sys/netinet/in_pcb.c

   1 /*
   2  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
   3  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to The DragonFly Project
   6  * by Jeffrey M. Hsu.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. Neither the name of The DragonFly Project nor the names of its
  17  *    contributors may be used to endorse or promote products derived
  18  *    from this software without specific, prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  */
  33
  34 /*
  35  * Copyright (c) 1982, 1986, 1991, 1993, 1995
  36  *      The Regents of the University of California.  All rights reserved.
  37  *
  38  * Redistribution and use in source and binary forms, with or without
  39  * modification, are permitted provided that the following conditions
  40  * are met:
  41  * 1. Redistributions of source code must retain the above copyright
  42  *    notice, this list of conditions and the following disclaimer.
  43  * 2. Redistributions in binary form must reproduce the above copyright
  44  *    notice, this list of conditions and the following disclaimer in the
  45  *    documentation and/or other materials provided with the distribution.
  46  * 4. Neither the name of the University nor the names of its contributors
  47  *    may be used to endorse or promote products derived from this software
  48  *    without specific prior written permission.
  49  *
  50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  60  * SUCH DAMAGE.
  61  *
  62  *      @(#)in_pcb.c    8.4 (Berkeley) 5/24/95
  63  * $FreeBSD: src/sys/netinet/in_pcb.c,v 1.59.2.27 2004/01/02 04:06:42 ambrisko Exp $
  64  */
  65
  66 #include "opt_ipsec.h"
  67 #include "opt_inet6.h"
  68
  69 #include <sys/param.h>
  70 #include <sys/systm.h>
  71 #include <sys/malloc.h>
  72 #include <sys/mbuf.h>
  73 #include <sys/domain.h>
  74 #include <sys/protosw.h>
  75 #include <sys/socket.h>
  76 #include <sys/socketvar.h>
  77 #include <sys/proc.h>
  78 #include <sys/priv.h>
  79 #include <sys/jail.h>
  80 #include <sys/kernel.h>
  81 #include <sys/sysctl.h>
  82
  83 #include <sys/thread2.h>
  84 #include <sys/socketvar2.h>
  85 #include <sys/msgport2.h>
  86
  87 #include <machine/limits.h>
  88
  89 #include <net/if.h>
  90 #include <net/if_types.h>
  91 #include <net/route.h>
  92
  93 #include <netinet/in.h>
  94 #include <netinet/in_pcb.h>
  95 #include <netinet/in_var.h>
  96 #include <netinet/ip_var.h>
  97 #ifdef INET6
  98 #include <netinet/ip6.h>
  99 #include <netinet6/ip6_var.h>
 100 #endif /* INET6 */
 101
 102 #ifdef IPSEC
 103 #include <netinet6/ipsec.h>
 104 #include <netproto/key/key.h>
 105 #include <netproto/ipsec/esp_var.h>
 106 #endif
 107
 108 #ifdef FAST_IPSEC
 109 #if defined(IPSEC) || defined(IPSEC_ESP)
 110 #error "Bad idea: don't compile with both IPSEC and FAST_IPSEC!"
 111 #endif
 112
 113 #include <netproto/ipsec/ipsec.h>
 114 #include <netproto/ipsec/key.h>
 115 #define IPSEC
 116 #endif /* FAST_IPSEC */
 117
 118 #define INP_LOCALGROUP_SIZMIN   8
 119 #define INP_LOCALGROUP_SIZMAX   256
 120
 121 struct in_addr zeroin_addr;
 122
 123 /*
 124  * These configure the range of local port addresses assigned to
 125  * "unspecified" outgoing connections/packets/whatever.
