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  * 4. Neither the name of the University 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 REGENTS 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 REGENTS 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  *      @(#)in_pcb.c    8.4 (Berkeley) 5/24/95
  30  * $FreeBSD$
  31  */
  32
  33 #include "opt_ipsec.h"
  34 #include "opt_inet6.h"
  35 #include "opt_mac.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/mac.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mbuf.h>
  42 #include <sys/domain.h>
  43 #include <sys/protosw.h>
  44 #include <sys/socket.h>
  45 #include <sys/socketvar.h>
  46 #include <sys/proc.h>
  47 #include <sys/jail.h>
  48 #include <sys/kernel.h>
  49 #include <sys/sysctl.h>
  50
  51 #include <vm/uma.h>
  52
  53 #include <net/if.h>
  54 #include <net/if_types.h>
  55 #include <net/route.h>
  56
  57 #include <netinet/in.h>
  58 #include <netinet/in_pcb.h>
  59 #include <netinet/in_var.h>
  60 #include <netinet/ip_var.h>
  61 #include <netinet/tcp_var.h>
  62 #include <netinet/udp.h>
  63 #include <netinet/udp_var.h>
  64 #ifdef INET6
  65 #include <netinet/ip6.h>
  66 #include <netinet6/ip6_var.h>
  67 #endif /* INET6 */
  68
  69 #ifdef IPSEC
  70 #include <netinet6/ipsec.h>
  71 #include <netkey/key.h>
  72 #endif /* IPSEC */
  73
  74 #ifdef FAST_IPSEC
  75 #if defined(IPSEC) || defined(IPSEC_ESP)
  76 #error "Bad idea: don't compile with both IPSEC and FAST_IPSEC!"
  77 #endif
  78
  79 #include <netipsec/ipsec.h>
  80 #include <netipsec/key.h>
  81 #endif /* FAST_IPSEC */
  82
  83 /*
  84  * These configure the range of local port addresses assigned to
  85  * "unspecified" outgoing connections/packets/whatever.
  86  */
  87 int     ipport_lowfirstauto  = IPPORT_RESERVED - 1;     /* 1023 */
  88 int     ipport_lowlastauto = IPPORT_RESERVEDSTART;      /* 600 */
  89 int     ipport_firstauto = IPPORT_HIFIRSTAUTO;          /* 49152 */
  90 int     ipport_lastauto  = IPPORT_HILASTAUTO;           /* 65535 */
  91 int     ipport_hifirstauto = IPPORT_HIFIRSTAUTO;        /* 49152 */
  92 int     ipport_hilastauto  = IPPORT_HILASTAUTO;         /* 65535 */
  93
  94 /*
  95  * Reserved ports accessible only to root. There are significant
  96  * security considerations that must be accounted for when changing these,
  97  * but the security benefits can be great. Please be careful.
  98  */
  99 int     ipport_reservedhigh = IPPORT_RESERVED - 1;      /* 1023 */
 100 int     ipport_reservedlow = 0;
 101
 102 /* Variables dealing with random ephemeral port allocation. */
 103 int     ipport_randomized = 1;  /* user controlled via sysctl */
 104 int     ipport_randomcps = 10;  /* user controlled via sysctl */
 105 int     ipport_randomtime = 45; /* user controlled via sysctl */
 106 int     ipport_stoprandom = 0;  /* toggled by ipport_tick */
 107 int     ipport_tcpallocs;
 108 int     ipport_tcplastcount;
 109
 110 #define RANGECHK(var, min, max) \
 111         if ((var) < (min)) { (var) = (min); } \
 112         else if ((var) > (max)) { (var) = (max); }
 113
 114 static int
 115 sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
 116 {
 117         int error;
 118
 119         error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
 120         if (error == 0) {
 121                 RANGECHK(ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
 122                 RANGECHK(ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
 123                 RANGECHK(ipport_firstauto, IPPORT_RESERVED, IPPORT_MAX);
 124                 RANGECHK(ipport_lastauto, IPPORT_RESERVED, IPPORT_MAX);
 125                 RANGECHK(ipport_hifirstauto, IPPORT_RESERVED, IPPORT_MAX);
 126                 RANGECHK(ipport_hilastauto, IPPORT_RESERVED, IPPORT_MAX);
 127         }
 128         return (error);
 129 }
 130
 131 #undef RANGECHK
 132
 133 SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0, "IP Ports");
 134
 135 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLTYPE_INT|CTLFLAG_RW,
 136            &ipport_lowfirstauto, 0, &sysctl_net_ipport_check, "I", "");
 137 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLTYPE_INT|CTLFLAG_RW,
 138            &ipport_lowlastauto, 0, &sysctl_net_ipport_check, "I", "");
 139 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLTYPE_INT|CTLFLAG_RW,
 140            &ipport_firstauto, 0, &sysctl_net_ipport_check, "I", "");
 141 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLTYPE_INT|CTLFLAG_RW,
 142            &ipport_lastauto, 0, &sysctl_net_ipport_check, "I", "");
 143 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLTYPE_INT|CTLFLAG_RW,
 144            &ipport_hifirstauto, 0, &sysctl_net_ipport_check, "I", "");
 145 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLTYPE_INT|CTLFLAG_RW,
 146            &ipport_hilastauto, 0, &sysctl_net_ipport_check, "I", "");
 147 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedhigh,
 148            CTLFLAG_RW|CTLFLAG_SECURE, &ipport_reservedhigh, 0, "");
 149 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedlow,
 150            CTLFLAG_RW|CTLFLAG_SECURE, &ipport_reservedlow, 0, "");
 151 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomized, CTLFLAG_RW,
 152            &ipport_randomized, 0, "Enable random port allocation");
 153 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomcps, CTLFLAG_RW,
 154            &ipport_randomcps, 0, "Maximum number of random port "
 155            "allocations before switching to a sequental one");
 156 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomtime, CTLFLAG_RW,
 157            &ipport_randomtime, 0, "Minimum time to keep sequental port "
 158            "allocation before switching to a random one");
 159
 160 /*
 161  * in_pcb.c: manage the Protocol Control Blocks.
