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 ifaddr *ifa;
 184
 185                         sin6->sin6_port = 0;            /* yech... */
 186                         NET_EPOCH_ENTER();
 187                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 188                             NULL &&
 189                             (inp->inp_flags & INP_BINDANY) == 0) {
 190                                 NET_EPOCH_EXIT();
 191                                 return (EADDRNOTAVAIL);
 192                         }
 193
 194                         /*
 195                          * XXX: bind to an anycast address might accidentally
 196                          * cause sending a packet with anycast source address.
 197                          * We should allow to bind to a deprecated address, since
 198                          * the application dares to use it.
 199                          */
 200                         if (ifa != NULL &&
 201                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 202                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 203                                 NET_EPOCH_EXIT();
 204                                 return (EADDRNOTAVAIL);
 205                         }
 206                         NET_EPOCH_EXIT();
 207                 }
 208                 if (lport) {
 209                         struct inpcb *t;
 210                         struct tcptw *tw;
 211
 212                         /* GROSS */
 213                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 214                             ntohs(lport) >= V_ipport_reservedlow &&
 215                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
 216                             0))
 217                                 return (EACCES);
 218                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 219                             priv_check_cred(inp->inp_cred,
 220                             PRIV_NETINET_REUSEPORT, 0) != 0) {
 221                                 t = in6_pcblookup_local(pcbinfo,
 222                                     &sin6->sin6_addr, lport,
 223                                     INPLOOKUP_WILDCARD, cred);
 224                                 if (t &&
 225                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 226                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 227                                     (so->so_type != SOCK_STREAM ||
 228                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 229                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 230                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 231                                      (t->inp_flags2 & INP_REUSEPORT) ||
 232                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 233                                     (inp->inp_cred->cr_uid !=
 234                                      t->inp_cred->cr_uid))
 235                                         return (EADDRINUSE);
 236
 237                                 /*
 238                                  * If the socket is a BINDMULTI socket, then
 239                                  * the credentials need to match and the
 240                                  * original socket also has to have been bound
 241                                  * with BINDMULTI.
 242                                  */
 243                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 244                                         return (EADDRINUSE);
 245
 246 #ifdef INET
 247                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 248                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 249                                         struct sockaddr_in sin;
 250
 251                                         in6_sin6_2_sin(&sin, sin6);
 252                                         t = in_pcblookup_local(pcbinfo,
 253                                             sin.sin_addr, lport,
 254                                             INPLOOKUP_WILDCARD, cred);
 255                                         if (t &&
 256                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 257                                             ((t->inp_flags &
 258                                               INP_TIMEWAIT) == 0) &&
 259                                             (so->so_type != SOCK_STREAM ||
 260                                              ntohl(t->inp_faddr.s_addr) ==
 261                                               INADDR_ANY) &&
 262                                             (inp->inp_cred->cr_uid !=
 263                                              t->inp_cred->cr_uid))
 264                                                 return (EADDRINUSE);
 265
 266                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 267                                                 return (EADDRINUSE);
 268                                 }
 269 #endif
 270                         }
 271                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 272                             lport, lookupflags, cred);
 273                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 274                                 /*
 275                                  * XXXRW: If an incpb has had its timewait
 276                                  * state recycled, we treat the address as
 277                                  * being in use (for now).  This is better
 278                                  * than a panic, but not desirable.