 126  */
 127 int ipport_lowfirstauto = IPPORT_RESERVED - 1;  /* 1023 */
 128 int ipport_lowlastauto = IPPORT_RESERVEDSTART;  /* 600 */
 129
 130 int ipport_firstauto = IPPORT_RESERVED;         /* 1024 */
 131 int ipport_lastauto = IPPORT_USERRESERVED;      /* 5000 */
 132
 133 int ipport_hifirstauto = IPPORT_HIFIRSTAUTO;    /* 49152 */
 134 int ipport_hilastauto = IPPORT_HILASTAUTO;      /* 65535 */
 135
 136 #define RANGECHK(var, min, max) \
 137         if ((var) < (min)) { (var) = (min); } \
 138         else if ((var) > (max)) { (var) = (max); }
 139
 140 int udpencap_enable = 1;        /* enabled by default */
 141 int udpencap_port = 4500;       /* triggers decapsulation */
 142
 143 static int
 144 sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
 145 {
 146         int error;
 147
 148         error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
 149         if (!error) {
 150                 RANGECHK(ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
 151                 RANGECHK(ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
 152
 153                 RANGECHK(ipport_firstauto, IPPORT_RESERVED, USHRT_MAX);
 154                 RANGECHK(ipport_lastauto, IPPORT_RESERVED, USHRT_MAX);
 155
 156                 RANGECHK(ipport_hifirstauto, IPPORT_RESERVED, USHRT_MAX);
 157                 RANGECHK(ipport_hilastauto, IPPORT_RESERVED, USHRT_MAX);
 158         }
 159         return (error);
 160 }
 161
 162 #undef RANGECHK
 163
 164 SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0, "IP Ports");
 165
 166 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLTYPE_INT|CTLFLAG_RW,
 167            &ipport_lowfirstauto, 0, &sysctl_net_ipport_check, "I", "");
 168 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLTYPE_INT|CTLFLAG_RW,
 169            &ipport_lowlastauto, 0, &sysctl_net_ipport_check, "I", "");
 170 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLTYPE_INT|CTLFLAG_RW,
 171            &ipport_firstauto, 0, &sysctl_net_ipport_check, "I", "");
 172 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLTYPE_INT|CTLFLAG_RW,
 173            &ipport_lastauto, 0, &sysctl_net_ipport_check, "I", "");
 174 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLTYPE_INT|CTLFLAG_RW,
 175            &ipport_hifirstauto, 0, &sysctl_net_ipport_check, "I", "");
 176 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLTYPE_INT|CTLFLAG_RW,
 177            &ipport_hilastauto, 0, &sysctl_net_ipport_check, "I", "");
 178
 179 /*
 180  * in_pcb.c: manage the Protocol Control Blocks.
 181  *
 182  * NOTE: It is assumed that most of these functions will be called from
 183  * a critical section.  XXX - There are, unfortunately, a few exceptions
 184  * to this rule that should be fixed.
 185  *
 186  * NOTE: The caller should initialize the cpu field to the cpu running the
 187  * protocol stack associated with this inpcbinfo.
 188  */
 189
 190 void
 191 in_pcbinfo_init(struct inpcbinfo *pcbinfo)
 192 {
 193         LIST_INIT(&pcbinfo->pcblisthead);
 194         pcbinfo->cpu = -1;
 195         pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave), M_PCB,
 196                                     M_WAITOK | M_ZERO);
 197 }
 198
 199 struct baddynamicports baddynamicports;
 200
 201 /*
 202  * Check if the specified port is invalid for dynamic allocation.
 203  */
 204 int
 205 in_baddynamic(u_int16_t port, u_int16_t proto)
 206 {
 207         switch (proto) {
 208         case IPPROTO_TCP:
 209                 return (DP_ISSET(baddynamicports.tcp, port));
 210         case IPPROTO_UDP:
 211 #ifdef IPSEC
 212                 /* Cannot preset this as it is a sysctl */
 213                 if (port == udpencap_port)
 214                         return (1);
 215 #endif
 216                 return (DP_ISSET(baddynamicports.udp, port));
 217         default:
 218                 return (0);
 219         }
 220 }
 221
 222
 223 /*
 224  * Allocate a PCB and associate it with the socket.
 225  */
 226 int
 227 in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo)
 228 {
 229         struct inpcb *inp;
 230 #ifdef IPSEC
 231         int error;
 232 #endif
 233
 234         inp = kmalloc(pcbinfo->ipi_size, M_PCB, M_WAITOK|M_ZERO);
 235         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 236         inp->inp_pcbinfo = inp->inp_cpcbinfo = pcbinfo;
 237         inp->inp_socket = so;
 238 #ifdef IPSEC
 239         error = ipsec_init_policy(so, &inp->inp_sp);
 240         if (error != 0) {
 241                 kfree(inp, M_PCB);
 242                 return (error);
 243         }
 244 #endif
 245 #ifdef INET6
 246         if (INP_SOCKAF(so) == AF_INET6 && ip6_v6only)
 247                 inp->inp_flags |= IN6P_IPV6_V6ONLY;
 248         if (ip6_auto_flowlabel)
 249                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 250 #endif
 251         soreference(so);
 252         so->so_pcb = inp;
 253         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 254         pcbinfo->ipi_count++;
 255         return (0);
 256 }
 257
 258 /*
 259  * Unlink a pcb with the intention of moving it to another cpu with a
 260  * different pcbinfo.  While unlinked nothing should attempt to dereference
 261  * inp_pcbinfo, NULL it out so we assert if it does.
