sys/netinet6/in6_pcb.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by Robert N. M. Watson under
   9  * contract to Juniper Networks, Inc.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  * 3. Neither the name of the project nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
  36  */
  37
  38 /*-
  39  * Copyright (c) 1982, 1986, 1991, 1993
  40  *      The Regents of the University of California.  All rights reserved.
  41  *
  42  * Redistribution and use in source and binary forms, with or without
  43  * modification, are permitted provided that the following conditions
  44  * are met:
  45  * 1. Redistributions of source code must retain the above copyright
  46  *    notice, this list of conditions and the following disclaimer.
  47  * 2. Redistributions in binary form must reproduce the above copyright
  48  *    notice, this list of conditions and the following disclaimer in the
  49  *    documentation and/or other materials provided with the distribution.
  50  * 3. Neither the name of the University nor the names of its contributors
  51  *    may be used to endorse or promote products derived from this software
  52  *    without specific prior written permission.
  53  *
  54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  64  * SUCH DAMAGE.
  65  *
  66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
  67  */
  68
  69 #include <sys/cdefs.h>
  70 __FBSDID("$FreeBSD$");
  71
  72 #include "opt_inet.h"
  73 #include "opt_inet6.h"
  74 #include "opt_ipsec.h"
  75 #include "opt_pcbgroup.h"
  76 #include "opt_rss.h"
  77
  78 #include <sys/param.h>
  79 #include <sys/systm.h>
  80 #include <sys/malloc.h>
  81 #include <sys/mbuf.h>
  82 #include <sys/domain.h>
  83 #include <sys/protosw.h>
  84 #include <sys/socket.h>
  85 #include <sys/socketvar.h>
  86 #include <sys/sockio.h>
  87 #include <sys/errno.h>
  88 #include <sys/time.h>
  89 #include <sys/priv.h>
  90 #include <sys/proc.h>
  91 #include <sys/jail.h>
  92
  93 #include <vm/uma.h>
  94
  95 #include <net/if.h>
  96 #include <net/if_var.h>
  97 #include <net/if_llatbl.h>
  98 #include <net/if_types.h>
  99 #include <net/route.h>
 100 #include <net/route/nhop.h>
 101
 102 #include <netinet/in.h>
 103 #include <netinet/in_var.h>
 104 #include <netinet/in_systm.h>
 105 #include <netinet/tcp_var.h>
 106 #include <netinet/ip6.h>
 107 #include <netinet/ip_var.h>
 108
 109 #include <netinet6/ip6_var.h>
 110 #include <netinet6/nd6.h>
 111 #include <netinet/in_pcb.h>
 112 #include <netinet6/in6_pcb.h>
 113 #include <netinet6/in6_fib.h>
 114 #include <netinet6/scope6_var.h>
 115
 116 static struct inpcb *in6_pcblookup_hash_locked(struct inpcbinfo *,
 117     struct in6_addr *, u_int, struct in6_addr *, u_int, int, struct ifnet *);
 118
 119 int
 120 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
 121     struct ucred *cred)
 122 {
 123         struct socket *so = inp->inp_socket;
 124         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
 125         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 126         u_short lport = 0;
 127         int error, lookupflags = 0;
 128         int reuseport = (so->so_options & SO_REUSEPORT);
 129
 130         /*
 131          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
 132          * so that we don't have to add to the (already messy) code below.
 133          */
 134         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
 135
 136         INP_WLOCK_ASSERT(inp);
 137         INP_HASH_WLOCK_ASSERT(pcbinfo);
 138
 139         if (CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) /* XXX broken! */
 140                 return (EADDRNOTAVAIL);
 141         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
 142                 return (EINVAL);
 143         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
 144                 lookupflags = INPLOOKUP_WILDCARD;
 145         if (nam == NULL) {
 146                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
 147                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 148                         return (error);
 149         } else {
 150                 sin6 = (struct sockaddr_in6 *)nam;
 151                 if (nam->sa_len != sizeof(*sin6))
 152                         return (EINVAL);
 153                 /*
 154                  * family check.
 155                  */
 156                 if (nam->sa_family != AF_INET6)
 157                         return (EAFNOSUPPORT);
 158
 159                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 160                         return(error);
 161
 162                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
 163                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
 164                         return (error);
 165
 166                 lport = sin6->sin6_port;
 167                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
 168                         /*
 169                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 170                          * allow compepte duplication of binding if
 171                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 172                          * and a multicast address is bound on both
 173                          * new and duplicated sockets.
 174                          */
 175                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
 176                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 177                         /*
 178                          * XXX: How to deal with SO_REUSEPORT_LB here?
 179                          * Treat same as SO_REUSEPORT for now.
 180                          */
 181                         if ((so->so_options &
 182                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
 183                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
 184                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 185                         struct epoch_tracker et;
 186                         struct ifaddr *ifa;
 187
 188                         sin6->sin6_port = 0;            /* yech... */
 189                         NET_EPOCH_ENTER(et);
 190                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
 191                             NULL &&
 192                             (inp->inp_flags & INP_BINDANY) == 0) {
 193                                 NET_EPOCH_EXIT(et);
 194                                 return (EADDRNOTAVAIL);
 195                         }
 196
 197                         /*
 198                          * XXX: bind to an anycast address might accidentally
 199                          * cause sending a packet with anycast source address.
 200                          * We should allow to bind to a deprecated address, since
 201                          * the application dares to use it.
