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
  88 #include <net/if.h>
  89 #include <net/if_types.h>
  90 #include <net/route.h>
  91
  92 #include <netinet/in.h>
  93 #include <netinet/in_pcb.h>
  94 #include <netinet/in_systm.h>
  95 #include <netinet/in_var.h>
  96 #include <netinet/ip.h>
  97 #include <netinet/ip_icmp.h>
  98 #include <netinet/ip6.h>
  99 #include <netinet/icmp_var.h>
 100 #include <netinet/icmp6.h>
 101 #include <netinet/ip_var.h>
 102 #include <netinet/udp.h>
 103 #include <netinet/udp_var.h>
 104
 105 #include <netinet6/ip6protosw.h>
 106 #include <netinet6/ip6_var.h>
 107 #include <netinet6/in6_pcb.h>
 108 #include <netinet6/udp6_var.h>
 109 #include <netinet6/scope6_var.h>
 110
 111 #ifdef IPSEC
 112 #include <netipsec/ipsec.h>
 113 #include <netipsec/ipsec6.h>
 114 #endif /* IPSEC */
 115
 116 #include <security/mac/mac_framework.h>
 117
 118 /*
 119  * UDP protocol implementation.
 120  * Per RFC 768, August, 1980.
 121  */
 122
 123 extern struct protosw   inetsw[];
 124 static void             udp6_detach(struct socket *so);
 125
 126 static void
 127 udp6_append(struct inpcb *inp, struct mbuf *n, int off,
 128     struct sockaddr_in6 *fromsa)
 129 {
 130         struct socket *so;
 131         struct mbuf *opts;
 132
 133         INP_LOCK_ASSERT(inp);
 134
 135 #ifdef IPSEC
 136         /* Check AH/ESP integrity. */
 137         if (ipsec6_in_reject(n, inp)) {
 138                 m_freem(n);
 139                 V_ipsec6stat.in_polvio++;
 140                 return;
 141         }
 142 #endif /* IPSEC */
 143 #ifdef MAC
 144         if (mac_inpcb_check_deliver(inp, n) != 0) {
 145                 m_freem(n);
 146                 return;
 147         }
 148 #endif
 149         opts = NULL;
 150         if (inp->inp_flags & INP_CONTROLOPTS ||
 151             inp->inp_socket->so_options & SO_TIMESTAMP)
 152                 ip6_savecontrol(inp, n, &opts);
 153         m_adj(n, off + sizeof(struct udphdr));
 154
 155         so = inp->inp_socket;
 156         SOCKBUF_LOCK(&so->so_rcv);
 157         if (sbappendaddr_locked(&so->so_rcv, (struct sockaddr *)fromsa, n,
 158             opts) == 0) {
 159                 SOCKBUF_UNLOCK(&so->so_rcv);
 160                 m_freem(n);
 161                 if (opts)
 162                         m_freem(opts);
 163                 UDPSTAT_INC(udps_fullsock);
 164         } else
 165                 sorwakeup_locked(so);
 166 }
 167
 168 int
 169 udp6_input(struct mbuf **mp, int *offp, int proto)
 170 {
 171         struct mbuf *m = *mp;
 172         struct ifnet *ifp;
 173         struct ip6_hdr *ip6;
 174         struct udphdr *uh;
 175         struct inpcb *inp;
 176         struct udpcb *up;
 177         int off = *offp;
 178         int plen, ulen;
 179         struct sockaddr_in6 fromsa;
 180
 181         ifp = m->m_pkthdr.rcvif;
 182         ip6 = mtod(m, struct ip6_hdr *);
 183
 184         if (faithprefix_p != NULL && (*faithprefix_p)(&ip6->ip6_dst)) {
 185                 /* XXX send icmp6 host/port unreach? */
 186                 m_freem(m);
 187                 return (IPPROTO_DONE);
 188         }
 189
 190 #ifndef PULLDOWN_TEST
 191         IP6_EXTHDR_CHECK(m, off, sizeof(struct udphdr), IPPROTO_DONE);
 192         ip6 = mtod(m, struct ip6_hdr *);
 193         uh = (struct udphdr *)((caddr_t)ip6 + off);
 194 #else
 195         IP6_EXTHDR_GET(uh, struct udphdr *, m, off, sizeof(*uh));
 196         if (!uh)
 197                 return (IPPROTO_DONE);
 198 #endif
 199
 200         UDPSTAT_INC(udps_ipackets);
 201
 202         /*
 203          * Destination port of 0 is illegal, based on RFC768.
