sys/netinet6/udp6_usrreq.c

   1 /*-
   2  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the project nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      $KAME: udp6_usrreq.c,v 1.27 2001/05/21 05:45:10 jinmei Exp $
  30  *      $KAME: udp6_output.c,v 1.31 2001/05/21 16:39:15 jinmei Exp $
  31  */
  32
  33 /*-
  34  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
  35  *      The Regents of the University of California.
  36  * All rights reserved.
  37  *
  38  * Redistribution and use in source and binary forms, with or without
  39  * modification, are permitted provided that the following conditions
  40  * are met:
  41  * 1. Redistributions of source code must retain the above copyright
  42  *    notice, this list of conditions and the following disclaimer.
  43  * 2. Redistributions in binary form must reproduce the above copyright
  44  *    notice, this list of conditions and the following disclaimer in the
  45  *    documentation and/or other materials provided with the distribution.
  46  * 4. Neither the name of the University nor the names of its contributors
  47  *    may be used to endorse or promote products derived from this software
  48  *    without specific prior written permission.
  49  *
  50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  60  * SUCH DAMAGE.
  61  *
  62  *      @(#)udp_usrreq.c        8.6 (Berkeley) 5/23/95
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_inet.h"
  69 #include "opt_inet6.h"
  70 #include "opt_ipsec.h"
  71
  72 #include <sys/param.h>
  73 #include <sys/jail.h>
  74 #include <sys/kernel.h>
  75 #include <sys/lock.h>
  76 #include <sys/mbuf.h>
  77 #include <sys/priv.h>
  78 #include <sys/proc.h>
  79 #include <sys/protosw.h>
  80 #include <sys/signalvar.h>
  81 #include <sys/socket.h>
  82 #include <sys/socketvar.h>
  83 #include <sys/sx.h>
  84 #include <sys/sysctl.h>
  85 #include <sys/syslog.h>
  86 #include <sys/systm.h>
  87 #include <sys/vimage.h>
  88
  89 #include <net/if.h>
  90 #include <net/if_types.h>
  91 #include <net/route.h>
  92
  93 #include <netinet/in.h>
  94 #include <netinet/in_pcb.h>
  95 #include <netinet/in_systm.h>
  96 #include <netinet/in_var.h>
  97 #include <netinet/ip.h>
  98 #include <netinet/ip_icmp.h>
  99 #include <netinet/ip6.h>
 100 #include <netinet/icmp_var.h>
 101 #include <netinet/icmp6.h>
 102 #include <netinet/ip_var.h>
 103 #include <netinet/udp.h>
 104 #include <netinet/udp_var.h>
 105 #include <netinet/vinet.h>
 106
 107 #include <netinet6/ip6protosw.h>
 108 #include <netinet6/ip6_var.h>
 109 #include <netinet6/in6_pcb.h>
 110 #include <netinet6/udp6_var.h>
 111 #include <netinet6/scope6_var.h>
 112 #include <netinet6/vinet6.h>
 113
 114 #ifdef IPSEC
 115 #include <netipsec/ipsec.h>
 116 #include <netipsec/ipsec6.h>
 117 #endif /* IPSEC */
 118
 119 #include <security/mac/mac_framework.h>
 120
 121 /*
 122  * UDP protocol implementation.
 123  * Per RFC 768, August, 1980.
 124  */
 125
 126 extern struct protosw   inetsw[];
 127 static void             udp6_detach(struct socket *so);
 128
 129 static void
 130 udp6_append(struct inpcb *inp, struct mbuf *n, int off,
 131     struct sockaddr_in6 *fromsa)
 132 {
 133         INIT_VNET_INET(inp->inp_vnet);
 134         struct socket *so;
 135         struct mbuf *opts;
 136
 137         INP_LOCK_ASSERT(inp);
 138
 139 #ifdef IPSEC
 140         /* Check AH/ESP integrity. */
 141         if (ipsec6_in_reject(n, inp)) {
 142                 INIT_VNET_IPSEC(inp->inp_vnet);
 143                 m_freem(n);
 144                 V_ipsec6stat.in_polvio++;
 145                 return;
 146         }
 147 #endif /* IPSEC */
 148 #ifdef MAC
 149         if (mac_inpcb_check_deliver(inp, n) != 0) {
 150                 m_freem(n);
 151                 return;
 152         }
 153 #endif
 154         opts = NULL;
 155         if (inp->inp_flags & INP_CONTROLOPTS ||
 156             inp->inp_socket->so_options & SO_TIMESTAMP)
 157                 ip6_savecontrol(inp, n, &opts);
 158         m_adj(n, off + sizeof(struct udphdr));
 159
 160         so = inp->inp_socket;
 161         SOCKBUF_LOCK(&so->so_rcv);
 162         if (sbappendaddr_locked(&so->so_rcv, (struct sockaddr *)fromsa, n,
 163             opts) == 0) {
 164                 SOCKBUF_UNLOCK(&so->so_rcv);
 165                 m_freem(n);
 166                 if (opts)
 167                         m_freem(opts);
 168                 UDPSTAT_INC(udps_fullsock);
 169         } else
 170                 sorwakeup_locked(so);
 171 }
 172
 173 int
 174 udp6_input(struct mbuf **mp, int *offp, int proto)
 175 {
 176         INIT_VNET_INET(curvnet);
 177         INIT_VNET_INET6(curvnet);
 178         struct mbuf *m = *mp;
 179         struct ifnet *ifp;
 180         struct ip6_hdr *ip6;
 181         struct udphdr *uh;
 182         struct inpcb *inp;
 183         struct udpcb *up;
 184         int off = *offp;
 185         int plen, ulen;
 186         struct sockaddr_in6 fromsa;
 187
 188         ifp = m->m_pkthdr.rcvif;
 189         ip6 = mtod(m, struct ip6_hdr *);
 190
 191         if (faithprefix_p != NULL && (*faithprefix_p)(&ip6->ip6_dst)) {
 192                 /* XXX send icmp6 host/port unreach? */
 193                 m_freem(m);
 194                 return (IPPROTO_DONE);
 195         }
 196
 197 #ifndef PULLDOWN_TEST
 198         IP6_EXTHDR_CHECK(m, off, sizeof(struct udphdr), IPPROTO_DONE);
 199         ip6 = mtod(m, struct ip6_hdr *);
 200         uh = (struct udphdr *)((caddr_t)ip6 + off);
 201 #else
 202         IP6_EXTHDR_GET(uh, struct udphdr *, m, off, sizeof(*uh));
 203         if (!uh)
 204                 return (IPPROTO_DONE);
 205 #endif
 206
 207         UDPSTAT_INC(udps_ipackets);
 208
 209         /*
 210          * Destination port of 0 is illegal, based on RFC768.
