sys/netinet/in_pcb.c

   1 /*
   2  * Copyright (c) 1982, 1986, 1991, 1993, 1995
   3  *      The Regents of the University of California.  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. All advertising materials mentioning features or use of this software
  14  *    must display the following acknowledgement:
  15  *      This product includes software developed by the University of
  16  *      California, Berkeley and its contributors.
  17  * 4. Neither the name of the University nor the names of its contributors
  18  *    may be used to endorse or promote products derived from this software
  19  *    without specific prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      @(#)in_pcb.c    8.4 (Berkeley) 5/24/95
  34  * $FreeBSD: src/sys/netinet/in_pcb.c,v 1.59.2.26 2003/01/24 05:11:33 sam Exp $
  35  */
  36
  37 #include "opt_ipsec.h"
  38 #include "opt_inet6.h"
  39
  40 #include <sys/param.h>
  41 #include <sys/systm.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/domain.h>
  45 #include <sys/protosw.h>
  46 #include <sys/socket.h>
  47 #include <sys/socketvar.h>
  48 #include <sys/proc.h>
  49 #include <sys/jail.h>
  50 #include <sys/kernel.h>
  51 #include <sys/sysctl.h>
  52
  53 #include <machine/limits.h>
  54
  55 #include <vm/vm_zone.h>
  56
  57 #include <net/if.h>
  58 #include <net/if_types.h>
  59 #include <net/route.h>
  60
  61 #include <netinet/in.h>
  62 #include <netinet/in_pcb.h>
  63 #include <netinet/in_var.h>
  64 #include <netinet/ip_var.h>
  65 #ifdef INET6
  66 #include <netinet/ip6.h>
  67 #include <netinet6/ip6_var.h>
  68 #endif /* INET6 */
  69
  70 #ifdef IPSEC
  71 #include <netinet6/ipsec.h>
  72 #include <netkey/key.h>
  73 #endif /* IPSEC */
  74
  75 #ifdef FAST_IPSEC
  76 #if defined(IPSEC) || defined(IPSEC_ESP)
  77 #error "Bad idea: don't compile with both IPSEC and FAST_IPSEC!"
  78 #endif
  79
  80 #include <netipsec/ipsec.h>
  81 #include <netipsec/key.h>
  82 #define IPSEC
  83 #endif /* FAST_IPSEC */
  84
  85 struct  in_addr zeroin_addr;
  86
  87 /*
  88  * These configure the range of local port addresses assigned to
  89  * "unspecified" outgoing connections/packets/whatever.
  90  */
  91 int     ipport_lowfirstauto  = IPPORT_RESERVED - 1;     /* 1023 */
  92 int     ipport_lowlastauto = IPPORT_RESERVEDSTART;      /* 600 */
  93 int     ipport_firstauto = IPPORT_RESERVED;             /* 1024 */
  94 int     ipport_lastauto  = IPPORT_USERRESERVED;         /* 5000 */
  95 int     ipport_hifirstauto = IPPORT_HIFIRSTAUTO;        /* 49152 */
  96 int     ipport_hilastauto  = IPPORT_HILASTAUTO;         /* 65535 */
  97
  98 #define RANGECHK(var, min, max) \
  99         if ((var) < (min)) { (var) = (min); } \
 100         else if ((var) > (max)) { (var) = (max); }
 101
 102 static int
 103 sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
 104 {
 105         int error = sysctl_handle_int(oidp,
 106                 oidp->oid_arg1, oidp->oid_arg2, req);
 107         if (!error) {
 108                 RANGECHK(ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
 109                 RANGECHK(ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
 110                 RANGECHK(ipport_firstauto, IPPORT_RESERVED, USHRT_MAX);
 111                 RANGECHK(ipport_lastauto, IPPORT_RESERVED, USHRT_MAX);
 112                 RANGECHK(ipport_hifirstauto, IPPORT_RESERVED, USHRT_MAX);
 113                 RANGECHK(ipport_hilastauto, IPPORT_RESERVED, USHRT_MAX);
 114         }
 115         return error;
 116 }
 117
 118 #undef RANGECHK
 119
 120 SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0, "IP Ports");
 121
 122 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLTYPE_INT|CTLFLAG_RW,
 123            &ipport_lowfirstauto, 0, &sysctl_net_ipport_check, "I", "");
 124 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLTYPE_INT|CTLFLAG_RW,
 125            &ipport_lowlastauto, 0, &sysctl_net_ipport_check, "I", "");
 126 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLTYPE_INT|CTLFLAG_RW,
 127            &ipport_firstauto, 0, &sysctl_net_ipport_check, "I", "");
 128 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLTYPE_INT|CTLFLAG_RW,
 129            &ipport_lastauto, 0, &sysctl_net_ipport_check, "I", "");
 130 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLTYPE_INT|CTLFLAG_RW,
 131            &ipport_hifirstauto, 0, &sysctl_net_ipport_check, "I", "");
 132 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLTYPE_INT|CTLFLAG_RW,
 133            &ipport_hilastauto, 0, &sysctl_net_ipport_check, "I", "");
 134
 135 /*
 136  * in_pcb.c: manage the Protocol Control Blocks.