 279                                  */
 280                                 tw = intotw(t);
 281                                 if (tw == NULL ||
 282                                     ((reuseport & tw->tw_so_options) == 0 &&
 283                                          (reuseport_lb & tw->tw_so_options) == 0))
 284                                         return (EADDRINUSE);
 285                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 286                                            (reuseport_lb & inp_so_options(t)) == 0) {
 287                                 return (EADDRINUSE);
 288                         }
 289 #ifdef INET
 290                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 291                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 292                                 struct sockaddr_in sin;
 293
 294                                 in6_sin6_2_sin(&sin, sin6);
 295                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 296                                    lport, lookupflags, cred);
 297                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 298                                         tw = intotw(t);
 299                                         if (tw == NULL)
 300                                                 return (EADDRINUSE);
 301                                         if ((reuseport & tw->tw_so_options) == 0
 302                                             && (reuseport_lb & tw->tw_so_options) == 0
 303                                             && (ntohl(t->inp_laddr.s_addr) !=
 304                                                 INADDR_ANY || ((inp->inp_vflag &
 305                                                         INP_IPV6PROTO) ==
 306                                                     (t->inp_vflag & INP_IPV6PROTO))))
 307                                                 return (EADDRINUSE);
 308                                 } else if (t &&
 309                                     (reuseport & inp_so_options(t)) == 0 &&
 310                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 311                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 312                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 313                                         return (EADDRINUSE);
 314                                 }
 315                         }
 316 #endif
 317                 }
 318                 inp->in6p_laddr = sin6->sin6_addr;
 319         }
 320         if (lport == 0) {
 321                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 322                         /* Undo an address bind that may have occurred. */
 323                         inp->in6p_laddr = in6addr_any;
 324                         return (error);
 325                 }
 326         } else {
 327                 inp->inp_lport = lport;
 328                 if (in_pcbinshash(inp) != 0) {
 329                         inp->in6p_laddr = in6addr_any;
 330                         inp->inp_lport = 0;
 331                         return (EAGAIN);
 332                 }
 333         }
 334         return (0);
 335 }
 336
 337 /*
 338  *   Transform old in6_pcbconnect() into an inner subroutine for new
 339  *   in6_pcbconnect(): Do some validity-checking on the remote
 340  *   address (in mbuf 'nam') and then determine local host address
 341  *   (i.e., which interface) to use to access that remote host.
 342  *
 343  *   This preserves definition of in6_pcbconnect(), while supporting a
 344  *   slightly different version for T/TCP.  (This is more than
 345  *   a bit of a kludge, but cleaning up the internal interfaces would
 346  *   have forced minor changes in every protocol).
 347  */
 348 static int
 349 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 350     struct in6_addr *plocal_addr6)
 351 {
 352         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 353         int error = 0;
 354         int scope_ambiguous = 0;
 355         struct in6_addr in6a;
 356
 357         INP_WLOCK_ASSERT(inp);
 358         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 359
 360         if (nam->sa_len != sizeof (*sin6))
 361                 return (EINVAL);
 362         if (sin6->sin6_family != AF_INET6)
 363                 return (EAFNOSUPPORT);
 364         if (sin6->sin6_port == 0)
 365                 return (EADDRNOTAVAIL);
 366
 367         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 368                 scope_ambiguous = 1;
 369         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 370                 return(error);
 371
 372         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 373                 /*
 374                  * If the destination address is UNSPECIFIED addr,
 375                  * use the loopback addr, e.g ::1.
 376                  */
 377                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 378                         sin6->sin6_addr = in6addr_loopback;
 379         }
 380         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 381                 return (error);
 382
 383         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 384             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 385         if (error)
 386                 return (error);
 387
 388         /*
 389          * Do not update this earlier, in case we return with an error.
 390          *
 391          * XXX: this in6_selectsrc_socket result might replace the bound local
 392          * address with the address specified by setsockopt(IPV6_PKTINFO).
 393          * Is it the intended behavior?
 394          */
 395         *plocal_addr6 = in6a;
 396
 397         /*
 398          * Don't do pcblookup call here; return interface in
 399          * plocal_addr6
 400          * and exit to caller, that will do the lookup.
 401          */
 402
 403         return (0);
 404 }
 405
 406 /*
 407  * Outer subroutine:
 408  * Connect from a socket to a specified address.
 409  * Both address and port must be specified in argument sin.
 410  * If don't have a local address for this socket yet,
 411  * then pick one.