 211          */
 212         if (uh->uh_dport == 0)
 213                 goto badunlocked;
 214
 215         plen = ntohs(ip6->ip6_plen) - off + sizeof(*ip6);
 216         ulen = ntohs((u_short)uh->uh_ulen);
 217
 218         if (plen != ulen) {
 219                 UDPSTAT_INC(udps_badlen);
 220                 goto badunlocked;
 221         }
 222
 223         /*
 224          * Checksum extended UDP header and data.
 225          */
 226         if (uh->uh_sum == 0) {
 227                 UDPSTAT_INC(udps_nosum);
 228                 goto badunlocked;
 229         }
 230         if (in6_cksum(m, IPPROTO_UDP, off, ulen) != 0) {
 231                 UDPSTAT_INC(udps_badsum);
 232                 goto badunlocked;
 233         }
 234
 235         /*
 236          * Construct sockaddr format source address.
 237          */
 238         init_sin6(&fromsa, m);
 239         fromsa.sin6_port = uh->uh_sport;
 240
 241         INP_INFO_RLOCK(&V_udbinfo);
 242         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
 243                 struct inpcb *last;
 244                 struct ip6_moptions *imo;
 245
 246                 /*
 247                  * In the event that laddr should be set to the link-local
 248                  * address (this happens in RIPng), the multicast address
 249                  * specified in the received packet will not match laddr.  To
 250                  * handle this situation, matching is relaxed if the
 251                  * receiving interface is the same as one specified in the
 252                  * socket and if the destination multicast address matches
 253                  * one of the multicast groups specified in the socket.
 254                  */
 255
 256                 /*
 257                  * KAME note: traditionally we dropped udpiphdr from mbuf
 258                  * here.  We need udphdr for IPsec processing so we do that
 259                  * later.
 260                  */
 261                 last = NULL;
 262                 LIST_FOREACH(inp, &V_udb, inp_list) {
 263                         if ((inp->inp_vflag & INP_IPV6) == 0)
 264                                 continue;
 265                         if (inp->inp_lport != uh->uh_dport)
 266                                 continue;
 267                         if (inp->inp_fport != 0 &&
 268                             inp->inp_fport != uh->uh_sport)
 269                                 continue;
 270                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 271                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 272                                                         &ip6->ip6_dst))
 273                                         continue;
 274                         }
 275                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 276                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 277                                                         &ip6->ip6_src) ||
 278                                     inp->inp_fport != uh->uh_sport)
 279                                         continue;
 280                         }
 281
 282                         /*
 283                          * Handle socket delivery policy for any-source
 284                          * and source-specific multicast. [RFC3678]
 285                          */
 286                         imo = inp->in6p_moptions;
 287                         if (imo && IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
 288                                 struct sockaddr_in6      mcaddr;
 289                                 int                      blocked;
 290
 291                                 INP_RLOCK(inp);
 292
 293                                 bzero(&mcaddr, sizeof(struct sockaddr_in6));
 294                                 mcaddr.sin6_len = sizeof(struct sockaddr_in6);
 295                                 mcaddr.sin6_family = AF_INET6;
 296                                 mcaddr.sin6_addr = ip6->ip6_dst;
 297
 298                                 blocked = im6o_mc_filter(imo, ifp,
 299                                         (struct sockaddr *)&mcaddr,
 300                                         (struct sockaddr *)&fromsa);
 301                                 if (blocked != MCAST_PASS) {
 302                                         if (blocked == MCAST_NOTGMEMBER)
 303                                                 IP6STAT_INC(ip6s_notmember);
 304                                         if (blocked == MCAST_NOTSMEMBER ||
 305                                             blocked == MCAST_MUTED)
 306                                                 UDPSTAT_INC(udps_filtermcast);
 307                                         INP_RUNLOCK(inp); /* XXX */
 308                                         continue;
 309                                 }
 310
 311                                 INP_RUNLOCK(inp);
 312                         }
 313                         if (last != NULL) {
 314                                 struct mbuf *n;
 315
 316                                 if ((n = m_copy(m, 0, M_COPYALL)) != NULL) {
 317                                         INP_RLOCK(last);
 318                                         up = intoudpcb(last);
 319                                         if (up->u_tun_func == NULL) {
 320                                                 udp6_append(last, n, off, &fromsa);
 321                                         } else {
 322                                                 /*
 323                                                  * Engage the tunneling
 324                                                  * protocol we will have to
 325                                                  * leave the info_lock up,
 326                                                  * since we are hunting
 327                                                  * through multiple UDP's.
