sys/netinet6/in6_pcb.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Robert N. M. Watson under
   9  * contract to Juniper Networks, Inc.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  * 3. Neither the name of the project nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  36  */
  37
  38 /*-
  39  * Copyright (c) 1982, 1986, 1991, 1993
  40  *      The Regents of the University of California.  All rights reserved.
  41  *
  42  * Redistribution and use in source and binary forms, with or without
  43  * modification, are permitted provided that the following conditions
  44  * are met:
  45  * 1. Redistributions of source code must retain the above copyright
  46  *    notice, this list of conditions and the following disclaimer.
  47  * 2. Redistributions in binary form must reproduce the above copyright
  48  *    notice, this list of conditions and the following disclaimer in the
  49  *    documentation and/or other materials provided with the distribution.
  50  * 3. Neither the name of the University nor the names of its contributors
  51  *    may be used to endorse or promote products derived from this software
  52  *    without specific prior written permission.
  53  *
  54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  64  * SUCH DAMAGE.
  65  *
  66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  67  */
  68
  69 #include <sys/cdefs.h>
  70 __FBSDID("$FreeBSD$");
  71
  72 #include "opt_inet.h"
  73 #include "opt_inet6.h"
  74 #include "opt_ipsec.h"
  75 #include "opt_pcbgroup.h"
  76 #include "opt_rss.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/malloc.h>
  81 #include <sys/mbuf.h>
  82 #include <sys/domain.h>
  83 #include <sys/protosw.h>
  84 #include <sys/socket.h>
  85 #include <sys/socketvar.h>
  86 #include <sys/sockio.h>
  87 #include <sys/errno.h>
  88 #include <sys/time.h>
  89 #include <sys/priv.h>
  90 #include <sys/proc.h>
  91 #include <sys/jail.h>
  92
  93 #include <vm/uma.h>
  94
  95 #include <net/if.h>
  96 #include <net/if_var.h>
  97 #include <net/if_llatbl.h>
  98 #include <net/if_types.h>
  99 #include <net/route.h>
 100 #include <net/route/nhop.h>
 101
 102 #include <netinet/in.h>
 103 #include <netinet/in_var.h>
 104 #include <netinet/in_systm.h>
 105 #include <netinet/tcp_var.h>
 106 #include <netinet/ip6.h>
 107 #include <netinet/ip_var.h>
 108
 109 #include <netinet6/ip6_var.h>
 110 #include <netinet6/nd6.h>
 111 #include <netinet/in_pcb.h>
 112 #include <netinet6/in6_pcb.h>
 113 #include <netinet6/in6_fib.h>
 114 #include <netinet6/scope6_var.h>
 115
 116 int
 117 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 118     struct ucred *cred)
 119 {
 120         struct socket *so = inp->inp_socket;
 121         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 122         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 123         u_short lport = 0;
 124         int error, lookupflags = 0;
 125         int reuseport = (so->so_options & SO_REUSEPORT);
 126
 127         /*
 128          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
 129          * so that we don't have to add to the (already messy) code below.
 130          */
 131         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
 132
 133         INP_WLOCK_ASSERT(inp);
 134         INP_HASH_WLOCK_ASSERT(pcbinfo);
 135
 136         if (CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) /* XXX broken! */
 137                 return (EADDRNOTAVAIL);
 138         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 139                 return (EINVAL);
 140         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
 141                 lookupflags = INPLOOKUP_WILDCARD;
 142         if (nam == NULL) {
 143                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 144                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 145                         return (error);
 146         } else {
 147                 sin6 = (struct sockaddr_in6 *)nam;
 148                 if (nam->sa_len != sizeof(*sin6))
 149                         return (EINVAL);
 150                 /*
 151                  * family check.
 152                  */
 153                 if (nam->sa_family != AF_INET6)
 154                         return (EAFNOSUPPORT);
 155
 156                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 157                         return(error);
 158
 159                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 160                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 161                         return (error);
 162
 163                 lport = sin6->sin6_port;
 164                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 165                         /*
 166                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 167                          * allow compepte duplication of binding if
 168                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 169                          * and a multicast address is bound on both
 170                          * new and duplicated sockets.
 171                          */
 172                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
 173                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 174                         /*
 175                          * XXX: How to deal with SO_REUSEPORT_LB here?
 176                          * Treat same as SO_REUSEPORT for now.
 177                          */
 178                         if ((so->so_options &
 179                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
 180                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
 181                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 182                         struct epoch_tracker et;
 183                         struct ifaddr *ifa;
 184
 185                         sin6->sin6_port = 0;            /* yech... */
 186                         NET_EPOCH_ENTER(et);
 187                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 188                             NULL &&
 189                             (inp->inp_flags & INP_BINDANY) == 0) {
 190                                 NET_EPOCH_EXIT(et);
 191                                 return (EADDRNOTAVAIL);
 192                         }
 193
 194                         /*
 195                          * XXX: bind to an anycast address might accidentally
 196                          * cause sending a packet with anycast source address.
 197                          * We should allow to bind to a deprecated address, since
 198                          * the application dares to use it.