 412  */
 413 int
 414 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 415     struct ucred *cred, struct mbuf *m)
 416 {
 417         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 418         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 419         struct in6_addr addr6;
 420         int error;
 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, &addr6)) != 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                               ? &addr6 : &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                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 442                         if (error)
 443                                 return (error);
 444                 }
 445                 inp->in6p_laddr = addr6;
 446         }
 447         inp->in6p_faddr = sin6->sin6_addr;
 448         inp->inp_fport = sin6->sin6_port;
 449         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 450         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 451         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 452                 inp->inp_flow |=
 453                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 454
 455         in_pcbrehash_mbuf(inp, m);
 456
 457         return (0);
 458 }
 459
 460 int
 461 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 462 {
 463
 464         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL));
 465 }
 466
 467 void
 468 in6_pcbdisconnect(struct inpcb *inp)
 469 {
 470
 471         INP_WLOCK_ASSERT(inp);
 472         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 473
 474         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 475         inp->inp_fport = 0;
 476         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 477         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 478         in_pcbrehash(inp);
 479 }
 480
 481 struct sockaddr *
 482 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 483 {
 484         struct sockaddr_in6 *sin6;
 485
 486         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 487         bzero(sin6, sizeof *sin6);
 488         sin6->sin6_family = AF_INET6;
 489         sin6->sin6_len = sizeof(*sin6);
 490         sin6->sin6_port = port;
 491         sin6->sin6_addr = *addr_p;
 492         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 493
 494         return (struct sockaddr *)sin6;
 495 }
 496
 497 struct sockaddr *
 498 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 499 {
 500         struct sockaddr_in sin;
 501         struct sockaddr_in6 *sin6_p;
 502
 503         bzero(&sin, sizeof sin);
 504         sin.sin_family = AF_INET;
 505         sin.sin_len = sizeof(sin);
 506         sin.sin_port = port;
 507         sin.sin_addr = *addr_p;
 508
 509         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 510                 M_WAITOK);
 511         in6_sin_2_v4mapsin6(&sin, sin6_p);
 512
 513         return (struct sockaddr *)sin6_p;
 514 }
 515
 516 int
 517 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 518 {
 519         struct inpcb *inp;
 520         struct in6_addr addr;
 521         in_port_t port;
 522
 523         inp = sotoinpcb(so);
 524         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 525
 526         INP_RLOCK(inp);
 527         port = inp->inp_lport;
 528         addr = inp->in6p_laddr;
 529         INP_RUNLOCK(inp);
 530
 531         *nam = in6_sockaddr(port, &addr);
 532         return 0;
 533 }
 534
 535 int
 536 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 537 {
 538         struct inpcb *inp;
 539         struct in6_addr addr;
 540         in_port_t port;
 541
 542         inp = sotoinpcb(so);
 543         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 544
 545         INP_RLOCK(inp);
 546         port = inp->inp_fport;
 547         addr = inp->in6p_faddr;
 548         INP_RUNLOCK(inp);
 549
 550         *nam = in6_sockaddr(port, &addr);
 551         return 0;
 552 }
 553
 554 int
 555 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 556 {
 557         struct  inpcb *inp;
 558         int     error;
 559
 560         inp = sotoinpcb(so);
 561         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 562
 563 #ifdef INET
 564         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 565                 error = in_getsockaddr(so, nam);
 566                 if (error == 0)
 567                         in6_sin_2_v4mapsin6_in_sock(nam);
 568         } else
 569 #endif
 570         {
 571                 /* scope issues will be handled in in6_getsockaddr(). */
 572                 error = in6_getsockaddr(so, nam);
 573         }
 574
 575         return error;
 576 }
 577
 578 int
 579 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 580 {
 581         struct  inpcb *inp;
 582         int     error;
 583
 584         inp = sotoinpcb(so);
 585         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 586
 587 #ifdef INET
 588         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 589                 error = in_getpeeraddr(so, nam);
 590                 if (error == 0)
 591                         in6_sin_2_v4mapsin6_in_sock(nam);
 592         } else
 593 #endif
 594         /* scope issues will be handled in in6_getpeeraddr(). */
 595         error = in6_getpeeraddr(so, nam);
 596
 597         return error;
 598 }
 599
 600 /*
 601  * Pass some notification to all connections of a protocol
 602  * associated with address dst.  The local address and/or port numbers
 603  * may be specified to limit the search.  The "usual action" will be
 604  * taken, depending on the ctlinput cmd.  The caller must filter any
 605  * cmds that are uninteresting (e.g., no error in the map).
 606  * Call the protocol specific routine (if any) to report
 607  * any errors for each matching socket.