 328                                                  *
 329                                                  */
 330                                                 (*up->u_tun_func)(n, off, last);
 331                                         }
 332                                         INP_RUNLOCK(last);
 333                                 }
 334                         }
 335                         last = inp;
 336                         /*
 337                          * Don't look for additional matches if this one does
 338                          * not have either the SO_REUSEPORT or SO_REUSEADDR
 339                          * socket options set.  This heuristic avoids
 340                          * searching through all pcbs in the common case of a
 341                          * non-shared port.  It assumes that an application
 342                          * will never clear these options after setting them.
 343                          */
 344                         if ((last->inp_socket->so_options &
 345                              (SO_REUSEPORT|SO_REUSEADDR)) == 0)
 346                                 break;
 347                 }
 348
 349                 if (last == NULL) {
 350                         /*
 351                          * No matching pcb found; discard datagram.  (No need
 352                          * to send an ICMP Port Unreachable for a broadcast
 353                          * or multicast datgram.)
 354                          */
 355                         UDPSTAT_INC(udps_noport);
 356                         UDPSTAT_INC(udps_noportmcast);
 357                         goto badheadlocked;
 358                 }
 359                 INP_RLOCK(last);
 360                 INP_INFO_RUNLOCK(&V_udbinfo);
 361                 up = intoudpcb(last);
 362                 if (up->u_tun_func == NULL) {
 363                         udp6_append(last, m, off, &fromsa);
 364                 } else {
 365                         /*
 366                          * Engage the tunneling protocol.
 367                          */
 368                         (*up->u_tun_func)(m, off, last);
 369                 }
 370                 INP_RUNLOCK(last);
 371                 return (IPPROTO_DONE);
 372         }
 373         /*
 374          * Locate pcb for datagram.
 375          */
 376         inp = in6_pcblookup_hash(&V_udbinfo, &ip6->ip6_src, uh->uh_sport,
 377             &ip6->ip6_dst, uh->uh_dport, 1, m->m_pkthdr.rcvif);
 378         if (inp == NULL) {
 379                 if (udp_log_in_vain) {
 380                         char ip6bufs[INET6_ADDRSTRLEN];
 381                         char ip6bufd[INET6_ADDRSTRLEN];
 382
 383                         log(LOG_INFO,
 384                             "Connection attempt to UDP [%s]:%d from [%s]:%d\n",
 385                             ip6_sprintf(ip6bufd, &ip6->ip6_dst),
 386                             ntohs(uh->uh_dport),
 387                             ip6_sprintf(ip6bufs, &ip6->ip6_src),
 388                             ntohs(uh->uh_sport));
 389                 }
 390                 UDPSTAT_INC(udps_noport);
 391                 if (m->m_flags & M_MCAST) {
 392                         printf("UDP6: M_MCAST is set in a unicast packet.\n");
 393                         UDPSTAT_INC(udps_noportmcast);
 394                         goto badheadlocked;
 395                 }
 396                 INP_INFO_RUNLOCK(&V_udbinfo);
 397                 if (V_udp_blackhole)
 398                         goto badunlocked;
 399                 if (badport_bandlim(BANDLIM_ICMP6_UNREACH) < 0)
 400                         goto badunlocked;
 401                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_NOPORT, 0);
 402                 return (IPPROTO_DONE);
 403         }
 404         INP_RLOCK(inp);
 405         INP_INFO_RUNLOCK(&V_udbinfo);
 406         up = intoudpcb(inp);
 407         if (up->u_tun_func == NULL) {
 408                 udp6_append(inp, m, off, &fromsa);
 409         } else {
 410                 /*
 411                  * Engage the tunneling protocol.
