so);
 582         int     error;
 583
 584         if (inp == NULL)
 585                 return EINVAL;
 586         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 587                 error = in_setpeeraddr(so, nam, &tcbinfo);
 588                 if (error == 0)
 589                         in6_sin_2_v4mapsin6_in_sock(nam);
 590         } else
 591         /* scope issues will be handled in in6_setpeeraddr(). */
 592         error = in6_setpeeraddr(so, nam);
 593
 594         return error;
 595 }
 596
 597 /*
 598  * Pass some notification to all connections of a protocol
 599  * associated with address dst.  The local address and/or port numbers
 600  * may be specified to limit the search.  The "usual action" will be
 601  * taken, depending on the ctlinput cmd.  The caller must filter any
 602  * cmds that are uninteresting (e.g., no error in the map).
 603  * Call the protocol specific routine (if any) to report
 604  * any errors for each matching socket.
 605  *
 606  * Must be called at splnet.
 607  */
 608 void
 609 in6_pcbnotify(pcbinfo, dst, fport_arg, src, lport_arg, cmd, cmdarg, notify)
 610         struct inpcbinfo *pcbinfo;
 611         struct sockaddr *dst;
 612         const struct sockaddr *src;
 613         u_int fport_arg, lport_arg;
 614         int cmd;
 615         void *cmdarg;
 616         struct inpcb *(*notify) __P((struct inpcb *, int));
 617 {
 618         struct inpcbhead *head;
 619         struct inpcb *inp, *ninp;
 620         struct sockaddr_in6 sa6_src, *sa6_dst;
 621         u_short fport = fport_arg, lport = lport_arg;
 622         u_int32_t flowinfo;
 623         int errno, s;
 624
 625         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 626                 return;
 627
 628         sa6_dst = (struct sockaddr_in6 *)dst;
 629         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 630                 return;
 631
 632         /*
 633          * note that src can be NULL when we get notify by local fragmentation.
 634          */
 635         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 636         flowinfo = sa6_src.sin6_flowinfo;
 637
 638         /*
 639          * Redirects go to all references to the destination,
 640          * and use in6_rtchange to invalidate the route cache.
 641          * Dead host indications: also use in6_rtchange to invalidate
 642          * the cache, and deliver the error to all the sockets.
 643          * Otherwise, if we have knowledge of the local port and address,
 644          * deliver only to that socket.
 645          */
 646         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 647                 fport = 0;
 648                 lport = 0;
 649                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 650
 651                 if (cmd != PRC_HOSTDEAD)
 652                         notify = in6_rtchange;
 653         }
 654         errno = inet6ctlerrmap[cmd];
 655         s = splnet();
 656         head = pcbinfo->listhead;
 657         INP_INFO_WLOCK(pcbinfo);
 658         for (inp = LIST_FIRST(head); inp != NULL; inp = ninp) {
 659                 INP_LOCK(inp);
 660                 ninp = LIST_NEXT(inp, inp_list);
 661
 662                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 663                         INP_UNLOCK(inp);
 664                         continue;
 665                 }
 666
 667                 /*
 668                  * If the error designates a new path MTU for a destination
 669                  * and the application (associated with this socket) wanted to
 670                  * know the value, notify. Note that we notify for all
 671                  * disconnected sockets if the corresponding application
 672                  * wanted. This is because some UDP applications keep sending
 673                  * sockets disconnected.
 674                  * XXX: should we avoid to notify the value to TCP sockets?
 675                  */
 676                 if (cmd == PRC_MSGSIZE && (inp->inp_flags & IN6P_MTU) != 0 &&
 677                     (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 678                      IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr))) {
 679                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 680                                         (u_int32_t *)cmdarg);
 681                 }
 682
 683                 /*
 684                  * Detect if we should notify the error. If no source and
 685                  * destination ports are specifed, but non-zero flowinfo and
 686                  * local address match, notify the error. This is the case
 687                  * when the error is delivered with an encrypted buffer
 688                  * by ESP. Otherwise, just compare addresses and ports
 689                  * as usual.
