sys/netinet6/udp6_usrreq.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  * Copyright (c) 2014 Kevin Lo
   7  * All rights reserved.
   8  *
   9  * Portions of this software were developed by Robert N. M. Watson under
  10  * contract to Juniper Networks, Inc.
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  * 1. Redistributions of source code must retain the above copyright
  16  *    notice, this list of conditions and the following disclaimer.
  17  * 2. Redistributions in binary form must reproduce the above copyright
  18  *    notice, this list of conditions and the following disclaimer in the
  19  *    documentation and/or other materials provided with the distribution.
  20  * 3. Neither the name of the project nor the names of its contributors
  21  *    may be used to endorse or promote products derived from this software
  22  *    without specific prior written permission.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  *
  36  *      $KAME: udp6_usrreq.c,v 1.27 2001/05/21 05:45:10 jinmei Exp $
  37  *      $KAME: udp6_output.c,v 1.31 2001/05/21 16:39:15 jinmei Exp $
  38  */
  39
  40 /*-
  41  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
  42  *      The Regents of the University of California.
  43  * All rights reserved.
  44  *
  45  * Redistribution and use in source and binary forms, with or without
  46  * modification, are permitted provided that the following conditions
  47  * are met:
  48  * 1. Redistributions of source code must retain the above copyright
  49  *    notice, this list of conditions and the following disclaimer.
  50  * 2. Redistributions in binary form must reproduce the above copyright
  51  *    notice, this list of conditions and the following disclaimer in the
  52  *    documentation and/or other materials provided with the distribution.
  53  * 3. Neither the name of the University nor the names of its contributors
  54  *    may be used to endorse or promote products derived from this software
  55  *    without specific prior written permission.
  56  *
  57  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  58  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  59  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  60  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  61  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  62  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  63  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  64  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  65  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  66  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  67  * SUCH DAMAGE.
  68  *
  69  *      @(#)udp_usrreq.c        8.6 (Berkeley) 5/23/95
  70  */
  71
  72 #include <sys/cdefs.h>
  73 __FBSDID("$FreeBSD$");
  74
  75 #include "opt_inet.h"
  76 #include "opt_inet6.h"
  77 #include "opt_ipsec.h"
  78 #include "opt_rss.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/jail.h>
  82 #include <sys/kernel.h>
  83 #include <sys/lock.h>
  84 #include <sys/mbuf.h>
  85 #include <sys/priv.h>
  86 #include <sys/proc.h>
  87 #include <sys/protosw.h>
  88 #include <sys/sdt.h>
  89 #include <sys/signalvar.h>
  90 #include <sys/socket.h>
  91 #include <sys/socketvar.h>
  92 #include <sys/sx.h>
  93 #include <sys/sysctl.h>
  94 #include <sys/syslog.h>
  95 #include <sys/systm.h>
  96
  97 #include <net/if.h>
  98 #include <net/if_var.h>
  99 #include <net/if_types.h>
 100 #include <net/route.h>
 101 #include <net/rss_config.h>
 102
 103 #include <netinet/in.h>
 104 #include <netinet/in_kdtrace.h>
 105 #include <netinet/in_pcb.h>
 106 #include <netinet/in_systm.h>
 107 #include <netinet/in_var.h>
 108 #include <netinet/ip.h>
 109 #include <netinet/ip6.h>
 110 #include <netinet/icmp6.h>
 111 #include <netinet/ip_var.h>
 112 #include <netinet/udp.h>
 113 #include <netinet/udp_var.h>
 114 #include <netinet/udplite.h>
 115
 116 #include <netinet6/ip6protosw.h>
 117 #include <netinet6/ip6_var.h>
 118 #include <netinet6/in6_pcb.h>
 119 #include <netinet6/in6_rss.h>
 120 #include <netinet6/udp6_var.h>
 121 #include <netinet6/scope6_var.h>
 122
 123 #include <netipsec/ipsec_support.h>
 124
 125 #include <security/mac/mac_framework.h>
 126
 127 /*
 128  * UDP protocol implementation.
 129  * Per RFC 768, August, 1980.
 130  */
 131
 132 extern struct protosw   inetsw[];
 133 static void             udp6_detach(struct socket *so);
 134
 135 static int
 136 udp6_append(struct inpcb *inp, struct mbuf *n, int off,
 137     struct sockaddr_in6 *fromsa)
 138 {
 139         struct socket *so;
 140         struct mbuf *opts = NULL, *tmp_opts;
 141         struct udpcb *up;
 142
 143         INP_LOCK_ASSERT(inp);
 144
 145         /*
 146          * Engage the tunneling protocol.
 147          */
 148         up = intoudpcb(inp);
 149         if (up->u_tun_func != NULL) {
 150                 in_pcbref(inp);
 151                 INP_RUNLOCK(inp);
 152                 (*up->u_tun_func)(n, off, inp, (struct sockaddr *)&fromsa[0],
 153                     up->u_tun_ctx);
 154                 INP_RLOCK(inp);
 155                 return (in_pcbrele_rlocked(inp));
 156         }
 157 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 158         /* Check AH/ESP integrity. */
 159         if (IPSEC_ENABLED(ipv6)) {
 160                 if (IPSEC_CHECK_POLICY(ipv6, n, inp) != 0) {
 161                         m_freem(n);
 162                         return (0);
 163                 }
 164         }
 165 #endif /* IPSEC */
 166 #ifdef MAC
 167         if (mac_inpcb_check_deliver(inp, n) != 0) {
 168                 m_freem(n);
 169                 return (0);
 170         }
 171 #endif
 172         opts = NULL;
 173         if (inp->inp_flags & INP_CONTROLOPTS ||
 174             inp->inp_socket->so_options & SO_TIMESTAMP)
 175                 ip6_savecontrol(inp, n, &opts);
 176         if ((inp->inp_vflag & INP_IPV6) && (inp->inp_flags2 & INP_ORIGDSTADDR)) {
 177                 tmp_opts = sbcreatecontrol((caddr_t)&fromsa[1],
 178                         sizeof(struct sockaddr_in6), IPV6_ORIGDSTADDR, IPPROTO_IPV6);
 179                 if (tmp_opts) {
 180                         if (opts) {
 181                                 tmp_opts->m_next = opts;
 182                                 opts = tmp_opts;
 183                         } else
 184                                 opts = tmp_opts;
 185                 }
 186
 187         }
 188         m_adj(n, off + sizeof(struct udphdr));
 189
 190         so = inp->inp_socket;
 191         SOCKBUF_LOCK(&so->so_rcv);
 192         if (sbappendaddr_locked(&so->so_rcv, (struct sockaddr *)&fromsa[0], n,
 193             opts) == 0) {
 194                 SOCKBUF_UNLOCK(&so->so_rcv);
 195                 m_freem(n);
 196                 if (opts)
 197                         m_freem(opts);
 198                 UDPSTAT_INC(udps_fullsock);
 199         } else
 200                 sorwakeup_locked(so);
 201         return (0);
 202 }
 203
 204 int
 205 udp6_input(struct mbuf **mp, int *offp, int proto)
 206 {
 207         struct mbuf *m = *mp;
 208         struct ifnet *ifp;
 209         struct ip6_hdr *ip6;
 210         struct udphdr *uh;
 211         struct inpcb *inp;
 212         struct inpcbinfo *pcbinfo;
 213         struct udpcb *up;
 214         int off = *offp;
 215         int cscov_partial;
 216         int plen, ulen;
 217         struct epoch_tracker et;
 218         struct sockaddr_in6 fromsa[2];
 219         struct m_tag *fwd_tag;
 220         uint16_t uh_sum;
 221         uint8_t nxt;
 222
 223         ifp = m->m_pkthdr.rcvif;
 224
 225 #ifndef PULLDOWN_TEST
 226         IP6_EXTHDR_CHECK(m, off, sizeof(struct udphdr), IPPROTO_DONE);
 227         ip6 = mtod(m, struct ip6_hdr *);
 228         uh = (struct udphdr *)((caddr_t)ip6 + off);
 229 #else
 230         IP6_EXTHDR_GET(uh, struct udphdr *, m, off, sizeof(*uh));
 231         if (!uh)
 232                 return (IPPROTO_DONE);
 233         ip6 = mtod(m, struct ip6_hdr *);
 234 #endif
 235
 236         UDPSTAT_INC(udps_ipackets);
 237
 238         /*
 239          * Destination port of 0 is illegal, based on RFC768.