 412                  */
 413
 414                 (*up->u_tun_func)(m, off, inp);
 415         }
 416         INP_RUNLOCK(inp);
 417         return (IPPROTO_DONE);
 418
 419 badheadlocked:
 420         INP_INFO_RUNLOCK(&V_udbinfo);
 421 badunlocked:
 422         if (m)
 423                 m_freem(m);
 424         return (IPPROTO_DONE);
 425 }
 426
 427 void
 428 udp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
 429 {
 430         INIT_VNET_INET(curvnet);
 431         struct udphdr uh;
 432         struct ip6_hdr *ip6;
 433         struct mbuf *m;
 434         int off = 0;
 435         struct ip6ctlparam *ip6cp = NULL;
 436         const struct sockaddr_in6 *sa6_src = NULL;
 437         void *cmdarg;
 438         struct inpcb *(*notify)(struct inpcb *, int) = udp_notify;
 439         struct udp_portonly {
 440                 u_int16_t uh_sport;
 441                 u_int16_t uh_dport;
 442         } *uhp;
 443
 444         if (sa->sa_family != AF_INET6 ||
 445             sa->sa_len != sizeof(struct sockaddr_in6))
 446                 return;
 447
 448         if ((unsigned)cmd >= PRC_NCMDS)
 449                 return;
 450         if (PRC_IS_REDIRECT(cmd))
 451                 notify = in6_rtchange, d = NULL;
 452         else if (cmd == PRC_HOSTDEAD)
 453                 d = NULL;
 454         else if (inet6ctlerrmap[cmd] == 0)
 455                 return;
 456
 457         /* if the parameter is from icmp6, decode it. */
 458         if (d != NULL) {
 459                 ip6cp = (struct ip6ctlparam *)d;
 460                 m = ip6cp->ip6c_m;
 461                 ip6 = ip6cp->ip6c_ip6;
 462                 off = ip6cp->ip6c_off;
 463                 cmdarg = ip6cp->ip6c_cmdarg;
 464                 sa6_src = ip6cp->ip6c_src;
 465         } else {
 466                 m = NULL;
 467                 ip6 = NULL;
 468                 cmdarg = NULL;
 469                 sa6_src = &sa6_any;
 470         }
 471
 472         if (ip6) {
 473                 /*
 474                  * XXX: We assume that when IPV6 is non NULL,
 475                  * M and OFF are valid.
 476                  */
 477
 478                 /* Check if we can safely examine src and dst ports. */
 479                 if (m->m_pkthdr.len < off + sizeof(*uhp))
 480                         return;
 481
 482                 bzero(&uh, sizeof(uh));
 483                 m_copydata(m, off, sizeof(*uhp), (caddr_t)&uh);
 484
 485                 (void) in6_pcbnotify(&V_udbinfo, sa, uh.uh_dport,
 486                     (struct sockaddr *)ip6cp->ip6c_src, uh.uh_sport, cmd,
 487                     cmdarg, notify);
 488         } else
 489                 (void) in6_pcbnotify(&V_udbinfo, sa, 0,
 490                     (const struct sockaddr *)sa6_src, 0, cmd, cmdarg, notify);
 491 }
 492
 493 static int
 494 udp6_getcred(SYSCTL_HANDLER_ARGS)
 495 {
 496         INIT_VNET_INET(curvnet);
 497         INIT_VNET_INET6(curvnet);
 498         struct xucred xuc;
 499         struct sockaddr_in6 addrs[2];
 500         struct inpcb *inp;
 501         int error;
 502
 503         error = priv_check(req->td, PRIV_NETINET_GETCRED);
 504         if (error)
 505                 return (error);
 506
 507         if (req->newlen != sizeof(addrs))
 508                 return (EINVAL);
 509         if (req->oldlen != sizeof(struct xucred))
 510                 return (EINVAL);
 511         error = SYSCTL_IN(req, addrs, sizeof(addrs));
 512         if (error)
 513                 return (error);
 514         if ((error = sa6_embedscope(&addrs[0], V_ip6_use_defzone)) != 0 ||
 515             (error = sa6_embedscope(&addrs[1], V_ip6_use_defzone)) != 0) {
 516                 return (error);
 517         }
 518         INP_INFO_RLOCK(&V_udbinfo);
 519         inp = in6_pcblookup_hash(&V_udbinfo, &addrs[1].sin6_addr,
 520             addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port, 1,
 521             NULL);
 522         if (inp != NULL) {
 523                 INP_RLOCK(inp);
 524                 INP_INFO_RUNLOCK(&V_udbinfo);
 525                 if (inp->inp_socket == NULL)
 526                         error = ENOENT;
 527                 if (error == 0)
 528                         error = cr_canseesocket(req->td->td_ucred,
 529                             inp->inp_socket);
 530                 if (error == 0)
 531                         cru2x(inp->inp_cred, &xuc);
 532                 INP_RUNLOCK(inp);
 533         } else {
 534                 INP_INFO_RUNLOCK(&V_udbinfo);
 535                 error = ENOENT;
 536         }
 537         if (error == 0)
 538                 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
 539         return (error);
 540 }
 541
 542 SYSCTL_PROC(_net_inet6_udp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW, 0,
 543     0, udp6_getcred, "S,xucred", "Get the xucred of a UDP6 connection");
 544
 545 static int
 546 udp6_output(struct inpcb *inp, struct mbuf *m, struct sockaddr *addr6,
 547     struct mbuf *control, struct thread *td)
 548 {
 549         INIT_VNET_INET(curvnet);
 550         INIT_VNET_INET6(curvnet);
 551         u_int32_t ulen = m->m_pkthdr.len;
 552         u_int32_t plen = sizeof(struct udphdr) + ulen;
 553         struct ip6_hdr *ip6;
 554         struct udphdr *udp6;
 555         struct in6_addr *laddr, *faddr, in6a;
 556         struct sockaddr_in6 *sin6 = NULL;
 557         struct ifnet *oifp = NULL;
 558         int scope_ambiguous = 0;
 559         u_short fport;
 560         int error = 0;
 561         struct ip6_pktopts *optp, opt;
 562         int af = AF_INET6, hlen = sizeof(struct ip6_hdr);
 563         int flags;
 564         struct sockaddr_in6 tmp;
 565
 566         INP_WLOCK_ASSERT(inp);
 567
 568         if (addr6) {
 569                 /* addr6 has been validated in udp6_send(). */
 570                 sin6 = (struct sockaddr_in6 *)addr6;
 571
 572                 /* protect *sin6 from overwrites */
 573                 tmp = *sin6;
 574                 sin6 = &tmp;
 575
 576                 /*
 577                  * Application should provide a proper zone ID or the use of
 578                  * default zone IDs should be enabled.  Unfortunately, some
 579                  * applications do not behave as it should, so we need a
 580                  * workaround.  Even if an appropriate ID is not determined,
 581                  * we'll see if we can determine the outgoing interface.  If we
 582                  * can, determine the zone ID based on the interface below.
