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 #include "opt_inet.h"
  71 #include "opt_inet6.h"
  72 #include "opt_ipsec.h"
  73 #include "opt_pcbgroup.h"
  74 #include "opt_route.h"
  75 #include "opt_rss.h"
  76
  77 #include <sys/param.h>
  78 #include <sys/systm.h>
  79 #include <sys/malloc.h>
  80 #include <sys/mbuf.h>
  81 #include <sys/domain.h>
  82 #include <sys/protosw.h>
  83 #include <sys/socket.h>
  84 #include <sys/socketvar.h>
  85 #include <sys/sockio.h>
  86 #include <sys/errno.h>
  87 #include <sys/time.h>
  88 #include <sys/priv.h>
  89 #include <sys/proc.h>
  90 #include <sys/jail.h>
  91
  92 #include <vm/uma.h>
  93
  94 #include <net/if.h>
  95 #include <net/if_var.h>
  96 #include <net/if_llatbl.h>
  97 #include <net/if_types.h>
  98 #include <net/route.h>
  99 #include <net/route/nhop.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/in6_fib.h>
 113 #include <netinet6/scope6_var.h>
 114
 115 int
 116 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 117     struct ucred *cred)
 118 {
 119         struct socket *so = inp->inp_socket;
 120         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 121         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 122         u_short lport = 0;
 123         int error, lookupflags = 0;
 124         int reuseport = (so->so_options & SO_REUSEPORT);
 125
 126         /*
 127          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
 128          * so that we don't have to add to the (already messy) code below.
 129          */
 130         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
 131
 132         INP_WLOCK_ASSERT(inp);
 133         INP_HASH_WLOCK_ASSERT(pcbinfo);
 134
 135         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 136                 return (EINVAL);
 137         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
 138                 lookupflags = INPLOOKUP_WILDCARD;
 139         if (nam == NULL) {
 140                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 141                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 142                         return (error);
 143         } else {
 144                 sin6 = (struct sockaddr_in6 *)nam;
 145                 KASSERT(sin6->sin6_family == AF_INET6,
 146                     ("%s: invalid address family for %p", __func__, sin6));
 147                 KASSERT(sin6->sin6_len == sizeof(*sin6),
 148                     ("%s: invalid address length for %p", __func__, sin6));
 149
 150                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 151                         return(error);
 152
 153                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 154                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 155                         return (error);
 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|SO_REUSEPORT)) != 0)
 167                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 168                         /*
 169                          * XXX: How to deal with SO_REUSEPORT_LB here?
 170                          * Treat same as SO_REUSEPORT for now.
 171                          */
 172                         if ((so->so_options &
 173                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
 174                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
 175                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 176                         struct epoch_tracker et;
 177                         struct ifaddr *ifa;
 178
 179                         sin6->sin6_port = 0;            /* yech... */
 180                         NET_EPOCH_ENTER(et);
 181                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 182                             NULL &&
 183                             (inp->inp_flags & INP_BINDANY) == 0) {
 184                                 NET_EPOCH_EXIT(et);
 185                                 return (EADDRNOTAVAIL);
 186                         }
 187
 188                         /*
 189                          * XXX: bind to an anycast address might accidentally
 190                          * cause sending a packet with anycast source address.
 191                          * We should allow to bind to a deprecated address, since
 192                          * the application dares to use it.
 193                          */
 194                         if (ifa != NULL &&
 195                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 196                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 197                                 NET_EPOCH_EXIT(et);
 198                                 return (EADDRNOTAVAIL);
 199                         }
 200                         NET_EPOCH_EXIT(et);
 201                 }
 202                 if (lport) {
 203                         struct inpcb *t;
 204                         struct tcptw *tw;
 205
 206                         /* GROSS */
 207                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 208                             ntohs(lport) >= V_ipport_reservedlow &&
 209                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
 210                                 return (EACCES);
 211                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 212                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
 213                                 t = in6_pcblookup_local(pcbinfo,
 214                                     &sin6->sin6_addr, lport,
 215                                     INPLOOKUP_WILDCARD, cred);
 216                                 if (t &&
 217                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 218                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 219                                     (so->so_type != SOCK_STREAM ||
 220                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 221                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 222                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 223                                      (t->inp_flags2 & INP_REUSEPORT) ||
 224                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 225                                     (inp->inp_cred->cr_uid !=
 226                                      t->inp_cred->cr_uid))
 227                                         return (EADDRINUSE);
 228
 229                                 /*
 230                                  * If the socket is a BINDMULTI socket, then
 231                                  * the credentials need to match and the
 232                                  * original socket also has to have been bound
 233                                  * with BINDMULTI.
