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