 583                  */
 584                 if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 585                         scope_ambiguous = 1;
 586                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 587                         return (error);
 588         }
 589
 590         if (control) {
 591                 if ((error = ip6_setpktopts(control, &opt,
 592                     inp->in6p_outputopts, td->td_ucred, IPPROTO_UDP)) != 0)
 593                         goto release;
 594                 optp = &opt;
 595         } else
 596                 optp = inp->in6p_outputopts;
 597
 598         if (sin6) {
 599                 faddr = &sin6->sin6_addr;
 600
 601                 /*
 602                  * IPv4 version of udp_output calls in_pcbconnect in this case,
 603                  * which needs splnet and affects performance.
 604                  * Since we saw no essential reason for calling in_pcbconnect,
 605                  * we get rid of such kind of logic, and call in6_selectsrc
 606                  * and in6_pcbsetport in order to fill in the local address
 607                  * and the local port.
 608                  */
 609                 if (sin6->sin6_port == 0) {
 610                         error = EADDRNOTAVAIL;
 611                         goto release;
 612                 }
 613
 614                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 615                         /* how about ::ffff:0.0.0.0 case? */
 616                         error = EISCONN;
 617                         goto release;
 618                 }
 619
 620                 fport = sin6->sin6_port; /* allow 0 port */
 621
 622                 if (IN6_IS_ADDR_V4MAPPED(faddr)) {
 623                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY)) {
 624                                 /*
 625                                  * I believe we should explicitly discard the
 626                                  * packet when mapped addresses are disabled,
 627                                  * rather than send the packet as an IPv6 one.
 628                                  * If we chose the latter approach, the packet
 629                                  * might be sent out on the wire based on the
 630                                  * default route, the situation which we'd
 631                                  * probably want to avoid.
 632                                  * (20010421 jinmei@kame.net)
 633                                  */
 634                                 error = EINVAL;
 635                                 goto release;
 636                         }
 637                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) &&
 638                             !IN6_IS_ADDR_V4MAPPED(&inp->in6p_laddr)) {
 639                                 /*
 640                                  * when remote addr is an IPv4-mapped address,
 641                                  * local addr should not be an IPv6 address,
 642                                  * since you cannot determine how to map IPv6
 643                                  * source address to IPv4.
 644                                  */
 645                                 error = EINVAL;
 646                                 goto release;
 647                         }
 648
 649                         af = AF_INET;
 650                 }
 651
 652                 if (!IN6_IS_ADDR_V4MAPPED(faddr)) {
 653                         error = in6_selectsrc(sin6, optp, inp, NULL,
 654                             td->td_ucred, &oifp, &in6a);
 655                         if (error)
 656                                 goto release;
 657                         if (oifp && scope_ambiguous &&
 658                             (error = in6_setscope(&sin6->sin6_addr,
 659                             oifp, NULL))) {
 660                                 goto release;
 661                         }
 662                         laddr = &in6a;
 663                 } else
 664                         laddr = &inp->in6p_laddr;       /* XXX */
 665                 if (laddr == NULL) {
 666                         if (error == 0)
 667                                 error = EADDRNOTAVAIL;
 668                         goto release;
 669                 }
 670                 if (inp->inp_lport == 0 &&
 671                     (error = in6_pcbsetport(laddr, inp, td->td_ucred)) != 0)
 672                         goto release;
 673         } else {
 674                 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 675                         error = ENOTCONN;
 676                         goto release;
 677                 }
 678                 if (IN6_IS_ADDR_V4MAPPED(&inp->in6p_faddr)) {
 679                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY)) {
 680                                 /*
 681                                  * XXX: this case would happen when the
 682                                  * application sets the V6ONLY flag after
 683                                  * connecting the foreign address.
 684                                  * Such applications should be fixed,
 685                                  * so we bark here.
 686                                  */
 687                                 log(LOG_INFO, "udp6_output: IPV6_V6ONLY "
 688                                     "option was set for a connected socket\n");
 689                                 error = EINVAL;
 690                                 goto release;
 691                         } else
 692                                 af = AF_INET;
 693                 }
 694                 laddr = &inp->in6p_laddr;
 695                 faddr = &inp->in6p_faddr;
 696                 fport = inp->inp_fport;
 697         }
 698
 699         if (af == AF_INET)
 700                 hlen = sizeof(struct ip);
 701
 702         /*
 703          * Calculate data length and get a mbuf
 704          * for UDP and IP6 headers.