 162  *
 163  * NOTE: It is assumed that most of these functions will be called with
 164  * the pcbinfo lock held, and often, the inpcb lock held, as these utility
 165  * functions often modify hash chains or addresses in pcbs.
 166  */
 167
 168 /*
 169  * Allocate a PCB and associate it with the socket.
 170  */
 171 int
 172 in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo, const char *type)
 173 {
 174         struct inpcb *inp;
 175         int error;
 176
 177         INP_INFO_WLOCK_ASSERT(pcbinfo);
 178         error = 0;
 179         inp = uma_zalloc(pcbinfo->ipi_zone, M_NOWAIT | M_ZERO);
 180         if (inp == NULL)
 181                 return (ENOBUFS);
 182         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 183         inp->inp_pcbinfo = pcbinfo;
 184         inp->inp_socket = so;
 185 #ifdef MAC
 186         error = mac_init_inpcb(inp, M_NOWAIT);
 187         if (error != 0)
 188                 goto out;
 189         SOCK_LOCK(so);
 190         mac_create_inpcb_from_socket(so, inp);
 191         SOCK_UNLOCK(so);
 192 #endif
 193 #if defined(IPSEC) || defined(FAST_IPSEC)
 194 #ifdef FAST_IPSEC
 195         error = ipsec_init_policy(so, &inp->inp_sp);
 196 #else
 197         error = ipsec_init_pcbpolicy(so, &inp->inp_sp);
 198 #endif
 199         if (error != 0)
 200                 goto out;
 201 #endif /*IPSEC*/
 202 #if defined(INET6)
 203         if (INP_SOCKAF(so) == AF_INET6) {
 204                 inp->inp_vflag |= INP_IPV6PROTO;
 205                 if (ip6_v6only)
 206                         inp->inp_flags |= IN6P_IPV6_V6ONLY;
 207         }
 208 #endif
 209         LIST_INSERT_HEAD(pcbinfo->listhead, inp, inp_list);
 210         pcbinfo->ipi_count++;
 211         so->so_pcb = (caddr_t)inp;
 212         INP_LOCK_INIT(inp, "inp", type);
 213 #ifdef INET6
 214         if (ip6_auto_flowlabel)
 215                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 216 #endif
 217 #if defined(IPSEC) || defined(FAST_IPSEC) || defined(MAC)
 218 out:
 219         if (error != 0)
 220                 uma_zfree(pcbinfo->ipi_zone, inp);
 221 #endif
 222         return (error);
 223 }
 224
 225 int
 226 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 227 {
 228         int anonport, error;
 229
 230         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 231         INP_LOCK_ASSERT(inp);
 232
 233         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 234                 return (EINVAL);
 235         anonport = inp->inp_lport == 0 && (nam == NULL ||
 236             ((struct sockaddr_in *)nam)->sin_port == 0);
 237         error = in_pcbbind_setup(inp, nam, &inp->inp_laddr.s_addr,
 238             &inp->inp_lport, cred);
 239         if (error)
 240                 return (error);
 241         if (in_pcbinshash(inp) != 0) {
 242                 inp->inp_laddr.s_addr = INADDR_ANY;
 243                 inp->inp_lport = 0;
 244                 return (EAGAIN);
 245         }
 246         if (anonport)
 247                 inp->inp_flags |= INP_ANONPORT;
 248         return (0);
 249 }
 250
 251 /*
 252  * Set up a bind operation on a PCB, performing port allocation
 253  * as required, but do not actually modify the PCB. Callers can
 254  * either complete the bind by setting inp_laddr/inp_lport and
 255  * calling in_pcbinshash(), or they can just use the resulting
 256  * port and address to authorise the sending of a once-off packet.
 257  *
 258  * On error, the values of *laddrp and *lportp are not changed.
 259  */
 260 int
 261 in_pcbbind_setup(struct inpcb *inp, struct sockaddr *nam, in_addr_t *laddrp,
 262     u_short *lportp, struct ucred *cred)
 263 {
 264         struct socket *so = inp->inp_socket;
 265         unsigned short *lastport;
 266         struct sockaddr_in *sin;
 267         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 268         struct in_addr laddr;
 269         u_short lport = 0;
 270         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 271         int error, prison = 0;
 272         int dorandom;
 273
 274         INP_INFO_WLOCK_ASSERT(pcbinfo);
 275         INP_LOCK_ASSERT(inp);
 276
 277         if (TAILQ_EMPTY(&in_ifaddrhead)) /* XXX broken! */
 278                 return (EADDRNOTAVAIL);
 279         laddr.s_addr = *laddrp;
 280         if (nam != NULL && laddr.s_addr != INADDR_ANY)
 281                 return (EINVAL);
 282         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 283                 wild = 1;
 284         if (nam) {
 285                 sin = (struct sockaddr_in *)nam;
 286                 if (nam->sa_len != sizeof (*sin))
 287                         return (EINVAL);
 288 #ifdef notdef
 289                 /*
 290                  * We should check the family, but old programs
 291                  * incorrectly fail to initialize it.
 292                  */
 293                 if (sin->sin_family != AF_INET)
 294                         return (EAFNOSUPPORT);
 295 #endif
 296                 if (sin->sin_addr.s_addr != INADDR_ANY)
 297                         if (prison_ip(cred, 0, &sin->sin_addr.s_addr))
 298                                 return(EINVAL);
 299                 if (sin->sin_port != *lportp) {
 300                         /* Don't allow the port to change. */
 301                         if (*lportp != 0)
 302                                 return (EINVAL);
 303                         lport = sin->sin_port;
 304                 }
 305                 /* NB: lport is left as 0 if the port isn't being changed. */
 306                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 307                         /*
 308                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 309                          * allow complete duplication of binding if
 310                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 311                          * and a multicast address is bound on both
 312                          * new and duplicated sockets.
