sys/netinet6/in6_pcb.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Robert N. M. Watson under
   9  * contract to Juniper Networks, Inc.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  * 3. Neither the name of the project nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  36  */
  37
  38 /*-
  39  * Copyright (c) 1982, 1986, 1991, 1993
  40  *      The Regents of the University of California.  All rights reserved.
  41  *
  42  * Redistribution and use in source and binary forms, with or without
  43  * modification, are permitted provided that the following conditions
  44  * are met:
  45  * 1. Redistributions of source code must retain the above copyright
  46  *    notice, this list of conditions and the following disclaimer.
  47  * 2. Redistributions in binary form must reproduce the above copyright
  48  *    notice, this list of conditions and the following disclaimer in the
  49  *    documentation and/or other materials provided with the distribution.
  50  * 3. Neither the name of the University nor the names of its contributors
  51  *    may be used to endorse or promote products derived from this software
  52  *    without specific prior written permission.
  53  *
  54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  64  * SUCH DAMAGE.
  65  *
  66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  67  */
  68
  69 #include <sys/cdefs.h>
  70 __FBSDID("$FreeBSD$");
  71
  72 #include "opt_inet.h"
  73 #include "opt_inet6.h"
  74 #include "opt_ipsec.h"
  75 #include "opt_pcbgroup.h"
  76 #include "opt_route.h"
  77 #include "opt_rss.h"
  78
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/malloc.h>
  82 #include <sys/mbuf.h>
  83 #include <sys/domain.h>
  84 #include <sys/protosw.h>
  85 #include <sys/socket.h>
  86 #include <sys/socketvar.h>
  87 #include <sys/sockio.h>
  88 #include <sys/errno.h>
  89 #include <sys/time.h>
  90 #include <sys/priv.h>
  91 #include <sys/proc.h>
  92 #include <sys/jail.h>
  93
  94 #include <vm/uma.h>
  95
  96 #include <net/if.h>
  97 #include <net/if_var.h>
  98 #include <net/if_llatbl.h>
  99 #include <net/if_types.h>
 100 #include <net/route.h>
 101 #include <net/route/nhop.h>
 102
 103 #include <netinet/in.h>
 104 #include <netinet/in_var.h>
 105 #include <netinet/in_systm.h>
 106 #include <netinet/tcp_var.h>
 107 #include <netinet/ip6.h>
 108 #include <netinet/ip_var.h>
 109
 110 #include <netinet6/ip6_var.h>
 111 #include <netinet6/nd6.h>
 112 #include <netinet/in_pcb.h>
 113 #include <netinet6/in6_pcb.h>
 114 #include <netinet6/in6_fib.h>
 115 #include <netinet6/scope6_var.h>
 116
 117 int
 118 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 119     struct ucred *cred)
 120 {
 121         struct socket *so = inp->inp_socket;
 122         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 123         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 124         u_short lport = 0;
 125         int error, lookupflags = 0;
 126         int reuseport = (so->so_options & SO_REUSEPORT);
 127
 128         /*
 129          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
 130          * so that we don't have to add to the (already messy) code below.
 131          */
 132         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
 133
 134         INP_WLOCK_ASSERT(inp);
 135         INP_HASH_WLOCK_ASSERT(pcbinfo);
 136
 137         if (CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) /* XXX broken! */
 138                 return (EADDRNOTAVAIL);
 139         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 140                 return (EINVAL);
 141         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
 142                 lookupflags = INPLOOKUP_WILDCARD;
 143         if (nam == NULL) {
 144                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 145                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 146                         return (error);
 147         } else {
 148                 sin6 = (struct sockaddr_in6 *)nam;
 149                 if (nam->sa_len != sizeof(*sin6))
 150                         return (EINVAL);
 151                 /*
 152                  * family check.
 153                  */
 154                 if (nam->sa_family != AF_INET6)
 155                         return (EAFNOSUPPORT);
 156
 157                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 158                         return(error);
 159
 160                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 161                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 162                         return (error);
 163
 164                 lport = sin6->sin6_port;
 165                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 166                         /*
 167                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 168                          * allow compepte duplication of binding if
 169                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 170                          * and a multicast address is bound on both
 171                          * new and duplicated sockets.
 172                          */
 173                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
 174                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 175                         /*
 176                          * XXX: How to deal with SO_REUSEPORT_LB here?
 177                          * Treat same as SO_REUSEPORT for now.
 178                          */
 179                         if ((so->so_options &
 180                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
 181                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
 182                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 183                         struct epoch_tracker et;
 184                         struct ifaddr *ifa;
 185
 186                         sin6->sin6_port = 0;            /* yech... */
 187                         NET_EPOCH_ENTER(et);
 188                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 189                             NULL &&
 190                             (inp->inp_flags & INP_BINDANY) == 0) {
 191                                 NET_EPOCH_EXIT(et);
 192                                 return (EADDRNOTAVAIL);
 193                         }
 194
 195                         /*
 196                          * XXX: bind to an anycast address might accidentally
 197                          * cause sending a packet with anycast source address.
 198                          * We should allow to bind to a deprecated address, since
 199                          * the application dares to use it.
