sys/netinet6/in6_pcb.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Robert N. M. Watson under
   9  * contract to Juniper Networks, Inc.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  * 3. Neither the name of the project nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  36  */
  37
  38 /*-
  39  * Copyright (c) 1982, 1986, 1991, 1993
  40  *      The Regents of the University of California.  All rights reserved.
  41  *
  42  * Redistribution and use in source and binary forms, with or without
  43  * modification, are permitted provided that the following conditions
  44  * are met:
  45  * 1. Redistributions of source code must retain the above copyright
  46  *    notice, this list of conditions and the following disclaimer.
  47  * 2. Redistributions in binary form must reproduce the above copyright
  48  *    notice, this list of conditions and the following disclaimer in the
  49  *    documentation and/or other materials provided with the distribution.
  50  * 3. Neither the name of the University nor the names of its contributors
  51  *    may be used to endorse or promote products derived from this software
  52  *    without specific prior written permission.
  53  *
  54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  64  * SUCH DAMAGE.
  65  *
  66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  67  */
  68
  69 #include <sys/cdefs.h>
  70 __FBSDID("$FreeBSD$");
  71
  72 #include "opt_inet.h"
  73 #include "opt_inet6.h"
  74 #include "opt_ipsec.h"
  75 #include "opt_pcbgroup.h"
  76 #include "opt_rss.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/malloc.h>
  81 #include <sys/mbuf.h>
  82 #include <sys/domain.h>
  83 #include <sys/protosw.h>
  84 #include <sys/socket.h>
  85 #include <sys/socketvar.h>
  86 #include <sys/sockio.h>
  87 #include <sys/errno.h>
  88 #include <sys/time.h>
  89 #include <sys/priv.h>
  90 #include <sys/proc.h>
  91 #include <sys/jail.h>
  92
  93 #include <vm/uma.h>
  94
  95 #include <net/if.h>
  96 #include <net/if_var.h>
  97 #include <net/if_llatbl.h>
  98 #include <net/if_types.h>
  99 #include <net/route.h>
 100
 101 #include <netinet/in.h>
 102 #include <netinet/in_var.h>
 103 #include <netinet/in_systm.h>
 104 #include <netinet/tcp_var.h>
 105 #include <netinet/ip6.h>
 106 #include <netinet/ip_var.h>
 107
 108 #include <netinet6/ip6_var.h>
 109 #include <netinet6/nd6.h>
 110 #include <netinet/in_pcb.h>
 111 #include <netinet6/in6_pcb.h>
 112 #include <netinet6/scope6_var.h>
 113
 114 static struct inpcb *in6_pcblookup_hash_locked(struct inpcbinfo *,
 115     struct in6_addr *, u_int, struct in6_addr *, u_int, int, struct ifnet *);
 116
 117 int
 118 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 119     struct ucred *cred)
 120 {
 121         struct socket *so = inp->inp_socket;
 122         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 123         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 124         u_short lport = 0;
 125         int error, lookupflags = 0;
 126         int reuseport = (so->so_options & SO_REUSEPORT);
 127
 128         /*
 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
 356         INP_WLOCK_ASSERT(inp);
 357         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 358
 359         if (nam->sa_len != sizeof (*sin6))
 360                 return (EINVAL);
 361         if (sin6->sin6_family != AF_INET6)
 362                 return (EAFNOSUPPORT);
 363         if (sin6->sin6_port == 0)
 364                 return (EADDRNOTAVAIL);
 365
 366         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 367                 scope_ambiguous = 1;
 368         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 369                 return(error);
 370
 371         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 372                 /*
 373                  * If the destination address is UNSPECIFIED addr,
 374                  * use the loopback addr, e.g ::1.
 375                  */
 376                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 377                         sin6->sin6_addr = in6addr_loopback;
 378         }
 379         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 380                 return (error);
 381
 382         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 383             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 384         if (error)
 385                 return (error);
 386
 387         /*
 388          * Do not update this earlier, in case we return with an error.
 389          *
 390          * XXX: this in6_selectsrc_socket result might replace the bound local
 391          * address with the address specified by setsockopt(IPV6_PKTINFO).
