sys/netinet/in_pcb.c

   1 /*-
   2  * Copyright (c) 1982, 1986, 1991, 1993, 1995
   3  *      The Regents of the University of California.
   4  * Copyright (c) 2007 Robert N. M. Watson
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 4. Neither the name of the University nor the names of its contributors
  16  *    may be used to endorse or promote products derived from this software
  17  *    without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      @(#)in_pcb.c    8.4 (Berkeley) 5/24/95
  32  * $FreeBSD$
  33  */
  34
  35 #include "opt_ddb.h"
  36 #include "opt_ipsec.h"
  37 #include "opt_inet6.h"
  38 #include "opt_mac.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/priv.h>
  49 #include <sys/proc.h>
  50 #include <sys/jail.h>
  51 #include <sys/kernel.h>
  52 #include <sys/sysctl.h>
  53
  54 #ifdef DDB
  55 #include <ddb/ddb.h>
  56 #endif
  57
  58 #include <vm/uma.h>
  59
  60 #include <net/if.h>
  61 #include <net/if_types.h>
  62 #include <net/route.h>
  63
  64 #include <netinet/in.h>
  65 #include <netinet/in_pcb.h>
  66 #include <netinet/in_var.h>
  67 #include <netinet/ip_var.h>
  68 #include <netinet/tcp_var.h>
  69 #include <netinet/udp.h>
  70 #include <netinet/udp_var.h>
  71 #ifdef INET6
  72 #include <netinet/ip6.h>
  73 #include <netinet6/ip6_var.h>
  74 #endif /* INET6 */
  75
  76 #ifdef IPSEC
  77 #include <netinet6/ipsec.h>
  78 #include <netkey/key.h>
  79 #endif /* IPSEC */
  80
  81 #ifdef FAST_IPSEC
  82 #if defined(IPSEC) || defined(IPSEC_ESP)
  83 #error "Bad idea: don't compile with both IPSEC and FAST_IPSEC!"
  84 #endif
  85
  86 #include <netipsec/ipsec.h>
  87 #include <netipsec/key.h>
  88 #endif /* FAST_IPSEC */
  89
  90 #include <security/mac/mac_framework.h>
  91
  92 /*
  93  * These configure the range of local port addresses assigned to
  94  * "unspecified" outgoing connections/packets/whatever.
  95  */
  96 int     ipport_lowfirstauto  = IPPORT_RESERVED - 1;     /* 1023 */
  97 int     ipport_lowlastauto = IPPORT_RESERVEDSTART;      /* 600 */
  98 int     ipport_firstauto = IPPORT_HIFIRSTAUTO;          /* 49152 */
  99 int     ipport_lastauto  = IPPORT_HILASTAUTO;           /* 65535 */
 100 int     ipport_hifirstauto = IPPORT_HIFIRSTAUTO;        /* 49152 */
 101 int     ipport_hilastauto  = IPPORT_HILASTAUTO;         /* 65535 */
 102
 103 /*
 104  * Reserved ports accessible only to root. There are significant
 105  * security considerations that must be accounted for when changing these,
 106  * but the security benefits can be great. Please be careful.
 107  */
 108 int     ipport_reservedhigh = IPPORT_RESERVED - 1;      /* 1023 */
 109 int     ipport_reservedlow = 0;
 110
 111 /* Variables dealing with random ephemeral port allocation. */
 112 int     ipport_randomized = 1;  /* user controlled via sysctl */
 113 int     ipport_randomcps = 10;  /* user controlled via sysctl */
 114 int     ipport_randomtime = 45; /* user controlled via sysctl */
 115 int     ipport_stoprandom = 0;  /* toggled by ipport_tick */
 116 int     ipport_tcpallocs;
 117 int     ipport_tcplastcount;
 118
 119 #define RANGECHK(var, min, max) \
 120         if ((var) < (min)) { (var) = (min); } \
 121         else if ((var) > (max)) { (var) = (max); }
 122
 123 static int
 124 sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
 125 {
 126         int error;
 127
 128         error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
 129         if (error == 0) {
 130                 RANGECHK(ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
 131                 RANGECHK(ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
 132                 RANGECHK(ipport_firstauto, IPPORT_RESERVED, IPPORT_MAX);
 133                 RANGECHK(ipport_lastauto, IPPORT_RESERVED, IPPORT_MAX);
 134                 RANGECHK(ipport_hifirstauto, IPPORT_RESERVED, IPPORT_MAX);
 135                 RANGECHK(ipport_hilastauto, IPPORT_RESERVED, IPPORT_MAX);
 136         }
 137         return (error);
 138 }
 139
 140 #undef RANGECHK
 141
 142 SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0, "IP Ports");
 143
 144 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLTYPE_INT|CTLFLAG_RW,
 145            &ipport_lowfirstauto, 0, &sysctl_net_ipport_check, "I", "");
 146 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLTYPE_INT|CTLFLAG_RW,
 147            &ipport_lowlastauto, 0, &sysctl_net_ipport_check, "I", "");
 148 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLTYPE_INT|CTLFLAG_RW,
 149            &ipport_firstauto, 0, &sysctl_net_ipport_check, "I", "");
 150 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLTYPE_INT|CTLFLAG_RW,
 151            &ipport_lastauto, 0, &sysctl_net_ipport_check, "I", "");
 152 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLTYPE_INT|CTLFLAG_RW,
 153            &ipport_hifirstauto, 0, &sysctl_net_ipport_check, "I", "");
 154 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLTYPE_INT|CTLFLAG_RW,
 155            &ipport_hilastauto, 0, &sysctl_net_ipport_check, "I", "");
 156 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedhigh,
 157            CTLFLAG_RW|CTLFLAG_SECURE, &ipport_reservedhigh, 0, "");
 158 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedlow,
 159            CTLFLAG_RW|CTLFLAG_SECURE, &ipport_reservedlow, 0, "");
 160 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomized, CTLFLAG_RW,
 161            &ipport_randomized, 0, "Enable random port allocation");
 162 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomcps, CTLFLAG_RW,
 163            &ipport_randomcps, 0, "Maximum number of random port "
 164            "allocations before switching to a sequental one");
 165 SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomtime, CTLFLAG_RW,
 166            &ipport_randomtime, 0, "Minimum time to keep sequental port "
 167            "allocation before switching to a random one");
 168
 169 /*
 170  * in_pcb.c: manage the Protocol Control Blocks.
 171  *
 172  * NOTE: It is assumed that most of these functions will be called with
 173  * the pcbinfo lock held, and often, the inpcb lock held, as these utility
 174  * functions often modify hash chains or addresses in pcbs.
 175  */
 176
 177 /*
 178  * Allocate a PCB and associate it with the socket.
 179  * On success return with the PCB locked.