 200                          */
 201                         if (ifa != NULL &&
 202                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 203                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 204                                 NET_EPOCH_EXIT(et);
 205                                 return (EADDRNOTAVAIL);
 206                         }
 207                         NET_EPOCH_EXIT(et);
 208                 }
 209                 if (lport) {
 210                         struct inpcb *t;
 211                         struct tcptw *tw;
 212
 213                         /* GROSS */
 214                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 215                             ntohs(lport) >= V_ipport_reservedlow &&
 216                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
 217                                 return (EACCES);
 218                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 219                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
 220                                 t = in6_pcblookup_local(pcbinfo,
 221                                     &sin6->sin6_addr, lport,
 222                                     INPLOOKUP_WILDCARD, cred);
 223                                 if (t &&
 224                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 225                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 226                                     (so->so_type != SOCK_STREAM ||
 227                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 228                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 229                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 230                                      (t->inp_flags2 & INP_REUSEPORT) ||
 231                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 232                                     (inp->inp_cred->cr_uid !=
 233                                      t->inp_cred->cr_uid))
 234                                         return (EADDRINUSE);
 235
 236                                 /*
 237                                  * If the socket is a BINDMULTI socket, then
 238                                  * the credentials need to match and the
 239                                  * original socket also has to have been bound
 240                                  * with BINDMULTI.
 241                                  */
 242                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 243                                         return (EADDRINUSE);
 244
 245 #ifdef INET
 246                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 247                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 248                                         struct sockaddr_in sin;
 249
 250                                         in6_sin6_2_sin(&sin, sin6);
 251                                         t = in_pcblookup_local(pcbinfo,
 252                                             sin.sin_addr, lport,
 253                                             INPLOOKUP_WILDCARD, cred);
 254                                         if (t &&
 255                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 256                                             ((t->inp_flags &
 257                                               INP_TIMEWAIT) == 0) &&
 258                                             (so->so_type != SOCK_STREAM ||
 259                                              ntohl(t->inp_faddr.s_addr) ==
 260                                               INADDR_ANY) &&
 261                                             (inp->inp_cred->cr_uid !=
 262                                              t->inp_cred->cr_uid))
 263                                                 return (EADDRINUSE);
 264
 265                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 266                                                 return (EADDRINUSE);
 267                                 }
 268 #endif
 269                         }
 270                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 271                             lport, lookupflags, cred);
 272                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 273                                 /*
 274                                  * XXXRW: If an incpb has had its timewait
 275                                  * state recycled, we treat the address as
 276                                  * being in use (for now).  This is better
 277                                  * than a panic, but not desirable.
 278                                  */
 279                                 tw = intotw(t);
 280                                 if (tw == NULL ||
 281                                     ((reuseport & tw->tw_so_options) == 0 &&
 282                                          (reuseport_lb & tw->tw_so_options) == 0))
 283                                         return (EADDRINUSE);
 284                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 285                                            (reuseport_lb & inp_so_options(t)) == 0) {
 286                                 return (EADDRINUSE);
 287                         }
 288 #ifdef INET
 289                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 290                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 291                                 struct sockaddr_in sin;
 292
 293                                 in6_sin6_2_sin(&sin, sin6);
 294                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 295                                    lport, lookupflags, cred);
 296                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 297                                         tw = intotw(t);
 298                                         if (tw == NULL)
 299                                                 return (EADDRINUSE);
 300                                         if ((reuseport & tw->tw_so_options) == 0
 301                                             && (reuseport_lb & tw->tw_so_options) == 0
 302                                             && (ntohl(t->inp_laddr.s_addr) !=
 303                                                 INADDR_ANY || ((inp->inp_vflag &
 304                                                         INP_IPV6PROTO) ==
 305                                                     (t->inp_vflag & INP_IPV6PROTO))))
 306                                                 return (EADDRINUSE);
 307                                 } else if (t &&
 308                                     (reuseport & inp_so_options(t)) == 0 &&
 309                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 310                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 311                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 312                                         return (EADDRINUSE);
 313                                 }
 314                         }
 315 #endif
 316                 }
 317                 inp->in6p_laddr = sin6->sin6_addr;
 318         }
 319         if (lport == 0) {
 320                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 321                         /* Undo an address bind that may have occurred. */
 322                         inp->in6p_laddr = in6addr_any;
 323                         return (error);
 324                 }
 325         } else {
 326                 inp->inp_lport = lport;
 327                 if (in_pcbinshash(inp) != 0) {
 328                         inp->in6p_laddr = in6addr_any;
 329                         inp->inp_lport = 0;
 330                         return (EAGAIN);
 331                 }
 332         }
 333         return (0);
 334 }
 335
 336 /*
 337  *   Transform old in6_pcbconnect() into an inner subroutine for new
 338  *   in6_pcbconnect(): Do some validity-checking on the remote
 339  *   address (in mbuf 'nam') and then determine local host address
 340  *   (i.e., which interface) to use to access that remote host.
 341  *
 342  *   This preserves definition of in6_pcbconnect(), while supporting a
 343  *   slightly different version for T/TCP.  (This is more than
 344  *   a bit of a kludge, but cleaning up the internal interfaces would
 345  *   have forced minor changes in every protocol).
