sys/netinet6/in6_pcb.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Robert N. M. Watson under
   9  * contract to Juniper Networks, Inc.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  * 3. Neither the name of the project nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  36  */
  37
  38 /*-
  39  * Copyright (c) 1982, 1986, 1991, 1993
  40  *      The Regents of the University of California.  All rights reserved.
  41  *
  42  * Redistribution and use in source and binary forms, with or without
  43  * modification, are permitted provided that the following conditions
  44  * are met:
  45  * 1. Redistributions of source code must retain the above copyright
  46  *    notice, this list of conditions and the following disclaimer.
  47  * 2. Redistributions in binary form must reproduce the above copyright
  48  *    notice, this list of conditions and the following disclaimer in the
  49  *    documentation and/or other materials provided with the distribution.
  50  * 3. Neither the name of the University nor the names of its contributors
  51  *    may be used to endorse or promote products derived from this software
  52  *    without specific prior written permission.
  53  *
  54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  64  * SUCH DAMAGE.
  65  *
  66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  67  */
  68
  69 #include <sys/cdefs.h>
  70 __FBSDID("$FreeBSD$");
  71
  72 #include "opt_inet.h"
  73 #include "opt_inet6.h"
  74 #include "opt_ipsec.h"
  75 #include "opt_pcbgroup.h"
  76 #include "opt_rss.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/malloc.h>
  81 #include <sys/mbuf.h>
  82 #include <sys/domain.h>
  83 #include <sys/protosw.h>
  84 #include <sys/socket.h>
  85 #include <sys/socketvar.h>
  86 #include <sys/sockio.h>
  87 #include <sys/errno.h>
  88 #include <sys/time.h>
  89 #include <sys/priv.h>
  90 #include <sys/proc.h>
  91 #include <sys/jail.h>
  92
  93 #include <vm/uma.h>
  94
  95 #include <net/if.h>
  96 #include <net/if_var.h>
  97 #include <net/if_llatbl.h>
  98 #include <net/if_types.h>
  99 #include <net/route.h>
 100
 101 #include <netinet/in.h>
 102 #include <netinet/in_var.h>
 103 #include <netinet/in_systm.h>
 104 #include <netinet/tcp_var.h>
 105 #include <netinet/ip6.h>
 106 #include <netinet/ip_var.h>
 107
 108 #include <netinet6/ip6_var.h>
 109 #include <netinet6/nd6.h>
 110 #include <netinet/in_pcb.h>
 111 #include <netinet6/in6_pcb.h>
 112 #include <netinet6/scope6_var.h>
 113
 114 static struct inpcb *in6_pcblookup_hash_locked(struct inpcbinfo *,
 115     struct in6_addr *, u_int, struct in6_addr *, u_int, int, struct ifnet *);
 116
 117 int
 118 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 119     struct ucred *cred)
 120 {
 121         struct socket *so = inp->inp_socket;
 122         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 123         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 124         u_short lport = 0;
 125         int error, lookupflags = 0;
 126         int reuseport = (so->so_options & SO_REUSEPORT);
 127
 128         INP_WLOCK_ASSERT(inp);
 129         INP_HASH_WLOCK_ASSERT(pcbinfo);
 130
 131         if (TAILQ_EMPTY(&V_in6_ifaddrhead))     /* XXX broken! */
 132                 return (EADDRNOTAVAIL);
 133         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 134                 return (EINVAL);
 135         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 136                 lookupflags = INPLOOKUP_WILDCARD;
 137         if (nam == NULL) {
 138                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 139                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 140                         return (error);
 141         } else {
 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                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 152                         return(error);
 153
 154                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 155                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 156                         return (error);
 157
 158                 lport = sin6->sin6_port;
 159                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 160                         /*
 161                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 162                          * allow compepte duplication of binding if
 163                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 164                          * and a multicast address is bound on both
 165                          * new and duplicated sockets.
 166                          */
 167                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
 168                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 169                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 170                         struct ifaddr *ifa;
 171
 172                         sin6->sin6_port = 0;            /* yech... */
 173                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 174                             NULL &&
 175                             (inp->inp_flags & INP_BINDANY) == 0) {
 176                                 return (EADDRNOTAVAIL);
 177                         }
 178
 179                         /*
 180                          * XXX: bind to an anycast address might accidentally
 181                          * cause sending a packet with anycast source address.
 182                          * We should allow to bind to a deprecated address, since
 183                          * the application dares to use it.