 392          * Is it the intended behavior?
 393          */
 394         *plocal_addr6 = in6a;
 395
 396         /*
 397          * Don't do pcblookup call here; return interface in
 398          * plocal_addr6
 399          * and exit to caller, that will do the lookup.
 400          */
 401
 402         return (0);
 403 }
 404
 405 /*
 406  * Outer subroutine:
 407  * Connect from a socket to a specified address.
 408  * Both address and port must be specified in argument sin.
 409  * If don't have a local address for this socket yet,
 410  * then pick one.
 411  */
 412 int
 413 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 414     struct ucred *cred, struct mbuf *m, bool rehash)
 415 {
 416         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 417         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 418         struct in6_addr addr6;
 419         int error;
 420
 421         INP_WLOCK_ASSERT(inp);
 422         INP_HASH_WLOCK_ASSERT(pcbinfo);
 423
 424         /*
 425          * Call inner routine, to assign local interface address.
 426          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 427          */
 428         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 429                 return (error);
 430
 431         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 432                                sin6->sin6_port,
 433                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 434                               ? &addr6 : &inp->in6p_laddr,
 435                               inp->inp_lport, 0, NULL) != NULL) {
 436                 return (EADDRINUSE);
 437         }
 438         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 439                 if (inp->inp_lport == 0) {
 440                         KASSERT(rehash == true,
 441                             ("Rehashing required for unbound inps"));
 442                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 443                         if (error)
 444                                 return (error);
 445                 }
 446                 inp->in6p_laddr = addr6;
 447         }
 448         inp->in6p_faddr = sin6->sin6_addr;
 449         inp->inp_fport = sin6->sin6_port;
 450         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 451         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 452         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 453                 inp->inp_flow |=
 454                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 455
 456         if (rehash) {
 457                 in_pcbrehash_mbuf(inp, m);
 458         } else {
 459                 in_pcbinshash_mbuf(inp, m);
 460         }
 461
 462         return (0);
 463 }
 464
 465 int
 466 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 467 {
 468
 469         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 470 }
 471
 472 void
 473 in6_pcbdisconnect(struct inpcb *inp)
 474 {
 475
 476         INP_WLOCK_ASSERT(inp);
 477         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 478
 479         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 480         inp->inp_fport = 0;
 481         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 482         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 483         in_pcbrehash(inp);
 484 }
 485
 486 struct sockaddr *
 487 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 488 {
 489         struct sockaddr_in6 *sin6;
 490
 491         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 492         bzero(sin6, sizeof *sin6);
 493         sin6->sin6_family = AF_INET6;
 494         sin6->sin6_len = sizeof(*sin6);
 495         sin6->sin6_port = port;
 496         sin6->sin6_addr = *addr_p;
 497         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 498
 499         return (struct sockaddr *)sin6;
 500 }
 501
 502 struct sockaddr *
 503 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 504 {
 505         struct sockaddr_in sin;
 506         struct sockaddr_in6 *sin6_p;
 507
 508         bzero(&sin, sizeof sin);
 509         sin.sin_family = AF_INET;
 510         sin.sin_len = sizeof(sin);
 511         sin.sin_port = port;
 512         sin.sin_addr = *addr_p;
 513
 514         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 515                 M_WAITOK);
 516         in6_sin_2_v4mapsin6(&sin, sin6_p);
 517
 518         return (struct sockaddr *)sin6_p;
 519 }
 520
 521 int
 522 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 523 {
 524         struct inpcb *inp;
 525         struct in6_addr addr;
 526         in_port_t port;
 527
 528         inp = sotoinpcb(so);
 529         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 530
 531         INP_RLOCK(inp);
 532         port = inp->inp_lport;
 533         addr = inp->in6p_laddr;
 534         INP_RUNLOCK(inp);
 535
 536         *nam = in6_sockaddr(port, &addr);
 537         return 0;
 538 }
 539
 540 int
 541 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 542 {
 543         struct inpcb *inp;
 544         struct in6_addr addr;
 545         in_port_t port;
 546
 547         inp = sotoinpcb(so);