 202                          */
 203                         if (ifa != NULL &&
 204                             ((struct in6_ifaddr *)ifa)->ia6_flags &
 205                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
 206                                 NET_EPOCH_EXIT(et);
 207                                 return (EADDRNOTAVAIL);
 208                         }
 209                         NET_EPOCH_EXIT(et);
 210                 }
 211                 if (lport) {
 212                         struct inpcb *t;
 213                         struct tcptw *tw;
 214
 215                         /* GROSS */
 216                         if (ntohs(lport) <= V_ipport_reservedhigh &&
 217                             ntohs(lport) >= V_ipport_reservedlow &&
 218                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
 219                                 return (EACCES);
 220                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
 221                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
 222                                 t = in6_pcblookup_local(pcbinfo,
 223                                     &sin6->sin6_addr, lport,
 224                                     INPLOOKUP_WILDCARD, cred);
 225                                 if (t &&
 226                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 227                                     ((t->inp_flags & INP_TIMEWAIT) == 0) &&
 228                                     (so->so_type != SOCK_STREAM ||
 229                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
 230                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 231                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
 232                                      (t->inp_flags2 & INP_REUSEPORT) ||
 233                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
 234                                     (inp->inp_cred->cr_uid !=
 235                                      t->inp_cred->cr_uid))
 236                                         return (EADDRINUSE);
 237
 238                                 /*
 239                                  * If the socket is a BINDMULTI socket, then
 240                                  * the credentials need to match and the
 241                                  * original socket also has to have been bound
 242                                  * with BINDMULTI.
 243                                  */
 244                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 245                                         return (EADDRINUSE);
 246
 247 #ifdef INET
 248                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 249                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 250                                         struct sockaddr_in sin;
 251
 252                                         in6_sin6_2_sin(&sin, sin6);
 253                                         t = in_pcblookup_local(pcbinfo,
 254                                             sin.sin_addr, lport,
 255                                             INPLOOKUP_WILDCARD, cred);
 256                                         if (t &&
 257                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
 258                                             ((t->inp_flags &
 259                                               INP_TIMEWAIT) == 0) &&
 260                                             (so->so_type != SOCK_STREAM ||
 261                                              ntohl(t->inp_faddr.s_addr) ==
 262                                               INADDR_ANY) &&
 263                                             (inp->inp_cred->cr_uid !=
 264                                              t->inp_cred->cr_uid))
 265                                                 return (EADDRINUSE);
 266
 267                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
 268                                                 return (EADDRINUSE);
 269                                 }
 270 #endif
 271                         }
 272                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
 273                             lport, lookupflags, cred);
 274                         if (t && (t->inp_flags & INP_TIMEWAIT)) {
 275                                 /*
 276                                  * XXXRW: If an incpb has had its timewait
 277                                  * state recycled, we treat the address as
 278                                  * being in use (for now).  This is better
 279                                  * than a panic, but not desirable.
 280                                  */
 281                                 tw = intotw(t);
 282                                 if (tw == NULL ||
 283                                     ((reuseport & tw->tw_so_options) == 0 &&
 284                                          (reuseport_lb & tw->tw_so_options) == 0))
 285                                         return (EADDRINUSE);
 286                         } else if (t && (reuseport & inp_so_options(t)) == 0 &&
 287                                            (reuseport_lb & inp_so_options(t)) == 0) {
 288                                 return (EADDRINUSE);
 289                         }
 290 #ifdef INET
 291                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 292                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
 293                                 struct sockaddr_in sin;
 294
 295                                 in6_sin6_2_sin(&sin, sin6);
 296                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
 297                                    lport, lookupflags, cred);
 298                                 if (t && t->inp_flags & INP_TIMEWAIT) {
 299                                         tw = intotw(t);
 300                                         if (tw == NULL)
 301                                                 return (EADDRINUSE);
 302                                         if ((reuseport & tw->tw_so_options) == 0
 303                                             && (reuseport_lb & tw->tw_so_options) == 0
 304                                             && (ntohl(t->inp_laddr.s_addr) !=
 305                                                 INADDR_ANY || ((inp->inp_vflag &
 306                                                         INP_IPV6PROTO) ==
 307                                                     (t->inp_vflag & INP_IPV6PROTO))))
 308                                                 return (EADDRINUSE);
 309                                 } else if (t &&
 310                                     (reuseport & inp_so_options(t)) == 0 &&
 311                                     (reuseport_lb & inp_so_options(t)) == 0 &&
 312                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 313                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
 314                                         return (EADDRINUSE);
 315                                 }
 316                         }
 317 #endif
 318                 }
 319                 inp->in6p_laddr = sin6->sin6_addr;
 320         }
 321         if (lport == 0) {
 322                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
 323                         /* Undo an address bind that may have occurred. */
 324                         inp->in6p_laddr = in6addr_any;
 325                         return (error);
 326                 }
 327         } else {
 328                 inp->inp_lport = lport;
 329                 if (in_pcbinshash(inp) != 0) {
 330                         inp->in6p_laddr = in6addr_any;
 331                         inp->inp_lport = 0;
 332                         return (EAGAIN);
 333                 }
 334         }
 335         return (0);
 336 }
 337
 338 /*
 339  *   Transform old in6_pcbconnect() into an inner subroutine for new
 340  *   in6_pcbconnect(): Do some validity-checking on the remote
 341  *   address (in mbuf 'nam') and then determine local host address
 342  *   (i.e., which interface) to use to access that remote host.
 343  *
 344  *   This preserves definition of in6_pcbconnect(), while supporting a
 345  *   slightly different version for T/TCP.  (This is more than
 346  *   a bit of a kludge, but cleaning up the internal interfaces would
 347  *   have forced minor changes in every protocol).