 137  *
 138  * NOTE: It is assumed that most of these functions will be called at
 139  * splnet(). XXX - There are, unfortunately, a few exceptions to this
 140  * rule that should be fixed.
 141  */
 142
 143 /*
 144  * Allocate a PCB and associate it with the socket.
 145  */
 146 int
 147 in_pcballoc(so, pcbinfo, p)
 148         struct socket *so;
 149         struct inpcbinfo *pcbinfo;
 150         struct proc *p;
 151 {
 152         register struct inpcb *inp;
 153 #ifdef IPSEC
 154         int error;
 155 #endif
 156
 157         inp = zalloci(pcbinfo->ipi_zone);
 158         if (inp == NULL)
 159                 return (ENOBUFS);
 160         bzero((caddr_t)inp, sizeof(*inp));
 161         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 162         inp->inp_pcbinfo = pcbinfo;
 163         inp->inp_socket = so;
 164 #ifdef IPSEC
 165         error = ipsec_init_policy(so, &inp->inp_sp);
 166         if (error != 0) {
 167                 zfreei(pcbinfo->ipi_zone, inp);
 168                 return error;
 169         }
 170 #endif /*IPSEC*/
 171 #if defined(INET6)
 172         if (INP_SOCKAF(so) == AF_INET6 && ip6_v6only)
 173                 inp->inp_flags |= IN6P_IPV6_V6ONLY;
 174 #endif
 175         LIST_INSERT_HEAD(pcbinfo->listhead, inp, inp_list);
 176         pcbinfo->ipi_count++;
 177         so->so_pcb = (caddr_t)inp;
 178 #ifdef INET6
 179         if (ip6_auto_flowlabel)
 180                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 181 #endif
 182         return (0);
 183 }
 184
 185 int
 186 in_pcbbind(inp, nam, p)
 187         register struct inpcb *inp;
 188         struct sockaddr *nam;
 189         struct proc *p;
 190 {
 191         register struct socket *so = inp->inp_socket;
 192         unsigned short *lastport;
 193         struct sockaddr_in *sin;
 194         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 195         u_short lport = 0;
 196         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 197         int error, prison = 0;
 198
 199         if (TAILQ_EMPTY(&in_ifaddrhead)) /* XXX broken! */
 200                 return (EADDRNOTAVAIL);
 201         if (inp->inp_lport || inp->inp_laddr.s_addr != INADDR_ANY)
 202                 return (EINVAL);
 203         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 204                 wild = 1;
 205         if (nam) {
 206                 sin = (struct sockaddr_in *)nam;
 207                 if (nam->sa_len != sizeof (*sin))
 208                         return (EINVAL);
 209 #ifdef notdef
 210                 /*
 211                  * We should check the family, but old programs
 212                  * incorrectly fail to initialize it.
 213                  */
 214                 if (sin->sin_family != AF_INET)
 215                         return (EAFNOSUPPORT);
 216 #endif
 217                 if (sin->sin_addr.s_addr != INADDR_ANY)
 218                         if (prison_ip(p, 0, &sin->sin_addr.s_addr))
 219                                 return(EINVAL);
 220                 lport = sin->sin_port;
 221                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 222                         /*
 223                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 224                          * allow complete duplication of binding if
 225                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 226                          * and a multicast address is bound on both
 227                          * new and duplicated sockets.