 204          */
 205         if (uh->uh_dport == 0)
 206                 goto badunlocked;
 207
 208         plen = ntohs(ip6->ip6_plen) - off + sizeof(*ip6);
 209         ulen = ntohs((u_short)uh->uh_ulen);
 210
 211         if (plen != ulen) {
 212                 UDPSTAT_INC(udps_badlen);
 213                 goto badunlocked;
 214         }
 215
 216         /*
 217          * Checksum extended UDP header and data.
 218          */
 219         if (uh->uh_sum == 0) {
 220                 UDPSTAT_INC(udps_nosum);
 221                 goto badunlocked;
 222         }
 223         if (in6_cksum(m, IPPROTO_UDP, off, ulen) != 0) {
 224                 UDPSTAT_INC(udps_badsum);
 225                 goto badunlocked;
 226         }
 227
 228         /*
 229          * Construct sockaddr format source address.
 230          */
 231         init_sin6(&fromsa, m);
 232         fromsa.sin6_port = uh->uh_sport;
 233
 234         INP_INFO_RLOCK(&V_udbinfo);
 235         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
 236                 struct inpcb *last;
 237                 struct ip6_moptions *imo;
 238
 239                 /*
 240                  * In the event that laddr should be set to the link-local
 241                  * address (this happens in RIPng), the multicast address
 242                  * specified in the received packet will not match laddr.  To
 243                  * handle this situation, matching is relaxed if the
 244                  * receiving interface is the same as one specified in the
 245                  * socket and if the destination multicast address matches
 246                  * one of the multicast groups specified in the socket.
 247                  */
 248
 249                 /*
 250                  * KAME note: traditionally we dropped udpiphdr from mbuf
 251                  * here.  We need udphdr for IPsec processing so we do that
 252                  * later.
 253                  */
 254                 last = NULL;
 255                 LIST_FOREACH(inp, &V_udb, inp_list) {
 256                         if ((inp->inp_vflag & INP_IPV6) == 0)
 257                                 continue;
 258                         if (inp->inp_lport != uh->uh_dport)
 259                                 continue;
 260                         if (inp->inp_fport != 0 &&
 261                             inp->inp_fport != uh->uh_sport)
 262                                 continue;
 263                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
 264                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
 265                                                         &ip6->ip6_dst))
 266                                         continue;
 267                         }
 268                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 269                                 if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
 270                                                         &ip6->ip6_src) ||
 271                                     inp->inp_fport != uh->uh_sport)
 272                                         continue;
 273                         }
 274
 275                         /*
 276                          * Detached PCBs can linger in the list if someone
 277                          * holds a reference. (e.g. udp_pcblist)
 278                          */
 279                         if (inp->inp_socket == NULL)
 280                                 continue;
 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
 407         /*
 408          * Detached PCBs can linger in the hash table if someone holds a
 409          * reference. (e.g. udp_pcblist)
 410          */
 411         if (inp->inp_socket == NULL) {
 412                 INP_RUNLOCK(inp);
 413                 goto badunlocked;
 414         }
 415         up = intoudpcb(inp);
 416         if (up->u_tun_func == NULL) {
 417                 udp6_append(inp, m, off, &fromsa);
 418         } else {
 419                 /*
 420                  * Engage the tunneling protocol.