 234                                  */
 235                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 236                                         return (EADDRINUSE);
 237
 238 #ifdef INET
 239                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 240                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 241                                         struct sockaddr_in sin;
 242
 243                                         in6_sin6_2_sin(&sin, sin6);
 244                                         t = in_pcblookup_local(pcbinfo,
 245                                             sin.sin_addr, lport,
 246                                             INPLOOKUP_WILDCARD, cred);
 247                                         if (t &&
 248                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 249                                             ((t->inp_flags &
 250                                               INP_TIMEWAIT) == 0) &&
 251                                             (so->so_type != SOCK_STREAM ||
 252                                              ntohl(t->inp_faddr.s_addr) ==
 253                                               INADDR_ANY) &&
 254                                             (inp->inp_cred->cr_uid !=
 255                                              t->inp_cred->cr_uid))
 256                                                 return (EADDRINUSE);
 257
 258                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 259                                                 return (EADDRINUSE);
 260                                 }
 261 #endif
 262                         }
 263                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 264                             lport, lookupflags, cred);
 265                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 266                                 /*
 267                                  * XXXRW: If an incpb has had its timewait
 268                                  * state recycled, we treat the address as
 269                                  * being in use (for now).  This is better
 270                                  * than a panic, but not desirable.
 271                                  */
 272                                 tw = intotw(t);
 273                                 if (tw == NULL ||
 274                                     ((reuseport & tw->tw_so_options) == 0 &&
 275                                          (reuseport_lb & tw->tw_so_options) == 0))
 276                                         return (EADDRINUSE);
 277                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 278                                            (reuseport_lb & inp_so_options(t)) == 0) {
 279                                 return (EADDRINUSE);
 280                         }
 281 #ifdef INET
 282                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 283                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 284                                 struct sockaddr_in sin;
 285
 286                                 in6_sin6_2_sin(&sin, sin6);
 287                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 288                                    lport, lookupflags, cred);
 289                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 290                                         tw = intotw(t);
 291                                         if (tw == NULL)
 292                                                 return (EADDRINUSE);
 293                                         if ((reuseport & tw->tw_so_options) == 0
 294                                             && (reuseport_lb & tw->tw_so_options) == 0
 295                                             && (ntohl(t->inp_laddr.s_addr) !=
 296                                                 INADDR_ANY || ((inp->inp_vflag &
 297                                                         INP_IPV6PROTO) ==
 298                                                     (t->inp_vflag & INP_IPV6PROTO))))
 299                                                 return (EADDRINUSE);
 300                                 } else if (t &&
 301                                     (reuseport & inp_so_options(t)) == 0 &&
 302                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 303                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 304                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 305                                         return (EADDRINUSE);
 306                                 }
 307                         }
 308 #endif
 309                 }
 310                 inp->in6p_laddr = sin6->sin6_addr;
 311         }
 312         if (lport == 0) {
 313                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 314                         /* Undo an address bind that may have occurred. */
 315                         inp->in6p_laddr = in6addr_any;
 316                         return (error);
 317                 }
 318         } else {
 319                 inp->inp_lport = lport;
 320                 if (in_pcbinshash(inp) != 0) {
 321                         inp->in6p_laddr = in6addr_any;
 322                         inp->inp_lport = 0;
 323                         return (EAGAIN);
 324                 }
 325         }
 326         return (0);
 327 }
 328
 329 /*
 330  *   Transform old in6_pcbconnect() into an inner subroutine for new
 331  *   in6_pcbconnect(): Do some validity-checking on the remote
 332  *   address (in mbuf 'nam') and then determine local host address
 333  *   (i.e., which interface) to use to access that remote host.
 334  *
 335  *   This preserves definition of in6_pcbconnect(), while supporting a
 336  *   slightly different version for T/TCP.  (This is more than
 337  *   a bit of a kludge, but cleaning up the internal interfaces would
 338  *   have forced minor changes in every protocol).
 339  */
 340 static int
 341 in6_pcbladdr(struct inpcb *inp, struct sockaddr_in6 *sin6,
 342     struct in6_addr *plocal_addr6)
 343 {
 344         int error = 0;
 345         int scope_ambiguous = 0;
 346         struct in6_addr in6a;
 347         struct epoch_tracker et;
 348
 349         INP_WLOCK_ASSERT(inp);
 350         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 351
 352         if (sin6->sin6_port == 0)
 353                 return (EADDRNOTAVAIL);
 354
 355         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 356                 scope_ambiguous = 1;
 357         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 358                 return(error);
 359
 360         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 361                 /*
 362                  * If the destination address is UNSPECIFIED addr,
 363                  * use the loopback addr, e.g ::1.
 364                  */
 365                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 366                         sin6->sin6_addr = in6addr_loopback;
 367         }
 368         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 369                 return (error);
 370
 371         NET_EPOCH_ENTER(et);
 372         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 373             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 374         NET_EPOCH_EXIT(et);
 375         if (error)
 376                 return (error);
 377         if (IN6_IS_ADDR_UNSPECIFIED(&in6a))
 378                 return (EHOSTUNREACH);
 379
 380         /*
 381          * Do not update this earlier, in case we return with an error.
 382          *
 383          * XXX: this in6_selectsrc_socket result might replace the bound local
 384          * address with the address specified by setsockopt(IPV6_PKTINFO).
 385          * Is it the intended behavior?
 386          */
 387         *plocal_addr6 = in6a;
 388
 389         /*
 390          * Don't do pcblookup call here; return interface in
 391          * plocal_addr6
 392          * and exit to caller, that will do the lookup.
 393          */
 394
 395         return (0);
 396 }
 397
 398 /*
 399  * Outer subroutine:
 400  * Connect from a socket to a specified address.