 346  */
 347 static int
 348 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 349     struct in6_addr *plocal_addr6)
 350 {
 351         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 352         int error = 0;
 353         int scope_ambiguous = 0;
 354         struct in6_addr in6a;
 355
 356         INP_WLOCK_ASSERT(inp);
 357         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 358
 359         if (nam->sa_len != sizeof (*sin6))
 360                 return (EINVAL);
 361         if (sin6->sin6_family != AF_INET6)
 362                 return (EAFNOSUPPORT);
 363         if (sin6->sin6_port == 0)
 364                 return (EADDRNOTAVAIL);
 365
 366         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 367                 scope_ambiguous = 1;
 368         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 369                 return(error);
 370
 371         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 372                 /*
 373                  * If the destination address is UNSPECIFIED addr,
 374                  * use the loopback addr, e.g ::1.
 375                  */
 376                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 377                         sin6->sin6_addr = in6addr_loopback;
 378         }
 379         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 380                 return (error);
 381
 382         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 383             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 384         if (error)
 385                 return (error);
 386
 387         /*
 388          * Do not update this earlier, in case we return with an error.
 389          *
 390          * XXX: this in6_selectsrc_socket result might replace the bound local
 391          * address with the address specified by setsockopt(IPV6_PKTINFO).
 392          * Is it the intended behavior?
 393          */
 394         *plocal_addr6 = in6a;
 395
 396         /*
 397          * Don't do pcblookup call here; return interface in
 398          * plocal_addr6
 399          * and exit to caller, that will do the lookup.
 400          */
 401
 402         return (0);
 403 }
 404
 405 /*
 406  * Outer subroutine:
 407  * Connect from a socket to a specified address.
 408  * Both address and port must be specified in argument sin.
 409  * If don't have a local address for this socket yet,
 410  * then pick one.
 411  */
 412 int
 413 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 414     struct ucred *cred, struct mbuf *m, bool rehash)
 415 {
 416         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 417         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 418         struct sockaddr_in6 laddr6;
 419         int error;
 420
 421         bzero(&laddr6, sizeof(laddr6));
 422         laddr6.sin6_family = AF_INET6;
 423
 424         INP_WLOCK_ASSERT(inp);
 425         INP_HASH_WLOCK_ASSERT(pcbinfo);
 426
 427 #ifdef ROUTE_MPATH
 428         if (CALC_FLOWID_OUTBOUND) {
 429                 uint32_t hash_type, hash_val;
 430
 431                 hash_val = fib6_calc_software_hash(&inp->in6p_laddr,
 432                     &sin6->sin6_addr, 0, sin6->sin6_port,
 433                     inp->inp_socket->so_proto->pr_protocol, &hash_type);
 434                 inp->inp_flowid = hash_val;
 435                 inp->inp_flowtype = hash_type;
 436         }
 437 #endif
 438         /*
 439          * Call inner routine, to assign local interface address.
 440          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 441          */
 442         if ((error = in6_pcbladdr(inp, nam, &laddr6.sin6_addr)) != 0)
 443                 return (error);
 444
 445         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 446                                sin6->sin6_port,
 447                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 448                               ? &laddr6.sin6_addr : &inp->in6p_laddr,
 449                               inp->inp_lport, 0, NULL) != NULL) {
 450                 return (EADDRINUSE);
 451         }
 452         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 453                 if (inp->inp_lport == 0) {
 454                         /*
 455                          * rehash was required to be true in the past for
 456                          * this case; retain that convention.  However,
 457                          * we now call in_pcb_lport_dest rather than
 458                          * in6_pcbbind; the former does not insert into
 459                          * the hash table, the latter does.  Change rehash
 460                          * to false to do the in_pcbinshash below.
 461                          */
 462                         KASSERT(rehash == true,
 463                             ("Rehashing required for unbound inps"));
 464                         rehash = false;
 465                         error = in_pcb_lport_dest(inp,
 466                             (struct sockaddr *) &laddr6, &inp->inp_lport,
 467                             (struct sockaddr *) sin6, sin6->sin6_port, cred, 0);
 468                         if (error)
 469                                 return (error);
 470                 }
 471                 inp->in6p_laddr = laddr6.sin6_addr;
 472         }
 473         inp->in6p_faddr = sin6->sin6_addr;
 474         inp->inp_fport = sin6->sin6_port;
 475         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 476         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 477         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 478                 inp->inp_flow |=
 479                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 480
 481         if (rehash) {
 482                 in_pcbrehash_mbuf(inp, m);
 483         } else {
 484                 in_pcbinshash_mbuf(inp, m);
 485         }
 486
 487         return (0);
 488 }
 489
 490 int
 491 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 492 {
 493
 494         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 495 }
 496
 497 void
 498 in6_pcbdisconnect(struct inpcb *inp)
 499 {
 500
 501         INP_WLOCK_ASSERT(inp);
 502         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 503
 504         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 505         inp->inp_fport = 0;
 506         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 507         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 508         in_pcbrehash(inp);
 509 }
 510
 511 struct sockaddr *
 512 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 513 {
 514         struct sockaddr_in6 *sin6;
 515
 516         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 517         bzero(sin6, sizeof *sin6);