 348  */
 349 static int
 350 in6_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 351     struct in6_addr *plocal_addr6)
 352 {
 353         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 354         int error = 0;
 355         int scope_ambiguous = 0;
 356         struct in6_addr in6a;
 357
 358         INP_WLOCK_ASSERT(inp);
 359         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
 360
 361         if (nam->sa_len != sizeof (*sin6))
 362                 return (EINVAL);
 363         if (sin6->sin6_family != AF_INET6)
 364                 return (EAFNOSUPPORT);
 365         if (sin6->sin6_port == 0)
 366                 return (EADDRNOTAVAIL);
 367
 368         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 369                 scope_ambiguous = 1;
 370         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 371                 return(error);
 372
 373         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
 374                 /*
 375                  * If the destination address is UNSPECIFIED addr,
 376                  * use the loopback addr, e.g ::1.
 377                  */
 378                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
 379                         sin6->sin6_addr = in6addr_loopback;
 380         }
 381         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
 382                 return (error);
 383
 384         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
 385             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
 386         if (error)
 387                 return (error);
 388
 389         /*
 390          * Do not update this earlier, in case we return with an error.
 391          *
 392          * XXX: this in6_selectsrc_socket result might replace the bound local
 393          * address with the address specified by setsockopt(IPV6_PKTINFO).
 394          * Is it the intended behavior?
 395          */
 396         *plocal_addr6 = in6a;
 397
 398         /*
 399          * Don't do pcblookup call here; return interface in
 400          * plocal_addr6
 401          * and exit to caller, that will do the lookup.
 402          */
 403
 404         return (0);
 405 }
 406
 407 /*
 408  * Outer subroutine:
 409  * Connect from a socket to a specified address.
 410  * Both address and port must be specified in argument sin.
 411  * If don't have a local address for this socket yet,
 412  * then pick one.
 413  */
 414 int
 415 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
 416     struct ucred *cred, struct mbuf *m, bool rehash)
 417 {
 418         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 419         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
 420         struct in6_addr addr6;
 421         int error;
 422
 423         INP_WLOCK_ASSERT(inp);
 424         INP_HASH_WLOCK_ASSERT(pcbinfo);
 425
 426         /*
 427          * Call inner routine, to assign local interface address.
 428          * in6_pcbladdr() may automatically fill in sin6_scope_id.
 429          */
 430         if ((error = in6_pcbladdr(inp, nam, &addr6)) != 0)
 431                 return (error);
 432
 433         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
 434                                sin6->sin6_port,
 435                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
 436                               ? &addr6 : &inp->in6p_laddr,
 437                               inp->inp_lport, 0, NULL) != NULL) {
 438                 return (EADDRINUSE);
 439         }
 440         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 441                 if (inp->inp_lport == 0) {
 442                         KASSERT(rehash == true,
 443                             ("Rehashing required for unbound inps"));
 444                         error = in6_pcbbind(inp, (struct sockaddr *)0, cred);
 445                         if (error)
 446                                 return (error);
 447                 }
 448                 inp->in6p_laddr = addr6;
 449         }
 450         inp->in6p_faddr = sin6->sin6_addr;
 451         inp->inp_fport = sin6->sin6_port;
 452         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 453         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 454         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
 455                 inp->inp_flow |=
 456                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
 457
 458         if (rehash) {
 459                 in_pcbrehash_mbuf(inp, m);
 460         } else {
 461                 in_pcbinshash_mbuf(inp, m);
 462         }
 463
 464         return (0);
 465 }
 466
 467 int
 468 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 469 {
 470
 471         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
 472 }
 473
 474 void
 475 in6_pcbdisconnect(struct inpcb *inp)
 476 {
 477
 478         INP_WLOCK_ASSERT(inp);
 479         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
 480
 481         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
 482         inp->inp_fport = 0;
 483         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
 484         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
 485         in_pcbrehash(inp);
 486 }
 487
 488 struct sockaddr *
 489 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
 490 {
 491         struct sockaddr_in6 *sin6;
 492
 493         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
 494         bzero(sin6, sizeof *sin6);
 495         sin6->sin6_family = AF_INET6;
 496         sin6->sin6_len = sizeof(*sin6);
 497         sin6->sin6_port = port;
 498         sin6->sin6_addr = *addr_p;
 499         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
 500
 501         return (struct sockaddr *)sin6;
 502 }
 503
 504 struct sockaddr *
 505 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
 506 {
 507         struct sockaddr_in sin;
 508         struct sockaddr_in6 *sin6_p;
 509
 510         bzero(&sin, sizeof sin);
 511         sin.sin_family = AF_INET;
 512         sin.sin_len = sizeof(sin);
 513         sin.sin_port = port;
 514         sin.sin_addr = *addr_p;
 515
 516         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
 517                 M_WAITOK);
 518         in6_sin_2_v4mapsin6(&sin, sin6_p);
 519
 520         return (struct sockaddr *)sin6_p;
 521 }
 522
 523 int
 524 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
 525 {
 526         struct inpcb *inp;
 527         struct in6_addr addr;
 528         in_port_t port;
 529
 530         inp = sotoinpcb(so);
 531         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
 532
 533         INP_RLOCK(inp);
 534         port = inp->inp_lport;
 535         addr = inp->in6p_laddr;
 536         INP_RUNLOCK(inp);
 537
 538         *nam = in6_sockaddr(port, &addr);
 539         return 0;
 540 }
 541
 542 int
 543 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