 313                          */
 314                         if (so->so_options & SO_REUSEADDR)
 315                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 316                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 317                         sin->sin_port = 0;              /* yech... */
 318                         bzero(&sin->sin_zero, sizeof(sin->sin_zero));
 319                         if (ifa_ifwithaddr((struct sockaddr *)sin) == 0)
 320                                 return (EADDRNOTAVAIL);
 321                 }
 322                 laddr = sin->sin_addr;
 323                 if (lport) {
 324                         struct inpcb *t;
 325                         struct tcptw *tw;
 326
 327                         /* GROSS */
 328                         if (ntohs(lport) <= ipport_reservedhigh &&
 329                             ntohs(lport) >= ipport_reservedlow &&
 330                             suser_cred(cred, SUSER_ALLOWJAIL))
 331                                 return (EACCES);
 332                         if (jailed(cred))
 333                                 prison = 1;
 334                         if (so->so_cred->cr_uid != 0 &&
 335                             !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 336                                 t = in_pcblookup_local(inp->inp_pcbinfo,
 337                                     sin->sin_addr, lport,
 338                                     prison ? 0 :  INPLOOKUP_WILDCARD);
 339         /*
 340          * XXX
 341          * This entire block sorely needs a rewrite.
 342          */
 343                                 if (t &&
 344                                     ((t->inp_vflag & INP_TIMEWAIT) == 0) &&
 345                                     (so->so_type != SOCK_STREAM ||
 346                                      ntohl(t->inp_faddr.s_addr) == INADDR_ANY) &&
 347                                     (ntohl(sin->sin_addr.s_addr) != INADDR_ANY ||
 348                                      ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 349                                      (t->inp_socket->so_options &
 350                                          SO_REUSEPORT) == 0) &&
 351                                     (so->so_cred->cr_uid !=
 352                                      t->inp_socket->so_cred->cr_uid))
 353                                         return (EADDRINUSE);
 354                         }
 355                         if (prison && prison_ip(cred, 0, &sin->sin_addr.s_addr))
 356                                 return (EADDRNOTAVAIL);
 357                         t = in_pcblookup_local(pcbinfo, sin->sin_addr,
 358                             lport, prison ? 0 : wild);
 359                         if (t && (t->inp_vflag & INP_TIMEWAIT)) {
 360                                 /*
 361                                  * XXXRW: If an incpb has had its timewait
 362                                  * state recycled, we treat the address as
 363                                  * being in use (for now).  This is better
 364                                  * than a panic, but not desirable.
 365                                  */
 366                                 tw = intotw(inp);
 367                                 if (tw == NULL ||
 368                                     (reuseport & tw->tw_so_options) == 0)
 369                                         return (EADDRINUSE);
 370                         } else if (t &&
 371                             (reuseport & t->inp_socket->so_options) == 0) {
 372 #if defined(INET6)
 373                                 if (ntohl(sin->sin_addr.s_addr) !=
 374                                     INADDR_ANY ||
 375                                     ntohl(t->inp_laddr.s_addr) !=
 376                                     INADDR_ANY ||
 377                                     INP_SOCKAF(so) ==
 378                                     INP_SOCKAF(t->inp_socket))
 379 #endif /* defined(INET6) */
 380                                 return (EADDRINUSE);
 381                         }
 382                 }
 383         }
 384         if (*lportp != 0)
 385                 lport = *lportp;
 386         if (lport == 0) {
 387                 u_short first, last;
 388                 int count;
 389
 390                 if (laddr.s_addr != INADDR_ANY)
 391                         if (prison_ip(cred, 0, &laddr.s_addr))
 392                                 return (EINVAL);
 393
 394                 if (inp->inp_flags & INP_HIGHPORT) {
 395                         first = ipport_hifirstauto;     /* sysctl */
 396                         last  = ipport_hilastauto;
 397                         lastport = &pcbinfo->lasthi;
 398                 } else if (inp->inp_flags & INP_LOWPORT) {
 399                         if ((error = suser_cred(cred, SUSER_ALLOWJAIL)) != 0)
 400                                 return error;
 401                         first = ipport_lowfirstauto;    /* 1023 */
 402                         last  = ipport_lowlastauto;     /* 600 */
 403                         lastport = &pcbinfo->lastlow;
 404                 } else {
 405                         first = ipport_firstauto;       /* sysctl */
 406                         last  = ipport_lastauto;
 407                         lastport = &pcbinfo->lastport;
 408                 }
 409                 /*
 410                  * For UDP, use random port allocation as long as the user
 411                  * allows it.  For TCP (and as of yet unknown) connections,
 412                  * use random port allocation only if the user allows it AND
 413                  * ipport_tick() allows it.
 414                  */
 415                 if (ipport_randomized &&
 416                         (!ipport_stoprandom || pcbinfo == &udbinfo))
 417                         dorandom = 1;
 418                 else
 419                         dorandom = 0;
 420                 /*
 421                  * It makes no sense to do random port allocation if
 422                  * we have the only port available.
 423                  */
 424                 if (first == last)
 425                         dorandom = 0;
 426                 /* Make sure to not include UDP packets in the count. */
 427                 if (pcbinfo != &udbinfo)
 428                         ipport_tcpallocs++;
 429                 /*
 430                  * Simple check to ensure all ports are not used up causing
 431                  * a deadlock here.
 432                  *
 433                  * We split the two cases (up and down) so that the direction
 434                  * is not being tested on each round of the loop.