 180  */
 181 int
 182 in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo)
 183 {
 184         struct inpcb *inp;
 185         int error;
 186
 187         INP_INFO_WLOCK_ASSERT(pcbinfo);
 188         error = 0;
 189         inp = uma_zalloc(pcbinfo->ipi_zone, M_NOWAIT);
 190         if (inp == NULL)
 191                 return (ENOBUFS);
 192         bzero(inp, inp_zero_size);
 193         inp->inp_pcbinfo = pcbinfo;
 194         inp->inp_socket = so;
 195 #ifdef MAC
 196         error = mac_init_inpcb(inp, M_NOWAIT);
 197         if (error != 0)
 198                 goto out;
 199         SOCK_LOCK(so);
 200         mac_create_inpcb_from_socket(so, inp);
 201         SOCK_UNLOCK(so);
 202 #endif
 203 #if defined(IPSEC) || defined(FAST_IPSEC)
 204 #ifdef FAST_IPSEC
 205         error = ipsec_init_policy(so, &inp->inp_sp);
 206 #else
 207         error = ipsec_init_pcbpolicy(so, &inp->inp_sp);
 208 #endif
 209         if (error != 0)
 210                 goto out;
 211 #endif /*IPSEC*/
 212 #ifdef INET6
 213         if (INP_SOCKAF(so) == AF_INET6) {
 214                 inp->inp_vflag |= INP_IPV6PROTO;
 215                 if (ip6_v6only)
 216                         inp->inp_flags |= IN6P_IPV6_V6ONLY;
 217         }
 218 #endif
 219         LIST_INSERT_HEAD(pcbinfo->listhead, inp, inp_list);
 220         pcbinfo->ipi_count++;
 221         so->so_pcb = (caddr_t)inp;
 222 #ifdef INET6
 223         if (ip6_auto_flowlabel)
 224                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 225 #endif
 226         INP_LOCK(inp);
 227         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 228
 229 #if defined(IPSEC) || defined(FAST_IPSEC) || defined(MAC)
 230 out:
 231         if (error != 0)
 232                 uma_zfree(pcbinfo->ipi_zone, inp);
 233 #endif
 234         return (error);
 235 }
 236
 237 int
 238 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 239 {
 240         int anonport, error;
 241
 242         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 243         INP_LOCK_ASSERT(inp);
 244
 245         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 246                 return (EINVAL);
 247         anonport = inp->inp_lport == 0 && (nam == NULL ||
 248             ((struct sockaddr_in *)nam)->sin_port == 0);
 249         error = in_pcbbind_setup(inp, nam, &inp->inp_laddr.s_addr,
 250             &inp->inp_lport, cred);
 251         if (error)
 252                 return (error);
 253         if (in_pcbinshash(inp) != 0) {
 254                 inp->inp_laddr.s_addr = INADDR_ANY;
 255                 inp->inp_lport = 0;
 256                 return (EAGAIN);
 257         }
 258         if (anonport)
 259                 inp->inp_flags |= INP_ANONPORT;
 260         return (0);
 261 }
 262
 263 /*
 264  * Set up a bind operation on a PCB, performing port allocation
 265  * as required, but do not actually modify the PCB. Callers can
 266  * either complete the bind by setting inp_laddr/inp_lport and
 267  * calling in_pcbinshash(), or they can just use the resulting
 268  * port and address to authorise the sending of a once-off packet.
 269  *
 270  * On error, the values of *laddrp and *lportp are not changed.
 271  */
 272 int
 273 in_pcbbind_setup(struct inpcb *inp, struct sockaddr *nam, in_addr_t *laddrp,
 274     u_short *lportp, struct ucred *cred)
 275 {
 276         struct socket *so = inp->inp_socket;
 277         unsigned short *lastport;
 278         struct sockaddr_in *sin;
 279         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 280         struct in_addr laddr;
 281         u_short lport = 0;
 282         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 283         int error, prison = 0;
 284         int dorandom;
 285
 286         INP_INFO_WLOCK_ASSERT(pcbinfo);
 287         INP_LOCK_ASSERT(inp);
 288
 289         if (TAILQ_EMPTY(&in_ifaddrhead)) /* XXX broken! */
 290                 return (EADDRNOTAVAIL);
 291         laddr.s_addr = *laddrp;
 292         if (nam != NULL && laddr.s_addr != INADDR_ANY)
 293                 return (EINVAL);
 294         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) == 0)
 295                 wild = INPLOOKUP_WILDCARD;
 296         if (nam) {
 297                 sin = (struct sockaddr_in *)nam;
 298                 if (nam->sa_len != sizeof (*sin))
 299                         return (EINVAL);
 300 #ifdef notdef
 301                 /*
 302                  * We should check the family, but old programs
 303                  * incorrectly fail to initialize it.
 304                  */
 305                 if (sin->sin_family != AF_INET)
 306                         return (EAFNOSUPPORT);
 307 #endif
 308                 if (sin->sin_addr.s_addr != INADDR_ANY)
 309                         if (prison_ip(cred, 0, &sin->sin_addr.s_addr))
 310                                 return(EINVAL);
 311                 if (sin->sin_port != *lportp) {
 312                         /* Don't allow the port to change. */
 313                         if (*lportp != 0)
 314                                 return (EINVAL);
 315                         lport = sin->sin_port;
 316                 }
 317                 /* NB: lport is left as 0 if the port isn't being changed. */
 318                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 319                         /*
 320                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 321                          * allow complete duplication of binding if
 322                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 323                          * and a multicast address is bound on both
 324                          * new and duplicated sockets.
 325                          */
 326                         if (so->so_options & SO_REUSEADDR)
 327                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
 328                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 329                         sin->sin_port = 0;              /* yech... */
 330                         bzero(&sin->sin_zero, sizeof(sin->sin_zero));
 331                         if (ifa_ifwithaddr((struct sockaddr *)sin) == 0)
 332                                 return (EADDRNOTAVAIL);
 333                 }
 334                 laddr = sin->sin_addr;
 335                 if (lport) {
 336                         struct inpcb *t;
 337                         struct tcptw *tw;
 338
 339                         /* GROSS */
 340                         if (ntohs(lport) <= ipport_reservedhigh &&
 341                             ntohs(lport) >= ipport_reservedlow &&
 342                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT,
 343                             SUSER_ALLOWJAIL))
 344                                 return (EACCES);
 345                         if (jailed(cred))
 346                                 prison = 1;
 347                         if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 348                             suser_cred(so->so_cred, SUSER_ALLOWJAIL) != 0) {
 349                                 t = in_pcblookup_local(inp->inp_pcbinfo,
 350                                     sin->sin_addr, lport,
 351                                     prison ? 0 :  INPLOOKUP_WILDCARD);
 352         /*
 353          * XXX
 354          * This entire block sorely needs a rewrite.
 355          */
 356                                 if (t &&
 357                                     ((t->inp_vflag & INP_TIMEWAIT) == 0) &&
 358                                     (so->so_type != SOCK_STREAM ||
 359                                      ntohl(t->inp_faddr.s_addr) == INADDR_ANY) &&
 360                                     (ntohl(sin->sin_addr.s_addr) != INADDR_ANY ||
 361                                      ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
 362                                      (t->inp_socket->so_options &
 363                                          SO_REUSEPORT) == 0) &&
 364                                     (so->so_cred->cr_uid !=
 365                                      t->inp_socket->so_cred->cr_uid))
 366                                         return (EADDRINUSE);
 367                         }
 368                         if (prison && prison_ip(cred, 0, &sin->sin_addr.s_addr))
 369                                 return (EADDRNOTAVAIL);
 370                         t = in_pcblookup_local(pcbinfo, sin->sin_addr,
 371                             lport, prison ? 0 : wild);
 372                         if (t && (t->inp_vflag & INP_TIMEWAIT)) {
 373                                 /*
 374                                  * XXXRW: If an incpb has had its timewait
 375                                  * state recycled, we treat the address as
 376                                  * being in use (for now).  This is better
 377                                  * than a panic, but not desirable.