 240          */
 241         if (uh->uh_dport == 0)
 242                 goto badunlocked;
 243
 244         plen = ntohs(ip6->ip6_plen) - off + sizeof(*ip6);
 245         ulen = ntohs((u_short)uh->uh_ulen);
 246
 247         nxt = proto;
 248         cscov_partial = (nxt == IPPROTO_UDPLITE) ? 1 : 0;
 249         if (nxt == IPPROTO_UDPLITE) {
 250                 /* Zero means checksum over the complete packet. */
 251                 if (ulen == 0)
 252                         ulen = plen;
 253                 if (ulen == plen)
 254                         cscov_partial = 0;
 255                 if ((ulen < sizeof(struct udphdr)) || (ulen > plen)) {
 256                         /* XXX: What is the right UDPLite MIB counter? */
 257                         goto badunlocked;
 258                 }
 259                 if (uh->uh_sum == 0) {
 260                         /* XXX: What is the right UDPLite MIB counter? */
 261                         goto badunlocked;
 262                 }
 263         } else {
 264                 if ((ulen < sizeof(struct udphdr)) || (plen != ulen)) {
 265                         UDPSTAT_INC(udps_badlen);
 266                         goto badunlocked;
 267                 }
 268                 if (uh->uh_sum == 0) {
 269                         UDPSTAT_INC(udps_nosum);
 270                         goto badunlocked;
 271                 }
 272         }
 273
 274         if ((m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) &&
 275             !cscov_partial) {
 276                 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
 277                         uh_sum = m->m_pkthdr.csum_data;
 278                 else
 279                         uh_sum = in6_cksum_pseudo(ip6, ulen, nxt,
 280                             m->m_pkthdr.csum_data);
 281                 uh_sum ^= 0xffff;
 282         } else
 283                 uh_sum = in6_cksum_partial(m, nxt, off, plen, ulen);
 284
 285         if (uh_sum != 0) {
 286                 UDPSTAT_INC(udps_badsum);
 287                 goto badunlocked;
 288         }
 289
 290         /*
 291          * Construct sockaddr format source address.
 292          */
 293         init_sin6(&fromsa[0], m, 0);
 294         fromsa[0].sin6_port = uh->uh_sport;
 295         init_sin6(&fromsa[1], m, 1);
 296         fromsa[1].sin6_port = uh->uh_dport;
 297
 298         pcbinfo = udp_get_inpcbinfo(nxt);
 299         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
 300                 struct inpcb *last;
 301                 struct inpcbhead *pcblist;
 302                 struct ip6_moptions *imo;
 303
 304                 INP_INFO_RLOCK_ET(pcbinfo, et);
 305                 /*
 306                  * In the event that laddr should be set to the link-local
 307                  * address (this happens in RIPng), the multicast address
 308                  * specified in the received packet will not match laddr.  To
 309                  * handle this situation, matching is relaxed if the
 310                  * receiving interface is the same as one specified in the
 311                  * socket and if the destination multicast address matches
 312                  * one of the multicast groups specified in the socket.
 313                  */
 314
 315                 /*
 316                  * KAME note: traditionally we dropped udpiphdr from mbuf
 317                  * here.  We need udphdr for IPsec processing so we do that
 318                  * later.
 319                  */
 320                 pcblist = udp_get_pcblist(nxt);
 321                 last = NULL;
 322                 CK_LIST_FOREACH(inp, pcblist, inp_list) {
 323                         if ((inp->inp_vflag & INP_IPV6) == 0)
 324                                 continue;
 325                         if (inp->inp_lport != uh->uh_dport)
 326                                 continue;
 327                         if (inp->inp_fport != 0 &&
 328                             inp->inp_fport != uh->uh_sport)
 329                                 continue;
 330                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 331                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 332                                                         &ip6->ip6_dst))
 333                                         continue;
 334                         }
 335                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 336                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 337                                                         &ip6->ip6_src) ||
 338                                     inp->inp_fport != uh->uh_sport)
 339                                         continue;
 340                         }
 341
 342                         /*
 343                          * XXXRW: Because we weren't holding either the inpcb
 344                          * or the hash lock when we checked for a match
 345                          * before, we should probably recheck now that the
 346                          * inpcb lock is (supposed to be) held.