 518         sin6->sin6_family = AF_INET6;
 519         sin6->sin6_len = sizeof(*sin6);
 520         sin6->sin6_port = port;
 521         sin6->sin6_addr = *addr_p;
 522         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 523
 524         return (struct sockaddr *)sin6;
 525 }
 526
 527 struct sockaddr *
 528 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 529 {
 530         struct sockaddr_in sin;
 531         struct sockaddr_in6 *sin6_p;
 532
 533         bzero(&sin, sizeof sin);
 534         sin.sin_family = AF_INET;
 535         sin.sin_len = sizeof(sin);
 536         sin.sin_port = port;
 537         sin.sin_addr = *addr_p;
 538
 539         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 540                 M_WAITOK);
 541         in6_sin_2_v4mapsin6(&sin, sin6_p);
 542
 543         return (struct sockaddr *)sin6_p;
 544 }
 545
 546 int
 547 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 548 {
 549         struct inpcb *inp;
 550         struct in6_addr addr;
 551         in_port_t port;
 552
 553         inp = sotoinpcb(so);
 554         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 555
 556         INP_RLOCK(inp);
 557         port = inp->inp_lport;
 558         addr = inp->in6p_laddr;
 559         INP_RUNLOCK(inp);
 560
 561         *nam = in6_sockaddr(port, &addr);
 562         return 0;
 563 }
 564
 565 int
 566 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 567 {
 568         struct inpcb *inp;
 569         struct in6_addr addr;
 570         in_port_t port;
 571
 572         inp = sotoinpcb(so);
 573         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 574
 575         INP_RLOCK(inp);
 576         port = inp->inp_fport;
 577         addr = inp->in6p_faddr;
 578         INP_RUNLOCK(inp);
 579
 580         *nam = in6_sockaddr(port, &addr);
 581         return 0;
 582 }
 583
 584 int
 585 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 586 {
 587         struct  inpcb *inp;
 588         int     error;
 589
 590         inp = sotoinpcb(so);
 591         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 592
 593 #ifdef INET
 594         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 595                 error = in_getsockaddr(so, nam);
 596                 if (error == 0)
 597                         in6_sin_2_v4mapsin6_in_sock(nam);
 598         } else
 599 #endif
 600         {
 601                 /* scope issues will be handled in in6_getsockaddr(). */
 602                 error = in6_getsockaddr(so, nam);
 603         }
 604
 605         return error;
 606 }
 607
 608 int
 609 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 610 {
 611         struct  inpcb *inp;
 612         int     error;
 613
 614         inp = sotoinpcb(so);
 615         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 616
 617 #ifdef INET
 618         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 619                 error = in_getpeeraddr(so, nam);
 620                 if (error == 0)
 621                         in6_sin_2_v4mapsin6_in_sock(nam);
 622         } else
 623 #endif
 624         /* scope issues will be handled in in6_getpeeraddr(). */
 625         error = in6_getpeeraddr(so, nam);
 626
 627         return error;
 628 }
 629
 630 /*
 631  * Pass some notification to all connections of a protocol
 632  * associated with address dst.  The local address and/or port numbers
 633  * may be specified to limit the search.  The "usual action" will be
 634  * taken, depending on the ctlinput cmd.  The caller must filter any
 635  * cmds that are uninteresting (e.g., no error in the map).
 636  * Call the protocol specific routine (if any) to report
 637  * any errors for each matching socket.
 638  */
 639 void
 640 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 641     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 642     int cmd, void *cmdarg,
 643     struct inpcb *(*notify)(struct inpcb *, int))
 644 {
 645         struct inpcb *inp, *inp_temp;
 646         struct sockaddr_in6 sa6_src, *sa6_dst;
 647         u_short fport = fport_arg, lport = lport_arg;
 648         u_int32_t flowinfo;
 649         int errno;
 650
 651         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 652                 return;
 653
 654         sa6_dst = (struct sockaddr_in6 *)dst;
 655         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 656                 return;
 657
 658         /*
 659          * note that src can be NULL when we get notify by local fragmentation.
 660          */
 661         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 662         flowinfo = sa6_src.sin6_flowinfo;
 663
 664         /*
 665          * Redirects go to all references to the destination,
 666          * and use in6_rtchange to invalidate the route cache.
 667          * Dead host indications: also use in6_rtchange to invalidate
 668          * the cache, and deliver the error to all the sockets.
 669          * Otherwise, if we have knowledge of the local port and address,
 670          * deliver only to that socket.
 671          */
 672         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 673                 fport = 0;
 674                 lport = 0;
 675                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 676
 677                 if (cmd != PRC_HOSTDEAD)
 678                         notify = in6_rtchange;
 679         }
 680         errno = inet6ctlerrmap[cmd];
 681         INP_INFO_WLOCK(pcbinfo);
 682         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 683                 INP_WLOCK(inp);
 684                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 685                         INP_WUNLOCK(inp);
 686                         continue;
 687                 }
 688
 689                 /*
 690                  * If the error designates a new path MTU for a destination
 691                  * and the application (associated with this socket) wanted to
 692                  * know the value, notify.
 693                  * XXX: should we avoid to notify the value to TCP sockets?
 694                  */
 695                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 696                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 697                                         *(u_int32_t *)cmdarg);
 698
 699                 /*
 700                  * Detect if we should notify the error. If no source and
 701                  * destination ports are specifed, but non-zero flowinfo and
 702                  * local address match, notify the error. This is the case
 703                  * when the error is delivered with an encrypted buffer
 704                  * by ESP. Otherwise, just compare addresses and ports
 705                  * as usual.