 421                  */
 422
 423                 (*up->u_tun_func)(m, off, inp);
 424         }
 425         INP_RUNLOCK(inp);
 426         return (IPPROTO_DONE);
 427
 428 badheadlocked:
 429         INP_INFO_RUNLOCK(&V_udbinfo);
 430 badunlocked:
 431         if (m)
 432                 m_freem(m);
 433         return (IPPROTO_DONE);
 434 }
 435
 436 void
 437 udp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
 438 {
 439         struct udphdr uh;
 440         struct ip6_hdr *ip6;
 441         struct mbuf *m;
 442         int off = 0;
 443         struct ip6ctlparam *ip6cp = NULL;
 444         const struct sockaddr_in6 *sa6_src = NULL;
 445         void *cmdarg;
 446         struct inpcb *(*notify)(struct inpcb *, int) = udp_notify;
 447         struct udp_portonly {
 448                 u_int16_t uh_sport;
 449                 u_int16_t uh_dport;
 450         } *uhp;
 451
 452         if (sa->sa_family != AF_INET6 ||
 453             sa->sa_len != sizeof(struct sockaddr_in6))
 454                 return;
 455
 456         if ((unsigned)cmd >= PRC_NCMDS)
 457                 return;
 458         if (PRC_IS_REDIRECT(cmd))
 459                 notify = in6_rtchange, d = NULL;
 460         else if (cmd == PRC_HOSTDEAD)
 461                 d = NULL;
 462         else if (inet6ctlerrmap[cmd] == 0)
 463                 return;
 464
 465         /* if the parameter is from icmp6, decode it. */
 466         if (d != NULL) {
 467                 ip6cp = (struct ip6ctlparam *)d;
 468                 m = ip6cp->ip6c_m;
 469                 ip6 = ip6cp->ip6c_ip6;
 470                 off = ip6cp->ip6c_off;
 471                 cmdarg = ip6cp->ip6c_cmdarg;
 472                 sa6_src = ip6cp->ip6c_src;
 473         } else {
 474                 m = NULL;
 475                 ip6 = NULL;
 476                 cmdarg = NULL;
 477                 sa6_src = &sa6_any;
 478         }
 479
 480         if (ip6) {
 481                 /*
 482                  * XXX: We assume that when IPV6 is non NULL,
 483                  * M and OFF are valid.
 484                  */
 485
 486                 /* Check if we can safely examine src and dst ports. */
 487                 if (m->m_pkthdr.len < off + sizeof(*uhp))
 488                         return;
 489
 490                 bzero(&uh, sizeof(uh));
 491                 m_copydata(m, off, sizeof(*uhp), (caddr_t)&uh);
 492
 493                 (void) in6_pcbnotify(&V_udbinfo, sa, uh.uh_dport,
 494                     (struct sockaddr *)ip6cp->ip6c_src, uh.uh_sport, cmd,
 495                     cmdarg, notify);
 496         } else
 497                 (void) in6_pcbnotify(&V_udbinfo, sa, 0,
 498                     (const struct sockaddr *)sa6_src, 0, cmd, cmdarg, notify);
 499 }
 500
 501 static int
 502 udp6_getcred(SYSCTL_HANDLER_ARGS)
 503 {
 504         struct xucred xuc;
 505         struct sockaddr_in6 addrs[2];
 506         struct inpcb *inp;
 507         int error;
 508
 509         error = priv_check(req->td, PRIV_NETINET_GETCRED);
 510         if (error)
 511                 return (error);
 512
 513         if (req->newlen != sizeof(addrs))
 514                 return (EINVAL);
 515         if (req->oldlen != sizeof(struct xucred))
 516                 return (EINVAL);
 517         error = SYSCTL_IN(req, addrs, sizeof(addrs));
 518         if (error)
 519                 return (error);
 520         if ((error = sa6_embedscope(&addrs[0], V_ip6_use_defzone)) != 0 ||
 521             (error = sa6_embedscope(&addrs[1], V_ip6_use_defzone)) != 0) {
 522                 return (error);
 523         }
 524         INP_INFO_RLOCK(&V_udbinfo);
 525         inp = in6_pcblookup_hash(&V_udbinfo, &addrs[1].sin6_addr,
 526             addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port, 1,
 527             NULL);
 528         if (inp != NULL) {
 529                 INP_RLOCK(inp);
 530                 INP_INFO_RUNLOCK(&V_udbinfo);
 531                 if (inp->inp_socket == NULL)
 532                         error = ENOENT;
 533                 if (error == 0)
 534                         error = cr_canseesocket(req->td->td_ucred,
 535                             inp->inp_socket);
 536                 if (error == 0)
 537                         cru2x(inp->inp_cred, &xuc);
 538                 INP_RUNLOCK(inp);
 539         } else {
 540                 INP_INFO_RUNLOCK(&V_udbinfo);
 541                 error = ENOENT;
 542         }
 543         if (error == 0)
 544                 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
 545         return (error);
 546 }
 547
 548 SYSCTL_PROC(_net_inet6_udp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW, 0,
 549     0, udp6_getcred, "S,xucred", "Get the xucred of a UDP6 connection");