 608  */
 609 void
 610 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 611     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 612     int cmd, void *cmdarg,
 613     struct inpcb *(*notify)(struct inpcb *, int))
 614 {
 615         struct inpcb *inp, *inp_temp;
 616         struct sockaddr_in6 sa6_src, *sa6_dst;
 617         u_short fport = fport_arg, lport = lport_arg;
 618         u_int32_t flowinfo;
 619         int errno;
 620
 621         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 622                 return;
 623
 624         sa6_dst = (struct sockaddr_in6 *)dst;
 625         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 626                 return;
 627
 628         /*
 629          * note that src can be NULL when we get notify by local fragmentation.
 630          */
 631         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 632         flowinfo = sa6_src.sin6_flowinfo;
 633
 634         /*
 635          * Redirects go to all references to the destination,
 636          * and use in6_rtchange to invalidate the route cache.
 637          * Dead host indications: also use in6_rtchange to invalidate
 638          * the cache, and deliver the error to all the sockets.
 639          * Otherwise, if we have knowledge of the local port and address,
 640          * deliver only to that socket.
 641          */
 642         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 643                 fport = 0;
 644                 lport = 0;
 645                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 646
 647                 if (cmd != PRC_HOSTDEAD)
 648                         notify = in6_rtchange;
 649         }
 650         errno = inet6ctlerrmap[cmd];
 651         INP_INFO_WLOCK(pcbinfo);
 652         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 653                 INP_WLOCK(inp);
 654                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 655                         INP_WUNLOCK(inp);
 656                         continue;
 657                 }
 658
 659                 /*
 660                  * If the error designates a new path MTU for a destination
 661                  * and the application (associated with this socket) wanted to
 662                  * know the value, notify.
 663                  * XXX: should we avoid to notify the value to TCP sockets?
 664                  */
 665                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 666                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 667                                         *(u_int32_t *)cmdarg);
 668
 669                 /*
 670                  * Detect if we should notify the error. If no source and
 671                  * destination ports are specifed, but non-zero flowinfo and
 672                  * local address match, notify the error. This is the case
 673                  * when the error is delivered with an encrypted buffer
 674                  * by ESP. Otherwise, just compare addresses and ports
 675                  * as usual.
 676                  */
 677                 if (lport == 0 && fport == 0 && flowinfo &&
 678                     inp->inp_socket != NULL &&
 679                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 680                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 681                         goto do_notify;
 682                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 683                                              &sa6_dst->sin6_addr) ||
 684                          inp->inp_socket == 0 ||
 685                          (lport && inp->inp_lport != lport) ||
 686                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 687                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 688                                               &sa6_src.sin6_addr)) ||
 689                          (fport && inp->inp_fport != fport)) {
 690                         INP_WUNLOCK(inp);
 691                         continue;
 692                 }
 693
 694           do_notify:
 695                 if (notify) {
 696                         if ((*notify)(inp, errno))
 697                                 INP_WUNLOCK(inp);
 698                 } else
 699                         INP_WUNLOCK(inp);
 700         }
 701         INP_INFO_WUNLOCK(pcbinfo);
 702 }
 703
 704 /*
 705  * Lookup a PCB based on the local address and port.  Caller must hold the
 706  * hash lock.  No inpcb locks or references are acquired.
 707  */
 708 struct inpcb *
 709 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 710     u_short lport, int lookupflags, struct ucred *cred)
 711 {
 712         struct inpcb *inp;
 713         int matchwild = 3, wildcard;
 714
 715         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 716             ("%s: invalid lookup flags %d", __func__, lookupflags));
 717
 718         INP_HASH_WLOCK_ASSERT(pcbinfo);
 719
 720         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 721                 struct inpcbhead *head;
 722                 /*
 723                  * Look for an unconnected (wildcard foreign addr) PCB that
 724                  * matches the local address and port we're looking for.
 725                  */
 726                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 727                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 728                     pcbinfo->ipi_hashmask)];
 729                 CK_LIST_FOREACH(inp, head, inp_hash) {
 730                         /* XXX inp locking */
 731                         if ((inp->inp_vflag & INP_IPV6) == 0)
 732                                 continue;
 733                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 734                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 735                             inp->inp_lport == lport) {
 736                                 /* Found. */
 737                                 if (cred == NULL ||
 738                                     prison_equal_ip6(cred->cr_prison,
 739                                         inp->inp_cred->cr_prison))
 740                                         return (inp);
 741                         }
 742                 }
 743                 /*
 744                  * Not found.