 435                  */
 436                 if (first > last) {
 437                         /*
 438                          * counting down
 439                          */
 440                         if (dorandom)
 441                                 *lastport = first -
 442                                             (arc4random() % (first - last));
 443                         count = first - last;
 444
 445                         do {
 446                                 if (count-- < 0)        /* completely used? */
 447                                         return (EADDRNOTAVAIL);
 448                                 --*lastport;
 449                                 if (*lastport > first || *lastport < last)
 450                                         *lastport = first;
 451                                 lport = htons(*lastport);
 452                         } while (in_pcblookup_local(pcbinfo, laddr, lport,
 453                             wild));
 454                 } else {
 455                         /*
 456                          * counting up
 457                          */
 458                         if (dorandom)
 459                                 *lastport = first +
 460                                             (arc4random() % (last - first));
 461                         count = last - first;
 462
 463                         do {
 464                                 if (count-- < 0)        /* completely used? */
 465                                         return (EADDRNOTAVAIL);
 466                                 ++*lastport;
 467                                 if (*lastport < first || *lastport > last)
 468                                         *lastport = first;
 469                                 lport = htons(*lastport);
 470                         } while (in_pcblookup_local(pcbinfo, laddr, lport,
 471                             wild));
 472                 }
 473         }
 474         if (prison_ip(cred, 0, &laddr.s_addr))
 475                 return (EINVAL);
 476         *laddrp = laddr.s_addr;
 477         *lportp = lport;
 478         return (0);
 479 }
 480
 481 /*
 482  * Connect from a socket to a specified address.
 483  * Both address and port must be specified in argument sin.
 484  * If don't have a local address for this socket yet,
 485  * then pick one.
 486  */
 487 int
 488 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 489 {
 490         u_short lport, fport;
 491         in_addr_t laddr, faddr;
 492         int anonport, error;
 493
 494         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 495         INP_LOCK_ASSERT(inp);
 496
 497         lport = inp->inp_lport;
 498         laddr = inp->inp_laddr.s_addr;
 499         anonport = (lport == 0);
 500         error = in_pcbconnect_setup(inp, nam, &laddr, &lport, &faddr, &fport,
 501             NULL, cred);
 502         if (error)
 503                 return (error);
 504
 505         /* Do the initial binding of the local address if required. */
 506         if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) {
 507                 inp->inp_lport = lport;
 508                 inp->inp_laddr.s_addr = laddr;
 509                 if (in_pcbinshash(inp) != 0) {
 510                         inp->inp_laddr.s_addr = INADDR_ANY;
 511                         inp->inp_lport = 0;
 512                         return (EAGAIN);
 513                 }
 514         }
 515
 516         /* Commit the remaining changes. */
 517         inp->inp_lport = lport;
 518         inp->inp_laddr.s_addr = laddr;
 519         inp->inp_faddr.s_addr = faddr;
 520         inp->inp_fport = fport;
 521         in_pcbrehash(inp);
 522 #ifdef IPSEC
 523         if (inp->inp_socket->so_type == SOCK_STREAM)
 524                 ipsec_pcbconn(inp->inp_sp);
 525 #endif
 526         if (anonport)
 527                 inp->inp_flags |= INP_ANONPORT;
 528         return (0);
 529 }
 530
 531 /*
 532  * Set up for a connect from a socket to the specified address.
 533  * On entry, *laddrp and *lportp should contain the current local
 534  * address and port for the PCB; these are updated to the values
 535  * that should be placed in inp_laddr and inp_lport to complete
 536  * the connect.
 537  *
 538  * On success, *faddrp and *fportp will be set to the remote address
 539  * and port. These are not updated in the error case.
 540  *
 541  * If the operation fails because the connection already exists,
 542  * *oinpp will be set to the PCB of that connection so that the
 543  * caller can decide to override it. In all other cases, *oinpp
 544  * is set to NULL.
 545  */
 546 int
 547 in_pcbconnect_setup(struct inpcb *inp, struct sockaddr *nam,
 548     in_addr_t *laddrp, u_short *lportp, in_addr_t *faddrp, u_short *fportp,
 549     struct inpcb **oinpp, struct ucred *cred)
 550 {
 551         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 552         struct in_ifaddr *ia;
 553         struct sockaddr_in sa;
 554         struct ucred *socred;
 555         struct inpcb *oinp;
 556         struct in_addr laddr, faddr;
 557         u_short lport, fport;
 558         int error;
 559
 560         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 561         INP_LOCK_ASSERT(inp);
 562
 563         if (oinpp != NULL)
 564                 *oinpp = NULL;
 565         if (nam->sa_len != sizeof (*sin))
 566                 return (EINVAL);
 567         if (sin->sin_family != AF_INET)
 568                 return (EAFNOSUPPORT);
 569         if (sin->sin_port == 0)
 570                 return (EADDRNOTAVAIL);
 571         laddr.s_addr = *laddrp;
 572         lport = *lportp;
 573         faddr = sin->sin_addr;
 574         fport = sin->sin_port;
 575         socred = inp->inp_socket->so_cred;
 576         if (laddr.s_addr == INADDR_ANY && jailed(socred)) {
 577                 bzero(&sa, sizeof(sa));
 578                 sa.sin_addr.s_addr = htonl(prison_getip(socred));
 579                 sa.sin_len = sizeof(sa);
 580                 sa.sin_family = AF_INET;
 581                 error = in_pcbbind_setup(inp, (struct sockaddr *)&sa,
 582                     &laddr.s_addr, &lport, cred);
 583                 if (error)
 584                         return (error);
 585         }
 586         if (!TAILQ_EMPTY(&in_ifaddrhead)) {
 587                 /*
 588                  * If the destination address is INADDR_ANY,
 589                  * use the primary local address.
 590                  * If the supplied address is INADDR_BROADCAST,
 591                  * and the primary interface supports broadcast,
 592                  * choose the broadcast address for that interface.