 262  */
 263 void
 264 in_pcbunlink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 265 {
 266         KKASSERT(inp->inp_pcbinfo == pcbinfo);
 267
 268         LIST_REMOVE(inp, inp_list);
 269         pcbinfo->ipi_count--;
 270         inp->inp_pcbinfo = NULL;
 271 }
 272
 273 /*
 274  * Relink a pcb into a new pcbinfo.
 275  */
 276 void
 277 in_pcblink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 278 {
 279         KKASSERT(inp->inp_pcbinfo == NULL);
 280         inp->inp_pcbinfo = pcbinfo;
 281         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 282         pcbinfo->ipi_count++;
 283 }
 284
 285 int
 286 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 287 {
 288         struct socket *so = inp->inp_socket;
 289         unsigned short *lastport;
 290         struct sockaddr_in *sin;
 291         struct sockaddr_in jsin;
 292         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 293         struct ucred *cred = NULL;
 294         u_short lport = 0;
 295         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 296         int error;
 297
 298         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 299                 return (EADDRNOTAVAIL);
 300         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 301                 return (EINVAL);        /* already bound */
 302
 303         if (!(so->so_options & (SO_REUSEADDR|SO_REUSEPORT)))
 304                 wild = 1;    /* neither SO_REUSEADDR nor SO_REUSEPORT is set */
 305         if (td->td_proc)
 306                 cred = td->td_proc->p_ucred;
 307
 308         /*
 309          * This has to be atomic.  If the porthash is shared across multiple
 310          * protocol threads (aka tcp) then the token will be non-NULL.
 311          */
 312         if (pcbinfo->porttoken)
 313                 lwkt_gettoken(pcbinfo->porttoken);
 314
 315         if (nam != NULL) {
 316                 sin = (struct sockaddr_in *)nam;
 317                 if (nam->sa_len != sizeof *sin) {
 318                         error = EINVAL;
 319                         goto done;
 320                 }
 321 #ifdef notdef
 322                 /*
 323                  * We should check the family, but old programs
 324                  * incorrectly fail to initialize it.
 325                  */
 326                 if (sin->sin_family != AF_INET) {
 327                         error = EAFNOSUPPORT;
 328                         goto done;
 329                 }
 330 #endif
 331                 if (!prison_replace_wildcards(td, nam)) {
 332                         error = EINVAL;
 333                         goto done;
 334                 }
 335                 lport = sin->sin_port;
 336                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 337                         /*
 338                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 339                          * allow complete duplication of binding if
 340                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 341                          * and a multicast address is bound on both
 342                          * new and duplicated sockets.
 343                          */
 344                         if (so->so_options & SO_REUSEADDR)
 345                                 reuseport = SO_REUSEADDR | SO_REUSEPORT;
 346                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 347                         sin->sin_port = 0;              /* yech... */
 348                         bzero(&sin->sin_zero, sizeof sin->sin_zero);
 349                         if (ifa_ifwithaddr((struct sockaddr *)sin) == NULL) {
 350                                 error = EADDRNOTAVAIL;
 351                                 goto done;
 352                         }
 353                 }
 354                 if (lport != 0) {
 355                         struct inpcb *t;
 356
 357                         /* GROSS */
 358                         if (ntohs(lport) < IPPORT_RESERVED &&
 359                             cred &&
 360                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0)) {
 361                                 error = EACCES;
 362                                 goto done;
 363                         }
 364                         if (so->so_cred->cr_uid != 0 &&
 365                             !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 366                                 t = in_pcblookup_local(pcbinfo,
 367                                                        sin->sin_addr,
 368                                                        lport,
 369                                                        INPLOOKUP_WILDCARD,
 370                                                        cred);
 371                                 if (t &&
 372                                     (!in_nullhost(sin->sin_addr) ||
 373                                      !in_nullhost(t->inp_laddr) ||
 374                                      (t->inp_socket->so_options &
 375                                          SO_REUSEPORT) == 0) &&
 376                                     (so->so_cred->cr_uid !=
 377                                      t->inp_socket->so_cred->cr_uid)) {
 378 #ifdef INET6
 379                                         if (!in_nullhost(sin->sin_addr) ||
 380                                             !in_nullhost(t->inp_laddr) ||
 381                                             INP_SOCKAF(so) ==
 382                                             INP_SOCKAF(t->inp_socket))
 383 #endif
 384                                         {
 385                                                 error = EADDRINUSE;
 386                                                 goto done;
 387                                         }
 388                                 }
 389                         }
 390                         if (cred && !prison_replace_wildcards(td, nam)) {
 391                                 error = EADDRNOTAVAIL;
 392                                 goto done;
 393                         }