 548         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 549
 550         INP_RLOCK(inp);
 551         port = inp->inp_fport;
 552         addr = inp->in6p_faddr;
 553         INP_RUNLOCK(inp);
 554
 555         *nam = in6_sockaddr(port, &addr);
 556         return 0;
 557 }
 558
 559 int
 560 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 561 {
 562         struct  inpcb *inp;
 563         int     error;
 564
 565         inp = sotoinpcb(so);
 566         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 567
 568 #ifdef INET
 569         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 570                 error = in_getsockaddr(so, nam);
 571                 if (error == 0)
 572                         in6_sin_2_v4mapsin6_in_sock(nam);
 573         } else
 574 #endif
 575         {
 576                 /* scope issues will be handled in in6_getsockaddr(). */
 577                 error = in6_getsockaddr(so, nam);
 578         }
 579
 580         return error;
 581 }
 582
 583 int
 584 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 585 {
 586         struct  inpcb *inp;
 587         int     error;
 588
 589         inp = sotoinpcb(so);
 590         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 591
 592 #ifdef INET
 593         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 594                 error = in_getpeeraddr(so, nam);
 595                 if (error == 0)
 596                         in6_sin_2_v4mapsin6_in_sock(nam);
 597         } else
 598 #endif
 599         /* scope issues will be handled in in6_getpeeraddr(). */
 600         error = in6_getpeeraddr(so, nam);
 601
 602         return error;
 603 }
 604
 605 /*
 606  * Pass some notification to all connections of a protocol
 607  * associated with address dst.  The local address and/or port numbers
 608  * may be specified to limit the search.  The "usual action" will be
 609  * taken, depending on the ctlinput cmd.  The caller must filter any
 610  * cmds that are uninteresting (e.g., no error in the map).
 611  * Call the protocol specific routine (if any) to report
 612  * any errors for each matching socket.
 613  */
 614 void
 615 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 616     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 617     int cmd, void *cmdarg,
 618     struct inpcb *(*notify)(struct inpcb *, int))
 619 {
 620         struct inpcb *inp, *inp_temp;
 621         struct sockaddr_in6 sa6_src, *sa6_dst;
 622         u_short fport = fport_arg, lport = lport_arg;
 623         u_int32_t flowinfo;
 624         int errno;
 625
 626         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 627                 return;
 628
 629         sa6_dst = (struct sockaddr_in6 *)dst;
 630         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 631                 return;
 632
 633         /*
 634          * note that src can be NULL when we get notify by local fragmentation.
 635          */
 636         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 637         flowinfo = sa6_src.sin6_flowinfo;
 638
 639         /*
 640          * Redirects go to all references to the destination,
 641          * and use in6_rtchange to invalidate the route cache.
 642          * Dead host indications: also use in6_rtchange to invalidate
 643          * the cache, and deliver the error to all the sockets.
 644          * Otherwise, if we have knowledge of the local port and address,
 645          * deliver only to that socket.
 646          */
 647         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 648                 fport = 0;
 649                 lport = 0;
 650                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 651
 652                 if (cmd != PRC_HOSTDEAD)
 653                         notify = in6_rtchange;
 654         }
 655         errno = inet6ctlerrmap[cmd];
 656         INP_INFO_WLOCK(pcbinfo);
 657         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 658                 INP_WLOCK(inp);
 659                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 660                         INP_WUNLOCK(inp);
 661                         continue;
 662                 }
 663
 664                 /*
 665                  * If the error designates a new path MTU for a destination
 666                  * and the application (associated with this socket) wanted to
 667                  * know the value, notify.
 668                  * XXX: should we avoid to notify the value to TCP sockets?
 669                  */
 670                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 671                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 672                                         *(u_int32_t *)cmdarg);
 673
 674                 /*
 675                  * Detect if we should notify the error. If no source and
 676                  * destination ports are specifed, but non-zero flowinfo and
 677                  * local address match, notify the error. This is the case
 678                  * when the error is delivered with an encrypted buffer
 679                  * by ESP. Otherwise, just compare addresses and ports
 680                  * as usual.