 544 {
 545         struct inpcb *inp;
 546         struct in6_addr addr;
 547         in_port_t port;
 548
 549         inp = sotoinpcb(so);
 550         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
 551
 552         INP_RLOCK(inp);
 553         port = inp->inp_fport;
 554         addr = inp->in6p_faddr;
 555         INP_RUNLOCK(inp);
 556
 557         *nam = in6_sockaddr(port, &addr);
 558         return 0;
 559 }
 560
 561 int
 562 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
 563 {
 564         struct  inpcb *inp;
 565         int     error;
 566
 567         inp = sotoinpcb(so);
 568         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
 569
 570 #ifdef INET
 571         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 572                 error = in_getsockaddr(so, nam);
 573                 if (error == 0)
 574                         in6_sin_2_v4mapsin6_in_sock(nam);
 575         } else
 576 #endif
 577         {
 578                 /* scope issues will be handled in in6_getsockaddr(). */
 579                 error = in6_getsockaddr(so, nam);
 580         }
 581
 582         return error;
 583 }
 584
 585 int
 586 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
 587 {
 588         struct  inpcb *inp;
 589         int     error;
 590
 591         inp = sotoinpcb(so);
 592         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
 593
 594 #ifdef INET
 595         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
 596                 error = in_getpeeraddr(so, nam);
 597                 if (error == 0)
 598                         in6_sin_2_v4mapsin6_in_sock(nam);
 599         } else
 600 #endif
 601         /* scope issues will be handled in in6_getpeeraddr(). */
 602         error = in6_getpeeraddr(so, nam);
 603
 604         return error;
 605 }
 606
 607 /*
 608  * Pass some notification to all connections of a protocol
 609  * associated with address dst.  The local address and/or port numbers
 610  * may be specified to limit the search.  The "usual action" will be
 611  * taken, depending on the ctlinput cmd.  The caller must filter any
 612  * cmds that are uninteresting (e.g., no error in the map).
 613  * Call the protocol specific routine (if any) to report
 614  * any errors for each matching socket.
 615  */
 616 void
 617 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr *dst,
 618     u_int fport_arg, const struct sockaddr *src, u_int lport_arg,
 619     int cmd, void *cmdarg,
 620     struct inpcb *(*notify)(struct inpcb *, int))
 621 {
 622         struct inpcb *inp, *inp_temp;
 623         struct sockaddr_in6 sa6_src, *sa6_dst;
 624         u_short fport = fport_arg, lport = lport_arg;
 625         u_int32_t flowinfo;
 626         int errno;
 627
 628         if ((unsigned)cmd >= PRC_NCMDS || dst->sa_family != AF_INET6)
 629                 return;
 630
 631         sa6_dst = (struct sockaddr_in6 *)dst;
 632         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
 633                 return;
 634
 635         /*
 636          * note that src can be NULL when we get notify by local fragmentation.
 637          */
 638         sa6_src = (src == NULL) ? sa6_any : *(const struct sockaddr_in6 *)src;
 639         flowinfo = sa6_src.sin6_flowinfo;
 640
 641         /*
 642          * Redirects go to all references to the destination,
 643          * and use in6_rtchange to invalidate the route cache.
 644          * Dead host indications: also use in6_rtchange to invalidate
 645          * the cache, and deliver the error to all the sockets.
 646          * Otherwise, if we have knowledge of the local port and address,
 647          * deliver only to that socket.
 648          */
 649         if (PRC_IS_REDIRECT(cmd) || cmd == PRC_HOSTDEAD) {
 650                 fport = 0;
 651                 lport = 0;
 652                 bzero((caddr_t)&sa6_src.sin6_addr, sizeof(sa6_src.sin6_addr));
 653
 654                 if (cmd != PRC_HOSTDEAD)
 655                         notify = in6_rtchange;
 656         }
 657         errno = inet6ctlerrmap[cmd];
 658         INP_INFO_WLOCK(pcbinfo);
 659         CK_LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, inp_temp) {
 660                 INP_WLOCK(inp);
 661                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 662                         INP_WUNLOCK(inp);
 663                         continue;
 664                 }
 665
 666                 /*
 667                  * If the error designates a new path MTU for a destination
 668                  * and the application (associated with this socket) wanted to
 669                  * know the value, notify.
 670                  * XXX: should we avoid to notify the value to TCP sockets?
 671                  */
 672                 if (cmd == PRC_MSGSIZE && cmdarg != NULL)
 673                         ip6_notify_pmtu(inp, (struct sockaddr_in6 *)dst,
 674                                         *(u_int32_t *)cmdarg);
 675
 676                 /*
 677                  * Detect if we should notify the error. If no source and
 678                  * destination ports are specifed, but non-zero flowinfo and
 679                  * local address match, notify the error. This is the case
 680                  * when the error is delivered with an encrypted buffer
 681                  * by ESP. Otherwise, just compare addresses and ports
 682                  * as usual.
 683                  */
 684                 if (lport == 0 && fport == 0 && flowinfo &&
 685                     inp->inp_socket != NULL &&
 686                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
 687                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
 688                         goto do_notify;
 689                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 690                                              &sa6_dst->sin6_addr) ||
 691                          inp->inp_socket == 0 ||
 692                          (lport && inp->inp_lport != lport) ||
 693                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
 694                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 695                                               &sa6_src.sin6_addr)) ||
 696                          (fport && inp->inp_fport != fport)) {
 697                         INP_WUNLOCK(inp);
 698                         continue;
 699                 }
 700
 701           do_notify:
 702                 if (notify) {
 703                         if ((*notify)(inp, errno))
 704                                 INP_WUNLOCK(inp);
 705                 } else
 706                         INP_WUNLOCK(inp);
 707         }
 708         INP_INFO_WUNLOCK(pcbinfo);
 709 }
 710
 711 /*
 712  * Lookup a PCB based on the local address and port.  Caller must hold the
 713  * hash lock.  No inpcb locks or references are acquired.