 199                          */
 200                         if (ifa != NULL &&
 201                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 202                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 203                                 NET_EPOCH_EXIT(et);
 204                                 return (EADDRNOTAVAIL);
 205                         }
 206                         NET_EPOCH_EXIT(et);
 207                 }
 208                 if (lport) {
 209                         struct inpcb *t;
 210                         struct tcptw *tw;
 211
 212                         /* GROSS */
 213                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 214                             ntohs(lport) >= V_ipport_reservedlow &&
 215                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
 216                                 return (EACCES);
 217                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 218                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
 219                                 t = in6_pcblookup_local(pcbinfo,
 220                                     &sin6->sin6_addr, lport,
 221                                     INPLOOKUP_WILDCARD, cred);
 222                                 if (t &&
 223                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 224                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 225                                     (so->so_type != SOCK_STREAM ||
 226                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 227                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 228                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 229                                      (t->inp_flags2 & INP_REUSEPORT) ||
 230                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 231                                     (inp->inp_cred->cr_uid !=
 232                                      t->inp_cred->cr_uid))
 233                                         return (EADDRINUSE);
 234
 235                                 /*
 236                                  * If the socket is a BINDMULTI socket, then
 237                                  * the credentials need to match and the
 238                                  * original socket also has to have been bound
 239                                  * with BINDMULTI.
 240                                  */
 241                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 242                                         return (EADDRINUSE);
 243
 244 #ifdef INET
 245                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 246                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 247                                         struct sockaddr_in sin;
 248
 249                                         in6_sin6_2_sin(&sin, sin6);
 250                                         t = in_pcblookup_local(pcbinfo,
 251                                             sin.sin_addr, lport,
 252                                             INPLOOKUP_WILDCARD, cred);
 253                                         if (t &&
 254                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 255                                             ((t->inp_flags &
 256                                               INP_TIMEWAIT) == 0) &&
 257                                             (so->so_type != SOCK_STREAM ||
 258                                              ntohl(t->inp_faddr.s_addr) ==
 259                                               INADDR_ANY) &&
 260                                             (inp->inp_cred->cr_uid !=
 261                                              t->inp_cred->cr_uid))
 262                                                 return (EADDRINUSE);
 263
 264                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 265                                                 return (EADDRINUSE);
 266                                 }
 267 #endif
 268                         }
 269                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 270                             lport, lookupflags, cred);
 271                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 272                                 /*
 273                                  * XXXRW: If an incpb has had its timewait
 274                                  * state recycled, we treat the address as
 275                                  * being in use (for now).  This is better
 276                                  * than a panic, but not desirable.
 277                                  */
 278                                 tw = intotw(t);
 279                                 if (tw == NULL ||
 280                                     ((reuseport & tw->tw_so_options) == 0 &&
 281                                          (reuseport_lb & tw->tw_so_options) == 0))
 282                                         return (EADDRINUSE);
 283                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 284                                            (reuseport_lb & inp_so_options(t)) == 0) {
 285                                 return (EADDRINUSE);
 286                         }
 287 #ifdef INET
 288                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 289                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 290                                 struct sockaddr_in sin;
 291
 292                                 in6_sin6_2_sin(&sin, sin6);
 293                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 294                                    lport, lookupflags, cred);
 295                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 296                                         tw = intotw(t);
 297                                         if (tw == NULL)
 298                                                 return (EADDRINUSE);
 299                                         if ((reuseport & tw->tw_so_options) == 0
 300                                             && (reuseport_lb & tw->tw_so_options) == 0
 301                                             && (ntohl(t->inp_laddr.s_addr) !=
 302                                                 INADDR_ANY || ((inp->inp_vflag &
 303                                                         INP_IPV6PROTO) ==
 304                                                     (t->inp_vflag & INP_IPV6PROTO))))
 305                                                 return (EADDRINUSE);
 306                                 } else if (t &&
 307                                     (reuseport & inp_so_options(t)) == 0 &&
 308                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 309                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 310                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 311                                         return (EADDRINUSE);
 312                                 }
 313                         }
 314 #endif
 315                 }
 316                 inp->in6p_laddr = sin6->sin6_addr;
 317         }
 318         if (lport == 0) {
 319                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 320                         /* Undo an address bind that may have occurred. */
 321                         inp->in6p_laddr = in6addr_any;
 322                         return (error);
 323                 }
 324         } else {
 325                 inp->inp_lport = lport;
 326                 if (in_pcbinshash(inp) != 0) {
 327                         inp->in6p_laddr = in6addr_any;
 328                         inp->inp_lport = 0;
 329                         return (EAGAIN);
 330                 }
 331         }
 332         return (0);
 333 }
 334
 335 /*
 336  *   Transform old in6_pcbconnect() into an inner subroutine for new
 337  *   in6_pcbconnect(): Do some validity-checking on the remote
 338  *   address (in mbuf 'nam') and then determine local host address
 339  *   (i.e., which interface) to use to access that remote host.
 340  *
 341  *   This preserves definition of in6_pcbconnect(), while supporting a
 342  *   slightly different version for T/TCP.  (This is more than
 343  *   a bit of a kludge, but cleaning up the internal interfaces would
 344  *   have forced minor changes in every protocol).