 681                  */
 682                 if (lport == 0 && fport == 0 && flowinfo &&
 683                     inp->inp_socket != NULL &&
 684                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 685                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 686                         goto do_notify;
 687                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 688                                              &sa6_dst->sin6_addr) ||
 689                          inp->inp_socket == 0 ||
 690                          (lport && inp->inp_lport != lport) ||
 691                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 692                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 693                                               &sa6_src.sin6_addr)) ||
 694                          (fport && inp->inp_fport != fport)) {
 695                         INP_WUNLOCK(inp);
 696                         continue;
 697                 }
 698
 699           do_notify:
 700                 if (notify) {
 701                         if ((*notify)(inp, errno))
 702                                 INP_WUNLOCK(inp);
 703                 } else
 704                         INP_WUNLOCK(inp);
 705         }
 706         INP_INFO_WUNLOCK(pcbinfo);
 707 }
 708
 709 /*
 710  * Lookup a PCB based on the local address and port.  Caller must hold the
 711  * hash lock.  No inpcb locks or references are acquired.
 712  */
 713 struct inpcb *
 714 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 715     u_short lport, int lookupflags, struct ucred *cred)
 716 {
 717         struct inpcb *inp;
 718         int matchwild = 3, wildcard;
 719
 720         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 721             ("%s: invalid lookup flags %d", __func__, lookupflags));
 722
 723         INP_HASH_LOCK_ASSERT(pcbinfo);
 724
 725         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 726                 struct inpcbhead *head;
 727                 /*
 728                  * Look for an unconnected (wildcard foreign addr) PCB that
 729                  * matches the local address and port we're looking for.
 730                  */
 731                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 732                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 733                     pcbinfo->ipi_hashmask)];
 734                 CK_LIST_FOREACH(inp, head, inp_hash) {
 735                         /* XXX inp locking */
 736                         if ((inp->inp_vflag & INP_IPV6) == 0)
 737                                 continue;
 738                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 739                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 740                             inp->inp_lport == lport) {
 741                                 /* Found. */
 742                                 if (cred == NULL ||
 743                                     prison_equal_ip6(cred->cr_prison,
 744                                         inp->inp_cred->cr_prison))
 745                                         return (inp);
 746                         }
 747                 }
 748                 /*
 749                  * Not found.
 750                  */
 751                 return (NULL);
 752         } else {
 753                 struct inpcbporthead *porthash;
 754                 struct inpcbport *phd;
 755                 struct inpcb *match = NULL;
 756                 /*
 757                  * Best fit PCB lookup.
 758                  *
 759                  * First see if this local port is in use by looking on the
 760                  * port hash list.
 761                  */
 762                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 763                     pcbinfo->ipi_porthashmask)];
 764                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 765                         if (phd->phd_port == lport)
 766                                 break;
 767                 }
 768                 if (phd != NULL) {
 769                         /*
 770                          * Port is in use by one or more PCBs. Look for best
 771                          * fit.
 772                          */
 773                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 774                                 wildcard = 0;
 775                                 if (cred != NULL &&
 776                                     !prison_equal_ip6(cred->cr_prison,
 777                                         inp->inp_cred->cr_prison))
 778                                         continue;
 779                                 /* XXX inp locking */
 780                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 781                                         continue;
 782                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 783                                         wildcard++;
 784                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 785                                         &inp->in6p_laddr)) {
 786                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 787                                                 wildcard++;
 788                                         else if (!IN6_ARE_ADDR_EQUAL(
 789                                             &inp->in6p_laddr, laddr))
 790                                                 continue;
 791                                 } else {
 792                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 793                                                 wildcard++;
 794                                 }
 795                                 if (wildcard < matchwild) {
 796                                         match = inp;
 797                                         matchwild = wildcard;
 798                                         if (matchwild == 0)
 799                                                 break;
 800                                 }
 801                         }
 802                 }
 803                 return (match);
 804         }
 805 }
 806
 807 void
 808 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 809 {
 810         struct inpcb *inp;
 811         struct in6_multi *inm;
 812         struct in6_mfilter *imf;
 813         struct ip6_moptions *im6o;
 814
 815         INP_INFO_WLOCK(pcbinfo);
 816         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 817                 INP_WLOCK(inp);
 818                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 819                         INP_WUNLOCK(inp);
 820                         continue;
 821                 }
 822                 im6o = inp->in6p_moptions;
 823                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 824                         /*
 825                          * Unselect the outgoing ifp for multicast if it
 826                          * is being detached.