 550
 551 static int
 552 udp6_output(struct inpcb *inp, struct mbuf *m, struct sockaddr *addr6,
 553     struct mbuf *control, struct thread *td)
 554 {
 555         u_int32_t ulen = m->m_pkthdr.len;
 556         u_int32_t plen = sizeof(struct udphdr) + ulen;
 557         struct ip6_hdr *ip6;
 558         struct udphdr *udp6;
 559         struct in6_addr *laddr, *faddr, in6a;
 560         struct sockaddr_in6 *sin6 = NULL;
 561         struct ifnet *oifp = NULL;
 562         int scope_ambiguous = 0;
 563         u_short fport;
 564         int error = 0;
 565         struct ip6_pktopts *optp, opt;
 566         int af = AF_INET6, hlen = sizeof(struct ip6_hdr);
 567         int flags;
 568         struct sockaddr_in6 tmp;
 569
 570         INP_WLOCK_ASSERT(inp);
 571
 572         if (addr6) {
 573                 /* addr6 has been validated in udp6_send(). */
 574                 sin6 = (struct sockaddr_in6 *)addr6;
 575
 576                 /* protect *sin6 from overwrites */
 577                 tmp = *sin6;
 578                 sin6 = &tmp;
 579
 580                 /*
 581                  * Application should provide a proper zone ID or the use of
 582                  * default zone IDs should be enabled.  Unfortunately, some
 583                  * applications do not behave as it should, so we need a
 584                  * workaround.  Even if an appropriate ID is not determined,
 585                  * we'll see if we can determine the outgoing interface.  If we
 586                  * can, determine the zone ID based on the interface below.
 587                  */
 588                 if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
 589                         scope_ambiguous = 1;
 590                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
 591                         return (error);
 592         }
 593
 594         if (control) {
 595                 if ((error = ip6_setpktopts(control, &opt,
 596                     inp->in6p_outputopts, td->td_ucred, IPPROTO_UDP)) != 0)
 597                         goto release;
 598                 optp = &opt;
 599         } else
 600                 optp = inp->in6p_outputopts;
 601
 602         if (sin6) {
 603                 faddr = &sin6->sin6_addr;
 604
 605                 /*
 606                  * IPv4 version of udp_output calls in_pcbconnect in this case,
 607                  * which needs splnet and affects performance.
 608                  * Since we saw no essential reason for calling in_pcbconnect,
 609                  * we get rid of such kind of logic, and call in6_selectsrc
 610                  * and in6_pcbsetport in order to fill in the local address
 611                  * and the local port.
 612                  */
 613                 if (sin6->sin6_port == 0) {
 614                         error = EADDRNOTAVAIL;
 615                         goto release;
 616                 }
 617
 618                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 619                         /* how about ::ffff:0.0.0.0 case? */
 620                         error = EISCONN;
 621                         goto release;
 622                 }
 623
 624                 fport = sin6->sin6_port; /* allow 0 port */
 625
 626                 if (IN6_IS_ADDR_V4MAPPED(faddr)) {
 627                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY)) {
 628                                 /*
 629                                  * I believe we should explicitly discard the
 630                                  * packet when mapped addresses are disabled,
 631                                  * rather than send the packet as an IPv6 one.
 632                                  * If we chose the latter approach, the packet
 633                                  * might be sent out on the wire based on the
 634                                  * default route, the situation which we'd
 635                                  * probably want to avoid.
 636                                  * (20010421 jinmei@kame.net)
 637                                  */
 638                                 error = EINVAL;
 639                                 goto release;
 640                         }
 641                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) &&
 642                             !IN6_IS_ADDR_V4MAPPED(&inp->in6p_laddr)) {
 643                                 /*
 644                                  * when remote addr is an IPv4-mapped address,
 645                                  * local addr should not be an IPv6 address,
 646                                  * since you cannot determine how to map IPv6
 647                                  * source address to IPv4.