 228                          */
 229                         if (so->so_options & SO_REUSEADDR)
 230                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 231                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 232                         sin->sin_port = 0;              /* yech... */
 233                         bzero(&sin->sin_zero, sizeof(sin->sin_zero));
 234                         if (ifa_ifwithaddr((struct sockaddr *)sin) == 0)
 235                                 return (EADDRNOTAVAIL);
 236                 }
 237                 if (lport) {
 238                         struct inpcb *t;
 239
 240                         /* GROSS */
 241                         if (ntohs(lport) < IPPORT_RESERVED && p &&
 242                             suser_xxx(0, p, PRISON_ROOT))
 243                                 return (EACCES);
 244                         if (p && p->p_prison)
 245                                 prison = 1;
 246                         if (so->so_cred->cr_uid != 0 &&
 247                             !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 248                                 t = in_pcblookup_local(inp->inp_pcbinfo,
 249                                     sin->sin_addr, lport,
 250                                     prison ? 0 :  INPLOOKUP_WILDCARD);
 251                                 if (t &&
 252                                     (ntohl(sin->sin_addr.s_addr) != INADDR_ANY ||
 253                                      ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 254                                      (t->inp_socket->so_options &
 255                                          SO_REUSEPORT) == 0) &&
 256                                     (so->so_cred->cr_uid !=
 257                                      t->inp_socket->so_cred->cr_uid)) {
 258 #if defined(INET6)
 259                                         if (ntohl(sin->sin_addr.s_addr) !=
 260                                             INADDR_ANY ||
 261                                             ntohl(t->inp_laddr.s_addr) !=
 262                                             INADDR_ANY ||
 263                                             INP_SOCKAF(so) ==
 264                                             INP_SOCKAF(t->inp_socket))
 265 #endif /* defined(INET6) */
 266                                         return (EADDRINUSE);
 267                                 }
 268                         }
 269                         if (prison &&
 270                             prison_ip(p, 0, &sin->sin_addr.s_addr))
 271                                 return (EADDRNOTAVAIL);
 272                         t = in_pcblookup_local(pcbinfo, sin->sin_addr,
 273                             lport, prison ? 0 : wild);
 274                         if (t &&
 275                             (reuseport & t->inp_socket->so_options) == 0) {
 276 #if defined(INET6)
 277                                 if (ntohl(sin->sin_addr.s_addr) !=
 278                                     INADDR_ANY ||
 279                                     ntohl(t->inp_laddr.s_addr) !=
 280                                     INADDR_ANY ||
 281                                     INP_SOCKAF(so) ==
 282                                     INP_SOCKAF(t->inp_socket))
 283 #endif /* defined(INET6) */
 284                                 return (EADDRINUSE);
 285                         }
 286                 }
 287                 inp->inp_laddr = sin->sin_addr;
 288         }
 289         if (lport == 0) {
 290                 ushort first, last;
 291                 int count;
 292
 293                 if (inp->inp_laddr.s_addr != INADDR_ANY)
 294                         if (prison_ip(p, 0, &inp->inp_laddr.s_addr )) {
 295                                 inp->inp_laddr.s_addr = INADDR_ANY;
 296                                 return (EINVAL);
 297                         }
 298                 inp->inp_flags |= INP_ANONPORT;
 299
 300                 if (inp->inp_flags & INP_HIGHPORT) {
 301                         first = ipport_hifirstauto;     /* sysctl */
 302                         last  = ipport_hilastauto;
 303                         lastport = &pcbinfo->lasthi;
 304                 } else if (inp->inp_flags & INP_LOWPORT) {
 305                         if (p && (error = suser_xxx(0, p, PRISON_ROOT))) {
 306                                 inp->inp_laddr.s_addr = INADDR_ANY;
 307                                 return error;
 308                         }
 309                         first = ipport_lowfirstauto;    /* 1023 */
 310                         last  = ipport_lowlastauto;     /* 600 */
 311                         lastport = &pcbinfo->lastlow;
 312                 } else {
 313                         first = ipport_firstauto;       /* sysctl */
 314                         last  = ipport_lastauto;
 315                         lastport = &pcbinfo->lastport;
 316                 }
 317                 /*
 318                  * Simple check to ensure all ports are not used up causing
 319                  * a deadlock here.
 320                  *
 321                  * We split the two cases (up and down) so that the direction
 322                  * is not being tested on each round of the loop.
 323                  */
 324                 if (first > last) {
 325                         /*
 326                          * counting down
 327                          */
 328                         count = first - last;
 329
 330                         do {
 331                                 if (count-- < 0) {      /* completely used? */
 332                                         inp->inp_laddr.s_addr = INADDR_ANY;
 333                                         return (EADDRNOTAVAIL);
 334                                 }
 335                                 --*lastport;
 336                                 if (*lastport > first || *lastport < last)
 337                                         *lastport = first;
 338                                 lport = htons(*lastport);
 339                         } while (in_pcblookup_local(pcbinfo,
 340                                  inp->inp_laddr, lport, wild));
 341                 } else {
 342                         /*
 343                          * counting up
 344                          */
 345                         count = last - first;
 346
 347                         do {
 348                                 if (count-- < 0) {      /* completely used? */
 349                                         inp->inp_laddr.s_addr = INADDR_ANY;
 350                                         return (EADDRNOTAVAIL);
 351                                 }
 352                                 ++*lastport;
 353                                 if (*lastport < first || *lastport > last)
 354                                         *lastport = first;
 355                                 lport = htons(*lastport);
 356                         } while (in_pcblookup_local(pcbinfo,
 357                                  inp->inp_laddr, lport, wild));
 358                 }
 359         }
 360         inp->inp_lport = lport;
 361         if (prison_ip(p, 0, &inp->inp_laddr.s_addr)) {
 362                 inp->inp_laddr.s_addr = INADDR_ANY;
 363                 inp->inp_lport = 0;
 364                 return (EINVAL);
 365         }
 366         if (in_pcbinshash(inp) != 0) {
 367                 inp->inp_laddr.s_addr = INADDR_ANY;
 368                 inp->inp_lport = 0;
 369                 return (EAGAIN);
 370         }
 371         return (0);
 372 }
 373
 374 /*
 375  *   Transform old in_pcbconnect() into an inner subroutine for new
 376  *   in_pcbconnect(): Do some validity-checking on the remote
 377  *   address (in mbuf 'nam') and then determine local host address
 378  *   (i.e., which interface) to use to access that remote host.