 347                          */
 348
 349                         /*
 350                          * Handle socket delivery policy for any-source
 351                          * and source-specific multicast. [RFC3678]
 352                          */
 353                         imo = inp->in6p_moptions;
 354                         if (imo && IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
 355                                 struct sockaddr_in6      mcaddr;
 356                                 int                      blocked;
 357
 358                                 INP_RLOCK(inp);
 359                                 if (__predict_false(inp->inp_flags2 & INP_FREED)) {
 360                                         INP_RUNLOCK(inp);
 361                                         continue;
 362                                 }
 363
 364                                 bzero(&mcaddr, sizeof(struct sockaddr_in6));
 365                                 mcaddr.sin6_len = sizeof(struct sockaddr_in6);
 366                                 mcaddr.sin6_family = AF_INET6;
 367                                 mcaddr.sin6_addr = ip6->ip6_dst;
 368
 369                                 blocked = im6o_mc_filter(imo, ifp,
 370                                         (struct sockaddr *)&mcaddr,
 371                                         (struct sockaddr *)&fromsa[0]);
 372                                 if (blocked != MCAST_PASS) {
 373                                         if (blocked == MCAST_NOTGMEMBER)
 374                                                 IP6STAT_INC(ip6s_notmember);
 375                                         if (blocked == MCAST_NOTSMEMBER ||
 376                                             blocked == MCAST_MUTED)
 377                                                 UDPSTAT_INC(udps_filtermcast);
 378                                         INP_RUNLOCK(inp); /* XXX */
 379                                         continue;
 380                                 }
 381
 382                                 INP_RUNLOCK(inp);
 383                         }
 384                         if (last != NULL) {
 385                                 struct mbuf *n;
 386
 387                                 if ((n = m_copym(m, 0, M_COPYALL, M_NOWAIT)) !=
 388                                     NULL) {
 389                                         INP_RLOCK(last);
 390                                         if (__predict_true(last->inp_flags2 & INP_FREED) == 0) {
 391                                                 if (nxt == IPPROTO_UDPLITE)
 392                                                         UDPLITE_PROBE(receive, NULL, last,
 393                                                             ip6, last, uh);
 394                                                 else
 395                                                         UDP_PROBE(receive, NULL, last,
 396                                                             ip6, last, uh);
 397                                                 if (udp6_append(last, n, off, fromsa))
 398                                                         goto inp_lost;
 399                                         }
 400                                         INP_RUNLOCK(last);
 401                                 }
 402                         }
 403                         last = inp;
 404                         /*
 405                          * Don't look for additional matches if this one does
 406                          * not have either the SO_REUSEPORT or SO_REUSEADDR
 407                          * socket options set.  This heuristic avoids
 408                          * searching through all pcbs in the common case of a
 409                          * non-shared port.  It assumes that an application
 410                          * will never clear these options after setting them.
 411                          */
 412                         if ((last->inp_socket->so_options &
 413                              (SO_REUSEPORT|SO_REUSEPORT_LB|SO_REUSEADDR)) == 0)
 414                                 break;
 415                 }
 416
 417                 if (last == NULL) {
 418                         /*
 419                          * No matching pcb found; discard datagram.  (No need
 420                          * to send an ICMP Port Unreachable for a broadcast
 421                          * or multicast datgram.)
 422                          */
 423                         UDPSTAT_INC(udps_noport);
 424                         UDPSTAT_INC(udps_noportmcast);
 425                         goto badheadlocked;
 426                 }
 427                 INP_RLOCK(last);
 428                 if (__predict_true(last->inp_flags2 & INP_FREED) == 0) {
 429                         if (nxt == IPPROTO_UDPLITE)
 430                                 UDPLITE_PROBE(receive, NULL, last, ip6, last, uh);
 431                         else
 432                                 UDP_PROBE(receive, NULL, last, ip6, last, uh);
 433                         if (udp6_append(last, m, off, fromsa) == 0)
 434                                 INP_RUNLOCK(last);
 435                 } else
 436                         INP_RUNLOCK(last);
 437         inp_lost:
 438                 INP_INFO_RUNLOCK_ET(pcbinfo, et);
 439                 return (IPPROTO_DONE);
 440         }
 441         /*
 442          * Locate pcb for datagram.
 443          */
 444
 445         /*
 446          * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain.
 447          */
 448         if ((m->m_flags & M_IP6_NEXTHOP) &&
 449             (fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL)) != NULL) {
 450                 struct sockaddr_in6 *next_hop6;
 451
 452                 next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1);
 453
 454                 /*
 455                  * Transparently forwarded. Pretend to be the destination.
 456                  * Already got one like this?
 457                  */
 458                 inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_src,
 459                     uh->uh_sport, &ip6->ip6_dst, uh->uh_dport,
 460                     INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif, m);
 461                 if (!inp) {
 462                         /*
 463                          * It's new.  Try to find the ambushing socket.
 464                          * Because we've rewritten the destination address,
 465                          * any hardware-generated hash is ignored.
 466                          */
 467                         inp = in6_pcblookup(pcbinfo, &ip6->ip6_src,
 468                             uh->uh_sport, &next_hop6->sin6_addr,
 469                             next_hop6->sin6_port ? htons(next_hop6->sin6_port) :
 470                             uh->uh_dport, INPLOOKUP_WILDCARD |
 471                             INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif);
 472                 }
 473                 /* Remove the tag from the packet. We don't need it anymore. */
 474                 m_tag_delete(m, fwd_tag);
 475                 m->m_flags &= ~M_IP6_NEXTHOP;
 476         } else
 477                 inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_src,
 478                     uh->uh_sport, &ip6->ip6_dst, uh->uh_dport,
 479                     INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB,
 480                     m->m_pkthdr.rcvif, m);
 481         if (inp == NULL) {
 482                 if (udp_log_in_vain) {
 483                         char ip6bufs[INET6_ADDRSTRLEN];
 484                         char ip6bufd[INET6_ADDRSTRLEN];
 485
 486                         log(LOG_INFO,
 487                             "Connection attempt to UDP [%s]:%d from [%s]:%d\n",
 488                             ip6_sprintf(ip6bufd, &ip6->ip6_dst),
 489                             ntohs(uh->uh_dport),
 490                             ip6_sprintf(ip6bufs, &ip6->ip6_src),
 491                             ntohs(uh->uh_sport));
 492                 }
 493                 if (nxt == IPPROTO_UDPLITE)
 494                         UDPLITE_PROBE(receive, NULL, NULL, ip6, NULL, uh);
 495                 else
 496                         UDP_PROBE(receive, NULL, NULL, ip6, NULL, uh);
 497                 UDPSTAT_INC(udps_noport);
 498                 if (m->m_flags & M_MCAST) {
 499                         printf("UDP6: M_MCAST is set in a unicast packet.\n");
 500                         UDPSTAT_INC(udps_noportmcast);
 501                         goto badunlocked;
 502                 }
 503                 if (V_udp_blackhole)
 504                         goto badunlocked;
 505                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_NOPORT, 0);
 506                 return (IPPROTO_DONE);
 507         }
 508         INP_RLOCK_ASSERT(inp);
 509         up = intoudpcb(inp);
 510         if (cscov_partial) {
 511                 if (up->u_rxcslen == 0 || up->u_rxcslen > ulen) {
 512                         INP_RUNLOCK(inp);
 513                         m_freem(m);
 514                         return (IPPROTO_DONE);
 515                 }
 516         }
 517         if (nxt == IPPROTO_UDPLITE)
 518                 UDPLITE_PROBE(receive, NULL, inp, ip6, inp, uh);
 519         else
 520                 UDP_PROBE(receive, NULL, inp, ip6, inp, uh);
 521         if (udp6_append(inp, m, off, fromsa) == 0)
 522                 INP_RUNLOCK(inp);
 523         return (IPPROTO_DONE);
 524
 525 badheadlocked:
 526         INP_INFO_RUNLOCK_ET(pcbinfo, et);
 527 badunlocked:
 528         if (m)
 529                 m_freem(m);
 530         return (IPPROTO_DONE);
 531 }
 532
 533 static void
 534 udp6_common_ctlinput(int cmd, struct sockaddr *sa, void *d,
 535     struct inpcbinfo *pcbinfo)
 536 {
 537         struct udphdr uh;
 538         struct ip6_hdr *ip6;
 539         struct mbuf *m;
 540         int off = 0;
 541         struct ip6ctlparam *ip6cp = NULL;
 542         const struct sockaddr_in6 *sa6_src = NULL;
 543         void *cmdarg;
 544         struct inpcb *(*notify)(struct inpcb *, int) = udp_notify;
 545         struct udp_portonly {
 546                 u_int16_t uh_sport;
 547                 u_int16_t uh_dport;
 548         } *uhp;
 549
 550         if (sa->sa_family != AF_INET6 ||
 551             sa->sa_len != sizeof(struct sockaddr_in6))
 552                 return;
 553
 554         if ((unsigned)cmd >= PRC_NCMDS)
 555                 return;
 556         if (PRC_IS_REDIRECT(cmd))
 557                 notify = in6_rtchange, d = NULL;
 558         else if (cmd == PRC_HOSTDEAD)
 559                 d = NULL;
 560         else if (inet6ctlerrmap[cmd] == 0)
 561                 return;
 562
 563         /* if the parameter is from icmp6, decode it. */
 564         if (d != NULL) {
 565                 ip6cp = (struct ip6ctlparam *)d;
 566                 m = ip6cp->ip6c_m;
 567                 ip6 = ip6cp->ip6c_ip6;
 568                 off = ip6cp->ip6c_off;
 569                 cmdarg = ip6cp->ip6c_cmdarg;
 570                 sa6_src = ip6cp->ip6c_src;
 571         } else {
 572                 m = NULL;
 573                 ip6 = NULL;
 574                 cmdarg = NULL;
 575                 sa6_src = &sa6_any;
 576         }
 577
 578         if (ip6) {
 579                 /*
 580                  * XXX: We assume that when IPV6 is non NULL,
 581                  * M and OFF are valid.