 705          */
 706         M_PREPEND(m, hlen + sizeof(struct udphdr), M_DONTWAIT);
 707         if (m == 0) {
 708                 error = ENOBUFS;
 709                 goto release;
 710         }
 711
 712         /*
 713          * Stuff checksum and output datagram.
 714          */
 715         udp6 = (struct udphdr *)(mtod(m, caddr_t) + hlen);
 716         udp6->uh_sport = inp->inp_lport; /* lport is always set in the PCB */
 717         udp6->uh_dport = fport;
 718         if (plen <= 0xffff)
 719                 udp6->uh_ulen = htons((u_short)plen);
 720         else
 721                 udp6->uh_ulen = 0;
 722         udp6->uh_sum = 0;
 723
 724         switch (af) {
 725         case AF_INET6:
 726                 ip6 = mtod(m, struct ip6_hdr *);
 727                 ip6->ip6_flow   = inp->inp_flow & IPV6_FLOWINFO_MASK;
 728                 ip6->ip6_vfc    &= ~IPV6_VERSION_MASK;
 729                 ip6->ip6_vfc    |= IPV6_VERSION;
 730 #if 0                           /* ip6_plen will be filled in ip6_output. */
 731                 ip6->ip6_plen   = htons((u_short)plen);
 732 #endif
 733                 ip6->ip6_nxt    = IPPROTO_UDP;
 734                 ip6->ip6_hlim   = in6_selecthlim(inp, NULL);
 735                 ip6->ip6_src    = *laddr;
 736                 ip6->ip6_dst    = *faddr;
 737
 738                 if ((udp6->uh_sum = in6_cksum(m, IPPROTO_UDP,
 739                                 sizeof(struct ip6_hdr), plen)) == 0) {
 740                         udp6->uh_sum = 0xffff;
 741                 }
 742
 743                 flags = 0;
 744
 745                 UDPSTAT_INC(udps_opackets);
 746                 error = ip6_output(m, optp, NULL, flags, inp->in6p_moptions,
 747                     NULL, inp);
 748                 break;
 749         case AF_INET:
 750                 error = EAFNOSUPPORT;
 751                 goto release;
 752         }
 753         goto releaseopt;
 754
 755 release:
 756         m_freem(m);
 757
 758 releaseopt:
 759         if (control) {
 760                 ip6_clearpktopts(&opt, -1);
 761                 m_freem(control);
 762         }
 763         return (error);
 764 }
 765
 766 static void
 767 udp6_abort(struct socket *so)
 768 {
 769         INIT_VNET_INET(so->so_vnet);
 770         struct inpcb *inp;
 771
 772         inp = sotoinpcb(so);
 773         KASSERT(inp != NULL, ("udp6_abort: inp == NULL"));
 774
 775 #ifdef INET
 776         if (inp->inp_vflag & INP_IPV4) {
 777                 struct pr_usrreqs *pru;
 778
 779                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
 780                 (*pru->pru_abort)(so);
 781                 return;
 782         }
 783 #endif
 784
 785         INP_INFO_WLOCK(&V_udbinfo);
 786         INP_WLOCK(inp);
 787         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 788                 in6_pcbdisconnect(inp);
 789                 inp->in6p_laddr = in6addr_any;
 790                 soisdisconnected(so);
 791         }
 792         INP_WUNLOCK(inp);
 793         INP_INFO_WUNLOCK(&V_udbinfo);
 794 }
 795
 796 static int
 797 udp6_attach(struct socket *so, int proto, struct thread *td)
 798 {
 799         INIT_VNET_INET(so->so_vnet);
 800         struct inpcb *inp;
 801         int error;
 802
 803         inp = sotoinpcb(so);
 804         KASSERT(inp == NULL, ("udp6_attach: inp != NULL"));
 805
 806         if (so->so_snd.sb_hiwat == 0 || so->so_rcv.sb_hiwat == 0) {
 807                 error = soreserve(so, udp_sendspace, udp_recvspace);
 808                 if (error)
 809                         return (error);
 810         }
 811         INP_INFO_WLOCK(&V_udbinfo);
 812         error = in_pcballoc(so, &V_udbinfo);
 813         if (error) {
 814                 INP_INFO_WUNLOCK(&V_udbinfo);
 815                 return (error);
 816         }
 817         inp = (struct inpcb *)so->so_pcb;
 818         inp->inp_vflag |= INP_IPV6;
 819         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0)
 820                 inp->inp_vflag |= INP_IPV4;
 821         inp->in6p_hops = -1;    /* use kernel default */
 822         inp->in6p_cksum = -1;   /* just to be sure */
 823         /*
 824          * XXX: ugly!!
 825          * IPv4 TTL initialization is necessary for an IPv6 socket as well,
 826          * because the socket may be bound to an IPv6 wildcard address,
 827          * which may match an IPv4-mapped IPv6 address.