 394                         t = in_pcblookup_local(pcbinfo, sin->sin_addr, lport,
 395                                                wild, cred);
 396                         if (t && !(reuseport & t->inp_socket->so_options)) {
 397 #ifdef INET6
 398                                 if (!in_nullhost(sin->sin_addr) ||
 399                                     !in_nullhost(t->inp_laddr) ||
 400                                     INP_SOCKAF(so) == INP_SOCKAF(t->inp_socket))
 401 #endif
 402                                 {
 403                                         error = EADDRINUSE;
 404                                         goto done;
 405                                 }
 406                         }
 407                 }
 408                 inp->inp_laddr = sin->sin_addr;
 409         }
 410         if (lport == 0) {
 411                 ushort first, last;
 412                 int count;
 413
 414                 jsin.sin_family = AF_INET;
 415                 jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 416                 if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 417                         inp->inp_laddr.s_addr = INADDR_ANY;
 418                         error = EINVAL;
 419                         goto done;
 420                 }
 421                 inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 422
 423                 inp->inp_flags |= INP_ANONPORT;
 424
 425                 if (inp->inp_flags & INP_HIGHPORT) {
 426                         first = ipport_hifirstauto;     /* sysctl */
 427                         last  = ipport_hilastauto;
 428                         lastport = &pcbinfo->lasthi;
 429                 } else if (inp->inp_flags & INP_LOWPORT) {
 430                         if (cred &&
 431                             (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 432                                 inp->inp_laddr.s_addr = INADDR_ANY;
 433                                 goto done;
 434                         }
 435                         first = ipport_lowfirstauto;    /* 1023 */
 436                         last  = ipport_lowlastauto;     /* 600 */
 437                         lastport = &pcbinfo->lastlow;
 438                 } else {
 439                         first = ipport_firstauto;       /* sysctl */
 440                         last  = ipport_lastauto;
 441                         lastport = &pcbinfo->lastport;
 442                 }
 443                 /*
 444                  * Simple check to ensure all ports are not used up causing
 445                  * a deadlock here.
 446                  *
 447                  * We split the two cases (up and down) so that the direction
 448                  * is not being tested on each round of the loop.
 449                  */
 450                 if (first > last) {
 451                         /*
 452                          * counting down
 453                          */
 454                         count = first - last;
 455
 456                         do {
 457                                 if (count-- < 0) {      /* completely used? */
 458                                         inp->inp_laddr.s_addr = INADDR_ANY;
 459                                         error = EADDRNOTAVAIL;
 460                                         goto done;
 461                                 }
 462                                 --*lastport;
 463                                 if (*lastport > first || *lastport < last)
 464                                         *lastport = first;
 465                                 lport = htons(*lastport);
 466                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 467                                                     lport, wild, cred));
 468                 } else {
 469                         /*
 470                          * counting up
 471                          */
 472                         count = last - first;
 473
 474                         do {
 475                                 if (count-- < 0) {      /* completely used? */
 476                                         inp->inp_laddr.s_addr = INADDR_ANY;
 477                                         error = EADDRNOTAVAIL;
 478                                         goto done;
 479                                 }
 480                                 ++*lastport;
 481                                 if (*lastport < first || *lastport > last)
 482                                         *lastport = first;
 483                                 lport = htons(*lastport);
 484                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 485                                                     lport, wild, cred));
 486                 }
 487         }
 488         inp->inp_lport = lport;
 489
 490         jsin.sin_family = AF_INET;
 491         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 492         if (!prison_replace_wildcards(td, (struct sockaddr*)&jsin)) {
 493                 inp->inp_laddr.s_addr = INADDR_ANY;
 494                 inp->inp_lport = 0;
 495                 error = EINVAL;
 496                 goto done;
 497         }
 498         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 499
 500         if (in_pcbinsporthash(inp) != 0) {
 501                 inp->inp_laddr.s_addr = INADDR_ANY;
 502                 inp->inp_lport = 0;
 503                 error = EAGAIN;
 504                 goto done;
 505         }
 506         error = 0;
 507 done:
 508         if (pcbinfo->porttoken)
 509                 lwkt_reltoken(pcbinfo->porttoken);
 510         return error;
 511 }
 512
 513 static struct inpcb *
 514 in_pcblookup_addrport(struct inpcbinfo *pcbinfo, struct in_addr laddr,
 515     u_short lport, struct in_addr faddr, u_short fport, struct ucred *cred)
 516 {
 517         struct inpcb *inp;
 518         struct inpcbporthead *porthash;
 519         struct inpcbport *phd;
 520         struct inpcb *match = NULL;
 521
 522         /*
 523          * If the porthashbase is shared across several cpus we need
 524          * to lock.