 345  */
 346 static int
 347 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 348     struct in6_addr *plocal_addr6)
 349 {
 350         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 351         int error = 0;
 352         int scope_ambiguous = 0;
 353         struct in6_addr in6a;
 354
 355         INP_WLOCK_ASSERT(inp);
 356         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 357
 358         if (nam->sa_len != sizeof (*sin6))
 359                 return (EINVAL);
 360         if (sin6->sin6_family != AF_INET6)
 361                 return (EAFNOSUPPORT);
 362         if (sin6->sin6_port == 0)
 363                 return (EADDRNOTAVAIL);
 364
 365         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 366                 scope_ambiguous = 1;
 367         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 368                 return(error);
 369
 370         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 371                 /*
 372                  * If the destination address is UNSPECIFIED addr,
 373                  * use the loopback addr, e.g ::1.
 374                  */
 375                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 376                         sin6->sin6_addr = in6addr_loopback;
 377         }
 378         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 379                 return (error);
 380
 381         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 382             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 383         if (error)
 384                 return (error);
 385
 386         /*
 387          * Do not update this earlier, in case we return with an error.
 388          *
 389          * XXX: this in6_selectsrc_socket result might replace the bound local
 390          * address with the address specified by setsockopt(IPV6_PKTINFO).
 391          * Is it the intended behavior?
 392          */
 393         *plocal_addr6 = in6a;
 394
 395         /*
 396          * Don't do pcblookup call here; return interface in
 397          * plocal_addr6
 398          * and exit to caller, that will do the lookup.
 399          */
 400
 401         return (0);
 402 }
 403
 404 /*
 405  * Outer subroutine:
 406  * Connect from a socket to a specified address.
 407  * Both address and port must be specified in argument sin.
 408  * If don't have a local address for this socket yet,
 409  * then pick one.
 410  */
 411 int
 412 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 413     struct ucred *cred, struct mbuf *m, bool rehash)
 414 {
 415         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 416         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 417         struct sockaddr_in6 laddr6;
 418         int error;
 419
 420         bzero(&laddr6, sizeof(laddr6));
 421         laddr6.sin6_family = AF_INET6;
 422
 423         INP_WLOCK_ASSERT(inp);
 424         INP_HASH_WLOCK_ASSERT(pcbinfo);
 425
 426         /*
 427          * Call inner routine, to assign local interface address.
 428          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 429          */
 430         if ((error = in6_pcbladdr(inp, nam, &laddr6.sin6_addr)) != 0)
 431                 return (error);
 432
 433         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 434                                sin6->sin6_port,
 435                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 436                               ? &laddr6.sin6_addr : &inp->in6p_laddr,
 437                               inp->inp_lport, 0, NULL) != NULL) {
 438                 return (EADDRINUSE);
 439         }
 440         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 441                 if (inp->inp_lport == 0) {
 442                         /*
 443                          * rehash was required to be true in the past for
 444                          * this case; retain that convention.  However,
 445                          * we now call in_pcb_lport_dest rather than
 446                          * in6_pcbbind; the former does not insert into
 447                          * the hash table, the latter does.  Change rehash
 448                          * to false to do the in_pcbinshash below.
 449                          */
 450                         KASSERT(rehash == true,
 451                             ("Rehashing required for unbound inps"));
 452                         rehash = false;
 453                         error = in_pcb_lport_dest(inp,
 454                             (struct sockaddr *) &laddr6, &inp->inp_lport,
 455                             (struct sockaddr *) sin6, sin6->sin6_port, cred, 0);
 456                         if (error)
 457                                 return (error);
 458                 }
 459                 inp->in6p_laddr = laddr6.sin6_addr;
 460         }
 461         inp->in6p_faddr = sin6->sin6_addr;
 462         inp->inp_fport = sin6->sin6_port;
 463         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 464         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 465         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 466                 inp->inp_flow |=
 467                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 468
 469         if (rehash) {
 470                 in_pcbrehash_mbuf(inp, m);
 471         } else {
 472                 in_pcbinshash_mbuf(inp, m);
 473         }
 474
 475         return (0);
 476 }
 477
 478 int
 479 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 480 {
 481
 482         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 483 }
 484
 485 void
 486 in6_pcbdisconnect(struct inpcb *inp)
 487 {
 488
 489         INP_WLOCK_ASSERT(inp);
 490         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 491
 492         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 493         inp->inp_fport = 0;
 494         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 495         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 496         in_pcbrehash(inp);
 497 }
 498
 499 struct sockaddr *
 500 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 501 {
 502         struct sockaddr_in6 *sin6;
 503
 504         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 505         bzero(sin6, sizeof *sin6);
 506         sin6->sin6_family = AF_INET6;
 507         sin6->sin6_len = sizeof(*sin6);
 508         sin6->sin6_port = port;
 509         sin6->sin6_addr = *addr_p;
 510         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 511