 582                  */
 583
 584                 /* Check if we can safely examine src and dst ports. */
 585                 if (m->m_pkthdr.len < off + sizeof(*uhp))
 586                         return;
 587
 588                 bzero(&uh, sizeof(uh));
 589                 m_copydata(m, off, sizeof(*uhp), (caddr_t)&uh);
 590
 591                 if (!PRC_IS_REDIRECT(cmd)) {
 592                         /* Check to see if its tunneled */
 593                         struct inpcb *inp;
 594                         inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_dst,
 595                             uh.uh_dport, &ip6->ip6_src, uh.uh_sport,
 596                             INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB,
 597                             m->m_pkthdr.rcvif, m);
 598                         if (inp != NULL) {
 599                                 struct udpcb *up;
 600
 601                                 up = intoudpcb(inp);
 602                                 if (up->u_icmp_func) {
 603                                         /* Yes it is. */
 604                                         INP_RUNLOCK(inp);
 605                                         (*up->u_icmp_func)(cmd, (struct sockaddr *)ip6cp->ip6c_src,
 606                                               d, up->u_tun_ctx);
 607                                         return;
 608                                 } else {
 609                                         /* Can't find it. */
 610                                         INP_RUNLOCK(inp);
 611                                 }
 612                         }
 613                 }
 614                 (void)in6_pcbnotify(pcbinfo, sa, uh.uh_dport,
 615                     (struct sockaddr *)ip6cp->ip6c_src, uh.uh_sport, cmd,
 616                     cmdarg, notify);
 617         } else
 618                 (void)in6_pcbnotify(pcbinfo, sa, 0,
 619                     (const struct sockaddr *)sa6_src, 0, cmd, cmdarg, notify);
 620 }
 621
 622 void
 623 udp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
 624 {
 625
 626         return (udp6_common_ctlinput(cmd, sa, d, &V_udbinfo));
 627 }
 628
 629 void
 630 udplite6_ctlinput(int cmd, struct sockaddr *sa, void *d)
 631 {
 632
 633         return (udp6_common_ctlinput(cmd, sa, d, &V_ulitecbinfo));
 634 }
 635
 636 static int
 637 udp6_getcred(SYSCTL_HANDLER_ARGS)
 638 {
 639         struct xucred xuc;
 640         struct sockaddr_in6 addrs[2];
 641         struct inpcb *inp;
 642         int error;
 643
 644         error = priv_check(req->td, PRIV_NETINET_GETCRED);
 645         if (error)
 646                 return (error);
 647
 648         if (req->newlen != sizeof(addrs))
 649                 return (EINVAL);
 650         if (req->oldlen != sizeof(struct xucred))
 651                 return (EINVAL);
 652         error = SYSCTL_IN(req, addrs, sizeof(addrs));
 653         if (error)
 654                 return (error);
 655         if ((error = sa6_embedscope(&addrs[0], V_ip6_use_defzone)) != 0 ||
 656             (error = sa6_embedscope(&addrs[1], V_ip6_use_defzone)) != 0) {
 657                 return (error);
 658         }
 659         inp = in6_pcblookup(&V_udbinfo, &addrs[1].sin6_addr,
 660             addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port,
 661             INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, NULL);
 662         if (inp != NULL) {
 663                 INP_RLOCK_ASSERT(inp);
 664                 if (inp->inp_socket == NULL)
 665                         error = ENOENT;
 666                 if (error == 0)
 667                         error = cr_canseesocket(req->td->td_ucred,
 668                             inp->inp_socket);
 669                 if (error == 0)
 670                         cru2x(inp->inp_cred, &xuc);
 671                 INP_RUNLOCK(inp);
 672         } else
 673                 error = ENOENT;
 674         if (error == 0)
 675                 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
 676         return (error);
 677 }
 678
 679 SYSCTL_PROC(_net_inet6_udp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW, 0,
 680     0, udp6_getcred, "S,xucred", "Get the xucred of a UDP6 connection");
 681
 682 #define UH_WLOCKED      2
 683 #define UH_RLOCKED      1
 684 #define UH_UNLOCKED     0
 685 static int
 686 udp6_output(struct socket *so, int flags_arg, struct mbuf *m,
 687     struct sockaddr *addr6, struct mbuf *control, struct thread *td)
 688 {
 689         struct inpcbinfo *pcbinfo;
 690         struct inpcb *inp;
 691         struct ip6_hdr *ip6;
 692         struct udphdr *udp6;
 693         struct in6_addr *laddr, *faddr, in6a;
 694         struct ip6_pktopts *optp, opt;
 695         struct sockaddr_in6 *sin6, tmp;
 696         struct epoch_tracker et;
 697         int cscov_partial, error, flags, hlen, scope_ambiguous;
 698         u_int32_t ulen, plen;
 699         uint16_t cscov;
 700         u_short fport;
 701         uint8_t nxt, unlock_inp, unlock_udbinfo;
 702
 703         /* addr6 has been validated in udp6_send(). */
 704         sin6 = (struct sockaddr_in6 *)addr6;
 705
 706         /*
 707          * In contrast to to IPv4 we do not validate the max. packet length
 708          * here due to IPv6 Jumbograms (RFC2675).