 690                  */
 691                 if (lport == 0 && fport == 0 && flowinfo &&
 692                     inp->inp_socket != NULL &&
 693                     flowinfo == (inp->in6p_flowinfo & IPV6_FLOWLABEL_MASK) &&
 694                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 695                         goto do_notify;
 696                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 697                                              &sa6_dst->sin6_addr) ||
 698                          inp->inp_socket == 0 ||
 699                          (lport && inp->inp_lport != lport) ||
 700                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 701                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 702                                               &sa6_src.sin6_addr)) ||
 703                          (fport && inp->inp_fport != fport)) {
 704                         INP_UNLOCK(inp);
 705                         continue;
 706                 }
 707
 708           do_notify:
 709                 if (notify) {
 710                         if ((*notify)(inp, errno))
 711                                 INP_UNLOCK(inp);
 712                 } else
 713                         INP_UNLOCK(inp);
 714         }
 715         INP_INFO_WUNLOCK(pcbinfo);
 716         splx(s);
 717 }
 718
 719 /*
 720  * Lookup a PCB based on the local address and port.
 721  */
 722 struct inpcb *
 723 in6_pcblookup_local(pcbinfo, laddr, lport_arg, wild_okay)
 724         struct inpcbinfo *pcbinfo;
 725         struct in6_addr *laddr;
 726         u_int lport_arg;
 727         int wild_okay;
 728 {
 729         register struct inpcb *inp;
 730         int matchwild = 3, wildcard;
 731         u_short lport = lport_arg;
 732
 733         if (!wild_okay) {
 734                 struct inpcbhead *head;
 735                 /*
 736                  * Look for an unconnected (wildcard foreign addr) PCB that
 737                  * matches the local address and port we're looking for.
 738                  */
 739                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0,
 740                                                       pcbinfo->hashmask)];
 741                 LIST_FOREACH(inp, head, inp_hash) {
 742                         if ((inp->inp_vflag & INP_IPV6) == 0)
 743                                 continue;
 744                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 745                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 746                             inp->inp_lport == lport) {
 747                                 /*
 748                                  * Found.
 749                                  */
 750                                 return (inp);
 751                         }
 752                 }
 753                 /*
 754                  * Not found.
 755                  */
 756                 return (NULL);
 757         } else {
 758                 struct inpcbporthead *porthash;
 759                 struct inpcbport *phd;
 760                 struct inpcb *match = NULL;
 761                 /*
 762                  * Best fit PCB lookup.
 763                  *
 764                  * First see if this local port is in use by looking on the
 765                  * port hash list.
 766                  */
 767                 porthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(lport,
 768                     pcbinfo->porthashmask)];
 769                 LIST_FOREACH(phd, porthash, phd_hash) {
 770                         if (phd->phd_port == lport)
 771                                 break;
 772                 }
 773                 if (phd != NULL) {
 774                         /*
 775                          * Port is in use by one or more PCBs. Look for best
 776                          * fit.
 777                          */
 778                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 779                                 wildcard = 0;
 780                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 781                                         continue;
 782                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 783                                         wildcard++;
 784                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 785                                         &inp->in6p_laddr)) {
 786                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 787                                                 wildcard++;
 788                                         else if (!IN6_ARE_ADDR_EQUAL(
 789                                                 &inp->in6p_laddr, laddr))
 790                                                 continue;
 791                                 } else {
 792                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 793                                                 wildcard++;
 794                                 }
 795                                 if (wildcard < matchwild) {
 796                                         match = inp;
 797                                         matchwild = wildcard;
 798                                         if (matchwild == 0) {
 799                                                 break;
 800                                         }
 801                                 }
 802                         }
 803                 }
 804                 return (match);
 805         }
 806 }
 807
 808 void
 809 in6_pcbpurgeif0(head, ifp)
 810         struct in6pcb *head;
 811         struct ifnet *ifp;
 812 {
 813         struct in6pcb *in6p;
 814         struct ip6_moptions *im6o;
 815         struct in6_multi_mship *imm, *nimm;
 816
 817         for (in6p = head; in6p != NULL; in6p = LIST_NEXT(in6p, inp_list)) {
 818                 im6o = in6p->in6p_moptions;
 819                 if ((in6p->inp_vflag & INP_IPV6) &&
 820                     im6o) {
 821                         /*
 822                          * Unselect the outgoing interface if it is being
 823                          * detached.
 824                          */
 825                         if (im6o->im6o_multicast_ifp == ifp)
 826                                 im6o->im6o_multicast_ifp = NULL;
 827
 828                         /*
 829                          * Drop multicast group membership if we joined
 830                          * through the interface being detached.