 593                  */
 594                 if (faddr.s_addr == INADDR_ANY)
 595                         faddr = IA_SIN(TAILQ_FIRST(&in_ifaddrhead))->sin_addr;
 596                 else if (faddr.s_addr == (u_long)INADDR_BROADCAST &&
 597                     (TAILQ_FIRST(&in_ifaddrhead)->ia_ifp->if_flags &
 598                     IFF_BROADCAST))
 599                         faddr = satosin(&TAILQ_FIRST(
 600                             &in_ifaddrhead)->ia_broadaddr)->sin_addr;
 601         }
 602         if (laddr.s_addr == INADDR_ANY) {
 603                 ia = (struct in_ifaddr *)0;
 604                 /*
 605                  * If route is known our src addr is taken from the i/f,
 606                  * else punt.
 607                  *
 608                  * Find out route to destination
 609                  */
 610                 if ((inp->inp_socket->so_options & SO_DONTROUTE) == 0)
 611                         ia = ip_rtaddr(faddr);
 612                 /*
 613                  * If we found a route, use the address corresponding to
 614                  * the outgoing interface.
 615                  *
 616                  * Otherwise assume faddr is reachable on a directly connected
 617                  * network and try to find a corresponding interface to take
 618                  * the source address from.
 619                  */
 620                 if (ia == 0) {
 621                         bzero(&sa, sizeof(sa));
 622                         sa.sin_addr = faddr;
 623                         sa.sin_len = sizeof(sa);
 624                         sa.sin_family = AF_INET;
 625
 626                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(&sa)));
 627                         if (ia == 0)
 628                                 ia = ifatoia(ifa_ifwithnet(sintosa(&sa)));
 629                         if (ia == 0)
 630                                 return (ENETUNREACH);
 631                 }
 632                 /*
 633                  * If the destination address is multicast and an outgoing
 634                  * interface has been set as a multicast option, use the
 635                  * address of that interface as our source address.
 636                  */
 637                 if (IN_MULTICAST(ntohl(faddr.s_addr)) &&
 638                     inp->inp_moptions != NULL) {
 639                         struct ip_moptions *imo;
 640                         struct ifnet *ifp;
 641
 642                         imo = inp->inp_moptions;
 643                         if (imo->imo_multicast_ifp != NULL) {
 644                                 ifp = imo->imo_multicast_ifp;
 645                                 TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link)
 646                                         if (ia->ia_ifp == ifp)
 647                                                 break;
 648                                 if (ia == 0)
 649                                         return (EADDRNOTAVAIL);
 650                         }
 651                 }
 652                 laddr = ia->ia_addr.sin_addr;
 653         }
 654
 655         oinp = in_pcblookup_hash(inp->inp_pcbinfo, faddr, fport, laddr, lport,
 656             0, NULL);
 657         if (oinp != NULL) {
 658                 if (oinpp != NULL)
 659                         *oinpp = oinp;
 660                 return (EADDRINUSE);
 661         }
 662         if (lport == 0) {
 663                 error = in_pcbbind_setup(inp, NULL, &laddr.s_addr, &lport,
 664                     cred);
 665                 if (error)
 666                         return (error);
 667         }
 668         *laddrp = laddr.s_addr;
 669         *lportp = lport;
 670         *faddrp = faddr.s_addr;
 671         *fportp = fport;
 672         return (0);
 673 }
 674
 675 void
 676 in_pcbdisconnect(struct inpcb *inp)
 677 {
 678
 679         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 680         INP_LOCK_ASSERT(inp);
 681
 682         inp->inp_faddr.s_addr = INADDR_ANY;
 683         inp->inp_fport = 0;
 684         in_pcbrehash(inp);
 685 #ifdef IPSEC
 686         ipsec_pcbdisconn(inp->inp_sp);
 687 #endif
 688 }
 689
 690 /*
 691  * In the old world order, in_pcbdetach() served two functions: to detach the
 692  * pcb from the socket/potentially free the socket, and to free the pcb
 693  * itself.  In the new world order, the protocol code is responsible for
 694  * managing the relationship with the socket, and this code simply frees the
 695  * pcb.
 696  */
 697 void
 698 in_pcbdetach(struct inpcb *inp)
 699 {
 700
 701         KASSERT(inp->inp_socket != NULL, ("in_pcbdetach: inp_socket == NULL"));
 702         inp->inp_socket->so_pcb = NULL;
 703         inp->inp_socket = NULL;
 704 }
 705
 706 void
 707 in_pcbfree(struct inpcb *inp)
 708 {
 709         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 710
 711         KASSERT(inp->inp_socket == NULL, ("in_pcbfree: inp_socket != NULL"));
 712         INP_INFO_WLOCK_ASSERT(ipi);
 713         INP_LOCK_ASSERT(inp);
 714
 715 #if defined(IPSEC) || defined(FAST_IPSEC)
 716         ipsec4_delete_pcbpolicy(inp);
 717 #endif /*IPSEC*/
 718         inp->inp_gencnt = ++ipi->ipi_gencnt;
 719         in_pcbremlists(inp);
 720         if (inp->inp_options)
 721                 (void)m_free(inp->inp_options);
 722         ip_freemoptions(inp->inp_moptions);
 723         inp->inp_vflag = 0;
 724         INP_LOCK_DESTROY(inp);
 725 #ifdef MAC
 726         mac_destroy_inpcb(inp);
 727 #endif
 728         uma_zfree(ipi->ipi_zone, inp);
 729 }
 730
 731 /*
 732  * TCP needs to maintain its inpcb structure after the TCP connection has
 733  * been torn down.  However, it must be disconnected from the inpcb hashes as
 734  * it must not prevent binding of future connections to the same port/ip
 735  * combination by other inpcbs.