 525          */
 526         if (pcbinfo->porttoken)
 527                 lwkt_gettoken(pcbinfo->porttoken);
 528
 529         /*
 530          * Best fit PCB lookup.
 531          *
 532          * First see if this local port is in use by looking on the
 533          * port hash list.
 534          */
 535         porthash = &pcbinfo->porthashbase[
 536                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
 537         LIST_FOREACH(phd, porthash, phd_hash) {
 538                 if (phd->phd_port == lport)
 539                         break;
 540         }
 541         if (phd != NULL) {
 542                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 543 #ifdef INET6
 544                         if ((inp->inp_vflag & INP_IPV4) == 0)
 545                                 continue;
 546 #endif
 547                         if (inp->inp_laddr.s_addr != INADDR_ANY &&
 548                             inp->inp_laddr.s_addr != laddr.s_addr)
 549                                 continue;
 550
 551                         if (inp->inp_faddr.s_addr != INADDR_ANY &&
 552                             inp->inp_faddr.s_addr != faddr.s_addr)
 553                                 continue;
 554
 555                         if (inp->inp_fport != 0 && inp->inp_fport != fport)
 556                                 continue;
 557
 558                         if (cred == NULL ||
 559                             cred->cr_prison ==
 560                             inp->inp_socket->so_cred->cr_prison) {
 561                                 match = inp;
 562                                 break;
 563                         }
 564                 }
 565         }
 566         if (pcbinfo->porttoken)
 567                 lwkt_reltoken(pcbinfo->porttoken);
 568         return (match);
 569 }
 570
 571 int
 572 in_pcbconn_bind(struct inpcb *inp, const struct sockaddr *nam,
 573     struct thread *td)
 574 {
 575         struct proc *p = td->td_proc;
 576         unsigned short *lastport;
 577         const struct sockaddr_in *sin = (const struct sockaddr_in *)nam;
 578         struct sockaddr_in jsin;
 579         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 580         struct ucred *cred = NULL;
 581         u_short lport = 0;
 582         ushort first, last;
 583         int count, error, dup = 0;
 584
 585         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 586                 return (EADDRNOTAVAIL);
 587
 588         KKASSERT(inp->inp_laddr.s_addr != INADDR_ANY);
 589         if (inp->inp_lport != 0)
 590                 return (EINVAL);        /* already bound */
 591
 592         KKASSERT(p);
 593         cred = p->p_ucred;
 594
 595         /*
 596          * This has to be atomic.  If the porthash is shared across multiple
 597          * protocol threads (aka tcp) then the token will be non-NULL.
 598          */
 599         if (pcbinfo->porttoken)
 600                 lwkt_gettoken(pcbinfo->porttoken);
 601
 602         jsin.sin_family = AF_INET;
 603         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 604         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 605                 inp->inp_laddr.s_addr = INADDR_ANY;
 606                 error = EINVAL;
 607                 goto done;
 608         }
 609         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 610
 611         inp->inp_flags |= INP_ANONPORT;
 612
 613         if (inp->inp_flags & INP_HIGHPORT) {
 614                 first = ipport_hifirstauto;     /* sysctl */
 615                 last  = ipport_hilastauto;
 616                 lastport = &pcbinfo->lasthi;
 617         } else if (inp->inp_flags & INP_LOWPORT) {
 618                 if (cred &&
 619                     (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 620                         inp->inp_laddr.s_addr = INADDR_ANY;
 621                         goto done;
 622                 }
 623                 first = ipport_lowfirstauto;    /* 1023 */
 624                 last  = ipport_lowlastauto;     /* 600 */
 625                 lastport = &pcbinfo->lastlow;
 626         } else {
 627                 first = ipport_firstauto;       /* sysctl */
 628                 last  = ipport_lastauto;
 629                 lastport = &pcbinfo->lastport;
 630         }
 631
 632 again:
 633         /*
 634          * Simple check to ensure all ports are not used up causing
 635          * a deadlock here.
 636          *
 637          * We split the two cases (up and down) so that the direction
 638          * is not being tested on each round of the loop.