 831                          * XXX controversial - is it really legal for kernel
 832                          * to force this?
 833                          */
 834                         for (imm = im6o->im6o_memberships.lh_first;
 835                              imm != NULL; imm = nimm) {
 836                                 nimm = imm->i6mm_chain.le_next;
 837                                 if (imm->i6mm_maddr->in6m_ifp == ifp) {
 838                                         LIST_REMOVE(imm, i6mm_chain);
 839                                         in6_delmulti(imm->i6mm_maddr);
 840                                         free(imm, M_IPMADDR);
 841                                 }
 842                         }
 843                 }
 844         }
 845 }
 846
 847 /*
 848  * Check for alternatives when higher level complains
 849  * about service problems.  For now, invalidate cached
 850  * routing information.  If the route was created dynamically
 851  * (by a redirect), time to try a default gateway again.
 852  */
 853 void
 854 in6_losing(in6p)
 855         struct inpcb *in6p;
 856 {
 857         /*
 858          * We don't store route pointers in the routing table anymore
 859          */
 860         return;
 861 }
 862
 863 /*
 864  * After a routing change, flush old routing
 865  * and allocate a (hopefully) better one.
 866  */
 867 struct inpcb *
 868 in6_rtchange(inp, errno)
 869         struct inpcb *inp;
 870         int errno;
 871 {
 872         /*
 873          * We don't store route pointers in the routing table anymore
 874          */
 875         return inp;
 876 }
 877
 878 /*
 879  * Lookup PCB in hash list.
 880  */
 881 struct inpcb *
 882 in6_pcblookup_hash(pcbinfo, faddr, fport_arg, laddr, lport_arg, wildcard, ifp)
 883         struct inpcbinfo *pcbinfo;
 884         struct in6_addr *faddr, *laddr;
 885         u_int fport_arg, lport_arg;
 886         int wildcard;
 887         struct ifnet *ifp;
 888 {
 889         struct inpcbhead *head;
 890         register struct inpcb *inp;
 891         u_short fport = fport_arg, lport = lport_arg;
 892         int faith;
 893
 894         if (faithprefix_p != NULL)
 895                 faith = (*faithprefix_p)(laddr);
 896         else
 897                 faith = 0;
 898
 899         /*
 900          * First look for an exact match.
 901          */
 902         head = &pcbinfo->hashbase[INP_PCBHASH(faddr->s6_addr32[3] /* XXX */,
 903                                               lport, fport,
 904                                               pcbinfo->hashmask)];
 905         LIST_FOREACH(inp, head, inp_hash) {
 906                 if ((inp->inp_vflag & INP_IPV6) == 0)
 907                         continue;
 908                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 909                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 910                     inp->inp_fport == fport &&
 911                     inp->inp_lport == lport) {
 912                         /*
 913                          * Found.
 914                          */
 915                         return (inp);
 916                 }
 917         }
 918         if (wildcard) {
 919                 struct inpcb *local_wild = NULL;
 920
 921                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0,
 922                                                       pcbinfo->hashmask)];
 923                 LIST_FOREACH(inp, head, inp_hash) {
 924                         if ((inp->inp_vflag & INP_IPV6) == 0)
 925                                 continue;
 926                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 927                             inp->inp_lport == lport) {
 928                                 if (faith && (inp->inp_flags & INP_FAITH) == 0)
 929                                         continue;
 930                                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 931                                                        laddr))
 932                                         return (inp);
 933                                 else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 934                                         local_wild = inp;
 935                         }
 936                 }
 937                 return (local_wild);
 938         }
 939
 940         /*
 941          * Not found.
 942          */
 943         return (NULL);
 944 }
 945
 946 void
 947 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m)
 948 {
 949         struct ip6_hdr *ip;
 950
 951         ip = mtod(m, struct ip6_hdr *);
 952         bzero(sin6, sizeof(*sin6));
 953         sin6->sin6_len = sizeof(*sin6);
 954         sin6->sin6_family = AF_INET6;
 955         sin6->sin6_addr = ip->ip6_src;
 956         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 957                 sin6->sin6_addr.s6_addr16[1] = 0;
 958         sin6->sin6_scope_id =
 959                 (m->m_pkthdr.rcvif && IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr))
 960                 ? m->m_pkthdr.rcvif->if_index : 0;
 961
 962         return;
 963 }