 184                          */
 185                         if (ifa != NULL &&
 186                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 187                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 188                                 ifa_free(ifa);
 189                                 return (EADDRNOTAVAIL);
 190                         }
 191                         if (ifa != NULL)
 192                                 ifa_free(ifa);
 193                 }
 194                 if (lport) {
 195                         struct inpcb *t;
 196                         struct tcptw *tw;
 197
 198                         /* GROSS */
 199                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 200                             ntohs(lport) >= V_ipport_reservedlow &&
 201                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
 202                             0))
 203                                 return (EACCES);
 204                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 205                             priv_check_cred(inp->inp_cred,
 206                             PRIV_NETINET_REUSEPORT, 0) != 0) {
 207                                 t = in6_pcblookup_local(pcbinfo,
 208                                     &sin6->sin6_addr, lport,
 209                                     INPLOOKUP_WILDCARD, cred);
 210                                 if (t &&
 211                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 212                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 213                                     (so->so_type != SOCK_STREAM ||
 214                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 215                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 216                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 217                                      (t->inp_flags2 & INP_REUSEPORT) == 0) &&
 218                                     (inp->inp_cred->cr_uid !=
 219                                      t->inp_cred->cr_uid))
 220                                         return (EADDRINUSE);
 221
 222                                 /*
 223                                  * If the socket is a BINDMULTI socket, then
 224                                  * the credentials need to match and the
 225                                  * original socket also has to have been bound
 226                                  * with BINDMULTI.
 227                                  */
 228                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 229                                         return (EADDRINUSE);
 230
 231 #ifdef INET
 232                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 233                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 234                                         struct sockaddr_in sin;
 235
 236                                         in6_sin6_2_sin(&sin, sin6);
 237                                         t = in_pcblookup_local(pcbinfo,
 238                                             sin.sin_addr, lport,
 239                                             INPLOOKUP_WILDCARD, cred);
 240                                         if (t &&
 241                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 242                                             ((t->inp_flags &
 243                                               INP_TIMEWAIT) == 0) &&
 244                                             (so->so_type != SOCK_STREAM ||
 245                                              ntohl(t->inp_faddr.s_addr) ==
 246                                               INADDR_ANY) &&
 247                                             (inp->inp_cred->cr_uid !=
 248                                              t->inp_cred->cr_uid))
 249                                                 return (EADDRINUSE);
 250
 251                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 252                                                 return (EADDRINUSE);
 253                                 }
 254 #endif
 255                         }
 256                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 257                             lport, lookupflags, cred);
 258                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 259                                 /*
 260                                  * XXXRW: If an incpb has had its timewait
 261                                  * state recycled, we treat the address as
 262                                  * being in use (for now).  This is better
 263                                  * than a panic, but not desirable.
 264                                  */
 265                                 tw = intotw(t);
 266                                 if (tw == NULL ||
 267                                     (reuseport & tw->tw_so_options) == 0)
 268                                         return (EADDRINUSE);
 269                         } else if (t && (reuseport & inp_so_options(t)) == 0) {
 270                                 return (EADDRINUSE);
 271                         }
 272 #ifdef INET
 273                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 274                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 275                                 struct sockaddr_in sin;
 276
 277                                 in6_sin6_2_sin(&sin, sin6);
 278                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 279                                     lport, lookupflags, cred);
 280                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 281                                         tw = intotw(t);
 282                                         if (tw == NULL)
 283                                                 return (EADDRINUSE);
 284                                         if ((reuseport & tw->tw_so_options) == 0
 285                                             && (ntohl(t->inp_laddr.s_addr) !=
 286                                              INADDR_ANY || ((inp->inp_vflag &
 287                                              INP_IPV6PROTO) ==
 288                                              (t->inp_vflag & INP_IPV6PROTO))))
 289                                                 return (EADDRINUSE);
 290                                 } else if (t &&
 291                                     (reuseport & inp_so_options(t)) == 0 &&
 292                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 293                                     (t->inp_vflag & INP_IPV6PROTO) != 0))
 294                                         return (EADDRINUSE);
 295                         }
 296 #endif
 297                 }
 298                 inp->in6p_laddr = sin6->sin6_addr;
 299         }
 300         if (lport == 0) {
 301                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 302                         /* Undo an address bind that may have occurred. */
 303                         inp->in6p_laddr = in6addr_any;
 304                         return (error);
 305                 }
 306         } else {
 307                 inp->inp_lport = lport;
 308                 if (in_pcbinshash(inp) != 0) {
 309                         inp->in6p_laddr = in6addr_any;
 310                         inp->inp_lport = 0;
 311                         return (EAGAIN);
 312                 }
 313         }
 314         return (0);
 315 }
 316
 317 /*
 318  *   Transform old in6_pcbconnect() into an inner subroutine for new
 319  *   in6_pcbconnect(): Do some validity-checking on the remote
 320  *   address (in mbuf 'nam') and then determine local host address
 321  *   (i.e., which interface) to use to access that remote host.
 322  *
 323  *   This preserves definition of in6_pcbconnect(), while supporting a
 324  *   slightly different version for T/TCP.  (This is more than
 325  *   a bit of a kludge, but cleaning up the internal interfaces would
 326  *   have forced minor changes in every protocol).
