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_route.h"
  77 #include "opt_rss.h"
  78
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/malloc.h>
  82 #include <sys/mbuf.h>
  83 #include <sys/domain.h>
  84 #include <sys/protosw.h>
  85 #include <sys/socket.h>
  86 #include <sys/socketvar.h>
  87 #include <sys/sockio.h>
  88 #include <sys/errno.h>
  89 #include <sys/time.h>
  90 #include <sys/priv.h>
  91 #include <sys/proc.h>
  92 #include <sys/jail.h>
  93
  94 #include <vm/uma.h>
  95
  96 #include <net/if.h>
  97 #include <net/if_var.h>
  98 #include <net/if_llatbl.h>
  99 #include <net/if_types.h>
 100 #include <net/route.h>
 101 #include <net/route/nhop.h>
 102
 103 #include <netinet/in.h>
 104 #include <netinet/in_var.h>
 105 #include <netinet/in_systm.h>
 106 #include <netinet/tcp_var.h>
 107 #include <netinet/ip6.h>
 108 #include <netinet/ip_var.h>
 109
 110 #include <netinet6/ip6_var.h>
 111 #include <netinet6/nd6.h>
 112 #include <netinet/in_pcb.h>
 113 #include <netinet6/in6_pcb.h>
 114 #include <netinet6/in6_fib.h>
 115 #include <netinet6/scope6_var.h>
 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         /*
 129          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
 130          * so that we don't have to add to the (already messy) code below.
 131          */
 132         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
 133
 134         INP_WLOCK_ASSERT(inp);
 135         INP_HASH_WLOCK_ASSERT(pcbinfo);
 136
 137         if (CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) /* XXX broken! */
 138                 return (EADDRNOTAVAIL);
 139         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 140                 return (EINVAL);
 141         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
 142                 lookupflags = INPLOOKUP_WILDCARD;
 143         if (nam == NULL) {
 144                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 145                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 146                         return (error);
 147         } else {
 148                 sin6 = (struct sockaddr_in6 *)nam;
 149                 if (nam->sa_len != sizeof(*sin6))
 150                         return (EINVAL);
 151                 /*
 152                  * family check.
 153                  */
 154                 if (nam->sa_family != AF_INET6)
 155                         return (EAFNOSUPPORT);
 156
 157                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 158                         return(error);
 159
 160                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 161                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 162                         return (error);
 163
 164                 lport = sin6->sin6_port;
 165                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 166                         /*
 167                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 168                          * allow compepte duplication of binding if
 169                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 170                          * and a multicast address is bound on both
 171                          * new and duplicated sockets.
 172                          */
 173                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
 174                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 175                         /*
 176                          * XXX: How to deal with SO_REUSEPORT_LB here?
 177                          * Treat same as SO_REUSEPORT for now.
 178                          */
 179                         if ((so->so_options &
 180                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
 181                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
 182                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 183                         struct epoch_tracker et;
 184                         struct ifaddr *ifa;
 185
 186                         sin6->sin6_port = 0;            /* yech... */
 187                         NET_EPOCH_ENTER(et);
 188                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 189                             NULL &&
 190                             (inp->inp_flags & INP_BINDANY) == 0) {
 191                                 NET_EPOCH_EXIT(et);
 192                                 return (EADDRNOTAVAIL);
 193                         }
 194
 195                         /*
 196                          * XXX: bind to an anycast address might accidentally
 197                          * cause sending a packet with anycast source address.
 198                          * We should allow to bind to a deprecated address, since
 199                          * the application dares to use it.
 200                          */
 201                         if (ifa != NULL &&
 202                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 203                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 204                                 NET_EPOCH_EXIT(et);
 205                                 return (EADDRNOTAVAIL);
 206                         }
 207                         NET_EPOCH_EXIT(et);
 208                 }
 209                 if (lport) {
 210                         struct inpcb *t;
 211                         struct tcptw *tw;
 212
 213                         /* GROSS */
 214                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 215                             ntohs(lport) >= V_ipport_reservedlow &&
 216                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
 217                                 return (EACCES);
 218                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 219                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
 220                                 t = in6_pcblookup_local(pcbinfo,
 221                                     &sin6->sin6_addr, lport,
 222                                     INPLOOKUP_WILDCARD, cred);
 223                                 if (t &&
 224                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 225                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 226                                     (so->so_type != SOCK_STREAM ||
 227                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 228                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 229                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 230                                      (t->inp_flags2 & INP_REUSEPORT) ||
 231                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 232                                     (inp->inp_cred->cr_uid !=
 233                                      t->inp_cred->cr_uid))
 234                                         return (EADDRINUSE);
 235
 236                                 /*
 237                                  * If the socket is a BINDMULTI socket, then
 238                                  * the credentials need to match and the
 239                                  * original socket also has to have been bound
 240                                  * with BINDMULTI.