 512         return (struct sockaddr *)sin6;
 513 }
 514
 515 struct sockaddr *
 516 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 517 {
 518         struct sockaddr_in sin;
 519         struct sockaddr_in6 *sin6_p;
 520
 521         bzero(&sin, sizeof sin);
 522         sin.sin_family = AF_INET;
 523         sin.sin_len = sizeof(sin);
 524         sin.sin_port = port;
 525         sin.sin_addr = *addr_p;
 526
 527         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 528                 M_WAITOK);
 529         in6_sin_2_v4mapsin6(&sin, sin6_p);
 530
 531         return (struct sockaddr *)sin6_p;
 532 }
 533
 534 int
 535 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 536 {
 537         struct inpcb *inp;
 538         struct in6_addr addr;
 539         in_port_t port;
 540
 541         inp = sotoinpcb(so);
 542         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 543
 544         INP_RLOCK(inp);
 545         port = inp->inp_lport;
 546         addr = inp->in6p_laddr;
 547         INP_RUNLOCK(inp);
 548
 549         *nam = in6_sockaddr(port, &addr);
 550         return 0;
 551 }
 552
 553 int
 554 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 555 {
 556         struct inpcb *inp;
 557         struct in6_addr addr;
 558         in_port_t port;
 559
 560         inp = sotoinpcb(so);
 561         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 562
 563         INP_RLOCK(inp);
 564         port = inp->inp_fport;
 565         addr = inp->in6p_faddr;
 566         INP_RUNLOCK(inp);
 567
 568         *nam = in6_sockaddr(port, &addr);
 569         return 0;
 570 }
 571
 572 int
 573 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 574 {
 575         struct  inpcb *inp;
 576         int     error;
 577
 578         inp = sotoinpcb(so);
 579         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 580
 581 #ifdef INET
 582         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 583                 error = in_getsockaddr(so, nam);
 584                 if (error == 0)
 585                         in6_sin_2_v4mapsin6_in_sock(nam);
 586         } else
 587 #endif
 588         {
 589                 /* scope issues will be handled in in6_getsockaddr(). */
 590                 error = in6_getsockaddr(so, nam);
 591         }
 592
 593         return error;
 594 }
 595
 596 int
 597 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 598 {
 599         struct  inpcb *inp;
 600         int     error;
 601
 602         inp = sotoinpcb(so);
 603         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 604
 605 #ifdef INET
 606         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 607                 error = in_getpeeraddr(so, nam);
 608                 if (error == 0)
 609                         in6_sin_2_v4mapsin6_in_sock(nam);
 610         } else
 611 #endif
 612         /* scope issues will be handled in in6_getpeeraddr(). */
 613         error = in6_getpeeraddr(so, nam);
 614
 615         return error;
 616 }
 617
 618 /*
 619  * Pass some notification to all connections of a protocol
 620  * associated with address dst.  The local address and/or port numbers
 621  * may be specified to limit the search.  The "usual action" will be
 622  * taken, depending on the ctlinput cmd.  The caller must filter any
 623  * cmds that are uninteresting (e.g., no error in the map).
 624  * Call the protocol specific routine (if any) to report
 625  * any errors for each matching socket.
 626  */
 627 void
 628 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 629     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 630     int cmd, void *cmdarg,
 631     struct inpcb *(*notify)(struct inpcb *, int))
 632 {
 633         struct inpcb *inp, *inp_temp;
 634         struct sockaddr_in6 sa6_src, *sa6_dst;
 635         u_short fport = fport_arg, lport = lport_arg;
 636         u_int32_t flowinfo;
 637         int errno;
 638
 639         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 640                 return;
 641
 642         sa6_dst = (struct sockaddr_in6 *)dst;
 643         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 644                 return;
 645
 646         /*
 647          * note that src can be NULL when we get notify by local fragmentation.
 648          */
 649         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 650         flowinfo = sa6_src.sin6_flowinfo;
 651
 652         /*
 653          * Redirects go to all references to the destination,
 654          * and use in6_rtchange to invalidate the route cache.
 655          * Dead host indications: also use in6_rtchange to invalidate
 656          * the cache, and deliver the error to all the sockets.
 657          * Otherwise, if we have knowledge of the local port and address,
 658          * deliver only to that socket.
 659          */
 660         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 661                 fport = 0;
 662                 lport = 0;
 663                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 664
 665                 if (cmd != PRC_HOSTDEAD)
 666                         notify = in6_rtchange;
 667         }
 668         errno = inet6ctlerrmap[cmd];
 669         INP_INFO_WLOCK(pcbinfo);
 670         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 671                 INP_WLOCK(inp);
 672                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 673                         INP_WUNLOCK(inp);
 674                         continue;
 675                 }
 676
 677                 /*
 678                  * If the error designates a new path MTU for a destination
 679                  * and the application (associated with this socket) wanted to
 680                  * know the value, notify.
 681                  * XXX: should we avoid to notify the value to TCP sockets?
 682                  */
 683                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 684                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 685                                         *(u_int32_t *)cmdarg);
 686
 687                 /*
 688                  * Detect if we should notify the error. If no source and
 689                  * destination ports are specifed, but non-zero flowinfo and
 690                  * local address match, notify the error. This is the case
 691                  * when the error is delivered with an encrypted buffer
 692                  * by ESP. Otherwise, just compare addresses and ports
 693                  * as usual.