 736  */
 737 void
 738 in_pcbdrop(struct inpcb *inp)
 739 {
 740         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 741
 742         INP_INFO_WLOCK_ASSERT(pcbinfo);
 743         INP_LOCK_ASSERT(inp);
 744
 745         inp->inp_vflag |= INP_DROPPED;
 746         if (inp->inp_lport) {
 747                 struct inpcbport *phd = inp->inp_phd;
 748
 749                 LIST_REMOVE(inp, inp_hash);
 750                 LIST_REMOVE(inp, inp_portlist);
 751                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
 752                         LIST_REMOVE(phd, phd_hash);
 753                         free(phd, M_PCB);
 754                 }
 755                 inp->inp_lport = 0;
 756         }
 757 }
 758
 759 struct sockaddr *
 760 in_sockaddr(in_port_t port, struct in_addr *addr_p)
 761 {
 762         struct sockaddr_in *sin;
 763
 764         MALLOC(sin, struct sockaddr_in *, sizeof *sin, M_SONAME,
 765                 M_WAITOK | M_ZERO);
 766         sin->sin_family = AF_INET;
 767         sin->sin_len = sizeof(*sin);
 768         sin->sin_addr = *addr_p;
 769         sin->sin_port = port;
 770
 771         return (struct sockaddr *)sin;
 772 }
 773
 774 /*
 775  * The wrapper function will pass down the pcbinfo for this function to lock.
 776  * The socket must have a valid
 777  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 778  * except through a kernel programming error, so it is acceptable to panic
 779  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 780  * because there actually /is/ a programming error somewhere... XXX)
 781  */
 782 int
 783 in_setsockaddr(struct socket *so, struct sockaddr **nam,
 784     struct inpcbinfo *pcbinfo)
 785 {
 786         struct inpcb *inp;
 787         struct in_addr addr;
 788         in_port_t port;
 789
 790         inp = sotoinpcb(so);
 791         KASSERT(inp != NULL, ("in_setsockaddr: inp == NULL"));
 792
 793         INP_LOCK(inp);
 794         port = inp->inp_lport;
 795         addr = inp->inp_laddr;
 796         INP_UNLOCK(inp);
 797
 798         *nam = in_sockaddr(port, &addr);
 799         return 0;
 800 }
 801
 802 /*
 803  * The wrapper function will pass down the pcbinfo for this function to lock.
 804  */
 805 int
 806 in_setpeeraddr(struct socket *so, struct sockaddr **nam,
 807     struct inpcbinfo *pcbinfo)
 808 {
 809         struct inpcb *inp;
 810         struct in_addr addr;
 811         in_port_t port;
 812
 813         inp = sotoinpcb(so);
 814         KASSERT(inp != NULL, ("in_setpeeraddr: inp == NULL"));
 815
 816         INP_LOCK(inp);
 817         port = inp->inp_fport;
 818         addr = inp->inp_faddr;
 819         INP_UNLOCK(inp);
 820
 821         *nam = in_sockaddr(port, &addr);
 822         return 0;
 823 }
 824
 825 void
 826 in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr faddr, int errno,
 827     struct inpcb *(*notify)(struct inpcb *, int))
 828 {
 829         struct inpcb *inp, *ninp;
 830         struct inpcbhead *head;
 831
 832         INP_INFO_WLOCK(pcbinfo);
 833         head = pcbinfo->listhead;
 834         for (inp = LIST_FIRST(head); inp != NULL; inp = ninp) {
 835                 INP_LOCK(inp);
 836                 ninp = LIST_NEXT(inp, inp_list);
 837 #ifdef INET6
 838                 if ((inp->inp_vflag & INP_IPV4) == 0) {
 839                         INP_UNLOCK(inp);
 840                         continue;
 841                 }
 842 #endif
 843                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
 844                     inp->inp_socket == NULL) {
 845                         INP_UNLOCK(inp);
 846                         continue;
 847                 }
 848                 if ((*notify)(inp, errno))
 849                         INP_UNLOCK(inp);
 850         }
 851         INP_INFO_WUNLOCK(pcbinfo);
 852 }
 853
 854 void
 855 in_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 856 {
 857         struct inpcb *inp;
 858         struct ip_moptions *imo;
 859         int i, gap;
 860
 861         INP_INFO_RLOCK(pcbinfo);
 862         LIST_FOREACH(inp, pcbinfo->listhead, inp_list) {
 863                 INP_LOCK(inp);
 864                 imo = inp->inp_moptions;
 865                 if ((inp->inp_vflag & INP_IPV4) &&
 866                     imo != NULL) {
 867                         /*
 868                          * Unselect the outgoing interface if it is being
 869                          * detached.
 870                          */
 871                         if (imo->imo_multicast_ifp == ifp)
 872                                 imo->imo_multicast_ifp = NULL;
 873
 874                         /*
 875                          * Drop multicast group membership if we joined
 876                          * through the interface being detached.
 877                          */
 878                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
 879                             i++) {
 880                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
 881                                         in_delmulti(imo->imo_membership[i]);
 882                                         gap++;
 883                                 } else if (gap != 0)
 884                                         imo->imo_membership[i - gap] =
 885                                             imo->imo_membership[i];
 886                         }
 887                         imo->imo_num_memberships -= gap;
 888                 }
 889                 INP_UNLOCK(inp);
 890         }
 891         INP_INFO_RUNLOCK(pcbinfo);
 892 }
 893
 894 /*
 895  * Lookup a PCB based on the local address and port.
 896  */
 897 #define INP_LOOKUP_MAPPED_PCB_COST      3
 898 struct inpcb *
 899 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
 900     u_int lport_arg, int wild_okay)
 901 {
 902         struct inpcb *inp;
 903         struct tcptw *tw;
 904 #ifdef INET6
 905         int matchwild = 3 + INP_LOOKUP_MAPPED_PCB_COST;
 906 #else
 907         int matchwild = 3;
 908 #endif
 909         int wildcard;
 910         u_short lport = lport_arg;
 911
 912         INP_INFO_WLOCK_ASSERT(pcbinfo);
 913
 914         if (!wild_okay) {
 915                 struct inpcbhead *head;
 916                 /*
 917                  * Look for an unconnected (wildcard foreign addr) PCB that
 918                  * matches the local address and port we're looking for.