 241                                  */
 242                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 243                                         return (EADDRINUSE);
 244
 245 #ifdef INET
 246                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 247                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 248                                         struct sockaddr_in sin;
 249
 250                                         in6_sin6_2_sin(&sin, sin6);
 251                                         t = in_pcblookup_local(pcbinfo,
 252                                             sin.sin_addr, lport,
 253                                             INPLOOKUP_WILDCARD, cred);
 254                                         if (t &&
 255                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 256                                             ((t->inp_flags &
 257                                               INP_TIMEWAIT) == 0) &&
 258                                             (so->so_type != SOCK_STREAM ||
 259                                              ntohl(t->inp_faddr.s_addr) ==
 260                                               INADDR_ANY) &&
 261                                             (inp->inp_cred->cr_uid !=
 262                                              t->inp_cred->cr_uid))
 263                                                 return (EADDRINUSE);
 264
 265                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 266                                                 return (EADDRINUSE);
 267                                 }
 268 #endif
 269                         }
 270                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 271                             lport, lookupflags, cred);
 272                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 273                                 /*
 274                                  * XXXRW: If an incpb has had its timewait
 275                                  * state recycled, we treat the address as
 276                                  * being in use (for now).  This is better
 277                                  * than a panic, but not desirable.
 278                                  */
 279                                 tw = intotw(t);
 280                                 if (tw == NULL ||
 281                                     ((reuseport & tw->tw_so_options) == 0 &&
 282                                          (reuseport_lb & tw->tw_so_options) == 0))
 283                                         return (EADDRINUSE);
 284                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 285                                            (reuseport_lb & inp_so_options(t)) == 0) {
 286                                 return (EADDRINUSE);
 287                         }
 288 #ifdef INET
 289                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 290                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 291                                 struct sockaddr_in sin;
 292
 293                                 in6_sin6_2_sin(&sin, sin6);
 294                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 295                                    lport, lookupflags, cred);
 296                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 297                                         tw = intotw(t);
 298                                         if (tw == NULL)
 299                                                 return (EADDRINUSE);
 300                                         if ((reuseport & tw->tw_so_options) == 0
 301                                             && (reuseport_lb & tw->tw_so_options) == 0
 302                                             && (ntohl(t->inp_laddr.s_addr) !=
 303                                                 INADDR_ANY || ((inp->inp_vflag &
 304                                                         INP_IPV6PROTO) ==
 305                                                     (t->inp_vflag & INP_IPV6PROTO))))
 306                                                 return (EADDRINUSE);
 307                                 } else if (t &&
 308                                     (reuseport & inp_so_options(t)) == 0 &&
 309                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 310                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 311                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 312                                         return (EADDRINUSE);
 313                                 }
 314                         }
 315 #endif
 316                 }
 317                 inp->in6p_laddr = sin6->sin6_addr;
 318         }
 319         if (lport == 0) {
 320                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 321                         /* Undo an address bind that may have occurred. */
 322                         inp->in6p_laddr = in6addr_any;
 323                         return (error);
 324                 }
 325         } else {
 326                 inp->inp_lport = lport;
 327                 if (in_pcbinshash(inp) != 0) {
 328                         inp->in6p_laddr = in6addr_any;
 329                         inp->inp_lport = 0;
 330                         return (EAGAIN);
 331                 }
 332         }
 333         return (0);
 334 }
 335
 336 /*
 337  *   Transform old in6_pcbconnect() into an inner subroutine for new
 338  *   in6_pcbconnect(): Do some validity-checking on the remote
 339  *   address (in mbuf 'nam') and then determine local host address
 340  *   (i.e., which interface) to use to access that remote host.
 341  *
 342  *   This preserves definition of in6_pcbconnect(), while supporting a
 343  *   slightly different version for T/TCP.  (This is more than
 344  *   a bit of a kludge, but cleaning up the internal interfaces would
 345  *   have forced minor changes in every protocol).
 346  */
 347 static int
 348 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 349     struct in6_addr *plocal_addr6)
 350 {
 351         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 352         int error = 0;
 353         int scope_ambiguous = 0;
 354         struct in6_addr in6a;
 355         struct epoch_tracker et;
 356
 357         INP_WLOCK_ASSERT(inp);
 358         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 359
 360         if (nam->sa_len != sizeof (*sin6))
 361                 return (EINVAL);
 362         if (sin6->sin6_family != AF_INET6)
 363                 return (EAFNOSUPPORT);
 364         if (sin6->sin6_port == 0)
 365                 return (EADDRNOTAVAIL);
 366
 367         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 368                 scope_ambiguous = 1;
 369         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 370                 return(error);
 371
 372         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 373                 /*
 374                  * If the destination address is UNSPECIFIED addr,
 375                  * use the loopback addr, e.g ::1.
 376                  */
 377                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 378                         sin6->sin6_addr = in6addr_loopback;
 379         }
 380         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 381                 return (error);
 382
 383         NET_EPOCH_ENTER(et);
 384         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 385             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 386         NET_EPOCH_EXIT(et);
 387         if (error)
 388                 return (error);
 389
 390         /*
 391          * Do not update this earlier, in case we return with an error.
 392          *
 393          * XXX: this in6_selectsrc_socket result might replace the bound local
 394          * address with the address specified by setsockopt(IPV6_PKTINFO).
 395          * Is it the intended behavior?
 396          */
 397         *plocal_addr6 = in6a;
 398
 399         /*
 400          * Don't do pcblookup call here; return interface in
 401          * plocal_addr6
 402          * and exit to caller, that will do the lookup.