 639          */
 640         if (first > last) {
 641                 /*
 642                  * counting down
 643                  */
 644                 count = first - last;
 645
 646                 do {
 647                         if (count-- < 0) {      /* completely used? */
 648                                 inp->inp_laddr.s_addr = INADDR_ANY;
 649                                 error = EADDRNOTAVAIL;
 650                                 goto done;
 651                         }
 652                         --*lastport;
 653                         if (*lastport > first || *lastport < last)
 654                                 *lastport = first;
 655                         lport = htons(*lastport);
 656                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 657                                 sin->sin_addr, sin->sin_port, cred));
 658         } else {
 659                 /*
 660                  * counting up
 661                  */
 662                 count = last - first;
 663
 664                 do {
 665                         if (count-- < 0) {      /* completely used? */
 666                                 inp->inp_laddr.s_addr = INADDR_ANY;
 667                                 error = EADDRNOTAVAIL;
 668                                 goto done;
 669                         }
 670                         ++*lastport;
 671                         if (*lastport < first || *lastport > last)
 672                                 *lastport = first;
 673                         lport = htons(*lastport);
 674                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 675                                 sin->sin_addr, sin->sin_port, cred));
 676         }
 677
 678         /* This could happen on loopback interface */
 679         if (sin->sin_port == lport &&
 680             sin->sin_addr.s_addr == inp->inp_laddr.s_addr) {
 681                 if (dup) {
 682                         /*
 683                          * Duplicate again; give up
 684                          */
 685                         inp->inp_laddr.s_addr = INADDR_ANY;
 686                         error = EADDRNOTAVAIL;
 687                         goto done;
 688                 }
 689                 dup = 1;
 690                 goto again;
 691         }
 692         inp->inp_lport = lport;
 693
 694         jsin.sin_family = AF_INET;
 695         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 696         if (!prison_replace_wildcards(td, (struct sockaddr*)&jsin)) {
 697                 inp->inp_laddr.s_addr = INADDR_ANY;
 698                 inp->inp_lport = 0;
 699                 error = EINVAL;
 700                 goto done;
 701         }
 702         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 703
 704         if (in_pcbinsporthash(inp) != 0) {
 705                 inp->inp_laddr.s_addr = INADDR_ANY;
 706                 inp->inp_lport = 0;
 707                 error = EAGAIN;
 708                 goto done;
 709         }
 710         error = 0;
 711 done:
 712         if (pcbinfo->porttoken)
 713                 lwkt_reltoken(pcbinfo->porttoken);
 714         return error;
 715 }
 716
 717 /*
 718  *   Transform old in_pcbconnect() into an inner subroutine for new
 719  *   in_pcbconnect(): Do some validity-checking on the remote
 720  *   address (in mbuf 'nam') and then determine local host address
 721  *   (i.e., which interface) to use to access that remote host.
 722  *
 723  *   This preserves definition of in_pcbconnect(), while supporting a
 724  *   slightly different version for T/TCP.  (This is more than
 725  *   a bit of a kludge, but cleaning up the internal interfaces would
 726  *   have forced minor changes in every protocol).
 727  */
 728 int
 729 in_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 730         struct sockaddr_in **plocal_sin, struct thread *td)
 731 {
 732         struct in_ifaddr *ia;
 733         struct ucred *cred = NULL;
 734         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 735         struct sockaddr *jsin;
 736         int jailed = 0, alloc_route = 0;
 737
 738         if (nam->sa_len != sizeof *sin)
 739                 return (EINVAL);
 740         if (sin->sin_family != AF_INET)
 741                 return (EAFNOSUPPORT);
 742         if (sin->sin_port == 0)
 743                 return (EADDRNOTAVAIL);
 744         if (td && td->td_proc && td->td_proc->p_ucred)
 745                 cred = td->td_proc->p_ucred;
 746         if (cred && cred->cr_prison)
 747                 jailed = 1;
 748         if (!TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) {
 749                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 750                 /*
 751                  * If the destination address is INADDR_ANY,
 752                  * use the primary local address.
 753                  * If the supplied address is INADDR_BROADCAST,
 754                  * and the primary interface supports broadcast,
 755                  * choose the broadcast address for that interface.
 756                  */
 757                 if (sin->sin_addr.s_addr == INADDR_ANY)
 758                         sin->sin_addr = IA_SIN(ia)->sin_addr;
 759                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 760                     (ia->ia_ifp->if_flags & IFF_BROADCAST))
 761                         sin->sin_addr = satosin(&ia->ia_broadaddr)->sin_addr;
 762         }
 763         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 764                 struct route *ro;
 765
 766                 ia = NULL;
 767                 /*
 768                  * If route is known or can be allocated now,
 769                  * our src addr is taken from the i/f, else punt.
 770                  * Note that we should check the address family of the cached
 771                  * destination, in case of sharing the cache with IPv6.