 648                                  */
 649                                 error = EINVAL;
 650                                 goto release;
 651                         }
 652
 653                         af = AF_INET;
 654                 }
 655
 656                 if (!IN6_IS_ADDR_V4MAPPED(faddr)) {
 657                         error = in6_selectsrc(sin6, optp, inp, NULL,
 658                             td->td_ucred, &oifp, &in6a);
 659                         if (error)
 660                                 goto release;
 661                         if (oifp && scope_ambiguous &&
 662                             (error = in6_setscope(&sin6->sin6_addr,
 663                             oifp, NULL))) {
 664                                 goto release;
 665                         }
 666                         laddr = &in6a;
 667                 } else
 668                         laddr = &inp->in6p_laddr;       /* XXX */
 669                 if (laddr == NULL) {
 670                         if (error == 0)
 671                                 error = EADDRNOTAVAIL;
 672                         goto release;
 673                 }
 674                 if (inp->inp_lport == 0 &&
 675                     (error = in6_pcbsetport(laddr, inp, td->td_ucred)) != 0) {
 676                         /* Undo an address bind that may have occurred. */
 677                         inp->in6p_laddr = in6addr_any;
 678                         goto release;
 679                 }
 680         } else {
 681                 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 682                         error = ENOTCONN;
 683                         goto release;
 684                 }
 685                 if (IN6_IS_ADDR_V4MAPPED(&inp->in6p_faddr)) {
 686                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY)) {
 687                                 /*
 688                                  * XXX: this case would happen when the
 689                                  * application sets the V6ONLY flag after
 690                                  * connecting the foreign address.
 691                                  * Such applications should be fixed,
 692                                  * so we bark here.
 693                                  */
 694                                 log(LOG_INFO, "udp6_output: IPV6_V6ONLY "
 695                                     "option was set for a connected socket\n");
 696                                 error = EINVAL;
 697                                 goto release;
 698                         } else
 699                                 af = AF_INET;
 700                 }
 701                 laddr = &inp->in6p_laddr;
 702                 faddr = &inp->in6p_faddr;
 703                 fport = inp->inp_fport;
 704         }
 705
 706         if (af == AF_INET)
 707                 hlen = sizeof(struct ip);
 708
 709         /*
 710          * Calculate data length and get a mbuf
 711          * for UDP and IP6 headers.
 712          */
 713         M_PREPEND(m, hlen + sizeof(struct udphdr), M_DONTWAIT);
 714         if (m == 0) {
 715                 error = ENOBUFS;
 716                 goto release;
 717         }
 718
 719         /*
 720          * Stuff checksum and output datagram.
 721          */
 722         udp6 = (struct udphdr *)(mtod(m, caddr_t) + hlen);
 723         udp6->uh_sport = inp->inp_lport; /* lport is always set in the PCB */
 724         udp6->uh_dport = fport;
 725         if (plen <= 0xffff)
 726                 udp6->uh_ulen = htons((u_short)plen);
 727         else
 728                 udp6->uh_ulen = 0;
 729         udp6->uh_sum = 0;
 730
 731         switch (af) {
 732         case AF_INET6:
 733                 ip6 = mtod(m, struct ip6_hdr *);
 734                 ip6->ip6_flow   = inp->inp_flow & IPV6_FLOWINFO_MASK;
 735                 ip6->ip6_vfc    &= ~IPV6_VERSION_MASK;
 736                 ip6->ip6_vfc    |= IPV6_VERSION;
 737 #if 0                           /* ip6_plen will be filled in ip6_output. */
 738                 ip6->ip6_plen   = htons((u_short)plen);
 739 #endif
 740                 ip6->ip6_nxt    = IPPROTO_UDP;
 741                 ip6->ip6_hlim   = in6_selecthlim(inp, NULL);
 742                 ip6->ip6_src    = *laddr;
 743                 ip6->ip6_dst    = *faddr;
 744