 827                          */
 828                         if (im6o->im6o_multicast_ifp == ifp)
 829                                 im6o->im6o_multicast_ifp = NULL;
 830                         /*
 831                          * Drop multicast group membership if we joined
 832                          * through the interface being detached.
 833                          */
 834 restart:
 835                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 836                                 if ((inm = imf->im6f_in6m) == NULL)
 837                                         continue;
 838                                 if (inm->in6m_ifp != ifp)
 839                                         continue;
 840                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 841                                 IN6_MULTI_LOCK_ASSERT();
 842                                 in6_leavegroup_locked(inm, NULL);
 843                                 ip6_mfilter_free(imf);
 844                                 goto restart;
 845                         }
 846                 }
 847                 INP_WUNLOCK(inp);
 848         }
 849         INP_INFO_WUNLOCK(pcbinfo);
 850 }
 851
 852 /*
 853  * Check for alternatives when higher level complains
 854  * about service problems.  For now, invalidate cached
 855  * routing information.  If the route was created dynamically
 856  * (by a redirect), time to try a default gateway again.
 857  */
 858 void
 859 in6_losing(struct inpcb *inp)
 860 {
 861
 862         RO_INVALIDATE_CACHE(&inp->inp_route6);
 863 }
 864
 865 /*
 866  * After a routing change, flush old routing
 867  * and allocate a (hopefully) better one.
 868  */
 869 struct inpcb *
 870 in6_rtchange(struct inpcb *inp, int errno __unused)
 871 {
 872
 873         RO_INVALIDATE_CACHE(&inp->inp_route6);
 874         return inp;
 875 }
 876
 877 static struct inpcb *
 878 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 879     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 880     uint16_t fport, int lookupflags)
 881 {
 882         struct inpcb *local_wild;
 883         const struct inpcblbgrouphead *hdr;
 884         struct inpcblbgroup *grp;
 885         uint32_t idx;
 886
 887         INP_HASH_LOCK_ASSERT(pcbinfo);
 888
 889         hdr = &pcbinfo->ipi_lbgrouphashbase[
 890             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 891
 892         /*
 893          * Order of socket selection:
 894          * 1. non-wild.
 895          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 896          *
 897          * NOTE:
 898          * - Load balanced group does not contain jailed sockets.
 899          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 900          */
 901         local_wild = NULL;
 902         CK_LIST_FOREACH(grp, hdr, il_list) {
 903 #ifdef INET
 904                 if (!(grp->il_vflag & INP_IPV6))
 905                         continue;
 906 #endif
 907                 if (grp->il_lport != lport)
 908                         continue;
 909
 910                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 911                     fport) % grp->il_inpcnt;
 912                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr))
 913                         return (grp->il_inp[idx]);
 914                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 915                     (lookupflags & INPLOOKUP_WILDCARD) != 0)
 916                         local_wild = grp->il_inp[idx];
 917         }
 918         return (local_wild);
 919 }
 920
 921 #ifdef PCBGROUP
 922 /*
 923  * Lookup PCB in hash list, using pcbgroup tables.
 924  */
 925 static struct inpcb *
 926 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 927     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 928     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 929 {
 930         struct inpcbhead *head;
 931         struct inpcb *inp, *tmpinp;
 932         u_short fport = fport_arg, lport = lport_arg;
 933         bool locked;
 934
 935         /*
 936          * First look for an exact match.
 937          */
 938         tmpinp = NULL;
 939         INP_GROUP_LOCK(pcbgroup);
 940         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 941             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 942         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 943                 /* XXX inp locking */
 944                 if ((inp->inp_vflag & INP_IPV6) == 0)
 945                         continue;
 946                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 947                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 948                     inp->inp_fport == fport &&
 949                     inp->inp_lport == lport) {
 950                         /*
 951                          * XXX We should be able to directly return
 952                          * the inp here, without any checks.
 953                          * Well unless both bound with SO_REUSEPORT?