 919                  */
 920                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0, pcbinfo->hashmask)];
 921                 LIST_FOREACH(inp, head, inp_hash) {
 922 #ifdef INET6
 923                         if ((inp->inp_vflag & INP_IPV4) == 0)
 924                                 continue;
 925 #endif
 926                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
 927                             inp->inp_laddr.s_addr == laddr.s_addr &&
 928                             inp->inp_lport == lport) {
 929                                 /*
 930                                  * Found.
 931                                  */
 932                                 return (inp);
 933                         }
 934                 }
 935                 /*
 936                  * Not found.
 937                  */
 938                 return (NULL);
 939         } else {
 940                 struct inpcbporthead *porthash;
 941                 struct inpcbport *phd;
 942                 struct inpcb *match = NULL;
 943                 /*
 944                  * Best fit PCB lookup.
 945                  *
 946                  * First see if this local port is in use by looking on the
 947                  * port hash list.
 948                  */
 949                 retrylookup:
 950                 porthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(lport,
 951                     pcbinfo->porthashmask)];
 952                 LIST_FOREACH(phd, porthash, phd_hash) {
 953                         if (phd->phd_port == lport)
 954                                 break;
 955                 }
 956                 if (phd != NULL) {
 957                         /*
 958                          * Port is in use by one or more PCBs. Look for best
 959                          * fit.
 960                          */
 961                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 962                                 wildcard = 0;
 963 #ifdef INET6
 964                                 if ((inp->inp_vflag & INP_IPV4) == 0)
 965                                         continue;
 966                                 /*
 967                                  * We never select the PCB that has
 968                                  * INP_IPV6 flag and is bound to :: if
 969                                  * we have another PCB which is bound
 970                                  * to 0.0.0.0.  If a PCB has the
 971                                  * INP_IPV6 flag, then we set its cost
 972                                  * higher than IPv4 only PCBs.
 973                                  *
 974                                  * Note that the case only happens
 975                                  * when a socket is bound to ::, under
 976                                  * the condition that the use of the
 977                                  * mapped address is allowed.
 978                                  */
 979                                 if ((inp->inp_vflag & INP_IPV6) != 0)
 980                                         wildcard += INP_LOOKUP_MAPPED_PCB_COST;
 981 #endif
 982                                 /*
 983                                  * Clean out old time_wait sockets if they
 984                                  * are clogging up needed local ports.
 985                                  */
 986                                 if ((inp->inp_vflag & INP_TIMEWAIT) != 0) {
 987                                         tw = intotw(inp);
 988                                         if (tw != NULL &&
 989                                             tcp_twrecycleable(tw)) {
 990                                                 INP_LOCK(inp);
 991                                                 tcp_twclose(tw, 0);
 992                                                 match = NULL;
 993                                                 goto retrylookup;
 994                                         }
 995                                 }
 996                                 if (inp->inp_faddr.s_addr != INADDR_ANY)
 997                                         wildcard++;
 998                                 if (inp->inp_laddr.s_addr != INADDR_ANY) {
 999                                         if (laddr.s_addr == INADDR_ANY)
1000                                                 wildcard++;
1001                                         else if (inp->inp_laddr.s_addr != laddr.s_addr)
1002                                                 continue;
1003                                 } else {
1004                                         if (laddr.s_addr != INADDR_ANY)
1005                                                 wildcard++;
1006                                 }
1007                                 if (wildcard < matchwild) {
1008                                         match = inp;
1009                                         matchwild = wildcard;
1010                                         if (matchwild == 0) {
1011                                                 break;
1012                                         }
1013                                 }
1014                         }
1015                 }
1016                 return (match);
1017         }
1018 }
1019 #undef INP_LOOKUP_MAPPED_PCB_COST
1020
1021 /*
1022  * Lookup PCB in hash list.
1023  */
1024 struct inpcb *
1025 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1026     u_int fport_arg, struct in_addr laddr, u_int lport_arg, int wildcard,
1027     struct ifnet *ifp)
1028 {
1029         struct inpcbhead *head;
1030         struct inpcb *inp;
1031         u_short fport = fport_arg, lport = lport_arg;
1032
1033         INP_INFO_RLOCK_ASSERT(pcbinfo);
1034
1035         /*
1036          * First look for an exact match.
1037          */
1038         head = &pcbinfo->hashbase[INP_PCBHASH(faddr.s_addr, lport, fport, pcbinfo->hashmask)];
1039         LIST_FOREACH(inp, head, inp_hash) {
1040 #ifdef INET6
1041                 if ((inp->inp_vflag & INP_IPV4) == 0)
1042                         continue;
1043 #endif
1044                 if (inp->inp_faddr.s_addr == faddr.s_addr &&
1045                     inp->inp_laddr.s_addr == laddr.s_addr &&
1046                     inp->inp_fport == fport &&
1047                     inp->inp_lport == lport) {
1048                         /*
1049                          * Found.
1050                          */
1051                         return (inp);
1052                 }
1053         }
1054         if (wildcard) {
1055                 struct inpcb *local_wild = NULL;
1056 #if defined(INET6)
1057                 struct inpcb *local_wild_mapped = NULL;
1058 #endif /* defined(INET6) */
1059
1060                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0, pcbinfo->hashmask)];
1061                 LIST_FOREACH(inp, head, inp_hash) {
1062 #ifdef INET6
1063                         if ((inp->inp_vflag & INP_IPV4) == 0)
1064                                 continue;
1065 #endif
1066                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
1067                             inp->inp_lport == lport) {
1068                                 if (ifp && ifp->if_type == IFT_FAITH &&
1069                                     (inp->inp_flags & INP_FAITH) == 0)
1070                                         continue;
1071                                 if (inp->inp_laddr.s_addr == laddr.s_addr)
1072                                         return (inp);
1073                                 else if (inp->inp_laddr.s_addr == INADDR_ANY) {
1074 #if defined(INET6)
1075                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1076                                                              AF_INET6))
1077                                                 local_wild_mapped = inp;
1078                                         else
1079 #endif /* defined(INET6) */
1080                                         local_wild = inp;
1081                                 }
1082                         }
1083                 }
1084 #if defined(INET6)
1085                 if (local_wild == NULL)
1086                         return (local_wild_mapped);
1087 #endif /* defined(INET6) */
1088                 return (local_wild);
1089         }
1090
1091         /*
1092          * Not found.