 378                                  */
 379                                 tw = intotw(inp);
 380                                 if (tw == NULL ||
 381                                     (reuseport & tw->tw_so_options) == 0)
 382                                         return (EADDRINUSE);
 383                         } else if (t &&
 384                             (reuseport & t->inp_socket->so_options) == 0) {
 385 #ifdef INET6
 386                                 if (ntohl(sin->sin_addr.s_addr) !=
 387                                     INADDR_ANY ||
 388                                     ntohl(t->inp_laddr.s_addr) !=
 389                                     INADDR_ANY ||
 390                                     INP_SOCKAF(so) ==
 391                                     INP_SOCKAF(t->inp_socket))
 392 #endif
 393                                 return (EADDRINUSE);
 394                         }
 395                 }
 396         }
 397         if (*lportp != 0)
 398                 lport = *lportp;
 399         if (lport == 0) {
 400                 u_short first, last;
 401                 int count;
 402
 403                 if (laddr.s_addr != INADDR_ANY)
 404                         if (prison_ip(cred, 0, &laddr.s_addr))
 405                                 return (EINVAL);
 406
 407                 if (inp->inp_flags & INP_HIGHPORT) {
 408                         first = ipport_hifirstauto;     /* sysctl */
 409                         last  = ipport_hilastauto;
 410                         lastport = &pcbinfo->lasthi;
 411                 } else if (inp->inp_flags & INP_LOWPORT) {
 412                         error = priv_check_cred(cred,
 413                             PRIV_NETINET_RESERVEDPORT, SUSER_ALLOWJAIL);
 414                         if (error)
 415                                 return error;
 416                         first = ipport_lowfirstauto;    /* 1023 */
 417                         last  = ipport_lowlastauto;     /* 600 */
 418                         lastport = &pcbinfo->lastlow;
 419                 } else {
 420                         first = ipport_firstauto;       /* sysctl */
 421                         last  = ipport_lastauto;
 422                         lastport = &pcbinfo->lastport;
 423                 }
 424                 /*
 425                  * For UDP, use random port allocation as long as the user
 426                  * allows it.  For TCP (and as of yet unknown) connections,
 427                  * use random port allocation only if the user allows it AND
 428                  * ipport_tick() allows it.
 429                  */
 430                 if (ipport_randomized &&
 431                         (!ipport_stoprandom || pcbinfo == &udbinfo))
 432                         dorandom = 1;
 433                 else
 434                         dorandom = 0;
 435                 /*
 436                  * It makes no sense to do random port allocation if
 437                  * we have the only port available.
 438                  */
 439                 if (first == last)
 440                         dorandom = 0;
 441                 /* Make sure to not include UDP packets in the count. */
 442                 if (pcbinfo != &udbinfo)
 443                         ipport_tcpallocs++;
 444                 /*
 445                  * Simple check to ensure all ports are not used up causing
 446                  * a deadlock here.
 447                  *
 448                  * We split the two cases (up and down) so that the direction
 449                  * is not being tested on each round of the loop.
 450                  */
 451                 if (first > last) {
 452                         /*
 453                          * counting down
 454                          */
 455                         if (dorandom)
 456                                 *lastport = first -
 457                                             (arc4random() % (first - last));
 458                         count = first - last;
 459
 460                         do {
 461                                 if (count-- < 0)        /* completely used? */
 462                                         return (EADDRNOTAVAIL);
 463                                 --*lastport;
 464                                 if (*lastport > first || *lastport < last)
 465                                         *lastport = first;
 466                                 lport = htons(*lastport);
 467                         } while (in_pcblookup_local(pcbinfo, laddr, lport,
 468                             wild));
 469                 } else {
 470                         /*
 471                          * counting up
 472                          */
 473                         if (dorandom)
 474                                 *lastport = first +
 475                                             (arc4random() % (last - first));
 476                         count = last - first;
 477
 478                         do {
 479                                 if (count-- < 0)        /* completely used? */
 480                                         return (EADDRNOTAVAIL);
 481                                 ++*lastport;
 482                                 if (*lastport < first || *lastport > last)
 483                                         *lastport = first;
 484                                 lport = htons(*lastport);
 485                         } while (in_pcblookup_local(pcbinfo, laddr, lport,
 486                             wild));
 487                 }
 488         }
 489         if (prison_ip(cred, 0, &laddr.s_addr))
 490                 return (EINVAL);
 491         *laddrp = laddr.s_addr;
 492         *lportp = lport;
 493         return (0);
 494 }
 495
 496 /*
 497  * Connect from a socket to a specified address.
 498  * Both address and port must be specified in argument sin.
 499  * If don't have a local address for this socket yet,
 500  * then pick one.
 501  */
 502 int
 503 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
 504 {
 505         u_short lport, fport;
 506         in_addr_t laddr, faddr;
 507         int anonport, error;
 508
 509         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 510         INP_LOCK_ASSERT(inp);
 511
 512         lport = inp->inp_lport;
 513         laddr = inp->inp_laddr.s_addr;
 514         anonport = (lport == 0);
 515         error = in_pcbconnect_setup(inp, nam, &laddr, &lport, &faddr, &fport,
 516             NULL, cred);
 517         if (error)
 518                 return (error);
 519
 520         /* Do the initial binding of the local address if required. */
 521         if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) {
 522                 inp->inp_lport = lport;
 523                 inp->inp_laddr.s_addr = laddr;
 524                 if (in_pcbinshash(inp) != 0) {
 525                         inp->inp_laddr.s_addr = INADDR_ANY;
 526                         inp->inp_lport = 0;
 527                         return (EAGAIN);
 528                 }
 529         }
 530
 531         /* Commit the remaining changes. */
 532         inp->inp_lport = lport;
 533         inp->inp_laddr.s_addr = laddr;
 534         inp->inp_faddr.s_addr = faddr;
 535         inp->inp_fport = fport;
 536         in_pcbrehash(inp);
 537 #ifdef IPSEC
 538         if (inp->inp_socket->so_type == SOCK_STREAM)
 539                 ipsec_pcbconn(inp->inp_sp);
 540 #endif
 541         if (anonport)
 542                 inp->inp_flags |= INP_ANONPORT;
 543         return (0);
 544 }
 545
 546 /*
 547  * Set up for a connect from a socket to the specified address.
 548  * On entry, *laddrp and *lportp should contain the current local
 549  * address and port for the PCB; these are updated to the values
 550  * that should be placed in inp_laddr and inp_lport to complete
 551  * the connect.