 706                  */
 707                 if (lport == 0 && fport == 0 && flowinfo &&
 708                     inp->inp_socket != NULL &&
 709                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 710                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 711                         goto do_notify;
 712                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 713                                              &sa6_dst->sin6_addr) ||
 714                          inp->inp_socket == 0 ||
 715                          (lport && inp->inp_lport != lport) ||
 716                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 717                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 718                                               &sa6_src.sin6_addr)) ||
 719                          (fport && inp->inp_fport != fport)) {
 720                         INP_WUNLOCK(inp);
 721                         continue;
 722                 }
 723
 724           do_notify:
 725                 if (notify) {
 726                         if ((*notify)(inp, errno))
 727                                 INP_WUNLOCK(inp);
 728                 } else
 729                         INP_WUNLOCK(inp);
 730         }
 731         INP_INFO_WUNLOCK(pcbinfo);
 732 }
 733
 734 /*
 735  * Lookup a PCB based on the local address and port.  Caller must hold the
 736  * hash lock.  No inpcb locks or references are acquired.
 737  */
 738 struct inpcb *
 739 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 740     u_short lport, int lookupflags, struct ucred *cred)
 741 {
 742         struct inpcb *inp;
 743         int matchwild = 3, wildcard;
 744
 745         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 746             ("%s: invalid lookup flags %d", __func__, lookupflags));
 747
 748         INP_HASH_LOCK_ASSERT(pcbinfo);
 749
 750         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 751                 struct inpcbhead *head;
 752                 /*
 753                  * Look for an unconnected (wildcard foreign addr) PCB that
 754                  * matches the local address and port we're looking for.
 755                  */
 756                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 757                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 758                     pcbinfo->ipi_hashmask)];
 759                 CK_LIST_FOREACH(inp, head, inp_hash) {
 760                         /* XXX inp locking */
 761                         if ((inp->inp_vflag & INP_IPV6) == 0)
 762                                 continue;
 763                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 764                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 765                             inp->inp_lport == lport) {
 766                                 /* Found. */
 767                                 if (cred == NULL ||
 768                                     prison_equal_ip6(cred->cr_prison,
 769                                         inp->inp_cred->cr_prison))
 770                                         return (inp);
 771                         }
 772                 }
 773                 /*
 774                  * Not found.
 775                  */
 776                 return (NULL);
 777         } else {
 778                 struct inpcbporthead *porthash;
 779                 struct inpcbport *phd;
 780                 struct inpcb *match = NULL;
 781                 /*
 782                  * Best fit PCB lookup.
 783                  *
 784                  * First see if this local port is in use by looking on the
 785                  * port hash list.
 786                  */
 787                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 788                     pcbinfo->ipi_porthashmask)];
 789                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 790                         if (phd->phd_port == lport)
 791                                 break;
 792                 }
 793                 if (phd != NULL) {
 794                         /*
 795                          * Port is in use by one or more PCBs. Look for best
 796                          * fit.
 797                          */
 798                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 799                                 wildcard = 0;
 800                                 if (cred != NULL &&
 801                                     !prison_equal_ip6(cred->cr_prison,
 802                                         inp->inp_cred->cr_prison))
 803                                         continue;
 804                                 /* XXX inp locking */
 805                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 806                                         continue;
 807                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 808                                         wildcard++;
 809                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 810                                         &inp->in6p_laddr)) {
 811                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 812                                                 wildcard++;
 813                                         else if (!IN6_ARE_ADDR_EQUAL(
 814                                             &inp->in6p_laddr, laddr))
 815                                                 continue;
 816                                 } else {
 817                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 818                                                 wildcard++;
 819                                 }
 820                                 if (wildcard < matchwild) {
 821                                         match = inp;
 822                                         matchwild = wildcard;
 823                                         if (matchwild == 0)
 824                                                 break;
 825                                 }
 826                         }
 827                 }
 828                 return (match);
 829         }
 830 }
 831
 832 void
 833 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 834 {
 835         struct inpcb *inp;
 836         struct in6_multi *inm;
 837         struct in6_mfilter *imf;
 838         struct ip6_moptions *im6o;
 839
 840         INP_INFO_WLOCK(pcbinfo);
 841         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 842                 INP_WLOCK(inp);
 843                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 844                         INP_WUNLOCK(inp);
 845                         continue;
 846                 }
 847                 im6o = inp->in6p_moptions;
 848                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 849                         /*
 850                          * Unselect the outgoing ifp for multicast if it
 851                          * is being detached.
 852                          */
 853                         if (im6o->im6o_multicast_ifp == ifp)
 854                                 im6o->im6o_multicast_ifp = NULL;
 855                         /*
 856                          * Drop multicast group membership if we joined
 857                          * through the interface being detached.