 709          */
 710
 711         scope_ambiguous = 0;
 712         if (sin6) {
 713                 /* Protect *addr6 from overwrites. */
 714                 tmp = *sin6;
 715                 sin6 = &tmp;
 716
 717                 /*
 718                  * Application should provide a proper zone ID or the use of
 719                  * default zone IDs should be enabled.  Unfortunately, some
 720                  * applications do not behave as it should, so we need a
 721                  * workaround.  Even if an appropriate ID is not determined,
 722                  * we'll see if we can determine the outgoing interface.  If we
 723                  * can, determine the zone ID based on the interface below.
 724                  */
 725                 if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 726                         scope_ambiguous = 1;
 727                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) {
 728                         if (control)
 729                                 m_freem(control);
 730                         m_freem(m);
 731                         return (error);
 732                 }
 733         }
 734
 735         inp = sotoinpcb(so);
 736         KASSERT(inp != NULL, ("%s: inp == NULL", __func__));
 737         /*
 738          * In the following cases we want a write lock on the inp for either
 739          * local operations or for possible route cache updates in the IPv6
 740          * output path:
 741          * - on connected sockets (sin6 is NULL) for route cache updates,
 742          * - when we are not bound to an address and source port (it is
 743          *   in6_pcbsetport() which will require the write lock).
 744          */
 745 retry:
 746         if (sin6 == NULL || (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) &&
 747             inp->inp_lport == 0)) {
 748                 INP_WLOCK(inp);
 749                 /*
 750                  * In case we lost a race and another thread bound addr/port
 751                  * on the inp we cannot keep the wlock (which still would be
 752                  * fine) as further down, based on these values we make
 753                  * decisions for the pcbinfo lock.  If the locks are not in
 754                  * synch the assertions on unlock will fire, hence we go for
 755                  * one retry loop.
 756                  */
 757                 if (sin6 != NULL &&
 758                     (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) ||
 759                     inp->inp_lport != 0)) {
 760                         INP_WUNLOCK(inp);
 761                         goto retry;
 762                 }
 763                 unlock_inp = UH_WLOCKED;
 764         } else {
 765                 INP_RLOCK(inp);
 766                 unlock_inp = UH_RLOCKED;
 767         }
 768         nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ?
 769             IPPROTO_UDP : IPPROTO_UDPLITE;
 770
 771 #ifdef INET
 772         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
 773                 int hasv4addr;
 774
 775                 if (sin6 == NULL)
 776                         hasv4addr = (inp->inp_vflag & INP_IPV4);
 777                 else
 778                         hasv4addr = IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)
 779                             ? 1 : 0;
 780                 if (hasv4addr) {
 781                         struct pr_usrreqs *pru;
 782
 783                         /*
 784                          * XXXRW: We release UDP-layer locks before calling
 785                          * udp_send() in order to avoid recursion.  However,
 786                          * this does mean there is a short window where inp's
 787                          * fields are unstable.  Could this lead to a
 788                          * potential race in which the factors causing us to
 789                          * select the UDPv4 output routine are invalidated?
 790                          */
 791                         if (unlock_inp == UH_WLOCKED)
 792                                 INP_WUNLOCK(inp);
 793                         else
 794                                 INP_RUNLOCK(inp);
 795                         if (sin6)
 796                                 in6_sin6_2_sin_in_sock((struct sockaddr *)sin6);
 797                         pru = inetsw[ip_protox[nxt]].pr_usrreqs;
 798                         /* addr will just be freed in sendit(). */
 799                         return ((*pru->pru_send)(so, flags_arg, m,
 800                             (struct sockaddr *)sin6, control, td));
 801                 }
 802         } else
 803 #endif
 804         if (sin6 && IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
 805                 /*
 806                  * Given this is either an IPv6-only socket or no INET is
 807                  * supported we will fail the send if the given destination
 808                  * address is a v4mapped address.
 809                  */
 810                 if (unlock_inp == UH_WLOCKED)
 811                         INP_WUNLOCK(inp);
 812                 else
 813                         INP_RUNLOCK(inp);
 814                 return (EINVAL);
 815         }
 816
 817         if (control) {
 818                 if ((error = ip6_setpktopts(control, &opt,
 819                     inp->in6p_outputopts, td->td_ucred, nxt)) != 0) {
 820                         if (unlock_inp == UH_WLOCKED)
 821                                 INP_WUNLOCK(inp);
 822                         else
 823                                 INP_RUNLOCK(inp);
 824                         ip6_clearpktopts(&opt, -1);
 825                         if (control)
 826                                 m_freem(control);
 827                         m_freem(m);
 828                         return (error);
 829                 }
 830                 optp = &opt;
 831         } else
 832                 optp = inp->in6p_outputopts;
 833
 834         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
 835         if (sin6 != NULL &&
 836             IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) && inp->inp_lport == 0) {
 837                 INP_HASH_WLOCK(pcbinfo);
 838                 unlock_udbinfo = UH_WLOCKED;
 839         } else if (sin6 != NULL &&
 840             (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
 841             IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) ||
 842             inp->inp_lport == 0)) {
 843                 INP_HASH_RLOCK_ET(pcbinfo, et);
 844                 unlock_udbinfo = UH_RLOCKED;
 845         } else
 846                 unlock_udbinfo = UH_UNLOCKED;
 847
 848         if (sin6) {
 849
 850                 /*
 851                  * Since we saw no essential reason for calling in_pcbconnect,
 852                  * we get rid of such kind of logic, and call in6_selectsrc
 853                  * and in6_pcbsetport in order to fill in the local address
 854                  * and the local port.
 855                  */
 856                 if (sin6->sin6_port == 0) {
 857                         error = EADDRNOTAVAIL;
 858                         goto release;
 859                 }
 860
 861                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 862                         /* how about ::ffff:0.0.0.0 case? */
 863                         error = EISCONN;
 864                         goto release;
 865                 }
 866
 867                 /*
 868                  * Given we handle the v4mapped case in the INET block above
 869                  * assert here that it must not happen anymore.