 714  */
 715 struct inpcb *
 716 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
 717     u_short lport, int lookupflags, struct ucred *cred)
 718 {
 719         struct inpcb *inp;
 720         int matchwild = 3, wildcard;
 721
 722         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
 723             ("%s: invalid lookup flags %d", __func__, lookupflags));
 724
 725         INP_HASH_LOCK_ASSERT(pcbinfo);
 726
 727         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
 728                 struct inpcbhead *head;
 729                 /*
 730                  * Look for an unconnected (wildcard foreign addr) PCB that
 731                  * matches the local address and port we're looking for.
 732                  */
 733                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
 734                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
 735                     pcbinfo->ipi_hashmask)];
 736                 CK_LIST_FOREACH(inp, head, inp_hash) {
 737                         /* XXX inp locking */
 738                         if ((inp->inp_vflag & INP_IPV6) == 0)
 739                                 continue;
 740                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
 741                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 742                             inp->inp_lport == lport) {
 743                                 /* Found. */
 744                                 if (cred == NULL ||
 745                                     prison_equal_ip6(cred->cr_prison,
 746                                         inp->inp_cred->cr_prison))
 747                                         return (inp);
 748                         }
 749                 }
 750                 /*
 751                  * Not found.
 752                  */
 753                 return (NULL);
 754         } else {
 755                 struct inpcbporthead *porthash;
 756                 struct inpcbport *phd;
 757                 struct inpcb *match = NULL;
 758                 /*
 759                  * Best fit PCB lookup.
 760                  *
 761                  * First see if this local port is in use by looking on the
 762                  * port hash list.
 763                  */
 764                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
 765                     pcbinfo->ipi_porthashmask)];
 766                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
 767                         if (phd->phd_port == lport)
 768                                 break;
 769                 }
 770                 if (phd != NULL) {
 771                         /*
 772                          * Port is in use by one or more PCBs. Look for best
 773                          * fit.
 774                          */
 775                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 776                                 wildcard = 0;
 777                                 if (cred != NULL &&
 778                                     !prison_equal_ip6(cred->cr_prison,
 779                                         inp->inp_cred->cr_prison))
 780                                         continue;
 781                                 /* XXX inp locking */
 782                                 if ((inp->inp_vflag & INP_IPV6) == 0)
 783                                         continue;
 784                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
 785                                         wildcard++;
 786                                 if (!IN6_IS_ADDR_UNSPECIFIED(
 787                                         &inp->in6p_laddr)) {
 788                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
 789                                                 wildcard++;
 790                                         else if (!IN6_ARE_ADDR_EQUAL(
 791                                             &inp->in6p_laddr, laddr))
 792                                                 continue;
 793                                 } else {
 794                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
 795                                                 wildcard++;
 796                                 }
 797                                 if (wildcard < matchwild) {
 798                                         match = inp;
 799                                         matchwild = wildcard;
 800                                         if (matchwild == 0)
 801                                                 break;
 802                                 }
 803                         }
 804                 }
 805                 return (match);
 806         }
 807 }
 808
 809 void
 810 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 811 {
 812         struct inpcb *inp;
 813         struct in6_multi *inm;
 814         struct in6_mfilter *imf;
 815         struct ip6_moptions *im6o;
 816
 817         INP_INFO_WLOCK(pcbinfo);
 818         CK_LIST_FOREACH(inp, pcbinfo->ipi_listhead, inp_list) {
 819                 INP_WLOCK(inp);
 820                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 821                         INP_WUNLOCK(inp);
 822                         continue;
 823                 }
 824                 im6o = inp->in6p_moptions;
 825                 if ((inp->inp_vflag & INP_IPV6) && im6o != NULL) {
 826                         /*
 827                          * Unselect the outgoing ifp for multicast if it
 828                          * is being detached.
 829                          */
 830                         if (im6o->im6o_multicast_ifp == ifp)
 831                                 im6o->im6o_multicast_ifp = NULL;
 832                         /*
 833                          * Drop multicast group membership if we joined
 834                          * through the interface being detached.
 835                          */
 836 restart:
 837                         IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
 838                                 if ((inm = imf->im6f_in6m) == NULL)
 839                                         continue;
 840                                 if (inm->in6m_ifp != ifp)
 841                                         continue;
 842                                 ip6_mfilter_remove(&im6o->im6o_head, imf);
 843                                 IN6_MULTI_LOCK_ASSERT();
 844                                 in6_leavegroup_locked(inm, NULL);
 845                                 ip6_mfilter_free(imf);
 846                                 goto restart;
 847                         }
 848                 }
 849                 INP_WUNLOCK(inp);
 850         }
 851         INP_INFO_WUNLOCK(pcbinfo);
 852 }
 853
 854 /*
 855  * Check for alternatives when higher level complains
 856  * about service problems.  For now, invalidate cached
 857  * routing information.  If the route was created dynamically
 858  * (by a redirect), time to try a default gateway again.
 859  */
 860 void
 861 in6_losing(struct inpcb *inp)
 862 {
 863
 864         RO_INVALIDATE_CACHE(&inp->inp_route6);
 865 }
 866
 867 /*
 868  * After a routing change, flush old routing
 869  * and allocate a (hopefully) better one.