 772                  */
 773                 ro = &inp->inp_route;
 774                 if (ro->ro_rt &&
 775                     (!(ro->ro_rt->rt_flags & RTF_UP) ||
 776                      ro->ro_dst.sa_family != AF_INET ||
 777                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 778                                       sin->sin_addr.s_addr ||
 779                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 780                         RTFREE(ro->ro_rt);
 781                         ro->ro_rt = NULL;
 782                 }
 783                 if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
 784                     (ro->ro_rt == NULL ||
 785                     ro->ro_rt->rt_ifp == NULL)) {
 786                         /* No route yet, so try to acquire one */
 787                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 788                         ro->ro_dst.sa_family = AF_INET;
 789                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 790                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 791                                 sin->sin_addr;
 792                         rtalloc(ro);
 793                         alloc_route = 1;
 794                 }
 795                 /*
 796                  * If we found a route, use the address
 797                  * corresponding to the outgoing interface
 798                  * unless it is the loopback (in case a route
 799                  * to our address on another net goes to loopback).
 800                  */
 801                 if (ro->ro_rt && !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK)) {
 802                         if (jailed) {
 803                                 if (jailed_ip(cred->cr_prison,
 804                                     ro->ro_rt->rt_ifa->ifa_addr)) {
 805                                         ia = ifatoia(ro->ro_rt->rt_ifa);
 806                                 }
 807                         } else {
 808                                 ia = ifatoia(ro->ro_rt->rt_ifa);
 809                         }
 810                 }
 811                 if (ia == NULL) {
 812                         u_short fport = sin->sin_port;
 813
 814                         sin->sin_port = 0;
 815                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 816                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 817                             sintosa(&ia->ia_addr)))
 818                                 ia = NULL;
 819                         if (ia == NULL)
 820                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 821                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 822                             sintosa(&ia->ia_addr)))
 823                                 ia = NULL;
 824                         sin->sin_port = fport;
 825                         if (ia == NULL &&
 826                             !TAILQ_EMPTY(&in_ifaddrheads[mycpuid]))
 827                                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 828                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 829                             sintosa(&ia->ia_addr)))
 830                                 ia = NULL;
 831
 832                         if (!jailed && ia == NULL)
 833                                 goto fail;
 834                 }
 835                 /*
 836                  * If the destination address is multicast and an outgoing
 837                  * interface has been set as a multicast option, use the
 838                  * address of that interface as our source address.
 839                  */
 840                 if (!jailed && IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 841                     inp->inp_moptions != NULL) {
 842                         struct ip_moptions *imo;
 843                         struct ifnet *ifp;
 844
 845                         imo = inp->inp_moptions;
 846                         if (imo->imo_multicast_ifp != NULL) {
 847                                 struct in_ifaddr_container *iac;
 848
 849                                 ifp = imo->imo_multicast_ifp;
 850                                 ia = NULL;
 851                                 TAILQ_FOREACH(iac,
 852                                 &in_ifaddrheads[mycpuid], ia_link) {
 853                                         if (iac->ia->ia_ifp == ifp) {
 854                                                 ia = iac->ia;
 855                                                 break;
 856                                         }
 857                                 }
 858                                 if (ia == NULL)
 859                                         goto fail;
 860                         }
 861                 }
 862                 /*
 863                  * Don't do pcblookup call here; return interface in plocal_sin
 864                  * and exit to caller, that will do the lookup.
 865                  */
 866                 if (ia == NULL && jailed) {
 867                         if ((jsin = prison_get_nonlocal(cred->cr_prison, AF_INET, NULL)) != NULL ||
 868                             (jsin = prison_get_local(cred->cr_prison, AF_INET, NULL)) != NULL) {
 869                                 *plocal_sin = satosin(jsin);
 870                         } else {
 871                                 /* IPv6 only Jail */
 872                                 goto fail;
 873                         }
 874                 } else {
 875                         *plocal_sin = &ia->ia_addr;
 876                 }
 877         }
 878         return (0);
 879 fail:
 880         if (alloc_route) {
 881                 struct route *ro = &inp->inp_route;
 882
 883                 if (ro->ro_rt != NULL)
 884                         RTFREE(ro->ro_rt);
 885                 bzero(ro, sizeof(*ro));
 886         }
 887         return (EADDRNOTAVAIL);
 888 }
 889
 890 /*
 891  * Outer subroutine:
 892  * Connect from a socket to a specified address.
 893  * Both address and port must be specified in argument sin.
 894  * If don't have a local address for this socket yet,
 895  * then pick one.
 896  */
 897 int
 898 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 899 {
 900         struct sockaddr_in *if_sin;
 901         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 902         int error;
 903
 904         /* Call inner routine to assign local interface address. */
 905         if ((error = in_pcbladdr(inp, nam, &if_sin, td)) != 0)
 906                 return (error);
 907
 908         if (in_pcblookup_hash(inp->inp_cpcbinfo, sin->sin_addr, sin->sin_port,
 909                               inp->inp_laddr.s_addr ?