 858                          */
 859 restart:
 860                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 861                                 if ((inm = imf->im6f_in6m) == NULL)
 862                                         continue;
 863                                 if (inm->in6m_ifp != ifp)
 864                                         continue;
 865                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 866                                 IN6_MULTI_LOCK_ASSERT();
 867                                 in6_leavegroup_locked(inm, NULL);
 868                                 ip6_mfilter_free(imf);
 869                                 goto restart;
 870                         }
 871                 }
 872                 INP_WUNLOCK(inp);
 873         }
 874         INP_INFO_WUNLOCK(pcbinfo);
 875 }
 876
 877 /*
 878  * Check for alternatives when higher level complains
 879  * about service problems.  For now, invalidate cached
 880  * routing information.  If the route was created dynamically
 881  * (by a redirect), time to try a default gateway again.
 882  */
 883 void
 884 in6_losing(struct inpcb *inp)
 885 {
 886
 887         RO_INVALIDATE_CACHE(&inp->inp_route6);
 888 }
 889
 890 /*
 891  * After a routing change, flush old routing
 892  * and allocate a (hopefully) better one.
 893  */
 894 struct inpcb *
 895 in6_rtchange(struct inpcb *inp, int errno __unused)
 896 {
 897
 898         RO_INVALIDATE_CACHE(&inp->inp_route6);
 899         return inp;
 900 }
 901
 902 static struct inpcb *
 903 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 904     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 905     uint16_t fport, int lookupflags)
 906 {
 907         struct inpcb *local_wild;
 908         const struct inpcblbgrouphead *hdr;
 909         struct inpcblbgroup *grp;
 910         uint32_t idx;
 911
 912         INP_HASH_LOCK_ASSERT(pcbinfo);
 913
 914         hdr = &pcbinfo->ipi_lbgrouphashbase[
 915             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 916
 917         /*
 918          * Order of socket selection:
 919          * 1. non-wild.
 920          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 921          *
 922          * NOTE:
 923          * - Load balanced group does not contain jailed sockets.
 924          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 925          */
 926         local_wild = NULL;
 927         CK_LIST_FOREACH(grp, hdr, il_list) {
 928 #ifdef INET
 929                 if (!(grp->il_vflag & INP_IPV6))
 930                         continue;
 931 #endif
 932                 if (grp->il_lport != lport)
 933                         continue;
 934
 935                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 936                     fport) % grp->il_inpcnt;
 937                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr))
 938                         return (grp->il_inp[idx]);
 939                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 940                     (lookupflags & INPLOOKUP_WILDCARD) != 0)
 941                         local_wild = grp->il_inp[idx];
 942         }
 943         return (local_wild);
 944 }
 945
 946 #ifdef PCBGROUP
 947 /*
 948  * Lookup PCB in hash list, using pcbgroup tables.
 949  */
 950 static struct inpcb *
 951 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 952     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 953     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 954 {
 955         struct inpcbhead *head;
 956         struct inpcb *inp, *tmpinp;
 957         u_short fport = fport_arg, lport = lport_arg;
 958         bool locked;
 959
 960         /*
 961          * First look for an exact match.
 962          */
 963         tmpinp = NULL;
 964         INP_GROUP_LOCK(pcbgroup);
 965         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 966             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 967         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 968                 /* XXX inp locking */
 969                 if ((inp->inp_vflag & INP_IPV6) == 0)
 970                         continue;
 971                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 972                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 973                     inp->inp_fport == fport &&
 974                     inp->inp_lport == lport) {
 975                         /*
 976                          * XXX We should be able to directly return
 977                          * the inp here, without any checks.
 978                          * Well unless both bound with SO_REUSEPORT?
 979                          */
 980                         if (prison_flag(inp->inp_cred, PR_IP6))
 981                                 goto found;
 982                         if (tmpinp == NULL)
 983                                 tmpinp = inp;
 984                 }
 985         }
 986         if (tmpinp != NULL) {
 987                 inp = tmpinp;
 988                 goto found;
 989         }
 990
 991         /*
 992          * Then look for a wildcard match in the pcbgroup.
 993          */
 994         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 995                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 996                 struct inpcb *jail_wild = NULL;
 997                 int injail;
 998
 999                 /*
1000                  * Order of socket selection - we always prefer jails.
1001                  *      1. jailed, non-wild.
1002                  *      2. jailed, wild.
1003                  *      3. non-jailed, non-wild.
1004                  *      4. non-jailed, wild.
1005                  */
1006                 head = &pcbgroup->ipg_hashbase[
1007                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
1008                 CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
1009                         /* XXX inp locking */
1010                         if ((inp->inp_vflag & INP_IPV6) == 0)
1011                                 continue;
1012
1013                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1014                             inp->inp_lport != lport) {
1015                                 continue;
1016                         }
1017
1018                         injail = prison_flag(inp->inp_cred, PR_IP6);
1019                         if (injail) {
1020                                 if (prison_check_ip6(inp->inp_cred,
1021                                     laddr) != 0)
1022                                         continue;
1023                         } else {
1024                                 if (local_exact != NULL)
1025                                         continue;
1026                         }
1027
1028                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1029                                 if (injail)
1030                                         goto found;
1031                                 else
1032                                         local_exact = inp;
1033                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1034                                 if (injail)
1035                                         jail_wild = inp;
1036                                 else
1037                                         local_wild = inp;
1038                         }
1039                 } /* LIST_FOREACH */
1040
1041                 inp = jail_wild;
1042                 if (inp == NULL)
1043                         inp = jail_wild;
1044                 if (inp == NULL)
1045                         inp = local_exact;
1046                 if (inp == NULL)
1047                         inp = local_wild;
1048                 if (inp != NULL)
1049                         goto found;
1050         }
1051
1052         /*
1053          * Then look for a wildcard match, if requested.