 745                 if ((udp6->uh_sum = in6_cksum(m, IPPROTO_UDP,
 746                                 sizeof(struct ip6_hdr), plen)) == 0) {
 747                         udp6->uh_sum = 0xffff;
 748                 }
 749
 750                 flags = 0;
 751
 752                 UDPSTAT_INC(udps_opackets);
 753                 error = ip6_output(m, optp, NULL, flags, inp->in6p_moptions,
 754                     NULL, inp);
 755                 break;
 756         case AF_INET:
 757                 error = EAFNOSUPPORT;
 758                 goto release;
 759         }
 760         goto releaseopt;
 761
 762 release:
 763         m_freem(m);
 764
 765 releaseopt:
 766         if (control) {
 767                 ip6_clearpktopts(&opt, -1);
 768                 m_freem(control);
 769         }
 770         return (error);
 771 }
 772
 773 static void
 774 udp6_abort(struct socket *so)
 775 {
 776         struct inpcb *inp;
 777
 778         inp = sotoinpcb(so);
 779         KASSERT(inp != NULL, ("udp6_abort: inp == NULL"));
 780
 781 #ifdef INET
 782         if (inp->inp_vflag & INP_IPV4) {
 783                 struct pr_usrreqs *pru;
 784
 785                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
 786                 (*pru->pru_abort)(so);
 787                 return;
 788         }
 789 #endif
 790
 791         INP_INFO_WLOCK(&V_udbinfo);
 792         INP_WLOCK(inp);
 793         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 794                 in6_pcbdisconnect(inp);
 795                 inp->in6p_laddr = in6addr_any;
 796                 soisdisconnected(so);
 797         }
 798         INP_WUNLOCK(inp);
 799         INP_INFO_WUNLOCK(&V_udbinfo);
 800 }
 801
 802 static int
 803 udp6_attach(struct socket *so, int proto, struct thread *td)
 804 {
 805         struct inpcb *inp;
 806         int error;
 807
 808         inp = sotoinpcb(so);
 809         KASSERT(inp == NULL, ("udp6_attach: inp != NULL"));
 810
 811         if (so->so_snd.sb_hiwat == 0 || so->so_rcv.sb_hiwat == 0) {
 812                 error = soreserve(so, udp_sendspace, udp_recvspace);
 813                 if (error)
 814                         return (error);
 815         }
 816         INP_INFO_WLOCK(&V_udbinfo);
 817         error = in_pcballoc(so, &V_udbinfo);
 818         if (error) {
 819                 INP_INFO_WUNLOCK(&V_udbinfo);
 820                 return (error);
 821         }
 822         inp = (struct inpcb *)so->so_pcb;
 823         inp->inp_vflag |= INP_IPV6;
 824         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0)
 825                 inp->inp_vflag |= INP_IPV4;
 826         inp->in6p_hops = -1;    /* use kernel default */
 827         inp->in6p_cksum = -1;   /* just to be sure */
 828         /*
 829          * XXX: ugly!!
 830          * IPv4 TTL initialization is necessary for an IPv6 socket as well,
 831          * because the socket may be bound to an IPv6 wildcard address,
 832          * which may match an IPv4-mapped IPv6 address.
 833          */
 834         inp->inp_ip_ttl = V_ip_defttl;
 835
 836         error = udp_newudpcb(inp);
 837         if (error) {
 838                 in_pcbdetach(inp);
 839                 in_pcbfree(inp);
 840                 INP_INFO_WUNLOCK(&V_udbinfo);
 841                 return (error);
 842         }
 843         INP_WUNLOCK(inp);
 844         INP_INFO_WUNLOCK(&V_udbinfo);
 845         return (0);
 846 }
 847
 848 static int
 849 udp6_bind(struct socket *so, struct sockaddr *nam, struct thread *td)
 850 {
 851         struct inpcb *inp;
 852         int error;
 853
 854         inp = sotoinpcb(so);
 855         KASSERT(inp != NULL, ("udp6_bind: inp == NULL"));
 856
 857         INP_INFO_WLOCK(&V_udbinfo);
 858         INP_WLOCK(inp);
 859         inp->inp_vflag &= ~INP_IPV4;
 860         inp->inp_vflag |= INP_IPV6;
 861         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
 862                 struct sockaddr_in6 *sin6_p;
 863
 864                 sin6_p = (struct sockaddr_in6 *)nam;
 865
 866                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6_p->sin6_addr))
 867                         inp->inp_vflag |= INP_IPV4;
 868                 else if (IN6_IS_ADDR_V4MAPPED(&sin6_p->sin6_addr)) {
 869                         struct sockaddr_in sin;
 870
 871                         in6_sin6_2_sin(&sin, sin6_p);