 870  */
 871 struct inpcb *
 872 in6_rtchange(struct inpcb *inp, int errno __unused)
 873 {
 874
 875         RO_INVALIDATE_CACHE(&inp->inp_route6);
 876         return inp;
 877 }
 878
 879 static struct inpcb *
 880 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
 881     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
 882     uint16_t fport, int lookupflags)
 883 {
 884         struct inpcb *local_wild;
 885         const struct inpcblbgrouphead *hdr;
 886         struct inpcblbgroup *grp;
 887         uint32_t idx;
 888
 889         INP_HASH_LOCK_ASSERT(pcbinfo);
 890
 891         hdr = &pcbinfo->ipi_lbgrouphashbase[
 892             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
 893
 894         /*
 895          * Order of socket selection:
 896          * 1. non-wild.
 897          * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD).
 898          *
 899          * NOTE:
 900          * - Load balanced group does not contain jailed sockets.
 901          * - Load balanced does not contain IPv4 mapped INET6 wild sockets.
 902          */
 903         local_wild = NULL;
 904         CK_LIST_FOREACH(grp, hdr, il_list) {
 905 #ifdef INET
 906                 if (!(grp->il_vflag & INP_IPV6))
 907                         continue;
 908 #endif
 909                 if (grp->il_lport != lport)
 910                         continue;
 911
 912                 idx = INP_PCBLBGROUP_PKTHASH(INP6_PCBHASHKEY(faddr), lport,
 913                     fport) % grp->il_inpcnt;
 914                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr))
 915                         return (grp->il_inp[idx]);
 916                 if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
 917                     (lookupflags & INPLOOKUP_WILDCARD) != 0)
 918                         local_wild = grp->il_inp[idx];
 919         }
 920         return (local_wild);
 921 }
 922
 923 #ifdef PCBGROUP
 924 /*
 925  * Lookup PCB in hash list, using pcbgroup tables.
 926  */
 927 static struct inpcb *
 928 in6_pcblookup_group(struct inpcbinfo *pcbinfo, struct inpcbgroup *pcbgroup,
 929     struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr,
 930     u_int lport_arg, int lookupflags, struct ifnet *ifp)
 931 {
 932         struct inpcbhead *head;
 933         struct inpcb *inp, *tmpinp;
 934         u_short fport = fport_arg, lport = lport_arg;
 935         bool locked;
 936
 937         /*
 938          * First look for an exact match.
 939          */
 940         tmpinp = NULL;
 941         INP_GROUP_LOCK(pcbgroup);
 942         head = &pcbgroup->ipg_hashbase[INP_PCBHASH(
 943             INP6_PCBHASHKEY(faddr), lport, fport, pcbgroup->ipg_hashmask)];
 944         CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 945                 /* XXX inp locking */
 946                 if ((inp->inp_vflag & INP_IPV6) == 0)
 947                         continue;
 948                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
 949                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
 950                     inp->inp_fport == fport &&
 951                     inp->inp_lport == lport) {
 952                         /*
 953                          * XXX We should be able to directly return
 954                          * the inp here, without any checks.
 955                          * Well unless both bound with SO_REUSEPORT?
 956                          */
 957                         if (prison_flag(inp->inp_cred, PR_IP6))
 958                                 goto found;
 959                         if (tmpinp == NULL)
 960                                 tmpinp = inp;
 961                 }
 962         }
 963         if (tmpinp != NULL) {
 964                 inp = tmpinp;
 965                 goto found;
 966         }
 967
 968         /*
 969          * Then look for a wildcard match in the pcbgroup.
 970          */
 971         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
 972                 struct inpcb *local_wild = NULL, *local_exact = NULL;
 973                 struct inpcb *jail_wild = NULL;
 974                 int injail;
 975
 976                 /*
 977                  * Order of socket selection - we always prefer jails.
 978                  *      1. jailed, non-wild.
 979                  *      2. jailed, wild.
 980                  *      3. non-jailed, non-wild.
 981                  *      4. non-jailed, wild.
 982                  */
 983                 head = &pcbgroup->ipg_hashbase[
 984                     INP_PCBHASH(INADDR_ANY, lport, 0, pcbgroup->ipg_hashmask)];
 985                 CK_LIST_FOREACH(inp, head, inp_pcbgrouphash) {
 986                         /* XXX inp locking */
 987                         if ((inp->inp_vflag & INP_IPV6) == 0)
 988                                 continue;
 989
 990                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
 991                             inp->inp_lport != lport) {
 992                                 continue;
 993                         }
 994
 995                         injail = prison_flag(inp->inp_cred, PR_IP6);
 996                         if (injail) {
 997                                 if (prison_check_ip6(inp->inp_cred,
 998                                     laddr) != 0)
 999                                         continue;
1000                         } else {
1001                                 if (local_exact != NULL)
1002                                         continue;
1003                         }
1004
1005                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1006                                 if (injail)
1007                                         goto found;
1008                                 else
1009                                         local_exact = inp;
1010                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1011                                 if (injail)
1012                                         jail_wild = inp;
1013                                 else
1014                                         local_wild = inp;
1015                         }
1016                 } /* LIST_FOREACH */
1017
1018                 inp = jail_wild;
1019                 if (inp == NULL)
1020                         inp = jail_wild;
1021                 if (inp == NULL)
1022                         inp = local_exact;
1023                 if (inp == NULL)
1024                         inp = local_wild;
1025                 if (inp != NULL)
1026                         goto found;
1027         }
1028
1029         /*
1030          * Then look for a wildcard match, if requested.
1031          */
1032         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1033                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1034                 struct inpcb *jail_wild = NULL;
1035                 int injail;
1036
1037                 /*
1038                  * Order of socket selection - we always prefer jails.
1039                  *      1. jailed, non-wild.