 552  *
 553  * On success, *faddrp and *fportp will be set to the remote address
 554  * and port. These are not updated in the error case.
 555  *
 556  * If the operation fails because the connection already exists,
 557  * *oinpp will be set to the PCB of that connection so that the
 558  * caller can decide to override it. In all other cases, *oinpp
 559  * is set to NULL.
 560  */
 561 int
 562 in_pcbconnect_setup(struct inpcb *inp, struct sockaddr *nam,
 563     in_addr_t *laddrp, u_short *lportp, in_addr_t *faddrp, u_short *fportp,
 564     struct inpcb **oinpp, struct ucred *cred)
 565 {
 566         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 567         struct in_ifaddr *ia;
 568         struct sockaddr_in sa;
 569         struct ucred *socred;
 570         struct inpcb *oinp;
 571         struct in_addr laddr, faddr;
 572         u_short lport, fport;
 573         int error;
 574
 575         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 576         INP_LOCK_ASSERT(inp);
 577
 578         if (oinpp != NULL)
 579                 *oinpp = NULL;
 580         if (nam->sa_len != sizeof (*sin))
 581                 return (EINVAL);
 582         if (sin->sin_family != AF_INET)
 583                 return (EAFNOSUPPORT);
 584         if (sin->sin_port == 0)
 585                 return (EADDRNOTAVAIL);
 586         laddr.s_addr = *laddrp;
 587         lport = *lportp;
 588         faddr = sin->sin_addr;
 589         fport = sin->sin_port;
 590         socred = inp->inp_socket->so_cred;
 591         if (laddr.s_addr == INADDR_ANY && jailed(socred)) {
 592                 bzero(&sa, sizeof(sa));
 593                 sa.sin_addr.s_addr = htonl(prison_getip(socred));
 594                 sa.sin_len = sizeof(sa);
 595                 sa.sin_family = AF_INET;
 596                 error = in_pcbbind_setup(inp, (struct sockaddr *)&sa,
 597                     &laddr.s_addr, &lport, cred);
 598                 if (error)
 599                         return (error);
 600         }
 601         if (!TAILQ_EMPTY(&in_ifaddrhead)) {
 602                 /*
 603                  * If the destination address is INADDR_ANY,
 604                  * use the primary local address.
 605                  * If the supplied address is INADDR_BROADCAST,
 606                  * and the primary interface supports broadcast,
 607                  * choose the broadcast address for that interface.
 608                  */
 609                 if (faddr.s_addr == INADDR_ANY)
 610                         faddr = IA_SIN(TAILQ_FIRST(&in_ifaddrhead))->sin_addr;
 611                 else if (faddr.s_addr == (u_long)INADDR_BROADCAST &&
 612                     (TAILQ_FIRST(&in_ifaddrhead)->ia_ifp->if_flags &
 613                     IFF_BROADCAST))
 614                         faddr = satosin(&TAILQ_FIRST(
 615                             &in_ifaddrhead)->ia_broadaddr)->sin_addr;
 616         }
 617         if (laddr.s_addr == INADDR_ANY) {
 618                 ia = (struct in_ifaddr *)0;
 619                 /*
 620                  * If route is known our src addr is taken from the i/f,
 621                  * else punt.
 622                  *
 623                  * Find out route to destination
 624                  */
 625                 if ((inp->inp_socket->so_options & SO_DONTROUTE) == 0)
 626                         ia = ip_rtaddr(faddr);
 627                 /*
 628                  * If we found a route, use the address corresponding to
 629                  * the outgoing interface.
 630                  *
 631                  * Otherwise assume faddr is reachable on a directly connected
 632                  * network and try to find a corresponding interface to take
 633                  * the source address from.
 634                  */
 635                 if (ia == 0) {
 636                         bzero(&sa, sizeof(sa));
 637                         sa.sin_addr = faddr;
 638                         sa.sin_len = sizeof(sa);
 639                         sa.sin_family = AF_INET;
 640
 641                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(&sa)));
 642                         if (ia == 0)
 643                                 ia = ifatoia(ifa_ifwithnet(sintosa(&sa)));
 644                         if (ia == 0)
 645                                 return (ENETUNREACH);
 646                 }
 647                 /*
 648                  * If the destination address is multicast and an outgoing
 649                  * interface has been set as a multicast option, use the
 650                  * address of that interface as our source address.
 651                  */
 652                 if (IN_MULTICAST(ntohl(faddr.s_addr)) &&
 653                     inp->inp_moptions != NULL) {
 654                         struct ip_moptions *imo;
 655                         struct ifnet *ifp;
 656
 657                         imo = inp->inp_moptions;
 658                         if (imo->imo_multicast_ifp != NULL) {
 659                                 ifp = imo->imo_multicast_ifp;
 660                                 TAILQ_FOREACH(ia, &in_ifaddrhead, ia_link)
 661                                         if (ia->ia_ifp == ifp)
 662                                                 break;
 663                                 if (ia == 0)
 664                                         return (EADDRNOTAVAIL);
 665                         }
 666                 }
 667                 laddr = ia->ia_addr.sin_addr;
 668         }
 669
 670         oinp = in_pcblookup_hash(inp->inp_pcbinfo, faddr, fport, laddr, lport,
 671             0, NULL);
 672         if (oinp != NULL) {
 673                 if (oinpp != NULL)
 674                         *oinpp = oinp;
 675                 return (EADDRINUSE);
 676         }
 677         if (lport == 0) {
 678                 error = in_pcbbind_setup(inp, NULL, &laddr.s_addr, &lport,
 679                     cred);
 680                 if (error)
 681                         return (error);
 682         }
 683         *laddrp = laddr.s_addr;
 684         *lportp = lport;
 685         *faddrp = faddr.s_addr;
 686         *fportp = fport;
 687         return (0);
 688 }
 689
 690 void
 691 in_pcbdisconnect(struct inpcb *inp)
 692 {
 693
 694         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 695         INP_LOCK_ASSERT(inp);
 696
 697         inp->inp_faddr.s_addr = INADDR_ANY;
 698         inp->inp_fport = 0;
 699         in_pcbrehash(inp);
 700 #ifdef IPSEC
 701         ipsec_pcbdisconn(inp->inp_sp);
 702 #endif
 703 }
 704
 705 /*
 706  * In the old world order, in_pcbdetach() served two functions: to detach the
 707  * pcb from the socket/potentially free the socket, and to free the pcb
 708  * itself.  In the new world order, the protocol code is responsible for
 709  * managing the relationship with the socket, and this code simply frees the
 710  * pcb.