 745                  */
 746                 return (NULL);
 747         } else {
 748                 struct inpcbporthead *porthash;
 749                 struct inpcbport *phd;
 750                 struct inpcb *match = NULL;
 751                 /*
 752                  * Best fit PCB lookup.
 753                  *
 754                  * First see if this local port is in use by looking on the
 755                  * port hash list.
 756                  */
 757                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 758                     pcbinfo->ipi_porthashmask)];
 759                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 760                         if (phd->phd_port == lport)
 761                                 break;
 762                 }
 763                 if (phd != NULL) {
 764                         /*
 765                          * Port is in use by one or more PCBs. Look for best
 766                          * fit.
 767                          */
 768                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 769                                 wildcard = 0;
 770                                 if (cred != NULL &&
 771                                     !prison_equal_ip6(cred->cr_prison,
 772                                         inp->inp_cred->cr_prison))
 773                                         continue;
 774                                 /* XXX inp locking */
 775                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 776                                         continue;
 777                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 778                                         wildcard++;
 779                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 780                                         &inp->in6p_laddr)) {
 781                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 782                                                 wildcard++;
 783                                         else if (!IN6_ARE_ADDR_EQUAL(
 784                                             &inp->in6p_laddr, laddr))
 785                                                 continue;
 786                                 } else {
 787                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 788                                                 wildcard++;
 789                                 }
 790                                 if (wildcard < matchwild) {
 791                                         match = inp;
 792                                         matchwild = wildcard;
 793                                         if (matchwild == 0)
 794                                                 break;
 795                                 }
 796                         }
 797                 }
 798                 return (match);
 799         }
 800 }
 801
 802 void
 803 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 804 {
 805         struct inpcb *in6p;
 806         struct ip6_moptions *im6o;
 807         int i, gap;
 808
 809         INP_INFO_WLOCK(pcbinfo);
 810         CK_LIST_FOREACH(in6p, pcbinfo->ipi_listhead, inp_list) {
 811                 INP_WLOCK(in6p);
 812                 if (__predict_false(in6p->inp_flags2 & INP_FREED)) {
 813                         INP_WUNLOCK(in6p);
 814                         continue;
 815                 }
 816                 im6o = in6p->in6p_moptions;
 817                 if ((in6p->inp_vflag & INP_IPV6) && im6o != NULL) {
 818                         /*
 819                          * Unselect the outgoing ifp for multicast if it
 820                          * is being detached.
 821                          */
 822                         if (im6o->im6o_multicast_ifp == ifp)
 823                                 im6o->im6o_multicast_ifp = NULL;
 824                         /*
 825                          * Drop multicast group membership if we joined
 826                          * through the interface being detached.
 827                          */
 828                         gap = 0;
 829                         for (i = 0; i < im6o->im6o_num_memberships; i++) {
 830                                 if (im6o->im6o_membership[i]->in6m_ifp ==
 831                                     ifp) {
 832                                         in6_leavegroup(im6o->im6o_membership[i], NULL);
 833                                         gap++;
 834                                 } else if (gap != 0) {
 835                                         im6o->im6o_membership[i - gap] =
 836                                             im6o->im6o_membership[i];
 837                                 }
 838                         }
 839                         im6o->im6o_num_memberships -= gap;
 840                 }
 841                 INP_WUNLOCK(in6p);
 842         }
 843         INP_INFO_WUNLOCK(pcbinfo);
 844 }
 845
 846 /*
 847  * Check for alternatives when higher level complains
 848  * about service problems.  For now, invalidate cached
 849  * routing information.  If the route was created dynamically
 850  * (by a redirect), time to try a default gateway again.
 851  */
 852 void
 853 in6_losing(struct inpcb *inp)
 854 {
 855
 856         RO_INVALIDATE_CACHE(&inp->inp_route6);
 857 }
 858
 859 /*
 860  * After a routing change, flush old routing
 861  * and allocate a (hopefully) better one.