 379  *
 380  *   This preserves definition of in_pcbconnect(), while supporting a
 381  *   slightly different version for T/TCP.  (This is more than
 382  *   a bit of a kludge, but cleaning up the internal interfaces would
 383  *   have forced minor changes in every protocol).
 384  */
 385
 386 int
 387 in_pcbladdr(inp, nam, plocal_sin)
 388         register struct inpcb *inp;
 389         struct sockaddr *nam;
 390         struct sockaddr_in **plocal_sin;
 391 {
 392         struct in_ifaddr *ia;
 393         register struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 394
 395         if (nam->sa_len != sizeof (*sin))
 396                 return (EINVAL);
 397         if (sin->sin_family != AF_INET)
 398                 return (EAFNOSUPPORT);
 399         if (sin->sin_port == 0)
 400                 return (EADDRNOTAVAIL);
 401         if (!TAILQ_EMPTY(&in_ifaddrhead)) {
 402                 /*
 403                  * If the destination address is INADDR_ANY,
 404                  * use the primary local address.
 405                  * If the supplied address is INADDR_BROADCAST,
 406                  * and the primary interface supports broadcast,
 407                  * choose the broadcast address for that interface.
 408                  */
 409                 if (sin->sin_addr.s_addr == INADDR_ANY)
 410                     sin->sin_addr = IA_SIN(TAILQ_FIRST(&in_ifaddrhead))->sin_addr;
 411                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 412                   (TAILQ_FIRST(&in_ifaddrhead)->ia_ifp->if_flags & IFF_BROADCAST))
 413                     sin->sin_addr = satosin(&TAILQ_FIRST(&in_ifaddrhead)->ia_broadaddr)->sin_addr;
 414         }
 415         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 416                 register struct route *ro;
 417
 418                 ia = (struct in_ifaddr *)0;
 419                 /*
 420                  * If route is known or can be allocated now,
 421                  * our src addr is taken from the i/f, else punt.
 422                  * Note that we should check the address family of the cached
 423                  * destination, in case of sharing the cache with IPv6.
 424                  */
 425                 ro = &inp->inp_route;
 426                 if (ro->ro_rt &&
 427                     (ro->ro_dst.sa_family != AF_INET ||
 428                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 429                      sin->sin_addr.s_addr ||
 430                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 431                         RTFREE(ro->ro_rt);
 432                         ro->ro_rt = (struct rtentry *)0;
 433                 }
 434                 if ((inp->inp_socket->so_options & SO_DONTROUTE) == 0 && /*XXX*/
 435                     (ro->ro_rt == (struct rtentry *)0 ||
 436                     ro->ro_rt->rt_ifp == (struct ifnet *)0)) {
 437                         /* No route yet, so try to acquire one */
 438                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 439                         ro->ro_dst.sa_family = AF_INET;
 440                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 441                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 442                                 sin->sin_addr;
 443                         rtalloc(ro);
 444                 }
 445                 /*
 446                  * If we found a route, use the address
 447                  * corresponding to the outgoing interface
 448                  * unless it is the loopback (in case a route
 449                  * to our address on another net goes to loopback).
 450                  */
 451                 if (ro->ro_rt && !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK))
 452                         ia = ifatoia(ro->ro_rt->rt_ifa);
 453                 if (ia == 0) {
 454                         u_short fport = sin->sin_port;
 455
 456                         sin->sin_port = 0;
 457                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 458                         if (ia == 0)
 459                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 460                         sin->sin_port = fport;
 461                         if (ia == 0)
 462                                 ia = TAILQ_FIRST(&in_ifaddrhead);
 463                         if (ia == 0)
 464                                 return (EADDRNOTAVAIL);
 465                 }
 466                 /*
 467                  * If the destination address is multicast and an outgoing
 468                  * interface has been set as a multicast option, use the
 469                  * address of that interface as our source address.
 470                  */
 471                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 472                     inp->inp_moptions != NULL) {
 473                         struct ip_moptions *imo;
 474                         struct ifnet *ifp;
 475
 476                         imo = inp->inp_moptions;
 477                         if (imo->imo_multicast_ifp != NULL) {
 478                                 ifp = imo->imo_multicast_ifp;
 479                                 TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link)
 480                                         if (ia->ia_ifp == ifp)
 481                                                 break;
 482                                 if (ia == 0)
 483                                         return (EADDRNOTAVAIL);
 484                         }
 485                 }
 486         /*
 487          * Don't do pcblookup call here; return interface in plocal_sin
 488          * and exit to caller, that will do the lookup.