1054          */
1055         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1056                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1057                 struct inpcb *jail_wild = NULL;
1058                 int injail;
1059
1060                 /*
1061                  * Order of socket selection - we always prefer jails.
1062                  *      1. jailed, non-wild.
1063                  *      2. jailed, wild.
1064                  *      3. non-jailed, non-wild.
1065                  *      4. non-jailed, wild.
1066                  */
1067                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1068                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1069                     pcbinfo->ipi_wildmask)];
1070                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1071                         /* XXX inp locking */
1072                         if ((inp->inp_vflag & INP_IPV6) == 0)
1073                                 continue;
1074
1075                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1076                             inp->inp_lport != lport) {
1077                                 continue;
1078                         }
1079
1080                         injail = prison_flag(inp->inp_cred, PR_IP6);
1081                         if (injail) {
1082                                 if (prison_check_ip6(inp->inp_cred,
1083                                     laddr) != 0)
1084                                         continue;
1085                         } else {
1086                                 if (local_exact != NULL)
1087                                         continue;
1088                         }
1089
1090                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1091                                 if (injail)
1092                                         goto found;
1093                                 else
1094                                         local_exact = inp;
1095                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1096                                 if (injail)
1097                                         jail_wild = inp;
1098                                 else
1099                                         local_wild = inp;
1100                         }
1101                 } /* LIST_FOREACH */
1102
1103                 inp = jail_wild;
1104                 if (inp == NULL)
1105                         inp = jail_wild;
1106                 if (inp == NULL)
1107                         inp = local_exact;
1108                 if (inp == NULL)
1109                         inp = local_wild;
1110                 if (inp != NULL)
1111                         goto found;
1112         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1113         INP_GROUP_UNLOCK(pcbgroup);
1114         return (NULL);
1115
1116 found:
1117         if (lookupflags & INPLOOKUP_WLOCKPCB)
1118                 locked = INP_TRY_WLOCK(inp);
1119         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1120                 locked = INP_TRY_RLOCK(inp);
1121         else
1122                 panic("%s: locking buf", __func__);
1123         if (!locked)
1124                 in_pcbref(inp);
1125         INP_GROUP_UNLOCK(pcbgroup);
1126         if (!locked) {
1127                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1128                         INP_WLOCK(inp);
1129                         if (in_pcbrele_wlocked(inp))
1130                                 return (NULL);
1131                 } else {
1132                         INP_RLOCK(inp);
1133                         if (in_pcbrele_rlocked(inp))
1134                                 return (NULL);
1135                 }
1136         }
1137 #ifdef INVARIANTS
1138         if (lookupflags & INPLOOKUP_WLOCKPCB)
1139                 INP_WLOCK_ASSERT(inp);
1140         else
1141                 INP_RLOCK_ASSERT(inp);
1142 #endif
1143         return (inp);
1144 }
1145 #endif /* PCBGROUP */
1146
1147 /*
1148  * Lookup PCB in hash list.  Used in in_pcb.c as well as here.
1149  */
1150 struct inpcb *
1151 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1152     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1153     int lookupflags, struct ifnet *ifp)
1154 {
1155         struct inpcbhead *head;
1156         struct inpcb *inp, *tmpinp;
1157         u_short fport = fport_arg, lport = lport_arg;
1158
1159         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1160             ("%s: invalid lookup flags %d", __func__, lookupflags));
1161
1162         INP_HASH_LOCK_ASSERT(pcbinfo);
1163
1164         /*
1165          * First look for an exact match.
1166          */
1167         tmpinp = NULL;
1168         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1169             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1170         CK_LIST_FOREACH(inp, head, inp_hash) {
1171                 /* XXX inp locking */
1172                 if ((inp->inp_vflag & INP_IPV6) == 0)
1173                         continue;
1174                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1175                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1176                     inp->inp_fport == fport &&
1177                     inp->inp_lport == lport) {
1178                         /*
1179                          * XXX We should be able to directly return
1180                          * the inp here, without any checks.
1181                          * Well unless both bound with SO_REUSEPORT?
1182                          */
1183                         if (prison_flag(inp->inp_cred, PR_IP6))
1184                                 return (inp);
1185                         if (tmpinp == NULL)
1186                                 tmpinp = inp;
1187                 }
1188         }
1189         if (tmpinp != NULL)
1190                 return (tmpinp);
1191
1192         /*
1193          * Then look in lb group (for wildcard match).
1194          */
1195         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1196                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1197                     fport, lookupflags);
1198                 if (inp != NULL)
1199                         return (inp);
1200         }
1201
1202         /*
1203          * Then look for a wildcard match, if requested.
1204          */
1205         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1206                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1207                 struct inpcb *jail_wild = NULL;
1208                 int injail;
1209
1210                 /*
1211                  * Order of socket selection - we always prefer jails.
1212                  *      1. jailed, non-wild.
1213                  *      2. jailed, wild.
1214                  *      3. non-jailed, non-wild.
1215                  *      4. non-jailed, wild.