 401  * Both address and port must be specified in argument sin.
 402  * If don't have a local address for this socket yet,
 403  * then pick one.
 404  */
 405 int
 406 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 407     struct ucred *cred, struct mbuf *m, bool rehash)
 408 {
 409         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 410         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 411         struct sockaddr_in6 laddr6;
 412         int error;
 413
 414         KASSERT(sin6->sin6_family == AF_INET6,
 415             ("%s: invalid address family for %p", __func__, sin6));
 416         KASSERT(sin6->sin6_len == sizeof(*sin6),
 417             ("%s: invalid address length for %p", __func__, sin6));
 418
 419         bzero(&laddr6, sizeof(laddr6));
 420         laddr6.sin6_family = AF_INET6;
 421
 422         INP_WLOCK_ASSERT(inp);
 423         INP_HASH_WLOCK_ASSERT(pcbinfo);
 424
 425 #ifdef ROUTE_MPATH
 426         if (CALC_FLOWID_OUTBOUND) {
 427                 uint32_t hash_type, hash_val;
 428
 429                 hash_val = fib6_calc_software_hash(&inp->in6p_laddr,
 430                     &sin6->sin6_addr, 0, sin6->sin6_port,
 431                     inp->inp_socket->so_proto->pr_protocol, &hash_type);
 432                 inp->inp_flowid = hash_val;
 433                 inp->inp_flowtype = hash_type;
 434         }
 435 #endif
 436         /*
 437          * Call inner routine, to assign local interface address.
 438          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 439          */
 440         if ((error = in6_pcbladdr(inp, sin6, &laddr6.sin6_addr)) != 0)
 441                 return (error);
 442
 443         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 444                                sin6->sin6_port,
 445                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 446                               ? &laddr6.sin6_addr : &inp->in6p_laddr,
 447                               inp->inp_lport, 0, NULL, M_NODOM) != NULL) {
 448                 return (EADDRINUSE);
 449         }
 450         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 451                 if (inp->inp_lport == 0) {
 452                         /*
 453                          * rehash was required to be true in the past for
 454                          * this case; retain that convention.  However,
 455                          * we now call in_pcb_lport_dest rather than
 456                          * in6_pcbbind; the former does not insert into
 457                          * the hash table, the latter does.  Change rehash
 458                          * to false to do the in_pcbinshash below.
 459                          */
 460                         KASSERT(rehash == true,
 461                             ("Rehashing required for unbound inps"));
 462                         rehash = false;
 463                         error = in_pcb_lport_dest(inp,
 464                             (struct sockaddr *) &laddr6, &inp->inp_lport,
 465                             (struct sockaddr *) sin6, sin6->sin6_port, cred,
 466                             INPLOOKUP_WILDCARD);
 467                         if (error)
 468                                 return (error);
 469                 }
 470                 inp->in6p_laddr = laddr6.sin6_addr;
 471         }
 472         inp->in6p_faddr = sin6->sin6_addr;
 473         inp->inp_fport = sin6->sin6_port;
 474         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 475         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 476         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 477                 inp->inp_flow |=
 478                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 479
 480         if (rehash) {
 481                 in_pcbrehash_mbuf(inp, m);
 482         } else {
 483                 in_pcbinshash_mbuf(inp, m);
 484         }
 485
 486         return (0);
 487 }
 488
 489 int
 490 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 491 {
 492
 493         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 494 }
 495
 496 void
 497 in6_pcbdisconnect(struct inpcb *inp)
 498 {
 499
 500         INP_WLOCK_ASSERT(inp);
 501         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 502
 503         memset(&inp->in6p_laddr, 0, sizeof(inp->in6p_laddr));
 504         memset(&inp->in6p_faddr, 0, sizeof(inp->in6p_faddr));
 505         inp->inp_fport = 0;
 506         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 507         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 508         in_pcbrehash(inp);
 509 }
 510
 511 struct sockaddr *
 512 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 513 {
 514         struct sockaddr_in6 *sin6;
 515
 516         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 517         bzero(sin6, sizeof *sin6);
 518         sin6->sin6_family = AF_INET6;
 519         sin6->sin6_len = sizeof(*sin6);
 520         sin6->sin6_port = port;
 521         sin6->sin6_addr = *addr_p;
 522         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 523
 524         return (struct sockaddr *)sin6;
 525 }
 526
 527 struct sockaddr *
 528 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 529 {
 530         struct sockaddr_in sin;
 531         struct sockaddr_in6 *sin6_p;
 532
 533         bzero(&sin, sizeof sin);
 534         sin.sin_family = AF_INET;
 535         sin.sin_len = sizeof(sin);
 536         sin.sin_port = port;
 537         sin.sin_addr = *addr_p;
 538
 539         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 540                 M_WAITOK);
 541         in6_sin_2_v4mapsin6(&sin, sin6_p);
 542
 543         return (struct sockaddr *)sin6_p;
 544 }
 545
 546 int
 547 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 548 {
 549         struct inpcb *inp;
 550         struct in6_addr addr;
 551         in_port_t port;
 552
 553         inp = sotoinpcb(so);
 554         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 555
 556         INP_RLOCK(inp);
 557         port = inp->inp_lport;
 558         addr = inp->in6p_laddr;
 559         INP_RUNLOCK(inp);
 560
 561         *nam = in6_sockaddr(port, &addr);
 562         return 0;
 563 }
 564
 565 int
 566 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 567 {
 568         struct inpcb *inp;
 569         struct in6_addr addr;
 570         in_port_t port;
 571
 572         inp = sotoinpcb(so);
 573         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 574
 575         INP_RLOCK(inp);
 576         port = inp->inp_fport;