 711  */
 712 void
 713 in_pcbdetach(struct inpcb *inp)
 714 {
 715
 716         KASSERT(inp->inp_socket != NULL, ("in_pcbdetach: inp_socket == NULL"));
 717         inp->inp_socket->so_pcb = NULL;
 718         inp->inp_socket = NULL;
 719 }
 720
 721 void
 722 in_pcbfree(struct inpcb *inp)
 723 {
 724         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 725
 726         KASSERT(inp->inp_socket == NULL, ("in_pcbfree: inp_socket != NULL"));
 727         INP_INFO_WLOCK_ASSERT(ipi);
 728         INP_LOCK_ASSERT(inp);
 729
 730 #if defined(IPSEC) || defined(FAST_IPSEC)
 731         ipsec4_delete_pcbpolicy(inp);
 732 #endif /*IPSEC*/
 733         inp->inp_gencnt = ++ipi->ipi_gencnt;
 734         in_pcbremlists(inp);
 735         if (inp->inp_options)
 736                 (void)m_free(inp->inp_options);
 737         ip_freemoptions(inp->inp_moptions);
 738         inp->inp_vflag = 0;
 739
 740 #ifdef MAC
 741         mac_destroy_inpcb(inp);
 742 #endif
 743         INP_UNLOCK(inp);
 744         uma_zfree(ipi->ipi_zone, inp);
 745 }
 746
 747 /*
 748  * TCP needs to maintain its inpcb structure after the TCP connection has
 749  * been torn down.  However, it must be disconnected from the inpcb hashes as
 750  * it must not prevent binding of future connections to the same port/ip
 751  * combination by other inpcbs.
 752  */
 753 void
 754 in_pcbdrop(struct inpcb *inp)
 755 {
 756
 757         INP_INFO_WLOCK_ASSERT(inp->inp_pcbinfo);
 758         INP_LOCK_ASSERT(inp);
 759
 760         inp->inp_vflag |= INP_DROPPED;
 761         if (inp->inp_lport) {
 762                 struct inpcbport *phd = inp->inp_phd;
 763
 764                 LIST_REMOVE(inp, inp_hash);
 765                 LIST_REMOVE(inp, inp_portlist);
 766                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
 767                         LIST_REMOVE(phd, phd_hash);
 768                         free(phd, M_PCB);
 769                 }
 770                 inp->inp_lport = 0;
 771         }
 772 }
 773
 774 struct sockaddr *
 775 in_sockaddr(in_port_t port, struct in_addr *addr_p)
 776 {
 777         struct sockaddr_in *sin;
 778
 779         MALLOC(sin, struct sockaddr_in *, sizeof *sin, M_SONAME,
 780                 M_WAITOK | M_ZERO);
 781         sin->sin_family = AF_INET;
 782         sin->sin_len = sizeof(*sin);
 783         sin->sin_addr = *addr_p;
 784         sin->sin_port = port;
 785
 786         return (struct sockaddr *)sin;
 787 }
 788
 789 /*
 790  * The wrapper function will pass down the pcbinfo for this function to lock.
 791  * The socket must have a valid
 792  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 793  * except through a kernel programming error, so it is acceptable to panic
 794  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 795  * because there actually /is/ a programming error somewhere... XXX)
 796  */
 797 int
 798 in_setsockaddr(struct socket *so, struct sockaddr **nam,
 799     struct inpcbinfo *pcbinfo)
 800 {
 801         struct inpcb *inp;
 802         struct in_addr addr;
 803         in_port_t port;
 804
 805         inp = sotoinpcb(so);
 806         KASSERT(inp != NULL, ("in_setsockaddr: inp == NULL"));
 807
 808         INP_LOCK(inp);
 809         port = inp->inp_lport;
 810         addr = inp->inp_laddr;
 811         INP_UNLOCK(inp);
 812
 813         *nam = in_sockaddr(port, &addr);
 814         return 0;
 815 }
 816
 817 /*
 818  * The wrapper function will pass down the pcbinfo for this function to lock.
 819  */
 820 int
 821 in_setpeeraddr(struct socket *so, struct sockaddr **nam,
 822     struct inpcbinfo *pcbinfo)
 823 {
 824         struct inpcb *inp;
 825         struct in_addr addr;
 826         in_port_t port;
 827
 828         inp = sotoinpcb(so);
 829         KASSERT(inp != NULL, ("in_setpeeraddr: inp == NULL"));
 830
 831         INP_LOCK(inp);
 832         port = inp->inp_fport;
 833         addr = inp->inp_faddr;
 834         INP_UNLOCK(inp);
 835
 836         *nam = in_sockaddr(port, &addr);
 837         return 0;
 838 }
 839
 840 void
 841 in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr faddr, int errno,
 842     struct inpcb *(*notify)(struct inpcb *, int))
 843 {
 844         struct inpcb *inp, *ninp;
 845         struct inpcbhead *head;
 846
 847         INP_INFO_WLOCK(pcbinfo);
 848         head = pcbinfo->listhead;
 849         for (inp = LIST_FIRST(head); inp != NULL; inp = ninp) {
 850                 INP_LOCK(inp);
 851                 ninp = LIST_NEXT(inp, inp_list);
 852 #ifdef INET6
 853                 if ((inp->inp_vflag & INP_IPV4) == 0) {
 854                         INP_UNLOCK(inp);
 855                         continue;
 856                 }
 857 #endif
 858                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
 859                     inp->inp_socket == NULL) {
 860                         INP_UNLOCK(inp);
 861                         continue;
 862                 }
 863                 if ((*notify)(inp, errno))
 864                         INP_UNLOCK(inp);
 865         }
 866         INP_INFO_WUNLOCK(pcbinfo);
 867 }
 868
 869 void
 870 in_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
 871 {
 872         struct inpcb *inp;
 873         struct ip_moptions *imo;
 874         int i, gap;
 875
 876         INP_INFO_RLOCK(pcbinfo);
 877         LIST_FOREACH(inp, pcbinfo->listhead, inp_list) {
 878                 INP_LOCK(inp);
 879                 imo = inp->inp_moptions;
 880                 if ((inp->inp_vflag & INP_IPV4) &&
 881                     imo != NULL) {
 882                         /*
 883                          * Unselect the outgoing interface if it is being
 884                          * detached.
 885                          */
 886                         if (imo->imo_multicast_ifp == ifp)
 887                                 imo->imo_multicast_ifp = NULL;
 888
 889                         /*
 890                          * Drop multicast group membership if we joined
 891                          * through the interface being detached.
 892                          */
 893                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
 894                             i++) {
 895                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
 896                                         in_delmulti(imo->imo_membership[i]);
 897                                         gap++;
 898                                 } else if (gap != 0)
 899                                         imo->imo_membership[i - gap] =
 900                                             imo->imo_membership[i];
 901                         }
 902                         imo->imo_num_memberships -= gap;
 903                 }
 904                 INP_UNLOCK(inp);
 905         }
 906         INP_INFO_RUNLOCK(pcbinfo);
 907 }
 908
 909 /*
 910  * Lookup a PCB based on the local address and port.
 911  */
 912 #define INP_LOOKUP_MAPPED_PCB_COST      3
 913 struct inpcb *
 914 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
 915     u_int lport_arg, int wild_okay)
 916 {
 917         struct inpcb *inp;
 918 #ifdef INET6
 919         int matchwild = 3 + INP_LOOKUP_MAPPED_PCB_COST;
 920 #else
 921         int matchwild = 3;
 922 #endif
 923         int wildcard;
 924         u_short lport = lport_arg;
 925
 926         INP_INFO_WLOCK_ASSERT(pcbinfo);
 927
 928         if (!wild_okay) {
 929                 struct inpcbhead *head;
 930                 /*
 931                  * Look for an unconnected (wildcard foreign addr) PCB that
 932                  * matches the local address and port we're looking for.