1040                  *      2. jailed, wild.
1041                  *      3. non-jailed, non-wild.
1042                  *      4. non-jailed, wild.
1043                  */
1044                 head = &pcbinfo->ipi_wildbase[INP_PCBHASH(
1045                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1046                     pcbinfo->ipi_wildmask)];
1047                 CK_LIST_FOREACH(inp, head, inp_pcbgroup_wild) {
1048                         /* XXX inp locking */
1049                         if ((inp->inp_vflag & INP_IPV6) == 0)
1050                                 continue;
1051
1052                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1053                             inp->inp_lport != lport) {
1054                                 continue;
1055                         }
1056
1057                         injail = prison_flag(inp->inp_cred, PR_IP6);
1058                         if (injail) {
1059                                 if (prison_check_ip6(inp->inp_cred,
1060                                     laddr) != 0)
1061                                         continue;
1062                         } else {
1063                                 if (local_exact != NULL)
1064                                         continue;
1065                         }
1066
1067                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1068                                 if (injail)
1069                                         goto found;
1070                                 else
1071                                         local_exact = inp;
1072                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1073                                 if (injail)
1074                                         jail_wild = inp;
1075                                 else
1076                                         local_wild = inp;
1077                         }
1078                 } /* LIST_FOREACH */
1079
1080                 inp = jail_wild;
1081                 if (inp == NULL)
1082                         inp = jail_wild;
1083                 if (inp == NULL)
1084                         inp = local_exact;
1085                 if (inp == NULL)
1086                         inp = local_wild;
1087                 if (inp != NULL)
1088                         goto found;
1089         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1090         INP_GROUP_UNLOCK(pcbgroup);
1091         return (NULL);
1092
1093 found:
1094         if (lookupflags & INPLOOKUP_WLOCKPCB)
1095                 locked = INP_TRY_WLOCK(inp);
1096         else if (lookupflags & INPLOOKUP_RLOCKPCB)
1097                 locked = INP_TRY_RLOCK(inp);
1098         else
1099                 panic("%s: locking buf", __func__);
1100         if (!locked)
1101                 in_pcbref(inp);
1102         INP_GROUP_UNLOCK(pcbgroup);
1103         if (!locked) {
1104                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1105                         INP_WLOCK(inp);
1106                         if (in_pcbrele_wlocked(inp))
1107                                 return (NULL);
1108                 } else {
1109                         INP_RLOCK(inp);
1110                         if (in_pcbrele_rlocked(inp))
1111                                 return (NULL);
1112                 }
1113         }
1114 #ifdef INVARIANTS
1115         if (lookupflags & INPLOOKUP_WLOCKPCB)
1116                 INP_WLOCK_ASSERT(inp);
1117         else
1118                 INP_RLOCK_ASSERT(inp);
1119 #endif
1120         return (inp);
1121 }
1122 #endif /* PCBGROUP */
1123
1124 /*
1125  * Lookup PCB in hash list.
1126  */
1127 static struct inpcb *
1128 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1129     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
1130     int lookupflags, struct ifnet *ifp)
1131 {
1132         struct inpcbhead *head;
1133         struct inpcb *inp, *tmpinp;
1134         u_short fport = fport_arg, lport = lport_arg;
1135
1136         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
1137             ("%s: invalid lookup flags %d", __func__, lookupflags));
1138
1139         INP_HASH_LOCK_ASSERT(pcbinfo);
1140
1141         /*
1142          * First look for an exact match.
1143          */
1144         tmpinp = NULL;
1145         head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1146             INP6_PCBHASHKEY(faddr), lport, fport, pcbinfo->ipi_hashmask)];
1147         CK_LIST_FOREACH(inp, head, inp_hash) {
1148                 /* XXX inp locking */
1149                 if ((inp->inp_vflag & INP_IPV6) == 0)
1150                         continue;
1151                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
1152                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
1153                     inp->inp_fport == fport &&
1154                     inp->inp_lport == lport) {
1155                         /*
1156                          * XXX We should be able to directly return
1157                          * the inp here, without any checks.
1158                          * Well unless both bound with SO_REUSEPORT?
1159                          */
1160                         if (prison_flag(inp->inp_cred, PR_IP6))
1161                                 return (inp);
1162                         if (tmpinp == NULL)
1163                                 tmpinp = inp;
1164                 }
1165         }
1166         if (tmpinp != NULL)
1167                 return (tmpinp);
1168
1169         /*
1170          * Then look in lb group (for wildcard match).
1171          */
1172         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1173                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1174                     fport, lookupflags);
1175                 if (inp != NULL)
1176                         return (inp);
1177         }
1178
1179         /*
1180          * Then look for a wildcard match, if requested.
1181          */
1182         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1183                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1184                 struct inpcb *jail_wild = NULL;
1185                 int injail;
1186
1187                 /*
1188                  * Order of socket selection - we always prefer jails.
1189                  *      1. jailed, non-wild.
1190                  *      2. jailed, wild.
1191                  *      3. non-jailed, non-wild.
1192                  *      4. non-jailed, wild.