 828          */
 829         inp->inp_ip_ttl = V_ip_defttl;
 830
 831         error = udp_newudpcb(inp);
 832         if (error) {
 833                 in_pcbdetach(inp);
 834                 in_pcbfree(inp);
 835                 INP_INFO_WUNLOCK(&V_udbinfo);
 836                 return (error);
 837         }
 838         INP_WUNLOCK(inp);
 839         INP_INFO_WUNLOCK(&V_udbinfo);
 840         return (0);
 841 }
 842
 843 static int
 844 udp6_bind(struct socket *so, struct sockaddr *nam, struct thread *td)
 845 {
 846         INIT_VNET_INET(so->so_vnet);
 847         struct inpcb *inp;
 848         int error;
 849
 850         inp = sotoinpcb(so);
 851         KASSERT(inp != NULL, ("udp6_bind: inp == NULL"));
 852
 853         INP_INFO_WLOCK(&V_udbinfo);
 854         INP_WLOCK(inp);
 855         inp->inp_vflag &= ~INP_IPV4;
 856         inp->inp_vflag |= INP_IPV6;
 857         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
 858                 struct sockaddr_in6 *sin6_p;
 859
 860                 sin6_p = (struct sockaddr_in6 *)nam;
 861
 862                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6_p->sin6_addr))
 863                         inp->inp_vflag |= INP_IPV4;
 864                 else if (IN6_IS_ADDR_V4MAPPED(&sin6_p->sin6_addr)) {
 865                         struct sockaddr_in sin;
 866
 867                         in6_sin6_2_sin(&sin, sin6_p);
 868                         inp->inp_vflag |= INP_IPV4;
 869                         inp->inp_vflag &= ~INP_IPV6;
 870                         error = in_pcbbind(inp, (struct sockaddr *)&sin,
 871                             td->td_ucred);
 872                         goto out;
 873                 }
 874         }
 875
 876         error = in6_pcbbind(inp, nam, td->td_ucred);
 877 out:
 878         INP_WUNLOCK(inp);
 879         INP_INFO_WUNLOCK(&V_udbinfo);
 880         return (error);
 881 }
 882
 883 static void
 884 udp6_close(struct socket *so)
 885 {
 886         INIT_VNET_INET(so->so_vnet);
 887         struct inpcb *inp;
 888
 889         inp = sotoinpcb(so);
 890         KASSERT(inp != NULL, ("udp6_close: inp == NULL"));
 891
 892 #ifdef INET
 893         if (inp->inp_vflag & INP_IPV4) {
 894                 struct pr_usrreqs *pru;
 895
 896                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
 897                 (*pru->pru_disconnect)(so);
 898                 return;
 899         }
 900 #endif
 901         INP_INFO_WLOCK(&V_udbinfo);
 902         INP_WLOCK(inp);
 903         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 904                 in6_pcbdisconnect(inp);
 905                 inp->in6p_laddr = in6addr_any;
 906                 soisdisconnected(so);
 907         }
 908         INP_WUNLOCK(inp);
 909         INP_INFO_WUNLOCK(&V_udbinfo);
 910 }
 911
 912 static int
 913 udp6_connect(struct socket *so, struct sockaddr *nam, struct thread *td)
 914 {
 915         INIT_VNET_INET(so->so_vnet);
 916         struct inpcb *inp;
 917         struct sockaddr_in6 *sin6;
 918         int error;
 919
 920         inp = sotoinpcb(so);
 921         sin6 = (struct sockaddr_in6 *)nam;
 922         KASSERT(inp != NULL, ("udp6_connect: inp == NULL"));
 923
 924         INP_INFO_WLOCK(&V_udbinfo);
 925         INP_WLOCK(inp);
 926         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
 927             IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
 928                 struct sockaddr_in sin;
 929
 930                 if (inp->inp_faddr.s_addr != INADDR_ANY) {
 931                         error = EISCONN;
 932                         goto out;
 933                 }
 934                 in6_sin6_2_sin(&sin, sin6);
 935                 error = prison_remote_ip4(td->td_ucred, &sin.sin_addr);
 936                 if (error != 0)
 937                         goto out;
 938                 error = in_pcbconnect(inp, (struct sockaddr *)&sin,
 939                     td->td_ucred);
 940                 if (error == 0) {
 941                         inp->inp_vflag |= INP_IPV4;
 942                         inp->inp_vflag &= ~INP_IPV6;
 943                         soisconnected(so);
 944                 }
 945                 goto out;
 946         }
 947         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 948                 error = EISCONN;
 949                 goto out;
 950         }
 951         error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr);
 952         if (error != 0)
 953                 goto out;
 954         error = in6_pcbconnect(inp, nam, td->td_ucred);
 955         if (error == 0) {
 956                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
 957                         /* should be non mapped addr */
 958                         inp->inp_vflag &= ~INP_IPV4;
 959                         inp->inp_vflag |= INP_IPV6;
 960                 }
 961                 soisconnected(so);
 962         }
 963 out:
 964         INP_WUNLOCK(inp);
 965         INP_INFO_WUNLOCK(&V_udbinfo);
 966         return (error);
 967 }
 968
 969 static void
 970 udp6_detach(struct socket *so)
 971 {
 972         INIT_VNET_INET(so->so_vnet);
 973         struct inpcb *inp;
 974         struct udpcb *up;
 975
 976         inp = sotoinpcb(so);
 977         KASSERT(inp != NULL, ("udp6_detach: inp == NULL"));
 978
 979         INP_INFO_WLOCK(&V_udbinfo);
 980         INP_WLOCK(inp);
 981         up = intoudpcb(inp);
 982         KASSERT(up != NULL, ("%s: up == NULL", __func__));
 983         in_pcbdetach(inp);
 984         in_pcbfree(inp);
 985         INP_INFO_WUNLOCK(&V_udbinfo);
 986         udp_discardcb(up);