 933                  */
 934                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0, pcbinfo->hashmask)];
 935                 LIST_FOREACH(inp, head, inp_hash) {
 936 #ifdef INET6
 937                         if ((inp->inp_vflag & INP_IPV4) == 0)
 938                                 continue;
 939 #endif
 940                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
 941                             inp->inp_laddr.s_addr == laddr.s_addr &&
 942                             inp->inp_lport == lport) {
 943                                 /*
 944                                  * Found.
 945                                  */
 946                                 return (inp);
 947                         }
 948                 }
 949                 /*
 950                  * Not found.
 951                  */
 952                 return (NULL);
 953         } else {
 954                 struct inpcbporthead *porthash;
 955                 struct inpcbport *phd;
 956                 struct inpcb *match = NULL;
 957                 /*
 958                  * Best fit PCB lookup.
 959                  *
 960                  * First see if this local port is in use by looking on the
 961                  * port hash list.
 962                  */
 963                 porthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(lport,
 964                     pcbinfo->porthashmask)];
 965                 LIST_FOREACH(phd, porthash, phd_hash) {
 966                         if (phd->phd_port == lport)
 967                                 break;
 968                 }
 969                 if (phd != NULL) {
 970                         /*
 971                          * Port is in use by one or more PCBs. Look for best
 972                          * fit.
 973                          */
 974                         LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 975                                 wildcard = 0;
 976 #ifdef INET6
 977                                 if ((inp->inp_vflag & INP_IPV4) == 0)
 978                                         continue;
 979                                 /*
 980                                  * We never select the PCB that has
 981                                  * INP_IPV6 flag and is bound to :: if
 982                                  * we have another PCB which is bound
 983                                  * to 0.0.0.0.  If a PCB has the
 984                                  * INP_IPV6 flag, then we set its cost
 985                                  * higher than IPv4 only PCBs.
 986                                  *
 987                                  * Note that the case only happens
 988                                  * when a socket is bound to ::, under
 989                                  * the condition that the use of the
 990                                  * mapped address is allowed.
 991                                  */
 992                                 if ((inp->inp_vflag & INP_IPV6) != 0)
 993                                         wildcard += INP_LOOKUP_MAPPED_PCB_COST;
 994 #endif
 995                                 if (inp->inp_faddr.s_addr != INADDR_ANY)
 996                                         wildcard++;
 997                                 if (inp->inp_laddr.s_addr != INADDR_ANY) {
 998                                         if (laddr.s_addr == INADDR_ANY)
 999                                                 wildcard++;
1000                                         else if (inp->inp_laddr.s_addr != laddr.s_addr)
1001                                                 continue;
1002                                 } else {
1003                                         if (laddr.s_addr != INADDR_ANY)
1004                                                 wildcard++;
1005                                 }
1006                                 if (wildcard < matchwild) {
1007                                         match = inp;
1008                                         matchwild = wildcard;
1009                                         if (matchwild == 0) {
1010                                                 break;
1011                                         }
1012                                 }
1013                         }
1014                 }
1015                 return (match);
1016         }
1017 }
1018 #undef INP_LOOKUP_MAPPED_PCB_COST
1019
1020 /*
1021  * Lookup PCB in hash list.
1022  */
1023 struct inpcb *
1024 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1025     u_int fport_arg, struct in_addr laddr, u_int lport_arg, int wildcard,
1026     struct ifnet *ifp)
1027 {
1028         struct inpcbhead *head;
1029         struct inpcb *inp;
1030         u_short fport = fport_arg, lport = lport_arg;
1031
1032         INP_INFO_RLOCK_ASSERT(pcbinfo);
1033
1034         /*
1035          * First look for an exact match.
1036          */
1037         head = &pcbinfo->hashbase[INP_PCBHASH(faddr.s_addr, lport, fport,
1038             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                         return (inp);
1049         }
1050
1051         /*
1052          * Then look for a wildcard match, if requested.
1053          */
1054         if (wildcard) {
1055                 struct inpcb *local_wild = NULL;
1056 #ifdef INET6
1057                 struct inpcb *local_wild_mapped = NULL;
1058 #endif
1059
1060                 head = &pcbinfo->hashbase[INP_PCBHASH(INADDR_ANY, lport, 0,
1061                     pcbinfo->hashmask)];
1062                 LIST_FOREACH(inp, head, inp_hash) {
1063 #ifdef INET6
1064                         if ((inp->inp_vflag & INP_IPV4) == 0)
1065                                 continue;
1066 #endif
1067                         if (inp->inp_faddr.s_addr == INADDR_ANY &&
1068                             inp->inp_lport == lport) {
1069                                 if (ifp && ifp->if_type == IFT_FAITH &&
1070                                     (inp->inp_flags & INP_FAITH) == 0)
1071                                         continue;
1072                                 if (inp->inp_laddr.s_addr == laddr.s_addr)
1073                                         return (inp);
1074                                 else if (inp->inp_laddr.s_addr == INADDR_ANY) {
1075 #ifdef INET6
1076                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1077                                                              AF_INET6))
1078                                                 local_wild_mapped = inp;
1079                                         else
1080 #endif
1081                                                 local_wild = inp;
1082                                 }
1083                         }
1084                 }
1085 #ifdef INET6
1086                 if (local_wild == NULL)
1087                         return (local_wild_mapped);
1088 #endif
1089                 return (local_wild);
1090         }
1091         return (NULL);
1092 }
1093
1094 /*
1095  * Insert PCB onto various hash lists.
1096  */
1097 int
1098 in_pcbinshash(struct inpcb *inp)
1099 {
1100         struct inpcbhead *pcbhash;
1101         struct inpcbporthead *pcbporthash;
1102         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1103         struct inpcbport *phd;
1104         u_int32_t hashkey_faddr;
1105
1106         INP_INFO_WLOCK_ASSERT(pcbinfo);
1107         INP_LOCK_ASSERT(inp);
1108
1109 #ifdef INET6
1110         if (inp->inp_vflag & INP_IPV6)
1111                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
1112         else
1113 #endif /* INET6 */
1114         hashkey_faddr = inp->inp_faddr.s_addr;
1115
1116         pcbhash = &pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
1117                  inp->inp_lport, inp->inp_fport, pcbinfo->hashmask)];
1118
1119         pcbporthash = &pcbinfo->porthashbase[INP_PCBPORTHASH(inp->inp_lport,
1120             pcbinfo->porthashmask)];
1121
1122         /*
1123          * Go through port list and look for a head for this lport.
1124          */
1125         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1126                 if (phd->phd_port == inp->inp_lport)
1127                         break;
1128         }
1129         /*
1130          * If none exists, malloc one and tack it on.