 954                          */
 955                         if (prison_flag(inp->inp_cred, PR_IP6))
 956                                 goto found;
 957                         if (tmpinp == NULL)
 958                                 tmpinp = inp;
 959                 }
 960         }
 961         if (tmpinp != NULL) {
 962                 inp = tmpinp;
 963                 goto found;
 964         }
 965
 966         /*
 967          * Then look for a wildcard match in the pcbgroup.
 968          */
 969         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 970                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 971                 struct inpcb *jail_wild = NULL;
 972                 int injail;
 973
 974                 /*
 975                  * Order of socket selection - we always prefer jails.
 976                  *      1. jailed, non-wild.
 977                  *      2. jailed, wild.
 978                  *      3. non-jailed, non-wild.
 979                  *      4. non-jailed, wild.
 980                  */
 981                 head = &pcbgroup->ipg_hashbase[
 982                     INP_PCBHASH(INADDR_ANY, lport, 0, 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
 988                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 989                             inp->inp_lport != lport) {
 990                                 continue;
 991                         }
 992
 993                         injail = prison_flag(inp->inp_cred, PR_IP6);
 994                         if (injail) {
 995                                 if (prison_check_ip6(inp->inp_cred,
 996                                     laddr) != 0)
 997                                         continue;
 998                         } else {
 999                                 if (local_exact != NULL)
1000                                         continue;
1001                         }
1002
1003                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1004                                 if (injail)
1005                                         goto found;
1006                                 else
1007                                         local_exact = inp;
1008                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1009                                 if (injail)
1010                                         jail_wild = inp;
1011                                 else
1012                                         local_wild = inp;
1013                         }
1014                 } /* LIST_FOREACH */
1015
1016                 inp = jail_wild;
1017                 if (inp == NULL)
1018                         inp = jail_wild;
1019                 if (inp == NULL)
1020                         inp = local_exact;
1021                 if (inp == NULL)
1022                         inp = local_wild;
1023                 if (inp != NULL)
1024                         goto found;
1025         }
1026
1027         /*
1028          * Then look for a wildcard match, if requested.
1029          */
1030         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1031                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1032                 struct inpcb *jail_wild = NULL;
1033                 int injail;
1034
1035                 /*
1036                  * Order of socket selection - we always prefer jails.
1037                  *      1. jailed, non-wild.
1038                  *      2. jailed, wild.
1039                  *      3. non-jailed, non-wild.
1040                  *      4. non-jailed, wild.
1041                  */
1042                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1043                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1044                     pcbinfo->ipi_wildmask)];
1045                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1046                         /* XXX inp locking */
1047                         if ((inp->inp_vflag & INP_IPV6) == 0)
1048                                 continue;
1049
1050                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1051                             inp->inp_lport != lport) {
1052                                 continue;
1053                         }
1054
1055                         injail = prison_flag(inp->inp_cred, PR_IP6);
1056                         if (injail) {
1057                                 if (prison_check_ip6(inp->inp_cred,
1058                                     laddr) != 0)
1059                                         continue;
1060                         } else {
1061                                 if (local_exact != NULL)
1062                                         continue;
1063                         }
1064
1065                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1066                                 if (injail)
1067                                         goto found;
1068                                 else
1069                                         local_exact = inp;
1070                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1071                                 if (injail)
1072                                         jail_wild = inp;
1073                                 else
1074                                         local_wild = inp;
1075                         }
1076                 } /* LIST_FOREACH */
1077
1078                 inp = jail_wild;
1079                 if (inp == NULL)
1080                         inp = jail_wild;
1081                 if (inp == NULL)
1082                         inp = local_exact;
1083                 if (inp == NULL)
1084                         inp = local_wild;
1085                 if (inp != NULL)
1086                         goto found;
1087         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1088         INP_GROUP_UNLOCK(pcbgroup);
1089         return (NULL);
1090
1091 found:
1092         if (lookupflags & INPLOOKUP_WLOCKPCB)
1093                 locked = INP_TRY_WLOCK(inp);
1094         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1095                 locked = INP_TRY_RLOCK(inp);
1096         else
1097                 panic("%s: locking buf", __func__);
1098         if (!locked)
1099                 in_pcbref(inp);
1100         INP_GROUP_UNLOCK(pcbgroup);
1101         if (!locked) {
1102                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1103                         INP_WLOCK(inp);
1104                         if (in_pcbrele_wlocked(inp))
1105                                 return (NULL);
1106                 } else {
1107                         INP_RLOCK(inp);
1108                         if (in_pcbrele_rlocked(inp))
1109                                 return (NULL);
1110                 }
1111         }
1112 #ifdef INVARIANTS
1113         if (lookupflags & INPLOOKUP_WLOCKPCB)
1114                 INP_WLOCK_ASSERT(inp);
1115         else
1116                 INP_RLOCK_ASSERT(inp);
1117 #endif
1118         return (inp);
1119 }
1120 #endif /* PCBGROUP */
1121
1122 /*
1123  * Lookup PCB in hash list.