 489          */
 490                 *plocal_sin = &ia->ia_addr;
 491
 492         }
 493         return(0);
 494 }
 495
 496 /*
 497  * Outer subroutine:
 498  * Connect from a socket to a specified address.
 499  * Both address and port must be specified in argument sin.
 500  * If don't have a local address for this socket yet,
 501  * then pick one.
 502  */
 503 int
 504 in_pcbconnect(inp, nam, p)
 505         register struct inpcb *inp;
 506         struct sockaddr *nam;
 507         struct proc *p;
 508 {
 509         struct sockaddr_in *ifaddr;
 510         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 511         struct sockaddr_in sa;
 512         int error;
 513
 514         if (inp->inp_laddr.s_addr == INADDR_ANY && p->p_prison != NULL) {
 515                 bzero(&sa, sizeof (sa));
 516                 sa.sin_addr.s_addr = htonl(p->p_prison->pr_ip);
 517                 sa.sin_len=sizeof (sa);
 518                 sa.sin_family = AF_INET;
 519                 error = in_pcbbind(inp, (struct sockaddr *)&sa, p);
 520                 if (error)
 521                         return (error);
 522         }
 523         /*
 524          *   Call inner routine, to assign local interface address.
 525          */
 526         if ((error = in_pcbladdr(inp, nam, &ifaddr)) != 0)
 527                 return(error);
 528
 529         if (in_pcblookup_hash(inp->inp_pcbinfo, sin->sin_addr, sin->sin_port,
 530             inp->inp_laddr.s_addr ? inp->inp_laddr : ifaddr->sin_addr,
 531             inp->inp_lport, 0, NULL) != NULL) {
 532                 return (EADDRINUSE);
 533         }
 534         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 535                 if (inp->inp_lport == 0) {
 536                         error = in_pcbbind(inp, (struct sockaddr *)0, p);
 537                         if (error)
 538                             return (error);
 539                 }
 540                 inp->inp_laddr = ifaddr->sin_addr;
 541         }
 542         inp->inp_faddr = sin->sin_addr;
 543         inp->inp_fport = sin->sin_port;
 544         in_pcbrehash(inp);
 545         return (0);
 546 }
 547
 548 void
 549 in_pcbdisconnect(inp)
 550         struct inpcb *inp;
 551 {
 552
 553         inp->inp_faddr.s_addr = INADDR_ANY;
 554         inp->inp_fport = 0;
 555         in_pcbrehash(inp);
 556         if (inp->inp_socket->so_state & SS_NOFDREF)
 557                 in_pcbdetach(inp);
 558 }
 559
 560 void
 561 in_pcbdetach(inp)
 562         struct inpcb *inp;
 563 {
 564         struct socket *so = inp->inp_socket;
 565         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 566
 567 #ifdef IPSEC
 568         ipsec4_delete_pcbpolicy(inp);
 569 #endif /*IPSEC*/
 570         inp->inp_gencnt = ++ipi->ipi_gencnt;
 571         in_pcbremlists(inp);
 572         so->so_pcb = 0;
 573         sofree(so);
 574         if (inp->inp_options)
 575                 (void)m_free(inp->inp_options);
 576         if (inp->inp_route.ro_rt)
 577                 rtfree(inp->inp_route.ro_rt);
 578         ip_freemoptions(inp->inp_moptions);
 579         inp->inp_vflag = 0;
 580         zfreei(ipi->ipi_zone, inp);
 581 }
 582
 583 /*
 584  * The calling convention of in_setsockaddr() and in_setpeeraddr() was
 585  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 586  * in struct pr_usrreqs, so that protocols can just reference then directly
 587  * without the need for a wrapper function.  The socket must have a valid
 588  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 589  * except through a kernel programming error, so it is acceptable to panic
 590  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 591  * because there actually /is/ a programming error somewhere... XXX)
 592  */
 593 int
 594 in_setsockaddr(so, nam)
 595         struct socket *so;
 596         struct sockaddr **nam;
 597 {
 598         int s;
 599         register struct inpcb *inp;
 600         register struct sockaddr_in *sin;
 601
 602         /*
 603          * Do the malloc first in case it blocks.
 604          */
 605         MALLOC(sin, struct sockaddr_in *, sizeof *sin, M_SONAME,
 606                 M_WAITOK | M_ZERO);
 607         sin->sin_family = AF_INET;
 608         sin->sin_len = sizeof(*sin);
 609
 610         s = splnet();
 611         inp = sotoinpcb(so);
 612         if (!inp) {
 613                 splx(s);
 614                 free(sin, M_SONAME);
 615                 return ECONNRESET;
 616         }
 617         sin->sin_port = inp->inp_lport;
 618         sin->sin_addr = inp->inp_laddr;
 619         splx(s);
 620
 621         *nam = (struct sockaddr *)sin;
 622         return 0;
 623 }
 624
 625 int
 626 in_setpeeraddr(so, nam)
 627         struct socket *so;
 628         struct sockaddr **nam;
 629 {
 630         int s;
 631         struct inpcb *inp;
 632         register struct sockaddr_in *sin;
 633
 634         /*
 635          * Do the malloc first in case it blocks.