 327  */
 328 static int
 329 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 330     struct in6_addr *plocal_addr6)
 331 {
 332         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 333         int error = 0;
 334         int scope_ambiguous = 0;
 335         struct in6_addr in6a;
 336
 337         INP_WLOCK_ASSERT(inp);
 338         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 339
 340         if (nam->sa_len != sizeof (*sin6))
 341                 return (EINVAL);
 342         if (sin6->sin6_family != AF_INET6)
 343                 return (EAFNOSUPPORT);
 344         if (sin6->sin6_port == 0)
 345                 return (EADDRNOTAVAIL);
 346
 347         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 348                 scope_ambiguous = 1;
 349         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 350                 return(error);
 351
 352         if (!TAILQ_EMPTY(&V_in6_ifaddrhead)) {
 353                 /*
 354                  * If the destination address is UNSPECIFIED addr,
 355                  * use the loopback addr, e.g ::1.
 356                  */
 357                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 358                         sin6->sin6_addr = in6addr_loopback;
 359         }
 360         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 361                 return (error);
 362
 363         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 364             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 365         if (error)
 366                 return (error);
 367
 368         /*
 369          * Do not update this earlier, in case we return with an error.
 370          *
 371          * XXX: this in6_selectsrc_socket result might replace the bound local
 372          * address with the address specified by setsockopt(IPV6_PKTINFO).
 373          * Is it the intended behavior?
 374          */
 375         *plocal_addr6 = in6a;
 376
 377         /*
 378          * Don't do pcblookup call here; return interface in
 379          * plocal_addr6
 380          * and exit to caller, that will do the lookup.
 381          */
 382
 383         return (0);
 384 }
 385
 386 /*
 387  * Outer subroutine:
 388  * Connect from a socket to a specified address.
 389  * Both address and port must be specified in argument sin.
 390  * If don't have a local address for this socket yet,
 391  * then pick one.
 392  */
 393 int
 394 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 395     struct ucred *cred, struct mbuf *m)
 396 {
 397         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 398         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 399         struct in6_addr addr6;
 400         int error;
 401
 402         INP_WLOCK_ASSERT(inp);
 403         INP_HASH_WLOCK_ASSERT(pcbinfo);
 404
 405         /*
 406          * Call inner routine, to assign local interface address.
 407          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 408          */
 409         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 410                 return (error);
 411
 412         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 413                                sin6->sin6_port,
 414                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 415                               ? &addr6 : &inp->in6p_laddr,
 416                               inp->inp_lport, 0, NULL) != NULL) {
 417                 return (EADDRINUSE);
 418         }
 419         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 420                 if (inp->inp_lport == 0) {
 421                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 422                         if (error)
 423                                 return (error);
 424                 }
 425                 inp->in6p_laddr = addr6;
 426         }
 427         inp->in6p_faddr = sin6->sin6_addr;
 428         inp->inp_fport = sin6->sin6_port;
 429         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 430         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 431         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 432                 inp->inp_flow |=
 433                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 434
 435         in_pcbrehash_mbuf(inp, m);
 436
 437         return (0);
 438 }
 439
 440 int
 441 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 442 {
 443
 444         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL));
 445 }
 446
 447 void
 448 in6_pcbdisconnect(struct inpcb *inp)
 449 {
 450
 451         INP_WLOCK_ASSERT(inp);
 452         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 453
 454         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 455         inp->inp_fport = 0;
 456         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 457         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 458         in_pcbrehash(inp);
 459 }
 460
 461 struct sockaddr *
 462 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 463 {
 464         struct sockaddr_in6 *sin6;
 465
 466         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 467         bzero(sin6, sizeof *sin6);
 468         sin6->sin6_family = AF_INET6;
 469         sin6->sin6_len = sizeof(*sin6);
 470         sin6->sin6_port = port;
 471         sin6->sin6_addr = *addr_p;
 472         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 473
 474         return (struct sockaddr *)sin6;
 475 }
 476
 477 struct sockaddr *
 478 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 479 {
 480         struct sockaddr_in sin;
 481         struct sockaddr_in6 *sin6_p;
 482
 483         bzero(&sin, sizeof sin);
 484         sin.sin_family = AF_INET;
 485         sin.sin_len = sizeof(sin);
 486         sin.sin_port = port;
 487         sin.sin_addr = *addr_p;
 488
 489         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 490                 M_WAITOK);
 491         in6_sin_2_v4mapsin6(&sin, sin6_p);
 492
 493         return (struct sockaddr *)sin6_p;
 494 }
 495
 496 int
 497 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 498 {
 499         struct inpcb *inp;
 500         struct in6_addr addr;
 501         in_port_t port;
 502
 503         inp = sotoinpcb(so);
 504         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 505
 506         INP_RLOCK(inp);
 507         port = inp->inp_lport;
 508         addr = inp->in6p_laddr;
 509         INP_RUNLOCK(inp);
 510
 511         *nam = in6_sockaddr(port, &addr);
 512         return 0;
 513 }
 514
 515 int
 516 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 517 {
 518         struct inpcb *inp;
 519         struct in6_addr addr;
 520         in_port_t port;