 577         addr = inp->in6p_faddr;
 578         INP_RUNLOCK(inp);
 579
 580         *nam = in6_sockaddr(port, &addr);
 581         return 0;
 582 }
 583
 584 int
 585 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 586 {
 587         struct  inpcb *inp;
 588         int     error;
 589
 590         inp = sotoinpcb(so);
 591         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 592
 593 #ifdef INET
 594         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 595                 error = in_getsockaddr(so, nam);
 596                 if (error == 0)
 597                         in6_sin_2_v4mapsin6_in_sock(nam);
 598         } else
 599 #endif
 600         {
 601                 /* scope issues will be handled in in6_getsockaddr(). */
 602                 error = in6_getsockaddr(so, nam);
 603         }
 604
 605         return error;
 606 }
 607
 608 int
 609 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 610 {
 611         struct  inpcb *inp;
 612         int     error;
 613
 614         inp = sotoinpcb(so);
 615         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 616
 617 #ifdef INET
 618         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 619                 error = in_getpeeraddr(so, nam);
 620                 if (error == 0)
 621                         in6_sin_2_v4mapsin6_in_sock(nam);
 622         } else
 623 #endif
 624         /* scope issues will be handled in in6_getpeeraddr(). */
 625         error = in6_getpeeraddr(so, nam);
 626
 627         return error;
 628 }
 629
 630 /*
 631  * Pass some notification to all connections of a protocol
 632  * associated with address dst.  The local address and/or port numbers
 633  * may be specified to limit the search.  The "usual action" will be
 634  * taken, depending on the ctlinput cmd.  The caller must filter any
 635  * cmds that are uninteresting (e.g., no error in the map).
 636  * Call the protocol specific routine (if any) to report
 637  * any errors for each matching socket.
 638  */
 639 void
 640 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 641     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 642     int cmd, void *cmdarg,
 643     struct inpcb *(*notify)(struct inpcb *, int))
 644 {
 645         struct inpcb *inp, *inp_temp;
 646         struct sockaddr_in6 sa6_src, *sa6_dst;
 647         u_short fport = fport_arg, lport = lport_arg;
 648         u_int32_t flowinfo;
 649         int errno;
 650
 651         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 652                 return;
 653
 654         sa6_dst = (struct sockaddr_in6 *)dst;
 655         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 656                 return;
 657
 658         /*
 659          * note that src can be NULL when we get notify by local fragmentation.
 660          */
 661         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 662         flowinfo = sa6_src.sin6_flowinfo;
 663
 664         /*
 665          * Redirects go to all references to the destination,
 666          * and use in6_rtchange to invalidate the route cache.
 667          * Dead host indications: also use in6_rtchange to invalidate
 668          * the cache, and deliver the error to all the sockets.
 669          * Otherwise, if we have knowledge of the local port and address,
 670          * deliver only to that socket.
 671          */
 672         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 673                 fport = 0;
 674                 lport = 0;
 675                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 676
 677                 if (cmd != PRC_HOSTDEAD)
 678                         notify = in6_rtchange;
 679         }
 680         errno = inet6ctlerrmap[cmd];
 681         INP_INFO_WLOCK(pcbinfo);
 682         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 683                 INP_WLOCK(inp);
 684                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 685                         INP_WUNLOCK(inp);
 686                         continue;
 687                 }
 688
 689                 /*
 690                  * If the error designates a new path MTU for a destination
 691                  * and the application (associated with this socket) wanted to
 692                  * know the value, notify.
 693                  * XXX: should we avoid to notify the value to TCP sockets?
 694                  */
 695                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 696                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 697                                         *(u_int32_t *)cmdarg);
 698
 699                 /*
 700                  * Detect if we should notify the error. If no source and
 701                  * destination ports are specifed, but non-zero flowinfo and
 702                  * local address match, notify the error. This is the case
 703                  * when the error is delivered with an encrypted buffer
 704                  * by ESP. Otherwise, just compare addresses and ports
 705                  * as usual.
 706                  */
 707                 if (lport == 0 && fport == 0 && flowinfo &&
 708                     inp->inp_socket != NULL &&
 709                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 710                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 711                         goto do_notify;
 712                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 713                                              &sa6_dst->sin6_addr) ||
 714                          inp->inp_socket == 0 ||
 715                          (lport && inp->inp_lport != lport) ||
 716                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 717                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 718                                               &sa6_src.sin6_addr)) ||
 719                          (fport && inp->inp_fport != fport)) {
 720                         INP_WUNLOCK(inp);
 721                         continue;
 722                 }
 723
 724           do_notify:
 725                 if (notify) {
 726                         if ((*notify)(inp, errno))
 727                                 INP_WUNLOCK(inp);
 728                 } else
 729                         INP_WUNLOCK(inp);
 730         }
 731         INP_INFO_WUNLOCK(pcbinfo);
 732 }
 733
 734 /*
 735  * Lookup a PCB based on the local address and port.  Caller must hold the
 736  * hash lock.  No inpcb locks or references are acquired.