 636          */
 637         MALLOC(sin, struct sockaddr_in *, sizeof *sin, M_SONAME,
 638                 M_WAITOK | M_ZERO);
 639         sin->sin_family = AF_INET;
 640         sin->sin_len = sizeof(*sin);
 641
 642         s = splnet();
 643         inp = sotoinpcb(so);
 644         if (!inp) {
 645                 splx(s);
 646                 free(sin, M_SONAME);
 647                 return ECONNRESET;
 648         }
 649         sin->sin_port = inp->inp_fport;
 650         sin->sin_addr = inp->inp_faddr;
 651         splx(s);
 652
 653         *nam = (struct sockaddr *)sin;
 654         return 0;
 655 }
 656
 657 void
 658 in_pcbnotifyall(head, faddr, errno, notify)
 659         struct inpcbhead *head;
 660         struct in_addr faddr;
 661         void (*notify) __P((struct inpcb *, int));
 662 {
 663         struct inpcb *inp, *ninp;
 664         int s;
 665
 666         s = splnet();
 667         for (inp = LIST_FIRST(head); inp != NULL; inp = ninp) {
 668                 ninp = LIST_NEXT(inp, inp_list);
 669 #ifdef INET6
 670                 if ((inp->inp_vflag & INP_IPV4) == 0)
 671                         continue;
 672 #endif
 673                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
 674                     inp->inp_socket == NULL)
 675                                 continue;
 676                 (*notify)(inp, errno);
 677         }
 678         splx(s);
 679 }
 680
 681 void
 682 in_pcbpurgeif0(head, ifp)
 683         struct inpcb *head;
 684         struct ifnet *ifp;
 685 {
 686         struct inpcb *inp;
 687         struct ip_moptions *imo;
 688         int i, gap;
 689
 690         for (inp = head; inp != NULL; inp = LIST_NEXT(inp, inp_list)) {
 691                 imo = inp->inp_moptions;
 692                 if ((inp->inp_vflag & INP_IPV4) &&
 693                     imo != NULL) {
 694                         /*
 695                          * Unselect the outgoing interface if it is being
 696                          * detached.
 697                          */
 698                         if (imo->imo_multicast_ifp == ifp)
 699                                 imo->imo_multicast_ifp = NULL;
 700
 701                         /*
 702                          * Drop multicast group membership if we joined
 703                          * through the interface being detached.
 704                          */
 705                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
 706                             i++) {
 707                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
 708                                         in_delmulti(imo->imo_membership[i]);
 709                                         gap++;
 710                                 } else if (gap != 0)
 711                                         imo->imo_membership[i - gap] =
 712                                             imo->imo_membership[i];
 713                         }
 714                         imo->imo_num_memberships -= gap;
 715                 }
 716         }
 717 }
 718
 719 /*
 720  * Check for alternatives when higher level complains
 721  * about service problems.  For now, invalidate cached
 722  * routing information.  If the route was created dynamically
 723  * (by a redirect), time to try a default gateway again.
 724  */
 725 void
 726 in_losing(inp)
 727         struct inpcb *inp;
 728 {
 729         register struct rtentry *rt;
 730         struct rt_addrinfo info;
 731
 732         if ((rt = inp->inp_route.ro_rt)) {
 733                 bzero((caddr_t)&info, sizeof(info));
 734                 info.rti_flags = rt->rt_flags;
 735                 info.rti_info[RTAX_DST] = rt_key(rt);
 736                 info.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
 737                 info.rti_info[RTAX_NETMASK] = rt_mask(rt);
 738                 rt_missmsg(RTM_LOSING, &info, rt->rt_flags, 0);
 739                 if (rt->rt_flags & RTF_DYNAMIC)
 740                         (void) rtrequest1(RTM_DELETE, &info, NULL);
 741                 inp->inp_route.ro_rt = NULL;
 742                 rtfree(rt);
 743                 /*
 744                  * A new route can be allocated
 745                  * the next time output is attempted.
 746                  */
 747         }
 748 }
 749
 750 /*
 751  * After a routing change, flush old routing
 752  * and allocate a (hopefully) better one.
 753  */
 754 void
 755 in_rtchange(inp, errno)
 756         register struct inpcb *inp;
 757         int errno;
 758 {
 759         if (inp->inp_route.ro_rt) {
 760                 rtfree(inp->inp_route.ro_rt);
 761                 inp->inp_route.ro_rt = 0;
 762                 /*
 763                  * A new route can be allocated the next time
 764                  * output is attempted.
 765                  */
 766         }
 767 }
 768
 769 /*
 770  * Lookup a PCB based on the local address and port.