 862  */
 863 struct inpcb *
 864 in6_rtchange(struct inpcb *inp, int errno __unused)
 865 {
 866
 867         RO_INVALIDATE_CACHE(&inp->inp_route6);
 868         return inp;
 869 }
 870
 871 static struct inpcb *
 872 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 873     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 874     uint16_t fport, int lookupflags)
 875 {
 876         struct inpcb *local_wild = NULL;
 877         const struct inpcblbgrouphead *hdr;
 878         struct inpcblbgroup *grp;
 879         struct inpcblbgroup *grp_local_wild;
 880         uint32_t idx;
 881
 882         INP_HASH_LOCK_ASSERT(pcbinfo);
 883
 884         hdr = &pcbinfo->ipi_lbgrouphashbase[INP_PCBLBGROUP_PORTHASH(
 885             lport, pcbinfo->ipi_lbgrouphashmask)];
 886
 887         /*
 888          * Order of socket selection:
 889          * 1. non-wild.
 890          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 891          *
 892          * NOTE:
 893          * - Load balanced group does not contain jailed sockets.
 894          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 895          */
 896         CK_LIST_FOREACH(grp, hdr, il_list) {
 897 #ifdef INET
 898                 if (!(grp->il_vflag & INP_IPV6))
 899                         continue;
 900 #endif
 901                 if (grp->il_lport == lport) {
 902                         idx = 0;
 903                         int pkt_hash = INP_PCBLBGROUP_PKTHASH(
 904                             INP6_PCBHASHKEY(faddr), lport, fport);
 905
 906                         idx = pkt_hash % grp->il_inpcnt;
 907
 908                         if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) {
 909                                 return (grp->il_inp[idx]);
 910                         } else {
 911                                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 912                                     (lookupflags & INPLOOKUP_WILDCARD)) {
 913                                         local_wild = grp->il_inp[idx];
 914                                         grp_local_wild = grp;
 915                                 }
 916                         }
 917                 }
 918         }
 919         if (local_wild != NULL) {
 920                 return (local_wild);
 921         }
 922         return (NULL);
 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.
1128  */
1129 static 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 (pcbinfo->ipi_lbgrouphashbase != NULL &&
1175             (lookupflags & INPLOOKUP_WILDCARD)) {
1176                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1177                     fport, lookupflags);
1178                 if (inp != NULL) {
1179                         return (inp);
1180                 }
1181         }
1182
1183         /*
1184          * Then look for a wildcard match, if requested.
1185          */
1186         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1187                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1188                 struct inpcb *jail_wild = NULL;
1189                 int injail;
1190
1191                 /*
1192                  * Order of socket selection - we always prefer jails.
1193                  *      1. jailed, non-wild.
1194                  *      2. jailed, wild.
1195                  *      3. non-jailed, non-wild.
1196                  *      4. non-jailed, wild.
1197                  */
1198                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1199                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1200                     pcbinfo->ipi_hashmask)];
1201                 CK_LIST_FOREACH(inp, head, inp_hash) {
1202                         /* XXX inp locking */
1203                         if ((inp->inp_vflag & INP_IPV6) == 0)
1204                                 continue;
1205
1206                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1207                             inp->inp_lport != lport) {
1208                                 continue;
1209                         }
1210
1211                         injail = prison_flag(inp->inp_cred, PR_IP6);
1212                         if (injail) {
1213                                 if (prison_check_ip6(inp->inp_cred,
1214                                     laddr) != 0)
1215                                         continue;
1216                         } else {
1217                                 if (local_exact != NULL)
1218                                         continue;
1219                         }
1220
1221                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1222                                 if (injail)
1223                                         return (inp);
1224                                 else
1225                                         local_exact = inp;
1226                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1227                                 if (injail)
1228                                         jail_wild = inp;
1229                                 else
1230                                         local_wild = inp;
1231                         }
1232                 } /* LIST_FOREACH */
1233
1234                 if (jail_wild != NULL)
1235                         return (jail_wild);
1236                 if (local_exact != NULL)
1237                         return (local_exact);
1238                 if (local_wild != NULL)
1239                         return (local_wild);
1240         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1241
1242         /*
1243          * Not found.