 521
 522         inp = sotoinpcb(so);
 523         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 524
 525         INP_RLOCK(inp);
 526         port = inp->inp_fport;
 527         addr = inp->in6p_faddr;
 528         INP_RUNLOCK(inp);
 529
 530         *nam = in6_sockaddr(port, &addr);
 531         return 0;
 532 }
 533
 534 int
 535 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 536 {
 537         struct  inpcb *inp;
 538         int     error;
 539
 540         inp = sotoinpcb(so);
 541         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 542
 543 #ifdef INET
 544         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 545                 error = in_getsockaddr(so, nam);
 546                 if (error == 0)
 547                         in6_sin_2_v4mapsin6_in_sock(nam);
 548         } else
 549 #endif
 550         {
 551                 /* scope issues will be handled in in6_getsockaddr(). */
 552                 error = in6_getsockaddr(so, nam);
 553         }
 554
 555         return error;
 556 }
 557
 558 int
 559 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 560 {
 561         struct  inpcb *inp;
 562         int     error;
 563
 564         inp = sotoinpcb(so);
 565         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 566
 567 #ifdef INET
 568         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 569                 error = in_getpeeraddr(so, nam);
 570                 if (error == 0)
 571                         in6_sin_2_v4mapsin6_in_sock(nam);
 572         } else
 573 #endif
 574         /* scope issues will be handled in in6_getpeeraddr(). */
 575         error = in6_getpeeraddr(so, nam);
 576
 577         return error;
 578 }
 579
 580 /*
 581  * Pass some notification to all connections of a protocol
 582  * associated with address dst.  The local address and/or port numbers
 583  * may be specified to limit the search.  The "usual action" will be
 584  * taken, depending on the ctlinput cmd.  The caller must filter any
 585  * cmds that are uninteresting (e.g., no error in the map).
 586  * Call the protocol specific routine (if any) to report
 587  * any errors for each matching socket.
 588  */
 589 void
 590 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 591     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 592     int cmd, void *cmdarg,
 593     struct inpcb *(*notify)(struct inpcb *, int))
 594 {
 595         struct inpcb *inp, *inp_temp;
 596         struct sockaddr_in6 sa6_src, *sa6_dst;
 597         u_short fport = fport_arg, lport = lport_arg;
 598         u_int32_t flowinfo;
 599         int errno;
 600
 601         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 602                 return;
 603
 604         sa6_dst = (struct sockaddr_in6 *)dst;
 605         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 606                 return;
 607
 608         /*
 609          * note that src can be NULL when we get notify by local fragmentation.
 610          */
 611         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 612         flowinfo = sa6_src.sin6_flowinfo;
 613
 614         /*
 615          * Redirects go to all references to the destination,
 616          * and use in6_rtchange to invalidate the route cache.
 617          * Dead host indications: also use in6_rtchange to invalidate
 618          * the cache, and deliver the error to all the sockets.
 619          * Otherwise, if we have knowledge of the local port and address,
 620          * deliver only to that socket.
 621          */
 622         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 623                 fport = 0;
 624                 lport = 0;
 625                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 626
 627                 if (cmd != PRC_HOSTDEAD)
 628                         notify = in6_rtchange;
 629         }
 630         errno = inet6ctlerrmap[cmd];
 631         INP_INFO_WLOCK(pcbinfo);
 632         LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 633                 INP_WLOCK(inp);
 634                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 635                         INP_WUNLOCK(inp);
 636                         continue;
 637                 }
 638
 639                 /*
 640                  * If the error designates a new path MTU for a destination
 641                  * and the application (associated with this socket) wanted to
 642                  * know the value, notify.
 643                  * XXX: should we avoid to notify the value to TCP sockets?
 644                  */
 645                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 646                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 647                                         *(u_int32_t *)cmdarg);
 648
 649                 /*
 650                  * Detect if we should notify the error. If no source and
 651                  * destination ports are specifed, but non-zero flowinfo and
 652                  * local address match, notify the error. This is the case
 653                  * when the error is delivered with an encrypted buffer
 654                  * by ESP. Otherwise, just compare addresses and ports
 655                  * as usual.
 656                  */
 657                 if (lport == 0 && fport == 0 && flowinfo &&
 658                     inp->inp_socket != NULL &&
 659                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 660                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 661                         goto do_notify;
 662                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 663                                              &sa6_dst->sin6_addr) ||
 664                          inp->inp_socket == 0 ||
 665                          (lport && inp->inp_lport != lport) ||
 666                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 667                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 668                                               &sa6_src.sin6_addr)) ||
 669                          (fport && inp->inp_fport != fport)) {
 670                         INP_WUNLOCK(inp);
 671                         continue;
 672                 }
 673
 674           do_notify:
 675                 if (notify) {
 676                         if ((*notify)(inp, errno))
 677                                 INP_WUNLOCK(inp);
 678                 } else
 679                         INP_WUNLOCK(inp);
 680         }
 681         INP_INFO_WUNLOCK(pcbinfo);
 682 }
 683
 684 /*
 685  * Lookup a PCB based on the local address and port.  Caller must hold the
 686  * hash lock.  No inpcb locks or references are acquired.