 771  */
 772 struct inpcb *
 773 in_pcblookup_local(pcbinfo, laddr, lport_arg, wild_okay)
 774         struct inpcbinfo *pcbinfo;
 775         struct in_addr laddr;
 776         u_int lport_arg;
 777         int wild_okay;
 778 {
 779         register struct inpcb *inp;
 780         int matchwild = 3, wildcard;
 781         u_short lport = lport_arg;
 782
 783         if (!wild_okay) {
 784                 struct inpcbhead *head;
 785                 /*
 786                  * Look for an unconnected (wildcard foreign addr) PCB that
 787                  * matches the local address and port we're looking for.
 788                  */
 789                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0, pcbinfo->hashmask)];
 790                 LIST_FOREACH(inp, head, inp_hash) {
 791 #ifdef INET6
 792                         if ((inp->inp_vflag & INP_IPV4) == 0)
 793                                 continue;
 794 #endif
 795                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
 796                             inp->inp_laddr.s_addr == laddr.s_addr &&
 797                             inp->inp_lport == lport) {
 798                                 /*
 799                                  * Found.
 800                                  */
 801                                 return (inp);
 802                         }
 803                 }
 804                 /*
 805                  * Not found.
 806                  */
 807                 return (NULL);
 808         } else {
 809                 struct inpcbporthead *porthash;
 810                 struct inpcbport *phd;
 811                 struct inpcb *match = NULL;
 812                 /*
 813                  * Best fit PCB lookup.
 814                  *
 815                  * First see if this local port is in use by looking on the
 816                  * port hash list.
 817                  */
 818                 porthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(lport,
 819                     pcbinfo->porthashmask)];
 820                 LIST_FOREACH(phd, porthash, phd_hash) {
 821                         if (phd->phd_port == lport)
 822                                 break;
 823                 }
 824                 if (phd != NULL) {
 825                         /*
 826                          * Port is in use by one or more PCBs. Look for best
 827                          * fit.
 828                          */
 829                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 830                                 wildcard = 0;
 831 #ifdef INET6
 832                                 if ((inp->inp_vflag & INP_IPV4) == 0)
 833                                         continue;
 834 #endif
 835                                 if (inp->inp_faddr.s_addr != INADDR_ANY)
 836                                         wildcard++;
 837                                 if (inp->inp_laddr.s_addr != INADDR_ANY) {
 838                                         if (laddr.s_addr == INADDR_ANY)
 839                                                 wildcard++;
 840                                         else if (inp->inp_laddr.s_addr != laddr.s_addr)
 841                                                 continue;
 842                                 } else {
 843                                         if (laddr.s_addr != INADDR_ANY)
 844                                                 wildcard++;
 845                                 }
 846                                 if (wildcard < matchwild) {
 847                                         match = inp;
 848                                         matchwild = wildcard;
 849                                         if (matchwild == 0) {
 850                                                 break;
 851                                         }
 852                                 }
 853                         }
 854                 }
 855                 return (match);
 856         }
 857 }
 858
 859 /*
 860  * Lookup PCB in hash list.
 861  */
 862 struct inpcb *
 863 in_pcblookup_hash(pcbinfo, faddr, fport_arg, laddr, lport_arg, wildcard,
 864                   ifp)
 865         struct inpcbinfo *pcbinfo;
 866         struct in_addr faddr, laddr;
 867         u_int fport_arg, lport_arg;
 868         int wildcard;
 869         struct ifnet *ifp;
 870 {
 871         struct inpcbhead *head;
 872         register struct inpcb *inp;
 873         u_short fport = fport_arg, lport = lport_arg;
 874
 875         /*
 876          * First look for an exact match.
 877          */
 878         head = &pcbinfo->hashbase[INP_PCBHASH(faddr.s_addr, lport, fport, pcbinfo->hashmask)];
 879         LIST_FOREACH(inp, head, inp_hash) {
 880 #ifdef INET6
 881                 if ((inp->inp_vflag & INP_IPV4) == 0)
 882                         continue;
 883 #endif
 884                 if (inp->inp_faddr.s_addr == faddr.s_addr &&
 885                     inp->inp_laddr.s_addr == laddr.s_addr &&
 886                     inp->inp_fport == fport &&
 887                     inp->inp_lport == lport) {
 888                         /*
 889                          * Found.