1093          */
1094         return (NULL);
1095 }
1096
1097 /*
1098  * Insert PCB onto various hash lists.
1099  */
1100 int
1101 in_pcbinshash(struct inpcb *inp)
1102 {
1103         struct inpcbhead *pcbhash;
1104         struct inpcbporthead *pcbporthash;
1105         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1106         struct inpcbport *phd;
1107         u_int32_t hashkey_faddr;
1108
1109         INP_INFO_WLOCK_ASSERT(pcbinfo);
1110         INP_LOCK_ASSERT(inp);
1111
1112 #ifdef INET6
1113         if (inp->inp_vflag & INP_IPV6)
1114                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
1115         else
1116 #endif /* INET6 */
1117         hashkey_faddr = inp->inp_faddr.s_addr;
1118
1119         pcbhash = &pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
1120                  inp->inp_lport, inp->inp_fport, pcbinfo->hashmask)];
1121
1122         pcbporthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(inp->inp_lport,
1123             pcbinfo->porthashmask)];
1124
1125         /*
1126          * Go through port list and look for a head for this lport.
1127          */
1128         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1129                 if (phd->phd_port == inp->inp_lport)
1130                         break;
1131         }
1132         /*
1133          * If none exists, malloc one and tack it on.
1134          */
1135         if (phd == NULL) {
1136                 MALLOC(phd, struct inpcbport *, sizeof(struct inpcbport), M_PCB, M_NOWAIT);
1137                 if (phd == NULL) {
1138                         return (ENOBUFS); /* XXX */
1139                 }
1140                 phd->phd_port = inp->inp_lport;
1141                 LIST_INIT(&phd->phd_pcblist);
1142                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1143         }
1144         inp->inp_phd = phd;
1145         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1146         LIST_INSERT_HEAD(pcbhash, inp, inp_hash);
1147         return (0);
1148 }
1149
1150 /*
1151  * Move PCB to the proper hash bucket when { faddr, fport } have  been
1152  * changed. NOTE: This does not handle the case of the lport changing (the
1153  * hashed port list would have to be updated as well), so the lport must
1154  * not change after in_pcbinshash() has been called.
1155  */
1156 void
1157 in_pcbrehash(struct inpcb *inp)
1158 {
1159         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1160         struct inpcbhead *head;
1161         u_int32_t hashkey_faddr;
1162
1163         INP_INFO_WLOCK_ASSERT(pcbinfo);
1164         INP_LOCK_ASSERT(inp);
1165
1166 #ifdef INET6
1167         if (inp->inp_vflag & INP_IPV6)
1168                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
1169         else
1170 #endif /* INET6 */
1171         hashkey_faddr = inp->inp_faddr.s_addr;
1172
1173         head = &pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
1174                 inp->inp_lport, inp->inp_fport, pcbinfo->hashmask)];
1175
1176         LIST_REMOVE(inp, inp_hash);
1177         LIST_INSERT_HEAD(head, inp, inp_hash);
1178 }
1179
1180 /*
1181  * Remove PCB from various lists.
1182  */
1183 void
1184 in_pcbremlists(struct inpcb *inp)
1185 {
1186         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1187
1188         INP_INFO_WLOCK_ASSERT(pcbinfo);
1189         INP_LOCK_ASSERT(inp);
1190
1191         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
1192         if (inp->inp_lport) {
1193                 struct inpcbport *phd = inp->inp_phd;
1194
1195                 LIST_REMOVE(inp, inp_hash);
1196                 LIST_REMOVE(inp, inp_portlist);
1197                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1198                         LIST_REMOVE(phd, phd_hash);
1199                         free(phd, M_PCB);
1200                 }
1201         }
1202         LIST_REMOVE(inp, inp_list);
1203         pcbinfo->ipi_count--;
1204 }
1205
1206 /*
1207  * A set label operation has occurred at the socket layer, propagate the
1208  * label change into the in_pcb for the socket.
1209  */
1210 void
1211 in_pcbsosetlabel(struct socket *so)
1212 {
1213 #ifdef MAC
1214         struct inpcb *inp;
1215
1216         inp = sotoinpcb(so);
1217         KASSERT(inp != NULL, ("in_pcbsosetlabel: so->so_pcb == NULL"));
1218
1219         INP_LOCK(inp);
1220         SOCK_LOCK(so);
1221         mac_inpcb_sosetlabel(so, inp);
1222         SOCK_UNLOCK(so);
1223         INP_UNLOCK(inp);
1224 #endif
1225 }
1226
1227 /*
1228  * ipport_tick runs once per second, determining if random port
1229  * allocation should be continued.  If more than ipport_randomcps
1230  * ports have been allocated in the last second, then we return to
1231  * sequential port allocation. We return to random allocation only
1232  * once we drop below ipport_randomcps for at least ipport_randomtime
1233  * seconds.
1234  */
1235
1236 void
1237 ipport_tick(void *xtp)
1238 {
1239         if (ipport_tcpallocs > ipport_tcplastcount + ipport_randomcps) {
1240                 ipport_stoprandom = ipport_randomtime;
1241         } else {
1242                 if (ipport_stoprandom > 0)
1243                         ipport_stoprandom--;
1244         }
1245         ipport_tcplastcount = ipport_tcpallocs;
1246         callout_reset(&ipport_tick_callout, hz, ipport_tick, NULL);
1247 }