 870                  */
 871                 KASSERT(!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr),
 872                     ("%s: sin6(%p)->sin6_addr is v4mapped which we "
 873                     "should have handled.", __func__, sin6));
 874
 875                 /* This only requires read-locking. */
 876                 error = in6_selectsrc_socket(sin6, optp, inp,
 877                     td->td_ucred, scope_ambiguous, &in6a, NULL);
 878                 if (error)
 879                         goto release;
 880                 laddr = &in6a;
 881
 882                 if (inp->inp_lport == 0) {
 883
 884                         INP_WLOCK_ASSERT(inp);
 885                         error = in6_pcbsetport(laddr, inp, td->td_ucred);
 886                         if (error != 0) {
 887                                 /* Undo an address bind that may have occurred. */
 888                                 inp->in6p_laddr = in6addr_any;
 889                                 goto release;
 890                         }
 891                 }
 892                 faddr = &sin6->sin6_addr;
 893                 fport = sin6->sin6_port; /* allow 0 port */
 894
 895         } else {
 896                 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 897                         error = ENOTCONN;
 898                         goto release;
 899                 }
 900                 laddr = &inp->in6p_laddr;
 901                 faddr = &inp->in6p_faddr;
 902                 fport = inp->inp_fport;
 903         }
 904
 905         ulen = m->m_pkthdr.len;
 906         plen = sizeof(struct udphdr) + ulen;
 907         hlen = sizeof(struct ip6_hdr);
 908
 909         /*
 910          * Calculate data length and get a mbuf
 911          * for UDP and IP6 headers.
 912          */
 913         M_PREPEND(m, hlen + sizeof(struct udphdr), M_NOWAIT);
 914         if (m == NULL) {
 915                 error = ENOBUFS;
 916                 goto release;
 917         }
 918
 919         /*
 920          * Stuff checksum and output datagram.
 921          */
 922         cscov = cscov_partial = 0;
 923         udp6 = (struct udphdr *)(mtod(m, caddr_t) + hlen);
 924         udp6->uh_sport = inp->inp_lport; /* lport is always set in the PCB */
 925         udp6->uh_dport = fport;
 926         if (nxt == IPPROTO_UDPLITE) {
 927                 struct udpcb *up;
 928
 929                 up = intoudpcb(inp);
 930                 cscov = up->u_txcslen;
 931                 if (cscov >= plen)
 932                         cscov = 0;
 933                 udp6->uh_ulen = htons(cscov);
 934                 /*
 935                  * For UDP-Lite, checksum coverage length of zero means
 936                  * the entire UDPLite packet is covered by the checksum.
 937                  */
 938                 cscov_partial = (cscov == 0) ? 0 : 1;
 939         } else if (plen <= 0xffff)
 940                 udp6->uh_ulen = htons((u_short)plen);
 941         else
 942                 udp6->uh_ulen = 0;
 943         udp6->uh_sum = 0;
 944
 945         ip6 = mtod(m, struct ip6_hdr *);
 946         ip6->ip6_flow   = inp->inp_flow & IPV6_FLOWINFO_MASK;
 947         ip6->ip6_vfc    &= ~IPV6_VERSION_MASK;
 948         ip6->ip6_vfc    |= IPV6_VERSION;
 949         ip6->ip6_plen   = htons((u_short)plen);
 950         ip6->ip6_nxt    = nxt;
 951         ip6->ip6_hlim   = in6_selecthlim(inp, NULL);
 952         ip6->ip6_src    = *laddr;
 953         ip6->ip6_dst    = *faddr;
 954
 955 #ifdef MAC
 956         mac_inpcb_create_mbuf(inp, m);
 957 #endif
 958
 959         if (cscov_partial) {
 960                 if ((udp6->uh_sum = in6_cksum_partial(m, nxt,
 961                     sizeof(struct ip6_hdr), plen, cscov)) == 0)
 962                         udp6->uh_sum = 0xffff;
 963         } else {
 964                 udp6->uh_sum = in6_cksum_pseudo(ip6, plen, nxt, 0);
 965                 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
 966                 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
 967         }
 968
 969         flags = 0;
 970 #ifdef  RSS
 971         {
 972                 uint32_t hash_val, hash_type;
 973                 uint8_t pr;
 974
 975                 pr = inp->inp_socket->so_proto->pr_protocol;
 976                 /*
 977                  * Calculate an appropriate RSS hash for UDP and
 978                  * UDP Lite.
 979                  *
 980                  * The called function will take care of figuring out
 981                  * whether a 2-tuple or 4-tuple hash is required based
 982                  * on the currently configured scheme.
 983                  *
 984                  * Later later on connected socket values should be
 985                  * cached in the inpcb and reused, rather than constantly
 986                  * re-calculating it.
 987                  *
 988                  * UDP Lite is a different protocol number and will
 989                  * likely end up being hashed as a 2-tuple until
 990                  * RSS / NICs grow UDP Lite protocol awareness.
 991                  */
 992                 if (rss_proto_software_hash_v6(faddr, laddr, fport,
 993                     inp->inp_lport, pr, &hash_val, &hash_type) == 0) {
 994                         m->m_pkthdr.flowid = hash_val;
 995                         M_HASHTYPE_SET(m, hash_type);
 996                 }
 997
 998                 /*
 999                  * Don't override with the inp cached flowid.
1000                  *
1001                  * Until the whole UDP path is vetted, it may actually
1002                  * be incorrect.
1003                  */
1004                 flags |= IP_NODEFAULTFLOWID;
1005         }
1006 #endif
1007
1008         UDPSTAT_INC(udps_opackets);
1009         if (unlock_udbinfo == UH_WLOCKED)
1010                 INP_HASH_WUNLOCK(pcbinfo);
1011         else if (unlock_udbinfo == UH_RLOCKED)
1012                 INP_HASH_RUNLOCK_ET(pcbinfo, et);
1013         if (nxt == IPPROTO_UDPLITE)
1014                 UDPLITE_PROBE(send, NULL, inp, ip6, inp, udp6);
1015         else
1016                 UDP_PROBE(send, NULL, inp, ip6, inp, udp6);
1017         error = ip6_output(m, optp,
1018             (unlock_inp == UH_WLOCKED) ? &inp->inp_route6 : NULL, flags,
1019             inp->in6p_moptions, NULL, inp);
1020         if (unlock_inp == UH_WLOCKED)
1021                 INP_WUNLOCK(inp);
1022         else
1023                 INP_RUNLOCK(inp);
1024
1025         if (control) {
1026                 ip6_clearpktopts(&opt, -1);
1027                 m_freem(control);
1028         }
1029         return (error);
1030
1031 release:
1032         if (unlock_udbinfo == UH_WLOCKED) {
1033                 KASSERT(unlock_inp == UH_WLOCKED, ("%s: excl udbinfo lock, "
1034                     "non-excl inp lock: pcbinfo %p %#x inp %p %#x",
1035                     __func__, pcbinfo, unlock_udbinfo, inp, unlock_inp));
1036                 INP_HASH_WUNLOCK(pcbinfo);
1037                 INP_WUNLOCK(inp);
1038         } else if (unlock_udbinfo == UH_RLOCKED) {
1039                 KASSERT(unlock_inp == UH_RLOCKED, ("%s: non-excl udbinfo lock, "
1040                     "excl inp lock: pcbinfo %p %#x inp %p %#x",
1041                     __func__, pcbinfo, unlock_udbinfo, inp, unlock_inp));
1042                 INP_HASH_RUNLOCK_ET(pcbinfo, et);
1043                 INP_RUNLOCK(inp);
1044         } else if (unlock_inp == UH_WLOCKED)
1045                 INP_WUNLOCK(inp);
1046         else
1047                 INP_RUNLOCK(inp);
1048         if (control) {
1049                 ip6_clearpktopts(&opt, -1);
1050                 m_freem(control);
1051         }
1052         m_freem(m);
1053
1054         return (error);
1055 }
1056
1057 static void
1058 udp6_abort(struct socket *so)
1059 {
1060         struct inpcb *inp;
1061         struct inpcbinfo *pcbinfo;
1062
1063         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1064         inp = sotoinpcb(so);
1065         KASSERT(inp != NULL, ("udp6_abort: inp == NULL"));
1066
1067         INP_WLOCK(inp);
1068 #ifdef INET
1069         if (inp->inp_vflag & INP_IPV4) {
1070                 struct pr_usrreqs *pru;
1071                 uint8_t nxt;
1072
1073                 nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ?