 687  */
 688 struct inpcb *
 689 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 690     u_short lport, int lookupflags, struct ucred *cred)
 691 {
 692         struct inpcb *inp;
 693         int matchwild = 3, wildcard;
 694
 695         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 696             ("%s: invalid lookup flags %d", __func__, lookupflags));
 697
 698         INP_HASH_WLOCK_ASSERT(pcbinfo);
 699
 700         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 701                 struct inpcbhead *head;
 702                 /*
 703                  * Look for an unconnected (wildcard foreign addr) PCB that
 704                  * matches the local address and port we're looking for.
 705                  */
 706                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 707                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 708                     pcbinfo->ipi_hashmask)];
 709                 LIST_FOREACH(inp, head, inp_hash) {
 710                         /* XXX inp locking */
 711                         if ((inp->inp_vflag & INP_IPV6) == 0)
 712                                 continue;
 713                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 714                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 715                             inp->inp_lport == lport) {
 716                                 /* Found. */
 717                                 if (cred == NULL ||
 718                                     prison_equal_ip6(cred->cr_prison,
 719                                         inp->inp_cred->cr_prison))
 720                                         return (inp);
 721                         }
 722                 }
 723                 /*
 724                  * Not found.
 725                  */
 726                 return (NULL);
 727         } else {
 728                 struct inpcbporthead *porthash;
 729                 struct inpcbport *phd;
 730                 struct inpcb *match = NULL;
 731                 /*
 732                  * Best fit PCB lookup.
 733                  *
 734                  * First see if this local port is in use by looking on the
 735                  * port hash list.
 736                  */
 737                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 738                     pcbinfo->ipi_porthashmask)];
 739                 LIST_FOREACH(phd, porthash, phd_hash) {
 740                         if (phd->phd_port == lport)
 741                                 break;
 742                 }
 743                 if (phd != NULL) {
 744                         /*
 745                          * Port is in use by one or more PCBs. Look for best
 746                          * fit.
 747                          */
 748                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 749                                 wildcard = 0;
 750                                 if (cred != NULL &&
 751                                     !prison_equal_ip6(cred->cr_prison,
 752                                         inp->inp_cred->cr_prison))
 753                                         continue;
 754                                 /* XXX inp locking */
 755                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 756                                         continue;
 757                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 758                                         wildcard++;
 759                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 760                                         &inp->in6p_laddr)) {
 761                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 762                                                 wildcard++;
 763                                         else if (!IN6_ARE_ADDR_EQUAL(
 764                                             &inp->in6p_laddr, laddr))
 765                                                 continue;
 766                                 } else {
 767                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 768                                                 wildcard++;
 769                                 }
 770                                 if (wildcard < matchwild) {
 771                                         match = inp;
 772                                         matchwild = wildcard;
 773                                         if (matchwild == 0)
 774                                                 break;
 775                                 }
 776                         }
 777                 }
 778                 return (match);
 779         }
 780 }
 781
 782 void
 783 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 784 {
 785         struct inpcb *in6p;
 786         struct ip6_moptions *im6o;
 787         int i, gap;
 788
 789         INP_INFO_WLOCK(pcbinfo);
 790         LIST_FOREACH(in6p, pcbinfo->ipi_listhead, inp_list) {
 791                 INP_WLOCK(in6p);
 792                 im6o = in6p->in6p_moptions;
 793                 if ((in6p->inp_vflag & INP_IPV6) && im6o != NULL) {
 794                         /*
 795                          * Unselect the outgoing ifp for multicast if it
 796                          * is being detached.
 797                          */
 798                         if (im6o->im6o_multicast_ifp == ifp)
 799                                 im6o->im6o_multicast_ifp = NULL;
 800                         /*
 801                          * Drop multicast group membership if we joined
 802                          * through the interface being detached.
 803                          */
 804                         gap = 0;
 805                         for (i = 0; i < im6o->im6o_num_memberships; i++) {
 806                                 if (im6o->im6o_membership[i]->in6m_ifp ==
 807                                     ifp) {
 808                                         in6_leavegroup(im6o->im6o_membership[i], NULL);
 809                                         gap++;
 810                                 } else if (gap != 0) {
 811                                         im6o->im6o_membership[i - gap] =
 812                                             im6o->im6o_membership[i];
 813                                 }
 814                         }
 815                         im6o->im6o_num_memberships -= gap;
 816                 }
 817                 INP_WUNLOCK(in6p);
 818         }
 819         INP_INFO_WUNLOCK(pcbinfo);
 820 }
 821
 822 /*
 823  * Check for alternatives when higher level complains
 824  * about service problems.  For now, invalidate cached
 825  * routing information.  If the route was created dynamically
 826  * (by a redirect), time to try a default gateway again.