1216                  */
1217                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1218                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1219                     pcbinfo->ipi_hashmask)];
1220                 CK_LIST_FOREACH(inp, head, inp_hash) {
1221                         /* XXX inp locking */
1222                         if ((inp->inp_vflag & INP_IPV6) == 0)
1223                                 continue;
1224
1225                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1226                             inp->inp_lport != lport) {
1227                                 continue;
1228                         }
1229
1230                         injail = prison_flag(inp->inp_cred, PR_IP6);
1231                         if (injail) {
1232                                 if (prison_check_ip6(inp->inp_cred,
1233                                     laddr) != 0)
1234                                         continue;
1235                         } else {
1236                                 if (local_exact != NULL)
1237                                         continue;
1238                         }
1239
1240                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1241                                 if (injail)
1242                                         return (inp);
1243                                 else
1244                                         local_exact = inp;
1245                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1246                                 if (injail)
1247                                         jail_wild = inp;
1248                                 else
1249                                         local_wild = inp;
1250                         }
1251                 } /* LIST_FOREACH */
1252
1253                 if (jail_wild != NULL)
1254                         return (jail_wild);
1255                 if (local_exact != NULL)
1256                         return (local_exact);
1257                 if (local_wild != NULL)
1258                         return (local_wild);
1259         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1260
1261         /*
1262          * Not found.
1263          */
1264         return (NULL);
1265 }
1266
1267 /*
1268  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1269  * hash list lock, and will return the inpcb locked (i.e., requires
1270  * INPLOOKUP_LOCKPCB).
1271  */
1272 static struct inpcb *
1273 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1274     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1275     struct ifnet *ifp)
1276 {
1277         struct inpcb *inp;
1278
1279         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1280             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1281         if (inp != NULL) {
1282                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1283                         INP_WLOCK(inp);
1284                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1285                                 INP_WUNLOCK(inp);
1286                                 inp = NULL;
1287                         }
1288                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1289                         INP_RLOCK(inp);
1290                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1291                                 INP_RUNLOCK(inp);
1292                                 inp = NULL;
1293                         }
1294                 } else
1295                         panic("%s: locking bug", __func__);
1296 #ifdef INVARIANTS
1297                 if (inp != NULL) {
1298                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1299                                 INP_WLOCK_ASSERT(inp);
1300                         else
1301                                 INP_RLOCK_ASSERT(inp);
1302                 }
1303 #endif
1304         }
1305         return (inp);
1306 }
1307
1308 /*
1309  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1310  * from which a pre-calculated hash value may be extracted.
1311  *
1312  * Possibly more of this logic should be in in6_pcbgroup.c.
1313  */
1314 struct inpcb *
1315 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1316     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1317 {
1318 #if defined(PCBGROUP) && !defined(RSS)
1319         struct inpcbgroup *pcbgroup;
1320 #endif
1321
1322         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1323             ("%s: invalid lookup flags %d", __func__, lookupflags));
1324         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1325             ("%s: LOCKPCB not set", __func__));
1326
1327         /*
1328          * When not using RSS, use connection groups in preference to the
1329          * reservation table when looking up 4-tuples.  When using RSS, just
1330          * use the reservation table, due to the cost of the Toeplitz hash
1331          * in software.
1332          *
1333          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1334          * we could be doing RSS with a non-Toeplitz hash that is affordable
1335          * in software.
1336          */
1337 #if defined(PCBGROUP) && !defined(RSS)
1338         if (in_pcbgroup_enabled(pcbinfo)) {
1339                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1340                     fport);
1341                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1342                     laddr, lport, lookupflags, ifp));
1343         }
1344 #endif
1345         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1346             lookupflags, ifp));
1347 }
1348
1349 struct inpcb *
1350 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1351     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1352     struct ifnet *ifp, struct mbuf *m)
1353 {
1354 #ifdef PCBGROUP
1355         struct inpcbgroup *pcbgroup;
1356 #endif
1357
1358         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1359             ("%s: invalid lookup flags %d", __func__, lookupflags));
1360         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1361             ("%s: LOCKPCB not set", __func__));
1362
1363 #ifdef PCBGROUP
1364         /*
1365          * If we can use a hardware-generated hash to look up the connection
1366          * group, use that connection group to find the inpcb.  Otherwise
1367          * fall back on a software hash -- or the reservation table if we're
1368          * using RSS.
1369          *
1370          * XXXRW: As above, that policy belongs in the pcbgroup code.
1371          */
1372         if (in_pcbgroup_enabled(pcbinfo) &&
1373             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1374                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1375                     m->m_pkthdr.flowid);
1376                 if (pcbgroup != NULL)
1377                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1378                             fport, laddr, lport, lookupflags, ifp));
1379 #ifndef RSS
1380                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1381                     fport);
1382                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1383                     laddr, lport, lookupflags, ifp));
1384 #endif
1385         }
1386 #endif
1387         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1388             lookupflags, ifp));
1389 }
1390
1391 void
1392 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1393 {
1394         struct ip6_hdr *ip;
1395
1396         ip = mtod(m, struct ip6_hdr *);
1397         bzero(sin6, sizeof(*sin6));
1398         sin6->sin6_len = sizeof(*sin6);
1399         sin6->sin6_family = AF_INET6;
1400         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1401
1402         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1403
1404         return;
1405 }