 403          */
 404
 405         return (0);
 406 }
 407
 408 /*
 409  * Outer subroutine:
 410  * Connect from a socket to a specified address.
 411  * Both address and port must be specified in argument sin.
 412  * If don't have a local address for this socket yet,
 413  * then pick one.
 414  */
 415 int
 416 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 417     struct ucred *cred, struct mbuf *m, bool rehash)
 418 {
 419         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 420         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 421         struct sockaddr_in6 laddr6;
 422         int error;
 423
 424         bzero(&laddr6, sizeof(laddr6));
 425         laddr6.sin6_family = AF_INET6;
 426
 427         INP_WLOCK_ASSERT(inp);
 428         INP_HASH_WLOCK_ASSERT(pcbinfo);
 429
 430 #ifdef ROUTE_MPATH
 431         if (CALC_FLOWID_OUTBOUND) {
 432                 uint32_t hash_type, hash_val;
 433
 434                 hash_val = fib6_calc_software_hash(&inp->in6p_laddr,
 435                     &sin6->sin6_addr, 0, sin6->sin6_port,
 436                     inp->inp_socket->so_proto->pr_protocol, &hash_type);
 437                 inp->inp_flowid = hash_val;
 438                 inp->inp_flowtype = hash_type;
 439         }
 440 #endif
 441         /*
 442          * Call inner routine, to assign local interface address.
 443          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 444          */
 445         if ((error = in6_pcbladdr(inp, nam, &laddr6.sin6_addr)) != 0)
 446                 return (error);
 447
 448         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 449                                sin6->sin6_port,
 450                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 451                               ? &laddr6.sin6_addr : &inp->in6p_laddr,
 452                               inp->inp_lport, 0, NULL, M_NODOM) != NULL) {
 453                 return (EADDRINUSE);
 454         }
 455         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 456                 if (inp->inp_lport == 0) {
 457                         /*
 458                          * rehash was required to be true in the past for
 459                          * this case; retain that convention.  However,
 460                          * we now call in_pcb_lport_dest rather than
 461                          * in6_pcbbind; the former does not insert into
 462                          * the hash table, the latter does.  Change rehash
 463                          * to false to do the in_pcbinshash below.
 464                          */
 465                         KASSERT(rehash == true,
 466                             ("Rehashing required for unbound inps"));
 467                         rehash = false;
 468                         error = in_pcb_lport_dest(inp,
 469                             (struct sockaddr *) &laddr6, &inp->inp_lport,
 470                             (struct sockaddr *) sin6, sin6->sin6_port, cred,
 471                             INPLOOKUP_WILDCARD);
 472                         if (error)
 473                                 return (error);
 474                 }
 475                 inp->in6p_laddr = laddr6.sin6_addr;
 476         }
 477         inp->in6p_faddr = sin6->sin6_addr;
 478         inp->inp_fport = sin6->sin6_port;
 479         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 480         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 481         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 482                 inp->inp_flow |=
 483                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 484
 485         if (rehash) {
 486                 in_pcbrehash_mbuf(inp, m);
 487         } else {
 488                 in_pcbinshash_mbuf(inp, m);
 489         }
 490
 491         return (0);
 492 }
 493
 494 int
 495 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 496 {
 497
 498         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 499 }
 500
 501 void
 502 in6_pcbdisconnect(struct inpcb *inp)
 503 {
 504
 505         INP_WLOCK_ASSERT(inp);
 506         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 507
 508         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 509         inp->inp_fport = 0;
 510         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 511         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 512         in_pcbrehash(inp);
 513 }
 514
 515 struct sockaddr *
 516 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 517 {
 518         struct sockaddr_in6 *sin6;
 519
 520         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 521         bzero(sin6, sizeof *sin6);
 522         sin6->sin6_family = AF_INET6;
 523         sin6->sin6_len = sizeof(*sin6);
 524         sin6->sin6_port = port;
 525         sin6->sin6_addr = *addr_p;
 526         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 527
 528         return (struct sockaddr *)sin6;
 529 }
 530
 531 struct sockaddr *
 532 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 533 {
 534         struct sockaddr_in sin;
 535         struct sockaddr_in6 *sin6_p;
 536
 537         bzero(&sin, sizeof sin);
 538         sin.sin_family = AF_INET;
 539         sin.sin_len = sizeof(sin);
 540         sin.sin_port = port;
 541         sin.sin_addr = *addr_p;
 542
 543         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 544                 M_WAITOK);
 545         in6_sin_2_v4mapsin6(&sin, sin6_p);
 546
 547         return (struct sockaddr *)sin6_p;
 548 }
 549
 550 int
 551 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 552 {
 553         struct inpcb *inp;
 554         struct in6_addr addr;
 555         in_port_t port;
 556
 557         inp = sotoinpcb(so);
 558         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 559
 560         INP_RLOCK(inp);
 561         port = inp->inp_lport;
 562         addr = inp->in6p_laddr;
 563         INP_RUNLOCK(inp);
 564
 565         *nam = in6_sockaddr(port, &addr);
 566         return 0;
 567 }
 568
 569 int
 570 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 571 {
 572         struct inpcb *inp;
 573         struct in6_addr addr;
 574         in_port_t port;
 575
 576         inp = sotoinpcb(so);
 577         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 578
 579         INP_RLOCK(inp);
 580         port = inp->inp_fport;
 581         addr = inp->in6p_faddr;
 582         INP_RUNLOCK(inp);
 583