 694                  */
 695                 if (lport == 0 && fport == 0 && flowinfo &&
 696                     inp->inp_socket != NULL &&
 697                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 698                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 699                         goto do_notify;
 700                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 701                                              &sa6_dst->sin6_addr) ||
 702                          inp->inp_socket == 0 ||
 703                          (lport && inp->inp_lport != lport) ||
 704                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 705                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 706                                               &sa6_src.sin6_addr)) ||
 707                          (fport && inp->inp_fport != fport)) {
 708                         INP_WUNLOCK(inp);
 709                         continue;
 710                 }
 711
 712           do_notify:
 713                 if (notify) {
 714                         if ((*notify)(inp, errno))
 715                                 INP_WUNLOCK(inp);
 716                 } else
 717                         INP_WUNLOCK(inp);
 718         }
 719         INP_INFO_WUNLOCK(pcbinfo);
 720 }
 721
 722 /*
 723  * Lookup a PCB based on the local address and port.  Caller must hold the
 724  * hash lock.  No inpcb locks or references are acquired.
 725  */
 726 struct inpcb *
 727 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 728     u_short lport, int lookupflags, struct ucred *cred)
 729 {
 730         struct inpcb *inp;
 731         int matchwild = 3, wildcard;
 732
 733         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 734             ("%s: invalid lookup flags %d", __func__, lookupflags));
 735
 736         INP_HASH_LOCK_ASSERT(pcbinfo);
 737
 738         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 739                 struct inpcbhead *head;
 740                 /*
 741                  * Look for an unconnected (wildcard foreign addr) PCB that
 742                  * matches the local address and port we're looking for.
 743                  */
 744                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 745                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 746                     pcbinfo->ipi_hashmask)];
 747                 CK_LIST_FOREACH(inp, head, inp_hash) {
 748                         /* XXX inp locking */
 749                         if ((inp->inp_vflag & INP_IPV6) == 0)
 750                                 continue;
 751                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 752                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 753                             inp->inp_lport == lport) {
 754                                 /* Found. */
 755                                 if (cred == NULL ||
 756                                     prison_equal_ip6(cred->cr_prison,
 757                                         inp->inp_cred->cr_prison))
 758                                         return (inp);
 759                         }
 760                 }
 761                 /*
 762                  * Not found.
 763                  */
 764                 return (NULL);
 765         } else {
 766                 struct inpcbporthead *porthash;
 767                 struct inpcbport *phd;
 768                 struct inpcb *match = NULL;
 769                 /*
 770                  * Best fit PCB lookup.
 771                  *
 772                  * First see if this local port is in use by looking on the
 773                  * port hash list.
 774                  */
 775                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 776                     pcbinfo->ipi_porthashmask)];
 777                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 778                         if (phd->phd_port == lport)
 779                                 break;
 780                 }
 781                 if (phd != NULL) {
 782                         /*
 783                          * Port is in use by one or more PCBs. Look for best
 784                          * fit.
 785                          */
 786                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 787                                 wildcard = 0;
 788                                 if (cred != NULL &&
 789                                     !prison_equal_ip6(cred->cr_prison,
 790                                         inp->inp_cred->cr_prison))
 791                                         continue;
 792                                 /* XXX inp locking */
 793                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 794                                         continue;
 795                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 796                                         wildcard++;
 797                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 798                                         &inp->in6p_laddr)) {
 799                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 800                                                 wildcard++;
 801                                         else if (!IN6_ARE_ADDR_EQUAL(
 802                                             &inp->in6p_laddr, laddr))
 803                                                 continue;
 804                                 } else {
 805                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 806                                                 wildcard++;
 807                                 }
 808                                 if (wildcard < matchwild) {
 809                                         match = inp;
 810                                         matchwild = wildcard;
 811                                         if (matchwild == 0)
 812                                                 break;
 813                                 }
 814                         }
 815                 }
 816                 return (match);
 817         }
 818 }
 819
 820 void
 821 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 822 {
 823         struct inpcb *inp;
 824         struct in6_multi *inm;
 825         struct in6_mfilter *imf;
 826         struct ip6_moptions *im6o;
 827
 828         INP_INFO_WLOCK(pcbinfo);
 829         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 830                 INP_WLOCK(inp);
 831                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 832                         INP_WUNLOCK(inp);
 833                         continue;
 834                 }
 835                 im6o = inp->in6p_moptions;
 836                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 837                         /*
 838                          * Unselect the outgoing ifp for multicast if it
 839                          * is being detached.
 840                          */
 841                         if (im6o->im6o_multicast_ifp == ifp)
 842                                 im6o->im6o_multicast_ifp = NULL;
 843                         /*
 844                          * Drop multicast group membership if we joined
 845                          * through the interface being detached.
 846                          */
 847 restart:
 848                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 849                                 if ((inm = imf->im6f_in6m) == NULL)
 850                                         continue;
 851                                 if (inm->in6m_ifp != ifp)
 852                                         continue;
 853                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 854                                 IN6_MULTI_LOCK_ASSERT();
 855                                 in6_leavegroup_locked(inm, NULL);
 856                                 ip6_mfilter_free(imf);
 857                                 goto restart;
 858                         }
 859                 }
 860                 INP_WUNLOCK(inp);
 861         }
 862         INP_INFO_WUNLOCK(pcbinfo);
 863 }
 864
 865 /*
 866  * Check for alternatives when higher level complains
 867  * about service problems.  For now, invalidate cached
 868  * routing information.  If the route was created dynamically
 869  * (by a redirect), time to try a default gateway again.