1124  */
1125 static struct inpcb *
1126 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1127     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1128     int lookupflags, struct ifnet *ifp)
1129 {
1130         struct inpcbhead *head;
1131         struct inpcb *inp, *tmpinp;
1132         u_short fport = fport_arg, lport = lport_arg;
1133
1134         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1135             ("%s: invalid lookup flags %d", __func__, lookupflags));
1136
1137         INP_HASH_LOCK_ASSERT(pcbinfo);
1138
1139         /*
1140          * First look for an exact match.
1141          */
1142         tmpinp = NULL;
1143         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1144             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1145         CK_LIST_FOREACH(inp, head, inp_hash) {
1146                 /* XXX inp locking */
1147                 if ((inp->inp_vflag & INP_IPV6) == 0)
1148                         continue;
1149                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1150                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1151                     inp->inp_fport == fport &&
1152                     inp->inp_lport == lport) {
1153                         /*
1154                          * XXX We should be able to directly return
1155                          * the inp here, without any checks.
1156                          * Well unless both bound with SO_REUSEPORT?
1157                          */
1158                         if (prison_flag(inp->inp_cred, PR_IP6))
1159                                 return (inp);
1160                         if (tmpinp == NULL)
1161                                 tmpinp = inp;
1162                 }
1163         }
1164         if (tmpinp != NULL)
1165                 return (tmpinp);
1166
1167         /*
1168          * Then look in lb group (for wildcard match).
1169          */
1170         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1171                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1172                     fport, lookupflags);
1173                 if (inp != NULL)
1174                         return (inp);
1175         }
1176
1177         /*
1178          * Then look for a wildcard match, if requested.
1179          */
1180         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1181                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1182                 struct inpcb *jail_wild = NULL;
1183                 int injail;
1184
1185                 /*
1186                  * Order of socket selection - we always prefer jails.
1187                  *      1. jailed, non-wild.
1188                  *      2. jailed, wild.
1189                  *      3. non-jailed, non-wild.
1190                  *      4. non-jailed, wild.
1191                  */
1192                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1193                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1194                     pcbinfo->ipi_hashmask)];
1195                 CK_LIST_FOREACH(inp, head, inp_hash) {
1196                         /* XXX inp locking */
1197                         if ((inp->inp_vflag & INP_IPV6) == 0)
1198                                 continue;
1199
1200                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1201                             inp->inp_lport != lport) {
1202                                 continue;
1203                         }
1204
1205                         injail = prison_flag(inp->inp_cred, PR_IP6);
1206                         if (injail) {
1207                                 if (prison_check_ip6(inp->inp_cred,
1208                                     laddr) != 0)
1209                                         continue;
1210                         } else {
1211                                 if (local_exact != NULL)
1212                                         continue;
1213                         }
1214
1215                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1216                                 if (injail)
1217                                         return (inp);
1218                                 else
1219                                         local_exact = inp;
1220                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1221                                 if (injail)
1222                                         jail_wild = inp;
1223                                 else
1224                                         local_wild = inp;
1225                         }
1226                 } /* LIST_FOREACH */
1227
1228                 if (jail_wild != NULL)
1229                         return (jail_wild);
1230                 if (local_exact != NULL)
1231                         return (local_exact);
1232                 if (local_wild != NULL)
1233                         return (local_wild);
1234         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1235
1236         /*
1237          * Not found.