 737  */
 738 struct inpcb *
 739 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 740     u_short lport, int lookupflags, struct ucred *cred)
 741 {
 742         struct inpcb *inp;
 743         int matchwild = 3, wildcard;
 744
 745         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 746             ("%s: invalid lookup flags %d", __func__, lookupflags));
 747
 748         INP_HASH_LOCK_ASSERT(pcbinfo);
 749
 750         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 751                 struct inpcbhead *head;
 752                 /*
 753                  * Look for an unconnected (wildcard foreign addr) PCB that
 754                  * matches the local address and port we're looking for.
 755                  */
 756                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 757                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 758                     pcbinfo->ipi_hashmask)];
 759                 CK_LIST_FOREACH(inp, head, inp_hash) {
 760                         /* XXX inp locking */
 761                         if ((inp->inp_vflag & INP_IPV6) == 0)
 762                                 continue;
 763                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 764                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 765                             inp->inp_lport == lport) {
 766                                 /* Found. */
 767                                 if (cred == NULL ||
 768                                     prison_equal_ip6(cred->cr_prison,
 769                                         inp->inp_cred->cr_prison))
 770                                         return (inp);
 771                         }
 772                 }
 773                 /*
 774                  * Not found.
 775                  */
 776                 return (NULL);
 777         } else {
 778                 struct inpcbporthead *porthash;
 779                 struct inpcbport *phd;
 780                 struct inpcb *match = NULL;
 781                 /*
 782                  * Best fit PCB lookup.
 783                  *
 784                  * First see if this local port is in use by looking on the
 785                  * port hash list.
 786                  */
 787                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 788                     pcbinfo->ipi_porthashmask)];
 789                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 790                         if (phd->phd_port == lport)
 791                                 break;
 792                 }
 793                 if (phd != NULL) {
 794                         /*
 795                          * Port is in use by one or more PCBs. Look for best
 796                          * fit.
 797                          */
 798                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 799                                 wildcard = 0;
 800                                 if (cred != NULL &&
 801                                     !prison_equal_ip6(cred->cr_prison,
 802                                         inp->inp_cred->cr_prison))
 803                                         continue;
 804                                 /* XXX inp locking */
 805                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 806                                         continue;
 807                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 808                                         wildcard++;
 809                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 810                                         &inp->in6p_laddr)) {
 811                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 812                                                 wildcard++;
 813                                         else if (!IN6_ARE_ADDR_EQUAL(
 814                                             &inp->in6p_laddr, laddr))
 815                                                 continue;
 816                                 } else {
 817                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 818                                                 wildcard++;
 819                                 }
 820                                 if (wildcard < matchwild) {
 821                                         match = inp;
 822                                         matchwild = wildcard;
 823                                         if (matchwild == 0)
 824                                                 break;
 825                                 }
 826                         }
 827                 }
 828                 return (match);
 829         }
 830 }
 831
 832 void
 833 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 834 {
 835         struct inpcb *inp;
 836         struct in6_multi *inm;
 837         struct in6_mfilter *imf;
 838         struct ip6_moptions *im6o;
 839
 840         INP_INFO_WLOCK(pcbinfo);
 841         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 842                 INP_WLOCK(inp);
 843                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 844                         INP_WUNLOCK(inp);
 845                         continue;
 846                 }
 847                 im6o = inp->in6p_moptions;
 848                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 849                         /*
 850                          * Unselect the outgoing ifp for multicast if it
 851                          * is being detached.
 852                          */
 853                         if (im6o->im6o_multicast_ifp == ifp)
 854                                 im6o->im6o_multicast_ifp = NULL;
 855                         /*
 856                          * Drop multicast group membership if we joined
 857                          * through the interface being detached.
 858                          */
 859 restart:
 860                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 861                                 if ((inm = imf->im6f_in6m) == NULL)
 862                                         continue;
 863                                 if (inm->in6m_ifp != ifp)
 864                                         continue;
 865                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 866                                 IN6_MULTI_LOCK_ASSERT();
 867                                 in6_leavegroup_locked(inm, NULL);
 868                                 ip6_mfilter_free(imf);
 869                                 goto restart;
 870                         }
 871                 }
 872                 INP_WUNLOCK(inp);
 873         }
 874         INP_INFO_WUNLOCK(pcbinfo);
 875 }
 876
 877 /*
 878  * Check for alternatives when higher level complains
 879  * about service problems.  For now, invalidate cached
 880  * routing information.  If the route was created dynamically
 881  * (by a redirect), time to try a default gateway again.
 882  */
 883 void
 884 in6_losing(struct inpcb *inp)
 885 {
 886
 887         RO_INVALIDATE_CACHE(&inp->inp_route6);
 888 }
 889
 890 /*
 891  * After a routing change, flush old routing
 892  * and allocate a (hopefully) better one.