 870  */
 871 void
 872 in6_losing(struct inpcb *inp)
 873 {
 874
 875         RO_INVALIDATE_CACHE(&inp->inp_route6);
 876 }
 877
 878 /*
 879  * After a routing change, flush old routing
 880  * and allocate a (hopefully) better one.
 881  */
 882 struct inpcb *
 883 in6_rtchange(struct inpcb *inp, int errno __unused)
 884 {
 885
 886         RO_INVALIDATE_CACHE(&inp->inp_route6);
 887         return inp;
 888 }
 889
 890 static struct inpcb *
 891 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 892     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 893     uint16_t fport, int lookupflags)
 894 {
 895         struct inpcb *local_wild;
 896         const struct inpcblbgrouphead *hdr;
 897         struct inpcblbgroup *grp;
 898         uint32_t idx;
 899
 900         INP_HASH_LOCK_ASSERT(pcbinfo);
 901
 902         hdr = &pcbinfo->ipi_lbgrouphashbase[
 903             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 904
 905         /*
 906          * Order of socket selection:
 907          * 1. non-wild.
 908          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 909          *
 910          * NOTE:
 911          * - Load balanced group does not contain jailed sockets.
 912          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 913          */
 914         local_wild = NULL;
 915         CK_LIST_FOREACH(grp, hdr, il_list) {
 916 #ifdef INET
 917                 if (!(grp->il_vflag & INP_IPV6))
 918                         continue;
 919 #endif
 920                 if (grp->il_lport != lport)
 921                         continue;
 922
 923                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 924                     fport) % grp->il_inpcnt;
 925                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr))
 926                         return (grp->il_inp[idx]);
 927                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 928                     (lookupflags & INPLOOKUP_WILDCARD) != 0)
 929                         local_wild = grp->il_inp[idx];
 930         }
 931         return (local_wild);
 932 }
 933
 934 #ifdef PCBGROUP
 935 /*
 936  * Lookup PCB in hash list, using pcbgroup tables.
 937  */
 938 static struct inpcb *
 939 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 940     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 941     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 942 {
 943         struct inpcbhead *head;
 944         struct inpcb *inp, *tmpinp;
 945         u_short fport = fport_arg, lport = lport_arg;
 946         bool locked;
 947
 948         /*
 949          * First look for an exact match.
 950          */
 951         tmpinp = NULL;
 952         INP_GROUP_LOCK(pcbgroup);
 953         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 954             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 955         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 956                 /* XXX inp locking */
 957                 if ((inp->inp_vflag & INP_IPV6) == 0)
 958                         continue;
 959                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 960                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 961                     inp->inp_fport == fport &&
 962                     inp->inp_lport == lport) {
 963                         /*
 964                          * XXX We should be able to directly return
 965                          * the inp here, without any checks.
 966                          * Well unless both bound with SO_REUSEPORT?
 967                          */
 968                         if (prison_flag(inp->inp_cred, PR_IP6))
 969                                 goto found;
 970                         if (tmpinp == NULL)
 971                                 tmpinp = inp;
 972                 }
 973         }
 974         if (tmpinp != NULL) {
 975                 inp = tmpinp;
 976                 goto found;
 977         }
 978
 979         /*
 980          * Then look for a wildcard match in the pcbgroup.
 981          */
 982         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 983                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 984                 struct inpcb *jail_wild = NULL;
 985                 int injail;
 986
 987                 /*
 988                  * Order of socket selection - we always prefer jails.
 989                  *      1. jailed, non-wild.
 990                  *      2. jailed, wild.
 991                  *      3. non-jailed, non-wild.
 992                  *      4. non-jailed, wild.
 993                  */
 994                 head = &pcbgroup->ipg_hashbase[
 995                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
 996                 CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 997                         /* XXX inp locking */
 998                         if ((inp->inp_vflag & INP_IPV6) == 0)
 999                                 continue;
1000
1001                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1002                             inp->inp_lport != lport) {
1003                                 continue;
1004                         }
1005
1006                         injail = prison_flag(inp->inp_cred, PR_IP6);
1007                         if (injail) {
1008                                 if (prison_check_ip6(inp->inp_cred,
1009                                     laddr) != 0)
1010                                         continue;
1011                         } else {
1012                                 if (local_exact != NULL)
1013                                         continue;
1014                         }
1015
1016                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1017                                 if (injail)
1018                                         goto found;
1019                                 else
1020                                         local_exact = inp;
1021                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1022                                 if (injail)
1023                                         jail_wild = inp;
1024                                 else
1025                                         local_wild = inp;
1026                         }
1027                 } /* LIST_FOREACH */
1028
1029                 inp = jail_wild;
1030                 if (inp == NULL)
1031                         inp = jail_wild;
1032                 if (inp == NULL)
1033                         inp = local_exact;
1034                 if (inp == NULL)
1035                         inp = local_wild;
1036                 if (inp != NULL)
1037                         goto found;
1038         }
1039
1040         /*
1041          * Then look for a wildcard match, if requested.