1074                     IPPROTO_UDP : IPPROTO_UDPLITE;
1075                 INP_WUNLOCK(inp);
1076                 pru = inetsw[ip_protox[nxt]].pr_usrreqs;
1077                 (*pru->pru_abort)(so);
1078                 return;
1079         }
1080 #endif
1081
1082         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1083                 INP_HASH_WLOCK(pcbinfo);
1084                 in6_pcbdisconnect(inp);
1085                 inp->in6p_laddr = in6addr_any;
1086                 INP_HASH_WUNLOCK(pcbinfo);
1087                 soisdisconnected(so);
1088         }
1089         INP_WUNLOCK(inp);
1090 }
1091
1092 static int
1093 udp6_attach(struct socket *so, int proto, struct thread *td)
1094 {
1095         struct inpcb *inp;
1096         struct inpcbinfo *pcbinfo;
1097         int error;
1098
1099         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1100         inp = sotoinpcb(so);
1101         KASSERT(inp == NULL, ("udp6_attach: inp != NULL"));
1102
1103         if (so->so_snd.sb_hiwat == 0 || so->so_rcv.sb_hiwat == 0) {
1104                 error = soreserve(so, udp_sendspace, udp_recvspace);
1105                 if (error)
1106                         return (error);
1107         }
1108         INP_INFO_WLOCK(pcbinfo);
1109         error = in_pcballoc(so, pcbinfo);
1110         if (error) {
1111                 INP_INFO_WUNLOCK(pcbinfo);
1112                 return (error);
1113         }
1114         inp = (struct inpcb *)so->so_pcb;
1115         inp->inp_vflag |= INP_IPV6;
1116         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0)
1117                 inp->inp_vflag |= INP_IPV4;
1118         inp->in6p_hops = -1;    /* use kernel default */
1119         inp->in6p_cksum = -1;   /* just to be sure */
1120         /*
1121          * XXX: ugly!!
1122          * IPv4 TTL initialization is necessary for an IPv6 socket as well,
1123          * because the socket may be bound to an IPv6 wildcard address,
1124          * which may match an IPv4-mapped IPv6 address.
1125          */
1126         inp->inp_ip_ttl = V_ip_defttl;
1127
1128         error = udp_newudpcb(inp);
1129         if (error) {
1130                 in_pcbdetach(inp);
1131                 in_pcbfree(inp);
1132                 INP_INFO_WUNLOCK(pcbinfo);
1133                 return (error);
1134         }
1135         INP_WUNLOCK(inp);
1136         INP_INFO_WUNLOCK(pcbinfo);
1137         return (0);
1138 }
1139
1140 static int
1141 udp6_bind(struct socket *so, struct sockaddr *nam, struct thread *td)
1142 {
1143         struct inpcb *inp;
1144         struct inpcbinfo *pcbinfo;
1145         int error;
1146         u_char vflagsav;
1147
1148         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1149         inp = sotoinpcb(so);
1150         KASSERT(inp != NULL, ("udp6_bind: inp == NULL"));
1151
1152         INP_WLOCK(inp);
1153         INP_HASH_WLOCK(pcbinfo);
1154         vflagsav = inp->inp_vflag;
1155         inp->inp_vflag &= ~INP_IPV4;
1156         inp->inp_vflag |= INP_IPV6;
1157         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
1158                 struct sockaddr_in6 *sin6_p;
1159
1160                 sin6_p = (struct sockaddr_in6 *)nam;
1161
1162                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6_p->sin6_addr))
1163                         inp->inp_vflag |= INP_IPV4;
1164 #ifdef INET
1165                 else if (IN6_IS_ADDR_V4MAPPED(&sin6_p->sin6_addr)) {
1166                         struct sockaddr_in sin;
1167
1168                         in6_sin6_2_sin(&sin, sin6_p);
1169                         inp->inp_vflag |= INP_IPV4;
1170                         inp->inp_vflag &= ~INP_IPV6;
1171                         error = in_pcbbind(inp, (struct sockaddr *)&sin,
1172                             td->td_ucred);
1173                         goto out;
1174                 }
1175 #endif
1176         }
1177
1178         error = in6_pcbbind(inp, nam, td->td_ucred);
1179 #ifdef INET
1180 out:
1181 #endif
1182         if (error != 0)
1183                 inp->inp_vflag = vflagsav;
1184         INP_HASH_WUNLOCK(pcbinfo);
1185         INP_WUNLOCK(inp);
1186         return (error);
1187 }
1188
1189 static void
1190 udp6_close(struct socket *so)
1191 {
1192         struct inpcb *inp;
1193         struct inpcbinfo *pcbinfo;
1194
1195         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1196         inp = sotoinpcb(so);
1197         KASSERT(inp != NULL, ("udp6_close: inp == NULL"));
1198
1199         INP_WLOCK(inp);
1200 #ifdef INET
1201         if (inp->inp_vflag & INP_IPV4) {
1202                 struct pr_usrreqs *pru;
1203                 uint8_t nxt;
1204
1205                 nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ?
1206                     IPPROTO_UDP : IPPROTO_UDPLITE;
1207                 INP_WUNLOCK(inp);
1208                 pru = inetsw[ip_protox[nxt]].pr_usrreqs;
1209                 (*pru->pru_disconnect)(so);
1210                 return;
1211         }
1212 #endif
1213         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1214                 INP_HASH_WLOCK(pcbinfo);
1215                 in6_pcbdisconnect(inp);
1216                 inp->in6p_laddr = in6addr_any;
1217                 INP_HASH_WUNLOCK(pcbinfo);
1218                 soisdisconnected(so);
1219         }
1220         INP_WUNLOCK(inp);
1221 }
1222
1223 static int
1224 udp6_connect(struct socket *so, struct sockaddr *nam, struct thread *td)
1225 {
1226         struct inpcb *inp;
1227         struct inpcbinfo *pcbinfo;
1228         struct sockaddr_in6 *sin6;
1229         int error;
1230         u_char vflagsav;
1231
1232         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1233         inp = sotoinpcb(so);
1234         sin6 = (struct sockaddr_in6 *)nam;
1235         KASSERT(inp != NULL, ("udp6_connect: inp == NULL"));
1236
1237         /*
1238          * XXXRW: Need to clarify locking of v4/v6 flags.