 890                          */
 891                         return (inp);
 892                 }
 893         }
 894         if (wildcard) {
 895                 struct inpcb *local_wild = NULL;
 896 #if defined(INET6)
 897                 struct inpcb *local_wild_mapped = NULL;
 898 #endif /* defined(INET6) */
 899
 900                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0, pcbinfo->hashmask)];
 901                 LIST_FOREACH(inp, head, inp_hash) {
 902 #ifdef INET6
 903                         if ((inp->inp_vflag & INP_IPV4) == 0)
 904                                 continue;
 905 #endif
 906                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
 907                             inp->inp_lport == lport) {
 908                                 if (ifp && ifp->if_type == IFT_FAITH &&
 909                                     (inp->inp_flags & INP_FAITH) == 0)
 910                                         continue;
 911                                 if (inp->inp_laddr.s_addr == laddr.s_addr)
 912                                         return (inp);
 913                                 else if (inp->inp_laddr.s_addr == INADDR_ANY) {
 914 #if defined(INET6)
 915                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
 916                                                              AF_INET6))
 917                                                 local_wild_mapped = inp;
 918                                         else
 919 #endif /* defined(INET6) */
 920                                         local_wild = inp;
 921                                 }
 922                         }
 923                 }
 924 #if defined(INET6)
 925                 if (local_wild == NULL)
 926                         return (local_wild_mapped);
 927 #endif /* defined(INET6) */
 928                 return (local_wild);
 929         }
 930
 931         /*
 932          * Not found.
 933          */
 934         return (NULL);
 935 }
 936
 937 /*
 938  * Insert PCB onto various hash lists.
 939  */
 940 int
 941 in_pcbinshash(inp)
 942         struct inpcb *inp;
 943 {
 944         struct inpcbhead *pcbhash;
 945         struct inpcbporthead *pcbporthash;
 946         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 947         struct inpcbport *phd;
 948         u_int32_t hashkey_faddr;
 949
 950 #ifdef INET6
 951         if (inp->inp_vflag & INP_IPV6)
 952                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
 953         else
 954 #endif /* INET6 */
 955         hashkey_faddr = inp->inp_faddr.s_addr;
 956
 957         pcbhash = &pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
 958                  inp->inp_lport, inp->inp_fport, pcbinfo->hashmask)];
 959
 960         pcbporthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(inp->inp_lport,
 961             pcbinfo->porthashmask)];
 962
 963         /*
 964          * Go through port list and look for a head for this lport.
 965          */
 966         LIST_FOREACH(phd, pcbporthash, phd_hash) {
 967                 if (phd->phd_port == inp->inp_lport)
 968                         break;
 969         }
 970         /*
 971          * If none exists, malloc one and tack it on.
 972          */
 973         if (phd == NULL) {
 974                 MALLOC(phd, struct inpcbport *, sizeof(struct inpcbport), M_PCB, M_NOWAIT);
 975                 if (phd == NULL) {
 976                         return (ENOBUFS); /* XXX */
 977                 }
 978                 phd->phd_port = inp->inp_lport;
 979                 LIST_INIT(&phd->phd_pcblist);
 980                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
 981         }
 982         inp->inp_phd = phd;
 983         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
 984         LIST_INSERT_HEAD(pcbhash, inp, inp_hash);
 985         return (0);
 986 }
 987
 988 /*
 989  * Move PCB to the proper hash bucket when { faddr, fport } have  been
 990  * changed. NOTE: This does not handle the case of the lport changing (the
 991  * hashed port list would have to be updated as well), so the lport must
 992  * not change after in_pcbinshash() has been called.
 993  */
 994 void
 995 in_pcbrehash(inp)
 996         struct inpcb *inp;
 997 {
 998         struct inpcbhead *head;
 999         u_int32_t hashkey_faddr;
1000
1001 #ifdef INET6
1002         if (inp->inp_vflag & INP_IPV6)
1003                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
1004         else
1005 #endif /* INET6 */
1006         hashkey_faddr = inp->inp_faddr.s_addr;
1007
1008         head = &inp->inp_pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
1009                 inp->inp_lport, inp->inp_fport, inp->inp_pcbinfo->hashmask)];
1010
1011         LIST_REMOVE(inp, inp_hash);
1012         LIST_INSERT_HEAD(head, inp, inp_hash);
1013 }
1014
1015 /*
1016  * Remove PCB from various lists.
1017  */
1018 void
1019 in_pcbremlists(inp)
1020         struct inpcb *inp;
1021 {
1022         inp->inp_gencnt = ++inp->inp_pcbinfo->ipi_gencnt;
1023         if (inp->inp_lport) {
1024                 struct inpcbport *phd = inp->inp_phd;
1025
1026                 LIST_REMOVE(inp, inp_hash);
1027                 LIST_REMOVE(inp, inp_portlist);
1028                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1029                         LIST_REMOVE(phd, phd_hash);
1030                         free(phd, M_PCB);
1031                 }
1032         }
1033         LIST_REMOVE(inp, inp_list);
1034         inp->inp_pcbinfo->ipi_count--;
1035 }
1036
1037 int
1038 prison_xinpcb(struct proc *p, struct inpcb *inp)
1039 {
1040         if (!p->p_prison)
1041                 return (0);
1042         if (ntohl(inp->inp_laddr.s_addr) == p->p_prison->pr_ip)
1043                 return (0);
1044         return (1);
1045 }