 872                         inp->inp_vflag |= INP_IPV4;
 873                         inp->inp_vflag &= ~INP_IPV6;
 874                         error = in_pcbbind(inp, (struct sockaddr *)&sin,
 875                             td->td_ucred);
 876                         goto out;
 877                 }
 878         }
 879
 880         error = in6_pcbbind(inp, nam, td->td_ucred);
 881 out:
 882         INP_WUNLOCK(inp);
 883         INP_INFO_WUNLOCK(&V_udbinfo);
 884         return (error);
 885 }
 886
 887 static void
 888 udp6_close(struct socket *so)
 889 {
 890         struct inpcb *inp;
 891
 892         inp = sotoinpcb(so);
 893         KASSERT(inp != NULL, ("udp6_close: inp == NULL"));
 894
 895 #ifdef INET
 896         if (inp->inp_vflag & INP_IPV4) {
 897                 struct pr_usrreqs *pru;
 898
 899                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
 900                 (*pru->pru_disconnect)(so);
 901                 return;
 902         }
 903 #endif
 904         INP_INFO_WLOCK(&V_udbinfo);
 905         INP_WLOCK(inp);
 906         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 907                 in6_pcbdisconnect(inp);
 908                 inp->in6p_laddr = in6addr_any;
 909                 soisdisconnected(so);
 910         }
 911         INP_WUNLOCK(inp);
 912         INP_INFO_WUNLOCK(&V_udbinfo);
 913 }
 914
 915 static int
 916 udp6_connect(struct socket *so, struct sockaddr *nam, struct thread *td)
 917 {
 918         struct inpcb *inp;
 919         struct sockaddr_in6 *sin6;
 920         int error;
 921
 922         inp = sotoinpcb(so);
 923         sin6 = (struct sockaddr_in6 *)nam;
 924         KASSERT(inp != NULL, ("udp6_connect: inp == NULL"));
 925
 926         INP_INFO_WLOCK(&V_udbinfo);
 927         INP_WLOCK(inp);
 928         if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
 929                 struct sockaddr_in sin;
 930
 931                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) {
 932                         error = EINVAL;
 933                         goto out;
 934                 }
 935                 if (inp->inp_faddr.s_addr != INADDR_ANY) {
 936                         error = EISCONN;
 937                         goto out;
 938                 }
 939                 in6_sin6_2_sin(&sin, sin6);
 940                 inp->inp_vflag |= INP_IPV4;
 941                 inp->inp_vflag &= ~INP_IPV6;
 942                 error = prison_remote_ip4(td->td_ucred, &sin.sin_addr);
 943                 if (error != 0)
 944                         goto out;
 945                 error = in_pcbconnect(inp, (struct sockaddr *)&sin,
 946                     td->td_ucred);
 947                 if (error == 0)
 948                         soisconnected(so);
 949                 goto out;
 950         }
 951         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
 952                 error = EISCONN;
 953                 goto out;
 954         }
 955         inp->inp_vflag &= ~INP_IPV4;
 956         inp->inp_vflag |= INP_IPV6;
 957         error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr);
 958         if (error != 0)
 959                 goto out;
 960         error = in6_pcbconnect(inp, nam, td->td_ucred);
 961         if (error == 0)
 962                 soisconnected(so);
 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         struct inpcb *inp;
 973         struct udpcb *up;
 974
 975         inp = sotoinpcb(so);
 976         KASSERT(inp != NULL, ("udp6_detach: inp == NULL"));
 977
 978         INP_INFO_WLOCK(&V_udbinfo);
 979         INP_WLOCK(inp);
 980         up = intoudpcb(inp);
 981         KASSERT(up != NULL, ("%s: up == NULL", __func__));
 982         in_pcbdetach(inp);
 983         in_pcbfree(inp);
 984         INP_INFO_WUNLOCK(&V_udbinfo);
 985         udp_discardcb(up);
 986 }
 987
 988 static int
 989 udp6_disconnect(struct socket *so)
 990 {
 991         struct inpcb *inp;
 992         int error;
 993
 994         inp = sotoinpcb(so);
 995         KASSERT(inp != NULL, ("udp6_disconnect: inp == NULL"));
 996
 997         INP_INFO_WLOCK(&V_udbinfo);
 998         INP_WLOCK(inp);
 999
1000 #ifdef INET
1001         if (inp->inp_vflag & INP_IPV4) {
1002                 struct pr_usrreqs *pru;
1003
1004                 pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