 584         *nam = in6_sockaddr(port, &addr);
 585         return 0;
 586 }
 587
 588 int
 589 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 590 {
 591         struct  inpcb *inp;
 592         int     error;
 593
 594         inp = sotoinpcb(so);
 595         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 596
 597 #ifdef INET
 598         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 599                 error = in_getsockaddr(so, nam);
 600                 if (error == 0)
 601                         in6_sin_2_v4mapsin6_in_sock(nam);
 602         } else
 603 #endif
 604         {
 605                 /* scope issues will be handled in in6_getsockaddr(). */
 606                 error = in6_getsockaddr(so, nam);
 607         }
 608
 609         return error;
 610 }
 611
 612 int
 613 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 614 {
 615         struct  inpcb *inp;
 616         int     error;
 617
 618         inp = sotoinpcb(so);
 619         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 620
 621 #ifdef INET
 622         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 623                 error = in_getpeeraddr(so, nam);
 624                 if (error == 0)
 625                         in6_sin_2_v4mapsin6_in_sock(nam);
 626         } else
 627 #endif
 628         /* scope issues will be handled in in6_getpeeraddr(). */
 629         error = in6_getpeeraddr(so, nam);
 630
 631         return error;
 632 }
 633
 634 /*
 635  * Pass some notification to all connections of a protocol
 636  * associated with address dst.  The local address and/or port numbers
 637  * may be specified to limit the search.  The "usual action" will be
 638  * taken, depending on the ctlinput cmd.  The caller must filter any
 639  * cmds that are uninteresting (e.g., no error in the map).
 640  * Call the protocol specific routine (if any) to report
 641  * any errors for each matching socket.
 642  */
 643 void
 644 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 645     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 646     int cmd, void *cmdarg,
 647     struct inpcb *(*notify)(struct inpcb *, int))
 648 {
 649         struct inpcb *inp, *inp_temp;
 650         struct sockaddr_in6 sa6_src, *sa6_dst;
 651         u_short fport = fport_arg, lport = lport_arg;
 652         u_int32_t flowinfo;
 653         int errno;
 654
 655         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 656                 return;
 657
 658         sa6_dst = (struct sockaddr_in6 *)dst;
 659         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 660                 return;
 661
 662         /*
 663          * note that src can be NULL when we get notify by local fragmentation.
 664          */
 665         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 666         flowinfo = sa6_src.sin6_flowinfo;
 667
 668         /*
 669          * Redirects go to all references to the destination,
 670          * and use in6_rtchange to invalidate the route cache.
 671          * Dead host indications: also use in6_rtchange to invalidate
 672          * the cache, and deliver the error to all the sockets.
 673          * Otherwise, if we have knowledge of the local port and address,
 674          * deliver only to that socket.
 675          */
 676         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 677                 fport = 0;
 678                 lport = 0;
 679                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 680
 681                 if (cmd != PRC_HOSTDEAD)
 682                         notify = in6_rtchange;
 683         }
 684         errno = inet6ctlerrmap[cmd];
 685         INP_INFO_WLOCK(pcbinfo);
 686         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 687                 INP_WLOCK(inp);
 688                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 689                         INP_WUNLOCK(inp);
 690                         continue;
 691                 }
 692
 693                 /*
 694                  * If the error designates a new path MTU for a destination
 695                  * and the application (associated with this socket) wanted to
 696                  * know the value, notify.
 697                  * XXX: should we avoid to notify the value to TCP sockets?
 698                  */
 699                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 700                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 701                                         *(u_int32_t *)cmdarg);
 702
 703                 /*
 704                  * Detect if we should notify the error. If no source and
 705                  * destination ports are specifed, but non-zero flowinfo and
 706                  * local address match, notify the error. This is the case
 707                  * when the error is delivered with an encrypted buffer
 708                  * by ESP. Otherwise, just compare addresses and ports
 709                  * as usual.
 710                  */
 711                 if (lport == 0 && fport == 0 && flowinfo &&
 712                     inp->inp_socket != NULL &&
 713                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 714                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 715                         goto do_notify;
 716                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 717                                              &sa6_dst->sin6_addr) ||
 718                          inp->inp_socket == 0 ||
 719                          (lport && inp->inp_lport != lport) ||
 720                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 721                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 722                                               &sa6_src.sin6_addr)) ||
 723                          (fport && inp->inp_fport != fport)) {
 724                         INP_WUNLOCK(inp);
 725                         continue;
 726                 }
 727
 728           do_notify:
 729                 if (notify) {
 730                         if ((*notify)(inp, errno))
 731                                 INP_WUNLOCK(inp);
 732                 } else
 733                         INP_WUNLOCK(inp);
 734         }
 735         INP_INFO_WUNLOCK(pcbinfo);
 736 }
 737
 738 /*
 739  * Lookup a PCB based on the local address and port.  Caller must hold the
 740  * hash lock.  No inpcb locks or references are acquired.