 987 }
 988
 989 static int
 990 udp6_disconnect(struct socket *so)
 991 {
 992         INIT_VNET_INET(so->so_vnet);
 993         struct inpcb *inp;
 994         int error;
 995
 996         inp = sotoinpcb(so);
 997         KASSERT(inp != NULL, ("udp6_disconnect: inp == NULL"));
 998
 999         INP_INFO_WLOCK(&V_udbinfo);
1000         INP_WLOCK(inp);
1001
1002 #ifdef INET
1003         if (inp->inp_vflag & INP_IPV4) {
1004                 struct pr_usrreqs *pru;
1005
1006                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
1007                 error = (*pru->pru_disconnect)(so);
1008                 goto out;
1009         }
1010 #endif
1011
1012         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1013                 error = ENOTCONN;
1014                 goto out;
1015         }
1016
1017         in6_pcbdisconnect(inp);
1018         inp->in6p_laddr = in6addr_any;
1019         SOCK_LOCK(so);
1020         so->so_state &= ~SS_ISCONNECTED;                /* XXX */
1021         SOCK_UNLOCK(so);
1022 out:
1023         INP_WUNLOCK(inp);
1024         INP_INFO_WUNLOCK(&V_udbinfo);
1025         return (0);
1026 }
1027
1028 static int
1029 udp6_send(struct socket *so, int flags, struct mbuf *m,
1030     struct sockaddr *addr, struct mbuf *control, struct thread *td)
1031 {
1032         INIT_VNET_INET(so->so_vnet);
1033         struct inpcb *inp;
1034         int error = 0;
1035
1036         inp = sotoinpcb(so);
1037         KASSERT(inp != NULL, ("udp6_send: inp == NULL"));
1038
1039         INP_INFO_WLOCK(&V_udbinfo);
1040         INP_WLOCK(inp);
1041         if (addr) {
1042                 if (addr->sa_len != sizeof(struct sockaddr_in6)) {
1043                         error = EINVAL;
1044                         goto bad;
1045                 }
1046                 if (addr->sa_family != AF_INET6) {
1047                         error = EAFNOSUPPORT;
1048                         goto bad;
1049                 }
1050         }
1051
1052 #ifdef INET
1053         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
1054                 int hasv4addr;
1055                 struct sockaddr_in6 *sin6 = 0;
1056
1057                 if (addr == 0)
1058                         hasv4addr = (inp->inp_vflag & INP_IPV4);
1059                 else {
1060                         sin6 = (struct sockaddr_in6 *)addr;
1061                         hasv4addr = IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)
1062                             ? 1 : 0;
1063                 }
1064                 if (hasv4addr) {
1065                         struct pr_usrreqs *pru;
1066
1067                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) &&
1068                             !IN6_IS_ADDR_V4MAPPED(&inp->in6p_laddr)) {
1069                                 /*
1070                                  * When remote addr is IPv4-mapped address,
1071                                  * local addr should not be an IPv6 address;
1072                                  * since you cannot determine how to map IPv6
1073                                  * source address to IPv4.
1074                                  */
1075                                 error = EINVAL;
1076                                 goto out;
1077                         }
1078
1079                         /*
1080                          * XXXRW: We release UDP-layer locks before calling
1081                          * udp_send() in order to avoid recursion.  However,
1082                          * this does mean there is a short window where inp's
1083                          * fields are unstable.  Could this lead to a
1084                          * potential race in which the factors causing us to
1085                          * select the UDPv4 output routine are invalidated?
1086                          */
1087                         INP_WUNLOCK(inp);
1088                         INP_INFO_WUNLOCK(&V_udbinfo);
1089                         if (sin6)
1090                                 in6_sin6_2_sin_in_sock(addr);
1091                         pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
1092                         /* addr will just be freed in sendit(). */
1093                         return ((*pru->pru_send)(so, flags, m, addr, control,
1094                             td));
1095                 }
1096         }
1097 #endif
1098 #ifdef MAC
1099         mac_inpcb_create_mbuf(inp, m);
1100 #endif
1101         error = udp6_output(inp, m, addr, control, td);
1102 out:
1103         INP_WUNLOCK(inp);
1104         INP_INFO_WUNLOCK(&V_udbinfo);
1105         return (error);
1106
1107 bad:
1108         INP_WUNLOCK(inp);
1109         INP_INFO_WUNLOCK(&V_udbinfo);
1110         m_freem(m);
1111         return (error);
1112 }
1113
1114 struct pr_usrreqs udp6_usrreqs = {
1115         .pru_abort =            udp6_abort,
1116         .pru_attach =           udp6_attach,
1117         .pru_bind =             udp6_bind,
1118         .pru_connect =          udp6_connect,
1119         .pru_control =          in6_control,
1120         .pru_detach =           udp6_detach,
1121         .pru_disconnect =       udp6_disconnect,
1122         .pru_peeraddr =         in6_mapped_peeraddr,
1123         .pru_send =             udp6_send,
1124         .pru_shutdown =         udp_shutdown,
1125         .pru_sockaddr =         in6_mapped_sockaddr,
1126         .pru_soreceive =        soreceive_dgram,
1127         .pru_sosend =           sosend_dgram,
1128         .pru_sosetlabel =       in_pcbsosetlabel,
1129         .pru_close =            udp6_close
1130 };