1005                 error = (*pru->pru_disconnect)(so);
1006                 goto out;
1007         }
1008 #endif
1009
1010         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) {
1011                 error = ENOTCONN;
1012                 goto out;
1013         }
1014
1015         in6_pcbdisconnect(inp);
1016         inp->in6p_laddr = in6addr_any;
1017         SOCK_LOCK(so);
1018         so->so_state &= ~SS_ISCONNECTED;                /* XXX */
1019         SOCK_UNLOCK(so);
1020 out:
1021         INP_WUNLOCK(inp);
1022         INP_INFO_WUNLOCK(&V_udbinfo);
1023         return (0);
1024 }
1025
1026 static int
1027 udp6_send(struct socket *so, int flags, struct mbuf *m,
1028     struct sockaddr *addr, struct mbuf *control, struct thread *td)
1029 {
1030         struct inpcb *inp;
1031         int error = 0;
1032
1033         inp = sotoinpcb(so);
1034         KASSERT(inp != NULL, ("udp6_send: inp == NULL"));
1035
1036         INP_INFO_WLOCK(&V_udbinfo);
1037         INP_WLOCK(inp);
1038         if (addr) {
1039                 if (addr->sa_len != sizeof(struct sockaddr_in6)) {
1040                         error = EINVAL;
1041                         goto bad;
1042                 }
1043                 if (addr->sa_family != AF_INET6) {
1044                         error = EAFNOSUPPORT;
1045                         goto bad;
1046                 }
1047         }
1048
1049 #ifdef INET
1050         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) {
1051                 int hasv4addr;
1052                 struct sockaddr_in6 *sin6 = 0;
1053
1054                 if (addr == 0)
1055                         hasv4addr = (inp->inp_vflag & INP_IPV4);
1056                 else {
1057                         sin6 = (struct sockaddr_in6 *)addr;
1058                         hasv4addr = IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)
1059                             ? 1 : 0;
1060                 }
1061                 if (hasv4addr) {
1062                         struct pr_usrreqs *pru;
1063
1064                         /*
1065                          * XXXRW: We release UDP-layer locks before calling
1066                          * udp_send() in order to avoid recursion.  However,
1067                          * this does mean there is a short window where inp's
1068                          * fields are unstable.  Could this lead to a
1069                          * potential race in which the factors causing us to
1070                          * select the UDPv4 output routine are invalidated?
1071                          */
1072                         INP_WUNLOCK(inp);
1073                         INP_INFO_WUNLOCK(&V_udbinfo);
1074                         if (sin6)
1075                                 in6_sin6_2_sin_in_sock(addr);
1076                         pru = inetsw[ip_protox[IPPROTO_UDP]].pr_usrreqs;
1077                         /* addr will just be freed in sendit(). */
1078                         return ((*pru->pru_send)(so, flags, m, addr, control,
1079                             td));
1080                 }
1081         }
1082 #endif
1083 #ifdef MAC
1084         mac_inpcb_create_mbuf(inp, m);
1085 #endif
1086         error = udp6_output(inp, m, addr, control, td);
1087         INP_WUNLOCK(inp);
1088         INP_INFO_WUNLOCK(&V_udbinfo);
1089         return (error);
1090
1091 bad:
1092         INP_WUNLOCK(inp);
1093         INP_INFO_WUNLOCK(&V_udbinfo);
1094         m_freem(m);
1095         return (error);
1096 }
1097
1098 struct pr_usrreqs udp6_usrreqs = {
1099         .pru_abort =            udp6_abort,
1100         .pru_attach =           udp6_attach,
1101         .pru_bind =             udp6_bind,
1102         .pru_connect =          udp6_connect,
1103         .pru_control =          in6_control,
1104         .pru_detach =           udp6_detach,
1105         .pru_disconnect =       udp6_disconnect,
1106         .pru_peeraddr =         in6_mapped_peeraddr,
1107         .pru_send =             udp6_send,
1108         .pru_shutdown =         udp_shutdown,
1109         .pru_sockaddr =         in6_mapped_sockaddr,
1110         .pru_soreceive =        soreceive_dgram,
1111         .pru_sosend =           sosend_dgram,
1112         .pru_sosetlabel =       in_pcbsosetlabel,
1113         .pru_close =            udp6_close
1114 };