 893  */
 894 struct inpcb *
 895 in6_rtchange(struct inpcb *inp, int errno __unused)
 896 {
 897
 898         RO_INVALIDATE_CACHE(&inp->inp_route6);
 899         return inp;
 900 }
 901
 902 static struct inpcb *
 903 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 904     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 905     uint16_t fport, int lookupflags, uint8_t numa_domain)
 906 {
 907         struct inpcb *local_wild, *numa_wild;
 908         const struct inpcblbgrouphead *hdr;
 909         struct inpcblbgroup *grp;
 910         uint32_t idx;
 911
 912         INP_HASH_LOCK_ASSERT(pcbinfo);
 913
 914         hdr = &pcbinfo->ipi_lbgrouphashbase[
 915             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 916
 917         /*
 918          * Order of socket selection:
 919          * 1. non-wild.
 920          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 921          *
 922          * NOTE:
 923          * - Load balanced group does not contain jailed sockets.
 924          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 925          */
 926         local_wild = NULL;
 927         numa_wild = NULL;
 928         CK_LIST_FOREACH(grp, hdr, il_list) {
 929 #ifdef INET
 930                 if (!(grp->il_vflag & INP_IPV6))
 931                         continue;
 932 #endif
 933                 if (grp->il_lport != lport)
 934                         continue;
 935
 936                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 937                     fport) % grp->il_inpcnt;
 938                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) {
 939                         if (numa_domain == M_NODOM ||
 940                             grp->il_numa_domain == numa_domain) {
 941                                 return (grp->il_inp[idx]);
 942                         }
 943                         else
 944                                 numa_wild = grp->il_inp[idx];
 945                 }
 946                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 947                     (lookupflags & INPLOOKUP_WILDCARD) != 0 &&
 948                     (local_wild == NULL || numa_domain == M_NODOM ||
 949                         grp->il_numa_domain == numa_domain)) {
 950                         local_wild = grp->il_inp[idx];
 951                 }
 952         }
 953         if (numa_wild != NULL)
 954                 return (numa_wild);
 955         return (local_wild);
 956 }
 957
 958 #ifdef PCBGROUP
 959 /*
 960  * Lookup PCB in hash list, using pcbgroup tables.
 961  */
 962 static struct inpcb *
 963 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 964     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 965     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 966 {
 967         struct inpcbhead *head;
 968         struct inpcb *inp, *tmpinp;
 969         u_short fport = fport_arg, lport = lport_arg;
 970         bool locked;
 971
 972         /*
 973          * First look for an exact match.
 974          */
 975         tmpinp = NULL;
 976         INP_GROUP_LOCK(pcbgroup);
 977         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 978             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 979         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 980                 /* XXX inp locking */
 981                 if ((inp->inp_vflag & INP_IPV6) == 0)
 982                         continue;
 983                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 984                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 985                     inp->inp_fport == fport &&
 986                     inp->inp_lport == lport) {
 987                         /*
 988                          * XXX We should be able to directly return
 989                          * the inp here, without any checks.
 990                          * Well unless both bound with SO_REUSEPORT?
 991                          */
 992                         if (prison_flag(inp->inp_cred, PR_IP6))
 993                                 goto found;
 994                         if (tmpinp == NULL)
 995                                 tmpinp = inp;
 996                 }
 997         }
 998         if (tmpinp != NULL) {
 999                 inp = tmpinp;
1000                 goto found;
1001         }
1002
1003         /*
1004          * Then look for a wildcard match in the pcbgroup.
1005          */
1006         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1007                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1008                 struct inpcb *jail_wild = NULL;
1009                 int injail;
1010
1011                 /*
1012                  * Order of socket selection - we always prefer jails.
1013                  *      1. jailed, non-wild.
1014                  *      2. jailed, wild.
1015                  *      3. non-jailed, non-wild.
1016                  *      4. non-jailed, wild.
1017                  */
1018                 head = &pcbgroup->ipg_hashbase[
1019                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
1020                 CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
1021                         /* XXX inp locking */
1022                         if ((inp->inp_vflag & INP_IPV6) == 0)
1023                                 continue;
1024
1025                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1026                             inp->inp_lport != lport) {
1027                                 continue;
1028                         }
1029
1030                         injail = prison_flag(inp->inp_cred, PR_IP6);
1031                         if (injail) {
1032                                 if (prison_check_ip6(inp->inp_cred,
1033                                     laddr) != 0)
1034                                         continue;
1035                         } else {
1036                                 if (local_exact != NULL)
1037                                         continue;
1038                         }
1039
1040                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1041                                 if (injail)
1042                                         goto found;
1043                                 else
1044                                         local_exact = inp;
1045                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1046                                 if (injail)
1047                                         jail_wild = inp;
1048                                 else
1049                                         local_wild = inp;
1050                         }
1051                 } /* LIST_FOREACH */
1052
1053                 inp = jail_wild;
1054                 if (inp == NULL)
1055                         inp = jail_wild;
1056                 if (inp == NULL)
1057                         inp = local_exact;
1058                 if (inp == NULL)
1059                         inp = local_wild;
1060                 if (inp != NULL)
1061                         goto found;
1062         }
1063
1064         /*
1065          * Then look for a wildcard match, if requested.