1131          */
1132         if (phd == NULL) {
1133                 MALLOC(phd, struct inpcbport *, sizeof(struct inpcbport), M_PCB, M_NOWAIT);
1134                 if (phd == NULL) {
1135                         return (ENOBUFS); /* XXX */
1136                 }
1137                 phd->phd_port = inp->inp_lport;
1138                 LIST_INIT(&phd->phd_pcblist);
1139                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1140         }
1141         inp->inp_phd = phd;
1142         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1143         LIST_INSERT_HEAD(pcbhash, inp, inp_hash);
1144         return (0);
1145 }
1146
1147 /*
1148  * Move PCB to the proper hash bucket when { faddr, fport } have  been
1149  * changed. NOTE: This does not handle the case of the lport changing (the
1150  * hashed port list would have to be updated as well), so the lport must
1151  * not change after in_pcbinshash() has been called.
1152  */
1153 void
1154 in_pcbrehash(struct inpcb *inp)
1155 {
1156         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1157         struct inpcbhead *head;
1158         u_int32_t hashkey_faddr;
1159
1160         INP_INFO_WLOCK_ASSERT(pcbinfo);
1161         INP_LOCK_ASSERT(inp);
1162
1163 #ifdef INET6
1164         if (inp->inp_vflag & INP_IPV6)
1165                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX */;
1166         else
1167 #endif /* INET6 */
1168         hashkey_faddr = inp->inp_faddr.s_addr;
1169
1170         head = &pcbinfo->hashbase[INP_PCBHASH(hashkey_faddr,
1171                 inp->inp_lport, inp->inp_fport, pcbinfo->hashmask)];
1172
1173         LIST_REMOVE(inp, inp_hash);
1174         LIST_INSERT_HEAD(head, inp, inp_hash);
1175 }
1176
1177 /*
1178  * Remove PCB from various lists.
1179  */
1180 void
1181 in_pcbremlists(struct inpcb *inp)
1182 {
1183         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1184
1185         INP_INFO_WLOCK_ASSERT(pcbinfo);
1186         INP_LOCK_ASSERT(inp);
1187
1188         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
1189         if (inp->inp_lport) {
1190                 struct inpcbport *phd = inp->inp_phd;
1191
1192                 LIST_REMOVE(inp, inp_hash);
1193                 LIST_REMOVE(inp, inp_portlist);
1194                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1195                         LIST_REMOVE(phd, phd_hash);
1196                         free(phd, M_PCB);
1197                 }
1198         }
1199         LIST_REMOVE(inp, inp_list);
1200         pcbinfo->ipi_count--;
1201 }
1202
1203 /*
1204  * A set label operation has occurred at the socket layer, propagate the
1205  * label change into the in_pcb for the socket.
1206  */
1207 void
1208 in_pcbsosetlabel(struct socket *so)
1209 {
1210 #ifdef MAC
1211         struct inpcb *inp;
1212
1213         inp = sotoinpcb(so);
1214         KASSERT(inp != NULL, ("in_pcbsosetlabel: so->so_pcb == NULL"));
1215
1216         INP_LOCK(inp);
1217         SOCK_LOCK(so);
1218         mac_inpcb_sosetlabel(so, inp);
1219         SOCK_UNLOCK(so);
1220         INP_UNLOCK(inp);
1221 #endif
1222 }
1223
1224 /*
1225  * ipport_tick runs once per second, determining if random port allocation
1226  * should be continued.  If more than ipport_randomcps ports have been
1227  * allocated in the last second, then we return to sequential port
1228  * allocation. We return to random allocation only once we drop below
1229  * ipport_randomcps for at least ipport_randomtime seconds.
1230  */
1231 void
1232 ipport_tick(void *xtp)
1233 {
1234
1235         if (ipport_tcpallocs <= ipport_tcplastcount + ipport_randomcps) {
1236                 if (ipport_stoprandom > 0)
1237                         ipport_stoprandom--;
1238         } else
1239                 ipport_stoprandom = ipport_randomtime;
1240         ipport_tcplastcount = ipport_tcpallocs;
1241         callout_reset(&ipport_tick_callout, hz, ipport_tick, NULL);
1242 }
1243
1244 #ifdef DDB
1245 static void
1246 db_print_indent(int indent)
1247 {
1248         int i;
1249
1250         for (i = 0; i < indent; i++)
1251                 db_printf(" ");
1252 }
1253
1254 static void
1255 db_print_inconninfo(struct in_conninfo *inc, const char *name, int indent)
1256 {
1257         char faddr_str[48], laddr_str[48];
1258
1259         db_print_indent(indent);
1260         db_printf("%s at %p\n", name, inc);
1261
1262         indent += 2;
1263
1264 #ifdef INET6
1265         if (inc->inc_flags == 1) {
1266                 /* IPv6. */
1267                 ip6_sprintf(laddr_str, &inc->inc6_laddr);
1268                 ip6_sprintf(faddr_str, &inc->inc6_faddr);
1269         } else {
1270 #endif
1271                 /* IPv4. */
1272                 inet_ntoa_r(inc->inc_laddr, laddr_str);
1273                 inet_ntoa_r(inc->inc_faddr, faddr_str);
1274 #ifdef INET6
1275         }
1276 #endif
1277         db_print_indent(indent);
1278         db_printf("inc_laddr %s   inc_lport %u\n", laddr_str,
1279             ntohs(inc->inc_lport));
1280         db_print_indent(indent);
1281         db_printf("inc_faddr %s   inc_fport %u\n", faddr_str,
1282             ntohs(inc->inc_fport));
1283 }
1284
1285 static void
1286 db_print_inpflags(int inp_flags)
1287 {
1288         int comma;
1289
1290         comma = 0;
1291         if (inp_flags & INP_RECVOPTS) {
1292                 db_printf("%sINP_RECVOPTS", comma ? ", " : "");
1293                 comma = 1;
1294         }
1295         if (inp_flags & INP_RECVRETOPTS) {
1296                 db_printf("%sINP_RECVRETOPTS", comma ? ", " : "");
1297                 comma = 1;
1298         }
1299         if (inp_flags & INP_RECVDSTADDR) {
1300                 db_printf("%sINP_RECVDSTADDR", comma ? ", " : "");
1301                 comma = 1;
1302         }
1303         if (inp_flags & INP_HDRINCL) {
1304                 db_printf("%sINP_HDRINCL", comma ? ", " : "");
1305                 comma = 1;
1306         }
1307         if (inp_flags & INP_HIGHPORT) {
1308                 db_printf("%sINP_HIGHPORT", comma ? ", " : "");
1309                 comma = 1;
1310         }
1311         if (inp_flags & INP_LOWPORT) {
1312                 db_printf("%sINP_LOWPORT", comma ? ", " : "");
1313                 comma = 1;
1314         }
1315         if (inp_flags & INP_ANONPORT) {
1316                 db_printf("%sINP_ANONPORT", comma ? ", " : "");