 741  */
 742 struct inpcb *
 743 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 744     u_short lport, int lookupflags, struct ucred *cred)
 745 {
 746         struct inpcb *inp;
 747         int matchwild = 3, wildcard;
 748
 749         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 750             ("%s: invalid lookup flags %d", __func__, lookupflags));
 751
 752         INP_HASH_LOCK_ASSERT(pcbinfo);
 753
 754         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 755                 struct inpcbhead *head;
 756                 /*
 757                  * Look for an unconnected (wildcard foreign addr) PCB that
 758                  * matches the local address and port we're looking for.
 759                  */
 760                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 761                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 762                     pcbinfo->ipi_hashmask)];
 763                 CK_LIST_FOREACH(inp, head, inp_hash) {
 764                         /* XXX inp locking */
 765                         if ((inp->inp_vflag & INP_IPV6) == 0)
 766                                 continue;
 767                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 768                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 769                             inp->inp_lport == lport) {
 770                                 /* Found. */
 771                                 if (cred == NULL ||
 772                                     prison_equal_ip6(cred->cr_prison,
 773                                         inp->inp_cred->cr_prison))
 774                                         return (inp);
 775                         }
 776                 }
 777                 /*
 778                  * Not found.
 779                  */
 780                 return (NULL);
 781         } else {
 782                 struct inpcbporthead *porthash;
 783                 struct inpcbport *phd;
 784                 struct inpcb *match = NULL;
 785                 /*
 786                  * Best fit PCB lookup.
 787                  *
 788                  * First see if this local port is in use by looking on the
 789                  * port hash list.
 790                  */
 791                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 792                     pcbinfo->ipi_porthashmask)];
 793                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 794                         if (phd->phd_port == lport)
 795                                 break;
 796                 }
 797                 if (phd != NULL) {
 798                         /*
 799                          * Port is in use by one or more PCBs. Look for best
 800                          * fit.
 801                          */
 802                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 803                                 wildcard = 0;
 804                                 if (cred != NULL &&
 805                                     !prison_equal_ip6(cred->cr_prison,
 806                                         inp->inp_cred->cr_prison))
 807                                         continue;
 808                                 /* XXX inp locking */
 809                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 810                                         continue;
 811                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 812                                         wildcard++;
 813                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 814                                         &inp->in6p_laddr)) {
 815                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 816                                                 wildcard++;
 817                                         else if (!IN6_ARE_ADDR_EQUAL(
 818                                             &inp->in6p_laddr, laddr))
 819                                                 continue;
 820                                 } else {
 821                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 822                                                 wildcard++;
 823                                 }
 824                                 if (wildcard < matchwild) {
 825                                         match = inp;
 826                                         matchwild = wildcard;
 827                                         if (matchwild == 0)
 828                                                 break;
 829                                 }
 830                         }
 831                 }
 832                 return (match);
 833         }
 834 }
 835
 836 void
 837 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 838 {
 839         struct inpcb *inp;
 840         struct in6_multi *inm;
 841         struct in6_mfilter *imf;
 842         struct ip6_moptions *im6o;
 843
 844         INP_INFO_WLOCK(pcbinfo);
 845         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 846                 INP_WLOCK(inp);
 847                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 848                         INP_WUNLOCK(inp);
 849                         continue;
 850                 }
 851                 im6o = inp->in6p_moptions;
 852                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 853                         /*
 854                          * Unselect the outgoing ifp for multicast if it
 855                          * is being detached.
 856                          */
 857                         if (im6o->im6o_multicast_ifp == ifp)
 858                                 im6o->im6o_multicast_ifp = NULL;
 859                         /*
 860                          * Drop multicast group membership if we joined
 861                          * through the interface being detached.
 862                          */
 863 restart:
 864                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 865                                 if ((inm = imf->im6f_in6m) == NULL)
 866                                         continue;
 867                                 if (inm->in6m_ifp != ifp)
 868                                         continue;
 869                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 870                                 IN6_MULTI_LOCK_ASSERT();
 871                                 in6_leavegroup_locked(inm, NULL);
 872                                 ip6_mfilter_free(imf);
 873                                 goto restart;
 874                         }
 875                 }
 876                 INP_WUNLOCK(inp);
 877         }
 878         INP_INFO_WUNLOCK(pcbinfo);
 879 }
 880
 881 /*
 882  * Check for alternatives when higher level complains
 883  * about service problems.  For now, invalidate cached
 884  * routing information.  If the route was created dynamically
 885  * (by a redirect), time to try a default gateway again.
 886  */
 887 void
 888 in6_losing(struct inpcb *inp)
 889 {
 890
 891         RO_INVALIDATE_CACHE(&inp->inp_route6);
 892 }
 893
 894 /*
 895  * After a routing change, flush old routing
 896  * and allocate a (hopefully) better one.