1244          */
1245         return (NULL);
1246 }
1247
1248 /*
1249  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1250  * hash list lock, and will return the inpcb locked (i.e., requires
1251  * INPLOOKUP_LOCKPCB).
1252  */
1253 static struct inpcb *
1254 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1255     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1256     struct ifnet *ifp)
1257 {
1258         struct inpcb *inp;
1259
1260         INP_HASH_RLOCK(pcbinfo);
1261         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1262             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1263         if (inp != NULL) {
1264                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1265                         INP_WLOCK(inp);
1266                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1267                                 INP_WUNLOCK(inp);
1268                                 inp = NULL;
1269                         }
1270                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1271                         INP_RLOCK(inp);
1272                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1273                                 INP_RUNLOCK(inp);
1274                                 inp = NULL;
1275                         }
1276                 } else
1277                         panic("%s: locking bug", __func__);
1278 #ifdef INVARIANTS
1279                 if (inp != NULL) {
1280                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1281                                 INP_WLOCK_ASSERT(inp);
1282                         else
1283                                 INP_RLOCK_ASSERT(inp);
1284                 }
1285 #endif
1286         }
1287         INP_HASH_RUNLOCK(pcbinfo);
1288         return (inp);
1289 }
1290
1291 /*
1292  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1293  * from which a pre-calculated hash value may be extracted.
1294  *
1295  * Possibly more of this logic should be in in6_pcbgroup.c.
1296  */
1297 struct inpcb *
1298 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1299     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1300 {
1301 #if defined(PCBGROUP) && !defined(RSS)
1302         struct inpcbgroup *pcbgroup;
1303 #endif
1304
1305         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1306             ("%s: invalid lookup flags %d", __func__, lookupflags));
1307         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1308             ("%s: LOCKPCB not set", __func__));
1309
1310         /*
1311          * When not using RSS, use connection groups in preference to the
1312          * reservation table when looking up 4-tuples.  When using RSS, just
1313          * use the reservation table, due to the cost of the Toeplitz hash
1314          * in software.
1315          *
1316          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1317          * we could be doing RSS with a non-Toeplitz hash that is affordable
1318          * in software.
1319          */
1320 #if defined(PCBGROUP) && !defined(RSS)
1321         if (in_pcbgroup_enabled(pcbinfo)) {
1322                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1323                     fport);
1324                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1325                     laddr, lport, lookupflags, ifp));
1326         }
1327 #endif
1328         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1329             lookupflags, ifp));
1330 }
1331
1332 struct inpcb *
1333 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1334     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1335     struct ifnet *ifp, struct mbuf *m)
1336 {
1337 #ifdef PCBGROUP
1338         struct inpcbgroup *pcbgroup;
1339 #endif
1340
1341         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1342             ("%s: invalid lookup flags %d", __func__, lookupflags));
1343         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1344             ("%s: LOCKPCB not set", __func__));
1345
1346 #ifdef PCBGROUP
1347         /*
1348          * If we can use a hardware-generated hash to look up the connection
1349          * group, use that connection group to find the inpcb.  Otherwise
1350          * fall back on a software hash -- or the reservation table if we're
1351          * using RSS.
1352          *
1353          * XXXRW: As above, that policy belongs in the pcbgroup code.
1354          */
1355         if (in_pcbgroup_enabled(pcbinfo) &&
1356             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1357                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1358                     m->m_pkthdr.flowid);
1359                 if (pcbgroup != NULL)
1360                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1361                             fport, laddr, lport, lookupflags, ifp));
1362 #ifndef RSS
1363                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1364                     fport);
1365                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1366                     laddr, lport, lookupflags, ifp));
1367 #endif
1368         }
1369 #endif
1370         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1371             lookupflags, ifp));
1372 }
1373
1374 void
1375 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1376 {
1377         struct ip6_hdr *ip;
1378
1379         ip = mtod(m, struct ip6_hdr *);
1380         bzero(sin6, sizeof(*sin6));
1381         sin6->sin6_len = sizeof(*sin6);
1382         sin6->sin6_family = AF_INET6;
1383         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1384
1385         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1386
1387         return;
1388 }