1042          */
1043         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1044                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1045                 struct inpcb *jail_wild = NULL;
1046                 int injail;
1047
1048                 /*
1049                  * Order of socket selection - we always prefer jails.
1050                  *      1. jailed, non-wild.
1051                  *      2. jailed, wild.
1052                  *      3. non-jailed, non-wild.
1053                  *      4. non-jailed, wild.
1054                  */
1055                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1056                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1057                     pcbinfo->ipi_wildmask)];
1058                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1059                         /* XXX inp locking */
1060                         if ((inp->inp_vflag & INP_IPV6) == 0)
1061                                 continue;
1062
1063                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1064                             inp->inp_lport != lport) {
1065                                 continue;
1066                         }
1067
1068                         injail = prison_flag(inp->inp_cred, PR_IP6);
1069                         if (injail) {
1070                                 if (prison_check_ip6(inp->inp_cred,
1071                                     laddr) != 0)
1072                                         continue;
1073                         } else {
1074                                 if (local_exact != NULL)
1075                                         continue;
1076                         }
1077
1078                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1079                                 if (injail)
1080                                         goto found;
1081                                 else
1082                                         local_exact = inp;
1083                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1084                                 if (injail)
1085                                         jail_wild = inp;
1086                                 else
1087                                         local_wild = inp;
1088                         }
1089                 } /* LIST_FOREACH */
1090
1091                 inp = jail_wild;
1092                 if (inp == NULL)
1093                         inp = jail_wild;
1094                 if (inp == NULL)
1095                         inp = local_exact;
1096                 if (inp == NULL)
1097                         inp = local_wild;
1098                 if (inp != NULL)
1099                         goto found;
1100         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1101         INP_GROUP_UNLOCK(pcbgroup);
1102         return (NULL);
1103
1104 found:
1105         if (lookupflags & INPLOOKUP_WLOCKPCB)
1106                 locked = INP_TRY_WLOCK(inp);
1107         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1108                 locked = INP_TRY_RLOCK(inp);
1109         else
1110                 panic("%s: locking buf", __func__);
1111         if (!locked)
1112                 in_pcbref(inp);
1113         INP_GROUP_UNLOCK(pcbgroup);
1114         if (!locked) {
1115                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1116                         INP_WLOCK(inp);
1117                         if (in_pcbrele_wlocked(inp))
1118                                 return (NULL);
1119                 } else {
1120                         INP_RLOCK(inp);
1121                         if (in_pcbrele_rlocked(inp))
1122                                 return (NULL);
1123                 }
1124         }
1125 #ifdef INVARIANTS
1126         if (lookupflags & INPLOOKUP_WLOCKPCB)
1127                 INP_WLOCK_ASSERT(inp);
1128         else
1129                 INP_RLOCK_ASSERT(inp);
1130 #endif
1131         return (inp);
1132 }
1133 #endif /* PCBGROUP */
1134
1135 /*
1136  * Lookup PCB in hash list.  Used in in_pcb.c as well as here.
1137  */
1138 struct inpcb *
1139 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1140     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1141     int lookupflags, struct ifnet *ifp)
1142 {
1143         struct inpcbhead *head;
1144         struct inpcb *inp, *tmpinp;
1145         u_short fport = fport_arg, lport = lport_arg;
1146
1147         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1148             ("%s: invalid lookup flags %d", __func__, lookupflags));
1149
1150         INP_HASH_LOCK_ASSERT(pcbinfo);
1151
1152         /*
1153          * First look for an exact match.
1154          */
1155         tmpinp = NULL;
1156         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1157             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1158         CK_LIST_FOREACH(inp, head, inp_hash) {
1159                 /* XXX inp locking */
1160                 if ((inp->inp_vflag & INP_IPV6) == 0)
1161                         continue;
1162                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1163                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1164                     inp->inp_fport == fport &&
1165                     inp->inp_lport == lport) {
1166                         /*
1167                          * XXX We should be able to directly return
1168                          * the inp here, without any checks.
1169                          * Well unless both bound with SO_REUSEPORT?
1170                          */
1171                         if (prison_flag(inp->inp_cred, PR_IP6))
1172                                 return (inp);
1173                         if (tmpinp == NULL)
1174                                 tmpinp = inp;
1175                 }
1176         }
1177         if (tmpinp != NULL)
1178                 return (tmpinp);
1179
1180         /*
1181          * Then look in lb group (for wildcard match).
1182          */
1183         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1184                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1185                     fport, lookupflags);
1186                 if (inp != NULL)
1187                         return (inp);
1188         }
1189
1190         /*
1191          * Then look for a wildcard match, if requested.
1192          */
1193         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1194                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1195                 struct inpcb *jail_wild = NULL;
1196                 int injail;
1197
1198                 /*
1199                  * Order of socket selection - we always prefer jails.
1200                  *      1. jailed, non-wild.
1201                  *      2. jailed, wild.
1202                  *      3. non-jailed, non-wild.
1203                  *      4. non-jailed, wild.