 897  */
 898 struct inpcb *
 899 in6_rtchange(struct inpcb *inp, int errno __unused)
 900 {
 901
 902         RO_INVALIDATE_CACHE(&inp->inp_route6);
 903         return inp;
 904 }
 905
 906 static struct inpcb *
 907 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 908     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 909     uint16_t fport, int lookupflags, uint8_t numa_domain)
 910 {
 911         struct inpcb *local_wild, *numa_wild;
 912         const struct inpcblbgrouphead *hdr;
 913         struct inpcblbgroup *grp;
 914         uint32_t idx;
 915
 916         INP_HASH_LOCK_ASSERT(pcbinfo);
 917
 918         hdr = &pcbinfo->ipi_lbgrouphashbase[
 919             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 920
 921         /*
 922          * Order of socket selection:
 923          * 1. non-wild.
 924          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 925          *
 926          * NOTE:
 927          * - Load balanced group does not contain jailed sockets.
 928          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 929          */
 930         local_wild = NULL;
 931         numa_wild = NULL;
 932         CK_LIST_FOREACH(grp, hdr, il_list) {
 933 #ifdef INET
 934                 if (!(grp->il_vflag & INP_IPV6))
 935                         continue;
 936 #endif
 937                 if (grp->il_lport != lport)
 938                         continue;
 939
 940                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 941                     fport) % grp->il_inpcnt;
 942                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) {
 943                         if (numa_domain == M_NODOM ||
 944                             grp->il_numa_domain == numa_domain) {
 945                                 return (grp->il_inp[idx]);
 946                         }
 947                         else
 948                                 numa_wild = grp->il_inp[idx];
 949                 }
 950                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 951                     (lookupflags & INPLOOKUP_WILDCARD) != 0 &&
 952                     (local_wild == NULL || numa_domain == M_NODOM ||
 953                         grp->il_numa_domain == numa_domain)) {
 954                         local_wild = grp->il_inp[idx];
 955                 }
 956         }
 957         if (numa_wild != NULL)
 958                 return (numa_wild);
 959         return (local_wild);
 960 }
 961
 962 #ifdef PCBGROUP
 963 /*
 964  * Lookup PCB in hash list, using pcbgroup tables.
 965  */
 966 static struct inpcb *
 967 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 968     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 969     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 970 {
 971         struct inpcbhead *head;
 972         struct inpcb *inp, *tmpinp;
 973         u_short fport = fport_arg, lport = lport_arg;
 974         bool locked;
 975
 976         /*
 977          * First look for an exact match.
 978          */
 979         tmpinp = NULL;
 980         INP_GROUP_LOCK(pcbgroup);
 981         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 982             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 983         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 984                 /* XXX inp locking */
 985                 if ((inp->inp_vflag & INP_IPV6) == 0)
 986                         continue;
 987                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 988                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 989                     inp->inp_fport == fport &&
 990                     inp->inp_lport == lport) {
 991                         /*
 992                          * XXX We should be able to directly return
 993                          * the inp here, without any checks.
 994                          * Well unless both bound with SO_REUSEPORT?
 995                          */
 996                         if (prison_flag(inp->inp_cred, PR_IP6))
 997                                 goto found;
 998                         if (tmpinp == NULL)
 999                                 tmpinp = inp;
1000                 }
1001         }
1002         if (tmpinp != NULL) {
1003                 inp = tmpinp;
1004                 goto found;
1005         }
1006
1007         /*
1008          * Then look for a wildcard match in the pcbgroup.
1009          */
1010         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1011                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1012                 struct inpcb *jail_wild = NULL;
1013                 int injail;
1014
1015                 /*
1016                  * Order of socket selection - we always prefer jails.
1017                  *      1. jailed, non-wild.
1018                  *      2. jailed, wild.
1019                  *      3. non-jailed, non-wild.
1020                  *      4. non-jailed, wild.
1021                  */
1022                 head = &pcbgroup->ipg_hashbase[
1023                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
1024                 CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
1025                         /* XXX inp locking */
1026                         if ((inp->inp_vflag & INP_IPV6) == 0)
1027                                 continue;
1028
1029                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1030                             inp->inp_lport != lport) {
1031                                 continue;
1032                         }
1033
1034                         injail = prison_flag(inp->inp_cred, PR_IP6);
1035                         if (injail) {
1036                                 if (prison_check_ip6(inp->inp_cred,
1037                                     laddr) != 0)
1038                                         continue;
1039                         } else {
1040                                 if (local_exact != NULL)
1041                                         continue;
1042                         }
1043
1044                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1045                                 if (injail)
1046                                         goto found;
1047                                 else
1048                                         local_exact = inp;
1049                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1050                                 if (injail)
1051                                         jail_wild = inp;
1052                                 else
1053                                         local_wild = inp;
1054                         }
1055                 } /* LIST_FOREACH */
1056
1057                 inp = jail_wild;
1058                 if (inp == NULL)
1059                         inp = jail_wild;
1060                 if (inp == NULL)
1061                         inp = local_exact;
1062                 if (inp == NULL)
1063                         inp = local_wild;
1064                 if (inp != NULL)
1065                         goto found;
1066         }
1067
1068         /*
1069          * Then look for a wildcard match, if requested.
1070          */
1071         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1072                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1073                 struct inpcb *jail_wild = NULL;
1074                 int injail;
1075
1076                 /*
1077                  * Order of socket selection - we always prefer jails.
1078                  *      1. jailed, non-wild.