 827  */
 828 void
 829 in6_losing(struct inpcb *in6p)
 830 {
 831
 832         if (in6p->inp_route6.ro_rt) {
 833                 RTFREE(in6p->inp_route6.ro_rt);
 834                 in6p->inp_route6.ro_rt = (struct rtentry *)NULL;
 835         }
 836         if (in6p->inp_route.ro_lle)
 837                 LLE_FREE(in6p->inp_route.ro_lle);       /* zeros ro_lle */
 838         return;
 839 }
 840
 841 /*
 842  * After a routing change, flush old routing
 843  * and allocate a (hopefully) better one.
 844  */
 845 struct inpcb *
 846 in6_rtchange(struct inpcb *inp, int errno)
 847 {
 848
 849         if (inp->inp_route6.ro_rt) {
 850                 RTFREE(inp->inp_route6.ro_rt);
 851                 inp->inp_route6.ro_rt = (struct rtentry *)NULL;
 852         }
 853         if (inp->inp_route.ro_lle)
 854                 LLE_FREE(inp->inp_route.ro_lle);        /* zeros ro_lle */
 855         return inp;
 856 }
 857
 858 #ifdef PCBGROUP
 859 /*
 860  * Lookup PCB in hash list, using pcbgroup tables.
 861  */
 862 static struct inpcb *
 863 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 864     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 865     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 866 {
 867         struct inpcbhead *head;
 868         struct inpcb *inp, *tmpinp;
 869         u_short fport = fport_arg, lport = lport_arg;
 870         bool locked;
 871
 872         /*
 873          * First look for an exact match.
 874          */
 875         tmpinp = NULL;
 876         INP_GROUP_LOCK(pcbgroup);
 877         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 878             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 879         LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 880                 /* XXX inp locking */
 881                 if ((inp->inp_vflag & INP_IPV6) == 0)
 882                         continue;
 883                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 884                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 885                     inp->inp_fport == fport &&
 886                     inp->inp_lport == lport) {
 887                         /*
 888                          * XXX We should be able to directly return
 889                          * the inp here, without any checks.
 890                          * Well unless both bound with SO_REUSEPORT?
 891                          */
 892                         if (prison_flag(inp->inp_cred, PR_IP6))
 893                                 goto found;
 894                         if (tmpinp == NULL)
 895                                 tmpinp = inp;
 896                 }
 897         }
 898         if (tmpinp != NULL) {
 899                 inp = tmpinp;
 900                 goto found;
 901         }
 902
 903         /*
 904          * Then look for a wildcard match in the pcbgroup.
 905          */
 906         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 907                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 908                 struct inpcb *jail_wild = NULL;
 909                 int injail;
 910
 911                 /*
 912                  * Order of socket selection - we always prefer jails.
 913                  *      1. jailed, non-wild.
 914                  *      2. jailed, wild.
 915                  *      3. non-jailed, non-wild.
 916                  *      4. non-jailed, wild.
 917                  */
 918                 head = &pcbgroup->ipg_hashbase[
 919                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
 920                 LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 921                         /* XXX inp locking */
 922                         if ((inp->inp_vflag & INP_IPV6) == 0)
 923                                 continue;
 924
 925                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 926                             inp->inp_lport != lport) {
 927                                 continue;
 928                         }
 929
 930                         injail = prison_flag(inp->inp_cred, PR_IP6);
 931                         if (injail) {
 932                                 if (prison_check_ip6(inp->inp_cred,
 933                                     laddr) != 0)
 934                                         continue;
 935                         } else {
 936                                 if (local_exact != NULL)
 937                                         continue;
 938                         }
 939
 940                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
 941                                 if (injail)
 942                                         goto found;
 943                                 else
 944                                         local_exact = inp;
 945                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 946                                 if (injail)
 947                                         jail_wild = inp;
 948                                 else
 949                                         local_wild = inp;
 950                         }
 951                 } /* LIST_FOREACH */
 952
 953                 inp = jail_wild;
 954                 if (inp == NULL)
 955                         inp = jail_wild;
 956                 if (inp == NULL)
 957                         inp = local_exact;
 958                 if (inp == NULL)
 959                         inp = local_wild;
 960                 if (inp != NULL)
 961                         goto found;
 962         }
 963
 964         /*
 965          * Then look for a wildcard match, if requested.
 966          */
 967         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 968                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 969                 struct inpcb *jail_wild = NULL;
 970                 int injail;
 971
 972                 /*
 973                  * Order of socket selection - we always prefer jails.
 974                  *      1. jailed, non-wild.
 975                  *      2. jailed, wild.
 976                  *      3. non-jailed, non-wild.
 977                  *      4. non-jailed, wild.