1317                 comma = 1;
1318         }
1319         if (inp_flags & INP_RECVIF) {
1320                 db_printf("%sINP_RECVIF", comma ? ", " : "");
1321                 comma = 1;
1322         }
1323         if (inp_flags & INP_MTUDISC) {
1324                 db_printf("%sINP_MTUDISC", comma ? ", " : "");
1325                 comma = 1;
1326         }
1327         if (inp_flags & INP_FAITH) {
1328                 db_printf("%sINP_FAITH", comma ? ", " : "");
1329                 comma = 1;
1330         }
1331         if (inp_flags & INP_RECVTTL) {
1332                 db_printf("%sINP_RECVTTL", comma ? ", " : "");
1333                 comma = 1;
1334         }
1335         if (inp_flags & INP_DONTFRAG) {
1336                 db_printf("%sINP_DONTFRAG", comma ? ", " : "");
1337                 comma = 1;
1338         }
1339         if (inp_flags & IN6P_IPV6_V6ONLY) {
1340                 db_printf("%sIN6P_IPV6_V6ONLY", comma ? ", " : "");
1341                 comma = 1;
1342         }
1343         if (inp_flags & IN6P_PKTINFO) {
1344                 db_printf("%sIN6P_PKTINFO", comma ? ", " : "");
1345                 comma = 1;
1346         }
1347         if (inp_flags & IN6P_HOPLIMIT) {
1348                 db_printf("%sIN6P_HOPLIMIT", comma ? ", " : "");
1349                 comma = 1;
1350         }
1351         if (inp_flags & IN6P_HOPOPTS) {
1352                 db_printf("%sIN6P_HOPOPTS", comma ? ", " : "");
1353                 comma = 1;
1354         }
1355         if (inp_flags & IN6P_DSTOPTS) {
1356                 db_printf("%sIN6P_DSTOPTS", comma ? ", " : "");
1357                 comma = 1;
1358         }
1359         if (inp_flags & IN6P_RTHDR) {
1360                 db_printf("%sIN6P_RTHDR", comma ? ", " : "");
1361                 comma = 1;
1362         }
1363         if (inp_flags & IN6P_RTHDRDSTOPTS) {
1364                 db_printf("%sIN6P_RTHDRDSTOPTS", comma ? ", " : "");
1365                 comma = 1;
1366         }
1367         if (inp_flags & IN6P_TCLASS) {
1368                 db_printf("%sIN6P_TCLASS", comma ? ", " : "");
1369                 comma = 1;
1370         }
1371         if (inp_flags & IN6P_AUTOFLOWLABEL) {
1372                 db_printf("%sIN6P_AUTOFLOWLABEL", comma ? ", " : "");
1373                 comma = 1;
1374         }
1375         if (inp_flags & IN6P_RFC2292) {
1376                 db_printf("%sIN6P_RFC2292", comma ? ", " : "");
1377                 comma = 1;
1378         }
1379         if (inp_flags & IN6P_MTU) {
1380                 db_printf("IN6P_MTU%s", comma ? ", " : "");
1381                 comma = 1;
1382         }
1383 }
1384
1385 static void
1386 db_print_inpvflag(u_char inp_vflag)
1387 {
1388         int comma;
1389
1390         comma = 0;
1391         if (inp_vflag & INP_IPV4) {
1392                 db_printf("%sINP_IPV4", comma ? ", " : "");
1393                 comma  = 1;
1394         }
1395         if (inp_vflag & INP_IPV6) {
1396                 db_printf("%sINP_IPV6", comma ? ", " : "");
1397                 comma  = 1;
1398         }
1399         if (inp_vflag & INP_IPV6PROTO) {
1400                 db_printf("%sINP_IPV6PROTO", comma ? ", " : "");
1401                 comma  = 1;
1402         }
1403         if (inp_vflag & INP_TIMEWAIT) {
1404                 db_printf("%sINP_TIMEWAIT", comma ? ", " : "");
1405                 comma  = 1;
1406         }
1407         if (inp_vflag & INP_ONESBCAST) {
1408                 db_printf("%sINP_ONESBCAST", comma ? ", " : "");
1409                 comma  = 1;
1410         }
1411         if (inp_vflag & INP_DROPPED) {
1412                 db_printf("%sINP_DROPPED", comma ? ", " : "");
1413                 comma  = 1;
1414         }
1415         if (inp_vflag & INP_SOCKREF) {
1416                 db_printf("%sINP_SOCKREF", comma ? ", " : "");
1417                 comma  = 1;
1418         }
1419 }
1420
1421 void
1422 db_print_inpcb(struct inpcb *inp, const char *name, int indent)
1423 {
1424
1425         db_print_indent(indent);
1426         db_printf("%s at %p\n", name, inp);
1427
1428         indent += 2;
1429
1430         db_print_indent(indent);
1431         db_printf("inp_flow: 0x%x\n", inp->inp_flow);
1432
1433         db_print_inconninfo(&inp->inp_inc, "inp_conninfo", indent);
1434
1435         db_print_indent(indent);
1436         db_printf("inp_ppcb: %p   inp_pcbinfo: %p   inp_socket: %p\n",
1437             inp->inp_ppcb, inp->inp_pcbinfo, inp->inp_socket);
1438
1439         db_print_indent(indent);
1440         db_printf("inp_label: %p   inp_flags: 0x%x (",
1441            inp->inp_label, inp->inp_flags);
1442         db_print_inpflags(inp->inp_flags);
1443         db_printf(")\n");
1444
1445         db_print_indent(indent);
1446         db_printf("inp_sp: %p   inp_vflag: 0x%x (", inp->inp_sp,
1447             inp->inp_vflag);
1448         db_print_inpvflag(inp->inp_vflag);
1449         db_printf(")\n");
1450
1451         db_print_indent(indent);
1452         db_printf("inp_ip_ttl: %d   inp_ip_p: %d   inp_ip_minttl: %d\n",
1453             inp->inp_ip_ttl, inp->inp_ip_p, inp->inp_ip_minttl);
1454
1455         db_print_indent(indent);
1456 #ifdef INET6
1457         if (inp->inp_vflag & INP_IPV6) {
1458                 db_printf("in6p_options: %p   in6p_outputopts: %p   "
1459                     "in6p_moptions: %p\n", inp->in6p_options,
1460                     inp->in6p_outputopts, inp->in6p_moptions);
1461                 db_printf("in6p_icmp6filt: %p   in6p_cksum %d   "
1462                     "in6p_hops %u\n", inp->in6p_icmp6filt, inp->in6p_cksum,
1463                     inp->in6p_hops);
1464         } else
1465 #endif
1466         {
1467                 db_printf("inp_ip_tos: %d   inp_ip_options: %p   "
1468                     "inp_ip_moptions: %p\n", inp->inp_ip_tos,
1469                     inp->inp_options, inp->inp_moptions);
1470         }
1471
1472         db_print_indent(indent);
1473         db_printf("inp_phd: %p   inp_gencnt: %ju\n", inp->inp_phd,
1474             (uintmax_t)inp->inp_gencnt);
1475 }
1476
1477 DB_SHOW_COMMAND(inpcb, db_show_inpcb)
1478 {
1479         struct inpcb *inp;
1480
1481         if (!have_addr) {
1482                 db_printf("usage: show inpcb <addr>\n");
1483                 return;
1484         }
1485         inp = (struct inpcb *)addr;
1486
1487         db_print_inpcb(inp, "inpcb", 0);
1488 }
1489 #endif