1066          */
1067         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1068                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1069                 struct inpcb *jail_wild = NULL;
1070                 int injail;
1071
1072                 /*
1073                  * Order of socket selection - we always prefer jails.
1074                  *      1. jailed, non-wild.
1075                  *      2. jailed, wild.
1076                  *      3. non-jailed, non-wild.
1077                  *      4. non-jailed, wild.
1078                  */
1079                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1080                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1081                     pcbinfo->ipi_wildmask)];
1082                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1083                         /* XXX inp locking */
1084                         if ((inp->inp_vflag & INP_IPV6) == 0)
1085                                 continue;
1086
1087                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1088                             inp->inp_lport != lport) {
1089                                 continue;
1090                         }
1091
1092                         injail = prison_flag(inp->inp_cred, PR_IP6);
1093                         if (injail) {
1094                                 if (prison_check_ip6(inp->inp_cred,
1095                                     laddr) != 0)
1096                                         continue;
1097                         } else {
1098                                 if (local_exact != NULL)
1099                                         continue;
1100                         }
1101
1102                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1103                                 if (injail)
1104                                         goto found;
1105                                 else
1106                                         local_exact = inp;
1107                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1108                                 if (injail)
1109                                         jail_wild = inp;
1110                                 else
1111                                         local_wild = inp;
1112                         }
1113                 } /* LIST_FOREACH */
1114
1115                 inp = jail_wild;
1116                 if (inp == NULL)
1117                         inp = jail_wild;
1118                 if (inp == NULL)
1119                         inp = local_exact;
1120                 if (inp == NULL)
1121                         inp = local_wild;
1122                 if (inp != NULL)
1123                         goto found;
1124         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1125         INP_GROUP_UNLOCK(pcbgroup);
1126         return (NULL);
1127
1128 found:
1129         if (lookupflags & INPLOOKUP_WLOCKPCB)
1130                 locked = INP_TRY_WLOCK(inp);
1131         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1132                 locked = INP_TRY_RLOCK(inp);
1133         else
1134                 panic("%s: locking buf", __func__);
1135         if (!locked)
1136                 in_pcbref(inp);
1137         INP_GROUP_UNLOCK(pcbgroup);
1138         if (!locked) {
1139                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1140                         INP_WLOCK(inp);
1141                         if (in_pcbrele_wlocked(inp))
1142                                 return (NULL);
1143                 } else {
1144                         INP_RLOCK(inp);
1145                         if (in_pcbrele_rlocked(inp))
1146                                 return (NULL);
1147                 }
1148         }
1149 #ifdef INVARIANTS
1150         if (lookupflags & INPLOOKUP_WLOCKPCB)
1151                 INP_WLOCK_ASSERT(inp);
1152         else
1153                 INP_RLOCK_ASSERT(inp);
1154 #endif
1155         return (inp);
1156 }
1157 #endif /* PCBGROUP */
1158
1159 /*
1160  * Lookup PCB in hash list.  Used in in_pcb.c as well as here.
1161  */
1162 struct inpcb *
1163 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1164     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1165     int lookupflags, struct ifnet *ifp, uint8_t numa_domain)
1166 {
1167         struct inpcbhead *head;
1168         struct inpcb *inp, *tmpinp;
1169         u_short fport = fport_arg, lport = lport_arg;
1170
1171         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1172             ("%s: invalid lookup flags %d", __func__, lookupflags));
1173
1174         INP_HASH_LOCK_ASSERT(pcbinfo);
1175
1176         /*
1177          * First look for an exact match.
1178          */
1179         tmpinp = NULL;
1180         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1181             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1182         CK_LIST_FOREACH(inp, head, inp_hash) {
1183                 /* XXX inp locking */
1184                 if ((inp->inp_vflag & INP_IPV6) == 0)
1185                         continue;
1186                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1187                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1188                     inp->inp_fport == fport &&
1189                     inp->inp_lport == lport) {
1190                         /*
1191                          * XXX We should be able to directly return
1192                          * the inp here, without any checks.
1193                          * Well unless both bound with SO_REUSEPORT?
1194                          */
1195                         if (prison_flag(inp->inp_cred, PR_IP6))
1196                                 return (inp);
1197                         if (tmpinp == NULL)
1198                                 tmpinp = inp;
1199                 }
1200         }
1201         if (tmpinp != NULL)
1202                 return (tmpinp);
1203
1204         /*
1205          * Then look in lb group (for wildcard match).
1206          */
1207         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1208                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1209                     fport, lookupflags, numa_domain);
1210                 if (inp != NULL)
1211                         return (inp);
1212         }
1213
1214         /*
1215          * Then look for a wildcard match, if requested.
1216          */
1217         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1218                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1219                 struct inpcb *jail_wild = NULL;
1220                 int injail;
1221
1222                 /*
1223                  * Order of socket selection - we always prefer jails.
1224                  *      1. jailed, non-wild.
1225                  *      2. jailed, wild.
1226                  *      3. non-jailed, non-wild.
1227                  *      4. non-jailed, wild.