 978                  */
 979                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
 980                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 981                     pcbinfo->ipi_wildmask)];
 982                 LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
 983                         /* XXX inp locking */
 984                         if ((inp->inp_vflag & INP_IPV6) == 0)
 985                                 continue;
 986
 987                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 988                             inp->inp_lport != lport) {
 989                                 continue;
 990                         }
 991
 992                         injail = prison_flag(inp->inp_cred, PR_IP6);
 993                         if (injail) {
 994                                 if (prison_check_ip6(inp->inp_cred,
 995                                     laddr) != 0)
 996                                         continue;
 997                         } else {
 998                                 if (local_exact != NULL)
 999                                         continue;
1000                         }
1001
1002                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1003                                 if (injail)
1004                                         goto found;
1005                                 else
1006                                         local_exact = inp;
1007                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1008                                 if (injail)
1009                                         jail_wild = inp;
1010                                 else
1011                                         local_wild = inp;
1012                         }
1013                 } /* LIST_FOREACH */
1014
1015                 inp = jail_wild;
1016                 if (inp == NULL)
1017                         inp = jail_wild;
1018                 if (inp == NULL)
1019                         inp = local_exact;
1020                 if (inp == NULL)
1021                         inp = local_wild;
1022                 if (inp != NULL)
1023                         goto found;
1024         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1025         INP_GROUP_UNLOCK(pcbgroup);
1026         return (NULL);
1027
1028 found:
1029         if (lookupflags & INPLOOKUP_WLOCKPCB)
1030                 locked = INP_TRY_WLOCK(inp);
1031         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1032                 locked = INP_TRY_RLOCK(inp);
1033         else
1034                 panic("%s: locking buf", __func__);
1035         if (!locked)
1036                 in_pcbref(inp);
1037         INP_GROUP_UNLOCK(pcbgroup);
1038         if (!locked) {
1039                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1040                         INP_WLOCK(inp);
1041                         if (in_pcbrele_wlocked(inp))
1042                                 return (NULL);
1043                 } else {
1044                         INP_RLOCK(inp);
1045                         if (in_pcbrele_rlocked(inp))
1046                                 return (NULL);
1047                 }
1048         }
1049 #ifdef INVARIANTS
1050         if (lookupflags & INPLOOKUP_WLOCKPCB)
1051                 INP_WLOCK_ASSERT(inp);
1052         else
1053                 INP_RLOCK_ASSERT(inp);
1054 #endif
1055         return (inp);
1056 }
1057 #endif /* PCBGROUP */
1058
1059 /*
1060  * Lookup PCB in hash list.
1061  */
1062 static struct inpcb *
1063 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1064     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1065     int lookupflags, struct ifnet *ifp)
1066 {
1067         struct inpcbhead *head;
1068         struct inpcb *inp, *tmpinp;
1069         u_short fport = fport_arg, lport = lport_arg;
1070
1071         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1072             ("%s: invalid lookup flags %d", __func__, lookupflags));
1073
1074         INP_HASH_LOCK_ASSERT(pcbinfo);
1075
1076         /*
1077          * First look for an exact match.
1078          */
1079         tmpinp = NULL;
1080         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1081             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1082         LIST_FOREACH(inp, head, inp_hash) {
1083                 /* XXX inp locking */
1084                 if ((inp->inp_vflag & INP_IPV6) == 0)
1085                         continue;
1086                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1087                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1088                     inp->inp_fport == fport &&
1089                     inp->inp_lport == lport) {
1090                         /*
1091                          * XXX We should be able to directly return
1092                          * the inp here, without any checks.
1093                          * Well unless both bound with SO_REUSEPORT?
1094                          */
1095                         if (prison_flag(inp->inp_cred, PR_IP6))
1096                                 return (inp);
1097                         if (tmpinp == NULL)
1098                                 tmpinp = inp;
1099                 }
1100         }
1101         if (tmpinp != NULL)
1102                 return (tmpinp);
1103
1104         /*
1105          * Then look for a wildcard match, if requested.
1106          */
1107         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1108                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1109                 struct inpcb *jail_wild = NULL;
1110                 int injail;
1111
1112                 /*
1113                  * Order of socket selection - we always prefer jails.
1114                  *      1. jailed, non-wild.
1115                  *      2. jailed, wild.
1116                  *      3. non-jailed, non-wild.
1117                  *      4. non-jailed, wild.
1118                  */
1119                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1120                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1121                     pcbinfo->ipi_hashmask)];
1122                 LIST_FOREACH(inp, head, inp_hash) {
1123                         /* XXX inp locking */
1124                         if ((inp->inp_vflag & INP_IPV6) == 0)
1125                                 continue;
1126
1127                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1128                             inp->inp_lport != lport) {
1129                                 continue;
1130                         }
1131
1132                         injail = prison_flag(inp->inp_cred, PR_IP6);
1133                         if (injail) {
1134                                 if (prison_check_ip6(inp->inp_cred,
1135                                     laddr) != 0)
1136                                         continue;
1137                         } else {
1138                                 if (local_exact != NULL)
1139                                         continue;
1140                         }
1141
1142                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1143                                 if (injail)
1144                                         return (inp);
1145                                 else
1146                                         local_exact = inp;
1147                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1148                                 if (injail)
1149                                         jail_wild = inp;
1150                                 else
1151                                         local_wild = inp;
1152                         }
1153                 } /* LIST_FOREACH */
1154
1155                 if (jail_wild != NULL)
1156                         return (jail_wild);
1157                 if (local_exact != NULL)
1158                         return (local_exact);
1159                 if (local_wild != NULL)
1160                         return (local_wild);
1161         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1162
1163         /*
1164          * Not found.