1079                  *      2. jailed, wild.
1080                  *      3. non-jailed, non-wild.
1081                  *      4. non-jailed, wild.
1082                  */
1083                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1084                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1085                     pcbinfo->ipi_wildmask)];
1086                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1087                         /* XXX inp locking */
1088                         if ((inp->inp_vflag & INP_IPV6) == 0)
1089                                 continue;
1090
1091                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1092                             inp->inp_lport != lport) {
1093                                 continue;
1094                         }
1095
1096                         injail = prison_flag(inp->inp_cred, PR_IP6);
1097                         if (injail) {
1098                                 if (prison_check_ip6(inp->inp_cred,
1099                                     laddr) != 0)
1100                                         continue;
1101                         } else {
1102                                 if (local_exact != NULL)
1103                                         continue;
1104                         }
1105
1106                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1107                                 if (injail)
1108                                         goto found;
1109                                 else
1110                                         local_exact = inp;
1111                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1112                                 if (injail)
1113                                         jail_wild = inp;
1114                                 else
1115                                         local_wild = inp;
1116                         }
1117                 } /* LIST_FOREACH */
1118
1119                 inp = jail_wild;
1120                 if (inp == NULL)
1121                         inp = jail_wild;
1122                 if (inp == NULL)
1123                         inp = local_exact;
1124                 if (inp == NULL)
1125                         inp = local_wild;
1126                 if (inp != NULL)
1127                         goto found;
1128         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1129         INP_GROUP_UNLOCK(pcbgroup);
1130         return (NULL);
1131
1132 found:
1133         if (lookupflags & INPLOOKUP_WLOCKPCB)
1134                 locked = INP_TRY_WLOCK(inp);
1135         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1136                 locked = INP_TRY_RLOCK(inp);
1137         else
1138                 panic("%s: locking buf", __func__);
1139         if (!locked)
1140                 in_pcbref(inp);
1141         INP_GROUP_UNLOCK(pcbgroup);
1142         if (!locked) {
1143                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1144                         INP_WLOCK(inp);
1145                         if (in_pcbrele_wlocked(inp))
1146                                 return (NULL);
1147                 } else {
1148                         INP_RLOCK(inp);
1149                         if (in_pcbrele_rlocked(inp))
1150                                 return (NULL);
1151                 }
1152         }
1153 #ifdef INVARIANTS
1154         if (lookupflags & INPLOOKUP_WLOCKPCB)
1155                 INP_WLOCK_ASSERT(inp);
1156         else
1157                 INP_RLOCK_ASSERT(inp);
1158 #endif
1159         return (inp);
1160 }
1161 #endif /* PCBGROUP */
1162
1163 /*
1164  * Lookup PCB in hash list.  Used in in_pcb.c as well as here.
1165  */
1166 struct inpcb *
1167 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1168     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1169     int lookupflags, struct ifnet *ifp, uint8_t numa_domain)
1170 {
1171         struct inpcbhead *head;
1172         struct inpcb *inp, *tmpinp;
1173         u_short fport = fport_arg, lport = lport_arg;
1174
1175         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1176             ("%s: invalid lookup flags %d", __func__, lookupflags));
1177
1178         INP_HASH_LOCK_ASSERT(pcbinfo);
1179
1180         /*
1181          * First look for an exact match.
1182          */
1183         tmpinp = NULL;
1184         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1185             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1186         CK_LIST_FOREACH(inp, head, inp_hash) {
1187                 /* XXX inp locking */
1188                 if ((inp->inp_vflag & INP_IPV6) == 0)
1189                         continue;
1190                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1191                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1192                     inp->inp_fport == fport &&
1193                     inp->inp_lport == lport) {
1194                         /*
1195                          * XXX We should be able to directly return
1196                          * the inp here, without any checks.
1197                          * Well unless both bound with SO_REUSEPORT?
1198                          */
1199                         if (prison_flag(inp->inp_cred, PR_IP6))
1200                                 return (inp);
1201                         if (tmpinp == NULL)
1202                                 tmpinp = inp;
1203                 }
1204         }
1205         if (tmpinp != NULL)
1206                 return (tmpinp);
1207
1208         /*
1209          * Then look in lb group (for wildcard match).
1210          */
1211         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1212                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1213                     fport, lookupflags, numa_domain);
1214                 if (inp != NULL)
1215                         return (inp);
1216         }
1217
1218         /*
1219          * Then look for a wildcard match, if requested.
1220          */
1221         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1222                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1223                 struct inpcb *jail_wild = NULL;
1224                 int injail;
1225
1226                 /*
1227                  * Order of socket selection - we always prefer jails.
1228                  *      1. jailed, non-wild.
1229                  *      2. jailed, wild.
1230                  *      3. non-jailed, non-wild.
1231                  *      4. non-jailed, wild.