1238          */
1239         return (NULL);
1240 }
1241
1242 /*
1243  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1244  * hash list lock, and will return the inpcb locked (i.e., requires
1245  * INPLOOKUP_LOCKPCB).
1246  */
1247 static struct inpcb *
1248 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1249     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1250     struct ifnet *ifp)
1251 {
1252         struct inpcb *inp;
1253
1254         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1255             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1256         if (inp != NULL) {
1257                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1258                         INP_WLOCK(inp);
1259                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1260                                 INP_WUNLOCK(inp);
1261                                 inp = NULL;
1262                         }
1263                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1264                         INP_RLOCK(inp);
1265                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1266                                 INP_RUNLOCK(inp);
1267                                 inp = NULL;
1268                         }
1269                 } else
1270                         panic("%s: locking bug", __func__);
1271 #ifdef INVARIANTS
1272                 if (inp != NULL) {
1273                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1274                                 INP_WLOCK_ASSERT(inp);
1275                         else
1276                                 INP_RLOCK_ASSERT(inp);
1277                 }
1278 #endif
1279         }
1280         return (inp);
1281 }
1282
1283 /*
1284  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1285  * from which a pre-calculated hash value may be extracted.
1286  *
1287  * Possibly more of this logic should be in in6_pcbgroup.c.
1288  */
1289 struct inpcb *
1290 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1291     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1292 {
1293 #if defined(PCBGROUP) && !defined(RSS)
1294         struct inpcbgroup *pcbgroup;
1295 #endif
1296
1297         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1298             ("%s: invalid lookup flags %d", __func__, lookupflags));
1299         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1300             ("%s: LOCKPCB not set", __func__));
1301
1302         /*
1303          * When not using RSS, use connection groups in preference to the
1304          * reservation table when looking up 4-tuples.  When using RSS, just
1305          * use the reservation table, due to the cost of the Toeplitz hash
1306          * in software.
1307          *
1308          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1309          * we could be doing RSS with a non-Toeplitz hash that is affordable
1310          * in software.
1311          */
1312 #if defined(PCBGROUP) && !defined(RSS)
1313         if (in_pcbgroup_enabled(pcbinfo)) {
1314                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1315                     fport);
1316                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1317                     laddr, lport, lookupflags, ifp));
1318         }
1319 #endif
1320         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1321             lookupflags, ifp));
1322 }
1323
1324 struct inpcb *
1325 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1326     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1327     struct ifnet *ifp, struct mbuf *m)
1328 {
1329 #ifdef PCBGROUP
1330         struct inpcbgroup *pcbgroup;
1331 #endif
1332
1333         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1334             ("%s: invalid lookup flags %d", __func__, lookupflags));
1335         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1336             ("%s: LOCKPCB not set", __func__));
1337
1338 #ifdef PCBGROUP
1339         /*
1340          * If we can use a hardware-generated hash to look up the connection
1341          * group, use that connection group to find the inpcb.  Otherwise
1342          * fall back on a software hash -- or the reservation table if we're
1343          * using RSS.
1344          *
1345          * XXXRW: As above, that policy belongs in the pcbgroup code.
1346          */
1347         if (in_pcbgroup_enabled(pcbinfo) &&
1348             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1349                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1350                     m->m_pkthdr.flowid);
1351                 if (pcbgroup != NULL)
1352                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1353                             fport, laddr, lport, lookupflags, ifp));
1354 #ifndef RSS
1355                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1356                     fport);
1357                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1358                     laddr, lport, lookupflags, ifp));
1359 #endif
1360         }
1361 #endif
1362         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1363             lookupflags, ifp));
1364 }
1365
1366 void
1367 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1368 {
1369         struct ip6_hdr *ip;
1370
1371         ip = mtod(m, struct ip6_hdr *);
1372         bzero(sin6, sizeof(*sin6));
1373         sin6->sin6_len = sizeof(*sin6);
1374         sin6->sin6_family = AF_INET6;
1375         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1376
1377         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1378
1379         return;
1380 }