1165          */
1166         return (NULL);
1167 }
1168
1169 /*
1170  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1171  * hash list lock, and will return the inpcb locked (i.e., requires
1172  * INPLOOKUP_LOCKPCB).
1173  */
1174 static struct inpcb *
1175 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1176     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1177     struct ifnet *ifp)
1178 {
1179         struct inpcb *inp;
1180         bool locked;
1181
1182         INP_HASH_RLOCK(pcbinfo);
1183         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1184             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1185         if (inp != NULL) {
1186                 if (lookupflags & INPLOOKUP_WLOCKPCB)
1187                         locked = INP_TRY_WLOCK(inp);
1188                 else if (lookupflags & INPLOOKUP_RLOCKPCB)
1189                         locked = INP_TRY_RLOCK(inp);
1190                 else
1191                         panic("%s: locking bug", __func__);
1192                 if (!locked)
1193                         in_pcbref(inp);
1194                 INP_HASH_RUNLOCK(pcbinfo);
1195                 if (!locked) {
1196                         if (lookupflags & INPLOOKUP_WLOCKPCB) {
1197                                 INP_WLOCK(inp);
1198                                 if (in_pcbrele_wlocked(inp))
1199                                         return (NULL);
1200                         } else {
1201                                 INP_RLOCK(inp);
1202                                 if (in_pcbrele_rlocked(inp))
1203                                         return (NULL);
1204                         }
1205                 }
1206 #ifdef INVARIANTS
1207                 if (lookupflags & INPLOOKUP_WLOCKPCB)
1208                         INP_WLOCK_ASSERT(inp);
1209                 else
1210                         INP_RLOCK_ASSERT(inp);
1211 #endif
1212         } else
1213                 INP_HASH_RUNLOCK(pcbinfo);
1214         return (inp);
1215 }
1216
1217 /*
1218  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1219  * from which a pre-calculated hash value may be extracted.
1220  *
1221  * Possibly more of this logic should be in in6_pcbgroup.c.
1222  */
1223 struct inpcb *
1224 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1225     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1226 {
1227 #if defined(PCBGROUP) && !defined(RSS)
1228         struct inpcbgroup *pcbgroup;
1229 #endif
1230
1231         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1232             ("%s: invalid lookup flags %d", __func__, lookupflags));
1233         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1234             ("%s: LOCKPCB not set", __func__));
1235
1236         /*
1237          * When not using RSS, use connection groups in preference to the
1238          * reservation table when looking up 4-tuples.  When using RSS, just
1239          * use the reservation table, due to the cost of the Toeplitz hash
1240          * in software.
1241          *
1242          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1243          * we could be doing RSS with a non-Toeplitz hash that is affordable
1244          * in software.
1245          */
1246 #if defined(PCBGROUP) && !defined(RSS)
1247         if (in_pcbgroup_enabled(pcbinfo)) {
1248                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1249                     fport);
1250                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1251                     laddr, lport, lookupflags, ifp));
1252         }
1253 #endif
1254         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1255             lookupflags, ifp));
1256 }
1257
1258 struct inpcb *
1259 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1260     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1261     struct ifnet *ifp, struct mbuf *m)
1262 {
1263 #ifdef PCBGROUP
1264         struct inpcbgroup *pcbgroup;
1265 #endif
1266
1267         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1268             ("%s: invalid lookup flags %d", __func__, lookupflags));
1269         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1270             ("%s: LOCKPCB not set", __func__));
1271
1272 #ifdef PCBGROUP
1273         /*
1274          * If we can use a hardware-generated hash to look up the connection
1275          * group, use that connection group to find the inpcb.  Otherwise
1276          * fall back on a software hash -- or the reservation table if we're
1277          * using RSS.
1278          *
1279          * XXXRW: As above, that policy belongs in the pcbgroup code.
1280          */
1281         if (in_pcbgroup_enabled(pcbinfo) &&
1282             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1283                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1284                     m->m_pkthdr.flowid);
1285                 if (pcbgroup != NULL)
1286                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1287                             fport, laddr, lport, lookupflags, ifp));
1288 #ifndef RSS
1289                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1290                     fport);
1291                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1292                     laddr, lport, lookupflags, ifp));
1293 #endif
1294         }
1295 #endif
1296         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1297             lookupflags, ifp));
1298 }
1299
1300 void
1301 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1302 {
1303         struct ip6_hdr *ip;
1304
1305         ip = mtod(m, struct ip6_hdr *);
1306         bzero(sin6, sizeof(*sin6));
1307         sin6->sin6_len = sizeof(*sin6);
1308         sin6->sin6_family = AF_INET6;
1309         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1310
1311         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1312
1313         return;
1314 }