1228                  */
1229                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1230                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1231                     pcbinfo->ipi_hashmask)];
1232                 CK_LIST_FOREACH(inp, head, inp_hash) {
1233                         /* XXX inp locking */
1234                         if ((inp->inp_vflag & INP_IPV6) == 0)
1235                                 continue;
1236
1237                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1238                             inp->inp_lport != lport) {
1239                                 continue;
1240                         }
1241
1242                         injail = prison_flag(inp->inp_cred, PR_IP6);
1243                         if (injail) {
1244                                 if (prison_check_ip6(inp->inp_cred,
1245                                     laddr) != 0)
1246                                         continue;
1247                         } else {
1248                                 if (local_exact != NULL)
1249                                         continue;
1250                         }
1251
1252                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1253                                 if (injail)
1254                                         return (inp);
1255                                 else
1256                                         local_exact = inp;
1257                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1258                                 if (injail)
1259                                         jail_wild = inp;
1260                                 else
1261                                         local_wild = inp;
1262                         }
1263                 } /* LIST_FOREACH */
1264
1265                 if (jail_wild != NULL)
1266                         return (jail_wild);
1267                 if (local_exact != NULL)
1268                         return (local_exact);
1269                 if (local_wild != NULL)
1270                         return (local_wild);
1271         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1272
1273         /*
1274          * Not found.
1275          */
1276         return (NULL);
1277 }
1278
1279 /*
1280  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1281  * hash list lock, and will return the inpcb locked (i.e., requires
1282  * INPLOOKUP_LOCKPCB).
1283  */
1284 static struct inpcb *
1285 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1286     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1287     struct ifnet *ifp, uint8_t numa_domain)
1288 {
1289         struct inpcb *inp;
1290
1291         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1292             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp,
1293             numa_domain);
1294         if (inp != NULL) {
1295                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1296                         INP_WLOCK(inp);
1297                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1298                                 INP_WUNLOCK(inp);
1299                                 inp = NULL;
1300                         }
1301                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1302                         INP_RLOCK(inp);
1303                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1304                                 INP_RUNLOCK(inp);
1305                                 inp = NULL;
1306                         }
1307                 } else
1308                         panic("%s: locking bug", __func__);
1309 #ifdef INVARIANTS
1310                 if (inp != NULL) {
1311                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1312                                 INP_WLOCK_ASSERT(inp);
1313                         else
1314                                 INP_RLOCK_ASSERT(inp);
1315                 }
1316 #endif
1317         }
1318         return (inp);
1319 }
1320
1321 /*
1322  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1323  * from which a pre-calculated hash value may be extracted.
1324  *
1325  * Possibly more of this logic should be in in6_pcbgroup.c.
1326  */
1327 struct inpcb *
1328 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1329     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1330 {
1331 #if defined(PCBGROUP) && !defined(RSS)
1332         struct inpcbgroup *pcbgroup;
1333 #endif
1334
1335         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1336             ("%s: invalid lookup flags %d", __func__, lookupflags));
1337         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1338             ("%s: LOCKPCB not set", __func__));
1339
1340         /*
1341          * When not using RSS, use connection groups in preference to the
1342          * reservation table when looking up 4-tuples.  When using RSS, just
1343          * use the reservation table, due to the cost of the Toeplitz hash
1344          * in software.
1345          *
1346          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1347          * we could be doing RSS with a non-Toeplitz hash that is affordable
1348          * in software.
1349          */
1350 #if defined(PCBGROUP) && !defined(RSS)
1351         if (in_pcbgroup_enabled(pcbinfo)) {
1352                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1353                     fport);
1354                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1355                     laddr, lport, lookupflags, ifp));
1356         }
1357 #endif
1358         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1359             lookupflags, ifp, M_NODOM));
1360 }
1361
1362 struct inpcb *
1363 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1364     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1365     struct ifnet *ifp, struct mbuf *m)
1366 {
1367 #ifdef PCBGROUP
1368         struct inpcbgroup *pcbgroup;
1369 #endif
1370
1371         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1372             ("%s: invalid lookup flags %d", __func__, lookupflags));
1373         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1374             ("%s: LOCKPCB not set", __func__));
1375
1376 #ifdef PCBGROUP
1377         /*
1378          * If we can use a hardware-generated hash to look up the connection
1379          * group, use that connection group to find the inpcb.  Otherwise
1380          * fall back on a software hash -- or the reservation table if we're
1381          * using RSS.
1382          *
1383          * XXXRW: As above, that policy belongs in the pcbgroup code.
1384          */
1385         if (in_pcbgroup_enabled(pcbinfo) &&
1386             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1387                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1388                     m->m_pkthdr.flowid);
1389                 if (pcbgroup != NULL)
1390                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1391                             fport, laddr, lport, lookupflags, ifp));
1392 #ifndef RSS
1393                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1394                     fport);
1395                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1396                     laddr, lport, lookupflags, ifp));
1397 #endif
1398         }
1399 #endif
1400         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1401             lookupflags, ifp, m->m_pkthdr.numa_domain));
1402 }
1403
1404 void
1405 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1406 {
1407         struct ip6_hdr *ip;
1408
1409         ip = mtod(m, struct ip6_hdr *);
1410         bzero(sin6, sizeof(*sin6));
1411         sin6->sin6_len = sizeof(*sin6);
1412         sin6->sin6_family = AF_INET6;
1413         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1414
1415         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1416
1417         return;
1418 }