1232                  */
1233                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1234                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1235                     pcbinfo->ipi_hashmask)];
1236                 CK_LIST_FOREACH(inp, head, inp_hash) {
1237                         /* XXX inp locking */
1238                         if ((inp->inp_vflag & INP_IPV6) == 0)
1239                                 continue;
1240
1241                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1242                             inp->inp_lport != lport) {
1243                                 continue;
1244                         }
1245
1246                         injail = prison_flag(inp->inp_cred, PR_IP6);
1247                         if (injail) {
1248                                 if (prison_check_ip6(inp->inp_cred,
1249                                     laddr) != 0)
1250                                         continue;
1251                         } else {
1252                                 if (local_exact != NULL)
1253                                         continue;
1254                         }
1255
1256                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1257                                 if (injail)
1258                                         return (inp);
1259                                 else
1260                                         local_exact = inp;
1261                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1262                                 if (injail)
1263                                         jail_wild = inp;
1264                                 else
1265                                         local_wild = inp;
1266                         }
1267                 } /* LIST_FOREACH */
1268
1269                 if (jail_wild != NULL)
1270                         return (jail_wild);
1271                 if (local_exact != NULL)
1272                         return (local_exact);
1273                 if (local_wild != NULL)
1274                         return (local_wild);
1275         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1276
1277         /*
1278          * Not found.
1279          */
1280         return (NULL);
1281 }
1282
1283 /*
1284  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1285  * hash list lock, and will return the inpcb locked (i.e., requires
1286  * INPLOOKUP_LOCKPCB).
1287  */
1288 static struct inpcb *
1289 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1290     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1291     struct ifnet *ifp, uint8_t numa_domain)
1292 {
1293         struct inpcb *inp;
1294
1295         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1296             lookupflags & INPLOOKUP_WILDCARD, ifp, numa_domain);
1297         if (inp != NULL) {
1298                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1299                         INP_WLOCK(inp);
1300                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1301                         INP_RLOCK(inp);
1302                 } else if (lookupflags & INPLOOKUP_RLOCKLISTEN) {
1303                         if (inp->inp_socket != NULL &&
1304                             SOLISTENING(inp->inp_socket))
1305                                 INP_RLOCK(inp);
1306                         else
1307                                 INP_WLOCK(inp);
1308                 } else
1309                         panic("%s: locking bug", __func__);
1310                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1311                         INP_UNLOCK(inp);
1312                         inp = NULL;
1313                 }
1314         }
1315         return (inp);
1316 }
1317
1318 /*
1319  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1320  * from which a pre-calculated hash value may be extracted.
1321  *
1322  * Possibly more of this logic should be in in6_pcbgroup.c.
1323  */
1324 struct inpcb *
1325 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1326     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1327 {
1328 #if defined(PCBGROUP) && !defined(RSS)
1329         struct inpcbgroup *pcbgroup;
1330 #endif
1331
1332         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1333             ("%s: invalid lookup flags %d", __func__, lookupflags));
1334         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB |
1335             INPLOOKUP_RLOCKLISTEN)) != 0, ("%s: LOCKPCB not set", __func__));
1336
1337         /*
1338          * When not using RSS, use connection groups in preference to the
1339          * reservation table when looking up 4-tuples.  When using RSS, just
1340          * use the reservation table, due to the cost of the Toeplitz hash
1341          * in software.
1342          *
1343          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1344          * we could be doing RSS with a non-Toeplitz hash that is affordable
1345          * in software.
1346          */
1347 #if defined(PCBGROUP) && !defined(RSS)
1348         if (in_pcbgroup_enabled(pcbinfo)) {
1349                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1350                     fport);
1351                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1352                     laddr, lport, lookupflags, ifp));
1353         }
1354 #endif
1355         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1356             lookupflags, ifp, M_NODOM));
1357 }
1358
1359 struct inpcb *
1360 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1361     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1362     struct ifnet *ifp, struct mbuf *m)
1363 {
1364 #ifdef PCBGROUP
1365         struct inpcbgroup *pcbgroup;
1366 #endif
1367
1368         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1369             ("%s: invalid lookup flags %d", __func__, lookupflags));
1370         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB |
1371             INPLOOKUP_RLOCKLISTEN)) != 0,
1372             ("%s: LOCKPCB not set", __func__));
1373
1374 #ifdef PCBGROUP
1375         /*
1376          * If we can use a hardware-generated hash to look up the connection
1377          * group, use that connection group to find the inpcb.  Otherwise
1378          * fall back on a software hash -- or the reservation table if we're
1379          * using RSS.
1380          *
1381          * XXXRW: As above, that policy belongs in the pcbgroup code.
1382          */
1383         if (in_pcbgroup_enabled(pcbinfo) &&
1384             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1385                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1386                     m->m_pkthdr.flowid);
1387                 if (pcbgroup != NULL)
1388                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1389                             fport, laddr, lport, lookupflags, ifp));
1390 #ifndef RSS
1391                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1392                     fport);
1393                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1394                     laddr, lport, lookupflags, ifp));
1395 #endif
1396         }
1397 #endif
1398         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1399             lookupflags, ifp, m->m_pkthdr.numa_domain));
1400 }
1401
1402 void
1403 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1404 {
1405         struct ip6_hdr *ip;
1406
1407         ip = mtod(m, struct ip6_hdr *);
1408         bzero(sin6, sizeof(*sin6));
1409         sin6->sin6_len = sizeof(*sin6);
1410         sin6->sin6_family = AF_INET6;
1411         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1412
1413         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1414
1415         return;
1416 }