1193                  */
1194                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH(
1195                     INP6_PCBHASHKEY(&in6addr_any), lport, 0,
1196                     pcbinfo->ipi_hashmask)];
1197                 CK_LIST_FOREACH(inp, head, inp_hash) {
1198                         /* XXX inp locking */
1199                         if ((inp->inp_vflag & INP_IPV6) == 0)
1200                                 continue;
1201
1202                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1203                             inp->inp_lport != lport) {
1204                                 continue;
1205                         }
1206
1207                         injail = prison_flag(inp->inp_cred, PR_IP6);
1208                         if (injail) {
1209                                 if (prison_check_ip6(inp->inp_cred,
1210                                     laddr) != 0)
1211                                         continue;
1212                         } else {
1213                                 if (local_exact != NULL)
1214                                         continue;
1215                         }
1216
1217                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1218                                 if (injail)
1219                                         return (inp);
1220                                 else
1221                                         local_exact = inp;
1222                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1223                                 if (injail)
1224                                         jail_wild = inp;
1225                                 else
1226                                         local_wild = inp;
1227                         }
1228                 } /* LIST_FOREACH */
1229
1230                 if (jail_wild != NULL)
1231                         return (jail_wild);
1232                 if (local_exact != NULL)
1233                         return (local_exact);
1234                 if (local_wild != NULL)
1235                         return (local_wild);
1236         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1237
1238         /*
1239          * Not found.
1240          */
1241         return (NULL);
1242 }
1243
1244 /*
1245  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1246  * hash list lock, and will return the inpcb locked (i.e., requires
1247  * INPLOOKUP_LOCKPCB).
1248  */
1249 static struct inpcb *
1250 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1251     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1252     struct ifnet *ifp)
1253 {
1254         struct inpcb *inp;
1255
1256         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1257             (lookupflags & ~(INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)), ifp);
1258         if (inp != NULL) {
1259                 if (lookupflags & INPLOOKUP_WLOCKPCB) {
1260                         INP_WLOCK(inp);
1261                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1262                                 INP_WUNLOCK(inp);
1263                                 inp = NULL;
1264                         }
1265                 } else if (lookupflags & INPLOOKUP_RLOCKPCB) {
1266                         INP_RLOCK(inp);
1267                         if (__predict_false(inp->inp_flags2 & INP_FREED)) {
1268                                 INP_RUNLOCK(inp);
1269                                 inp = NULL;
1270                         }
1271                 } else
1272                         panic("%s: locking bug", __func__);
1273 #ifdef INVARIANTS
1274                 if (inp != NULL) {
1275                         if (lookupflags & INPLOOKUP_WLOCKPCB)
1276                                 INP_WLOCK_ASSERT(inp);
1277                         else
1278                                 INP_RLOCK_ASSERT(inp);
1279                 }
1280 #endif
1281         }
1282         return (inp);
1283 }
1284
1285 /*
1286  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1287  * from which a pre-calculated hash value may be extracted.
1288  *
1289  * Possibly more of this logic should be in in6_pcbgroup.c.
1290  */
1291 struct inpcb *
1292 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1293     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1294 {
1295 #if defined(PCBGROUP) && !defined(RSS)
1296         struct inpcbgroup *pcbgroup;
1297 #endif
1298
1299         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1300             ("%s: invalid lookup flags %d", __func__, lookupflags));
1301         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1302             ("%s: LOCKPCB not set", __func__));
1303
1304         /*
1305          * When not using RSS, use connection groups in preference to the
1306          * reservation table when looking up 4-tuples.  When using RSS, just
1307          * use the reservation table, due to the cost of the Toeplitz hash
1308          * in software.
1309          *
1310          * XXXRW: This policy belongs in the pcbgroup code, as in principle
1311          * we could be doing RSS with a non-Toeplitz hash that is affordable
1312          * in software.
1313          */
1314 #if defined(PCBGROUP) && !defined(RSS)
1315         if (in_pcbgroup_enabled(pcbinfo)) {
1316                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1317                     fport);
1318                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1319                     laddr, lport, lookupflags, ifp));
1320         }
1321 #endif
1322         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1323             lookupflags, ifp));
1324 }
1325
1326 struct inpcb *
1327 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1328     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1329     struct ifnet *ifp, struct mbuf *m)
1330 {
1331 #ifdef PCBGROUP
1332         struct inpcbgroup *pcbgroup;
1333 #endif
1334
1335         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1336             ("%s: invalid lookup flags %d", __func__, lookupflags));
1337         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1338             ("%s: LOCKPCB not set", __func__));
1339
1340 #ifdef PCBGROUP
1341         /*
1342          * If we can use a hardware-generated hash to look up the connection
1343          * group, use that connection group to find the inpcb.  Otherwise
1344          * fall back on a software hash -- or the reservation table if we're
1345          * using RSS.
1346          *
1347          * XXXRW: As above, that policy belongs in the pcbgroup code.
1348          */
1349         if (in_pcbgroup_enabled(pcbinfo) &&
1350             M_HASHTYPE_TEST(m, M_HASHTYPE_NONE) == 0) {
1351                 pcbgroup = in6_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
1352                     m->m_pkthdr.flowid);
1353                 if (pcbgroup != NULL)
1354                         return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr,
1355                             fport, laddr, lport, lookupflags, ifp));
1356 #ifndef RSS
1357                 pcbgroup = in6_pcbgroup_bytuple(pcbinfo, laddr, lport, faddr,
1358                     fport);
1359                 return (in6_pcblookup_group(pcbinfo, pcbgroup, faddr, fport,
1360                     laddr, lport, lookupflags, ifp));
1361 #endif
1362         }
1363 #endif
1364         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1365             lookupflags, ifp));
1366 }
1367
1368 void
1369 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1370 {
1371         struct ip6_hdr *ip;
1372
1373         ip = mtod(m, struct ip6_hdr *);
1374         bzero(sin6, sizeof(*sin6));
1375         sin6->sin6_len = sizeof(*sin6);
1376         sin6->sin6_family = AF_INET6;
1377         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1378
1379         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1380
1381         return;
1382 }