1239          */
1240         INP_WLOCK(inp);
1241 #ifdef INET
1242         if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
1243                 struct sockaddr_in sin;
1244
1245                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) {
1246                         error = EINVAL;
1247                         goto out;
1248                 }
1249                 if ((inp->inp_vflag & INP_IPV4) == 0) {
1250                         error = EAFNOSUPPORT;
1251                         goto out;
1252                 }
1253                 if (inp->inp_faddr.s_addr != INADDR_ANY) {
1254                         error = EISCONN;
1255                         goto out;
1256                 }
1257                 in6_sin6_2_sin(&sin, sin6);
1258                 error = prison_remote_ip4(td->td_ucred, &sin.sin_addr);
1259                 if (error != 0)
1260                         goto out;
1261                 vflagsav = inp->inp_vflag;
1262                 inp->inp_vflag |= INP_IPV4;
1263                 inp->inp_vflag &= ~INP_IPV6;
1264                 INP_HASH_WLOCK(pcbinfo);
1265                 error = in_pcbconnect(inp, (struct sockaddr *)&sin,
1266                     td->td_ucred);
1267                 INP_HASH_WUNLOCK(pcbinfo);
1268                 /*
1269                  * If connect succeeds, mark socket as connected. If
1270                  * connect fails and socket is unbound, reset inp_vflag
1271                  * field.
1272                  */
1273                 if (error == 0)
1274                         soisconnected(so);
1275                 else if (inp->inp_laddr.s_addr == INADDR_ANY &&
1276                     inp->inp_lport == 0)
1277                         inp->inp_vflag = vflagsav;
1278                 goto out;
1279         } else {
1280                 if ((inp->inp_vflag & INP_IPV6) == 0) {
1281                         error = EAFNOSUPPORT;
1282                         goto out;
1283                 }
1284         }
1285 #endif
1286         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1287                 error = EISCONN;
1288                 goto out;
1289         }
1290         error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr);
1291         if (error != 0)
1292                 goto out;
1293         vflagsav = inp->inp_vflag;
1294         inp->inp_vflag &= ~INP_IPV4;
1295         inp->inp_vflag |= INP_IPV6;
1296         INP_HASH_WLOCK(pcbinfo);
1297         error = in6_pcbconnect(inp, nam, td->td_ucred);
1298         INP_HASH_WUNLOCK(pcbinfo);
1299         /*
1300          * If connect succeeds, mark socket as connected. If
1301          * connect fails and socket is unbound, reset inp_vflag
1302          * field.
1303          */
1304         if (error == 0)
1305                 soisconnected(so);
1306         else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) &&
1307             inp->inp_lport == 0)
1308                 inp->inp_vflag = vflagsav;
1309 out:
1310         INP_WUNLOCK(inp);
1311         return (error);
1312 }
1313
1314 static void
1315 udp6_detach(struct socket *so)
1316 {
1317         struct inpcb *inp;
1318         struct inpcbinfo *pcbinfo;
1319         struct udpcb *up;
1320
1321         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1322         inp = sotoinpcb(so);
1323         KASSERT(inp != NULL, ("udp6_detach: inp == NULL"));
1324
1325         INP_INFO_WLOCK(pcbinfo);
1326         INP_WLOCK(inp);
1327         up = intoudpcb(inp);
1328         KASSERT(up != NULL, ("%s: up == NULL", __func__));
1329         in_pcbdetach(inp);
1330         in_pcbfree(inp);
1331         INP_INFO_WUNLOCK(pcbinfo);
1332         udp_discardcb(up);
1333 }
1334
1335 static int
1336 udp6_disconnect(struct socket *so)
1337 {
1338         struct inpcb *inp;
1339         struct inpcbinfo *pcbinfo;
1340
1341         pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol);
1342         inp = sotoinpcb(so);
1343         KASSERT(inp != NULL, ("udp6_disconnect: inp == NULL"));
1344
1345         INP_WLOCK(inp);
1346 #ifdef INET
1347         if (inp->inp_vflag & INP_IPV4) {
1348                 struct pr_usrreqs *pru;
1349                 uint8_t nxt;
1350
1351                 nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ?
1352                     IPPROTO_UDP : IPPROTO_UDPLITE;
1353                 INP_WUNLOCK(inp);
1354                 pru = inetsw[ip_protox[nxt]].pr_usrreqs;
1355                 (void)(*pru->pru_disconnect)(so);
1356                 return (0);
1357         }
1358 #endif
1359
1360         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1361                 INP_WUNLOCK(inp);
1362                 return (ENOTCONN);
1363         }
1364
1365         INP_HASH_WLOCK(pcbinfo);
1366         in6_pcbdisconnect(inp);
1367         inp->in6p_laddr = in6addr_any;
1368         INP_HASH_WUNLOCK(pcbinfo);
1369         SOCK_LOCK(so);
1370         so->so_state &= ~SS_ISCONNECTED;                /* XXX */
1371         SOCK_UNLOCK(so);
1372         INP_WUNLOCK(inp);
1373         return (0);
1374 }
1375
1376 static int
1377 udp6_send(struct socket *so, int flags, struct mbuf *m,
1378     struct sockaddr *addr, struct mbuf *control, struct thread *td)
1379 {
1380         int error;
1381
1382         if (addr) {
1383                 if (addr->sa_len != sizeof(struct sockaddr_in6)) {
1384                         error = EINVAL;
1385                         goto bad;
1386                 }
1387                 if (addr->sa_family != AF_INET6) {
1388                         error = EAFNOSUPPORT;
1389                         goto bad;
1390                 }
1391         }
1392
1393         return (udp6_output(so, flags, m, addr, control, td));
1394
1395 bad:
1396         if (control)
1397                 m_freem(control);
1398         m_freem(m);
1399         return (error);
1400 }
1401
1402 struct pr_usrreqs udp6_usrreqs = {
1403         .pru_abort =            udp6_abort,
1404         .pru_attach =           udp6_attach,
1405         .pru_bind =             udp6_bind,
1406         .pru_connect =          udp6_connect,
1407         .pru_control =          in6_control,
1408         .pru_detach =           udp6_detach,
1409         .pru_disconnect =       udp6_disconnect,
1410         .pru_peeraddr =         in6_mapped_peeraddr,
1411         .pru_send =             udp6_send,
1412         .pru_shutdown =         udp_shutdown,
1413         .pru_sockaddr =         in6_mapped_sockaddr,
1414         .pru_soreceive =        soreceive_dgram,
1415         .pru_sosend =           sosend_dgram,
1416         .pru_sosetlabel =       in_pcbsosetlabel,
1417         .pru_close =            udp6_close
1418 };