1204                  */
1205                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1206                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1207                     pcbinfo->ipi_hashmask)];
1208                 CK_LIST_FOREACH(inp, head, inp_hash) {
1209                         /* XXX inp locking */
1210                         if ((inp->inp_vflag & INP_IPV6) == 0)
1211                                 continue;
1212
1213                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1214                             inp->inp_lport != lport) {
1215                                 continue;
1216                         }
1217
1218                         injail = prison_flag(inp->inp_cred, PR_IP6);
1219                         if (injail) {
1220                                 if (prison_check_ip6(inp->inp_cred,
1221                                     laddr) != 0)
1222                                         continue;
1223                         } else {
1224                                 if (local_exact != NULL)
1225                                         continue;
1226                         }
1227
1228                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1229                                 if (injail)
1230                                         return (inp);
1231                                 else
1232                                         local_exact = inp;
1233                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1234                                 if (injail)
1235                                         jail_wild = inp;
1236                                 else
1237                                         local_wild = inp;
1238                         }
1239                 } /* LIST_FOREACH */
1240
1241                 if (jail_wild != NULL)
1242                         return (jail_wild);
1243                 if (local_exact != NULL)
1244                         return (local_exact);
1245                 if (local_wild != NULL)
1246                         return (local_wild);
1247         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1248
1249         /*
1250          * Not found.
1251          */
1252         return (NULL);
1253 }
1254
1255 /*
1256  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1257  * hash list lock, and will return the inpcb locked (i.e., requires
1258  * INPLOOKUP_LOCKPCB).
1259  */
1260 static struct inpcb *
1261 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1262     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1263     struct ifnet *ifp)
1264 {
1265         struct inpcb *inp;
1266
1267         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1268             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1269         if (inp != NULL) {
1270                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1271                         INP_WLOCK(inp);
1272                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1273                                 INP_WUNLOCK(inp);
1274                                 inp = NULL;
1275                         }
1276                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1277                         INP_RLOCK(inp);
1278                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1279                                 INP_RUNLOCK(inp);
1280                                 inp = NULL;
1281                         }
1282                 } else
1283                         panic("%s: locking bug", __func__);
1284 #ifdef INVARIANTS
1285                 if (inp != NULL) {
1286                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1287                                 INP_WLOCK_ASSERT(inp);
1288                         else
1289                                 INP_RLOCK_ASSERT(inp);
1290                 }
1291 #endif
1292         }
1293         return (inp);
1294 }
1295
1296 /*
1297  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1298  * from which a pre-calculated hash value may be extracted.
1299  *
1300  * Possibly more of this logic should be in in6_pcbgroup.c.
1301  */
1302 struct inpcb *
1303 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1304     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1305 {
1306 #if defined(PCBGROUP) && !defined(RSS)
1307         struct inpcbgroup *pcbgroup;
1308 #endif
1309
1310         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1311             ("%s: invalid lookup flags %d", __func__, lookupflags));
1312         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1313             ("%s: LOCKPCB not set", __func__));
1314
1315         /*
1316          * When not using RSS, use connection groups in preference to the
1317          * reservation table when looking up 4-tuples.  When using RSS, just
1318          * use the reservation table, due to the cost of the Toeplitz hash
1319          * in software.
1320          *
1321          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1322          * we could be doing RSS with a non-Toeplitz hash that is affordable
1323          * in software.
1324          */
1325 #if defined(PCBGROUP) && !defined(RSS)
1326         if (in_pcbgroup_enabled(pcbinfo)) {
1327                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1328                     fport);
1329                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1330                     laddr, lport, lookupflags, ifp));
1331         }
1332 #endif
1333         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1334             lookupflags, ifp));
1335 }
1336
1337 struct inpcb *
1338 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1339     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1340     struct ifnet *ifp, struct mbuf *m)
1341 {
1342 #ifdef PCBGROUP
1343         struct inpcbgroup *pcbgroup;
1344 #endif
1345
1346         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1347             ("%s: invalid lookup flags %d", __func__, lookupflags));
1348         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1349             ("%s: LOCKPCB not set", __func__));
1350
1351 #ifdef PCBGROUP
1352         /*
1353          * If we can use a hardware-generated hash to look up the connection
1354          * group, use that connection group to find the inpcb.  Otherwise
1355          * fall back on a software hash -- or the reservation table if we're
1356          * using RSS.
1357          *
1358          * XXXRW: As above, that policy belongs in the pcbgroup code.
1359          */
1360         if (in_pcbgroup_enabled(pcbinfo) &&
1361             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1362                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1363                     m->m_pkthdr.flowid);
1364                 if (pcbgroup != NULL)
1365                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1366                             fport, laddr, lport, lookupflags, ifp));
1367 #ifndef RSS
1368                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1369                     fport);
1370                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1371                     laddr, lport, lookupflags, ifp));
1372 #endif
1373         }
1374 #endif
1375         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1376             lookupflags, ifp));
1377 }
1378
1379 void
1380 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1381 {
1382         struct ip6_hdr *ip;
1383
1384         ip = mtod(m, struct ip6_hdr *);
1385         bzero(sin6, sizeof(*sin6));
1386         sin6->sin6_len = sizeof(*sin6);
1387         sin6->sin6_family = AF_INET6;
1388         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1389
1390         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1391
1392         return;
1393 }