 910                                 inp->inp_laddr : if_sin->sin_addr,
 911                               inp->inp_lport, FALSE, NULL) != NULL) {
 912                 return (EADDRINUSE);
 913         }
 914         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 915                 if (inp->inp_lport == 0) {
 916                         error = in_pcbbind(inp, NULL, td);
 917                         if (error)
 918                                 return (error);
 919                 }
 920                 inp->inp_laddr = if_sin->sin_addr;
 921         }
 922         inp->inp_faddr = sin->sin_addr;
 923         inp->inp_fport = sin->sin_port;
 924         in_pcbinsconnhash(inp);
 925         return (0);
 926 }
 927
 928 void
 929 in_pcbdisconnect(struct inpcb *inp)
 930 {
 931
 932         inp->inp_faddr.s_addr = INADDR_ANY;
 933         inp->inp_fport = 0;
 934         in_pcbremconnhash(inp);
 935         if (inp->inp_socket->so_state & SS_NOFDREF)
 936                 in_pcbdetach(inp);
 937 }
 938
 939 void
 940 in_pcbdetach(struct inpcb *inp)
 941 {
 942         struct socket *so = inp->inp_socket;
 943         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 944
 945 #ifdef IPSEC
 946         ipsec4_delete_pcbpolicy(inp);
 947 #endif /*IPSEC*/
 948         inp->inp_gencnt = ++ipi->ipi_gencnt;
 949         KKASSERT((so->so_state & SS_ASSERTINPROG) == 0);
 950         in_pcbremlists(inp);
 951         so->so_pcb = NULL;
 952         sofree(so);                     /* remove pcb ref */
 953         if (inp->inp_options)
 954                 m_free(inp->inp_options);
 955         if (inp->inp_route.ro_rt)
 956                 rtfree(inp->inp_route.ro_rt);
 957         ip_freemoptions(inp->inp_moptions);
 958         inp->inp_vflag = 0;
 959         kfree(inp, M_PCB);
 960 }
 961
 962 /*
 963  * The calling convention of in_setsockaddr() and in_setpeeraddr() was
 964  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 965  * in struct pr_usrreqs, so that protocols can just reference then directly
 966  * without the need for a wrapper function.  The socket must have a valid
 967  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 968  * except through a kernel programming error, so it is acceptable to panic
 969  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 970  * because there actually /is/ a programming error somewhere... XXX)
 971  */
 972 int
 973 in_setsockaddr(struct socket *so, struct sockaddr **nam)
 974 {
 975         struct inpcb *inp;
 976         struct sockaddr_in *sin;
 977
 978         /*
 979          * Do the malloc first in case it blocks.
 980          */
 981         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
 982         sin->sin_family = AF_INET;
 983         sin->sin_len = sizeof *sin;
 984
 985         crit_enter();
 986         inp = so->so_pcb;
 987         if (!inp) {
 988                 crit_exit();
 989                 kfree(sin, M_SONAME);
 990                 return (ECONNRESET);
 991         }
 992         sin->sin_port = inp->inp_lport;
 993         sin->sin_addr = inp->inp_laddr;
 994         crit_exit();
 995
 996         *nam = (struct sockaddr *)sin;
 997         return (0);
 998 }
 999
1000 void
1001 in_setsockaddr_dispatch(netmsg_t msg)
1002 {
1003         int error;
1004
1005         error = in_setsockaddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1006         lwkt_replymsg(&msg->lmsg, error);
1007 }
1008
1009 int
1010 in_setpeeraddr(struct socket *so, struct sockaddr **nam)
1011 {
1012         struct inpcb *inp;
1013         struct sockaddr_in *sin;
1014
1015         /*
1016          * Do the malloc first in case it blocks.
1017          */
1018         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
1019         sin->sin_family = AF_INET;
1020         sin->sin_len = sizeof *sin;
1021
1022         crit_enter();
1023         inp = so->so_pcb;
1024         if (!inp) {
1025                 crit_exit();
1026                 kfree(sin, M_SONAME);
1027                 return (ECONNRESET);
1028         }
1029         sin->sin_port = inp->inp_fport;
1030         sin->sin_addr = inp->inp_faddr;
1031         crit_exit();
1032
1033         *nam = (struct sockaddr *)sin;
1034         return (0);
1035 }
1036
1037 void
1038 in_setpeeraddr_dispatch(netmsg_t msg)
1039 {
1040         int error;
1041
1042         error = in_setpeeraddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1043         lwkt_replymsg(&msg->lmsg, error);
1044 }
1045
1046 void
1047 in_pcbnotifyall(struct inpcbhead *head, struct in_addr faddr, int err,
1048                 void (*notify)(struct inpcb *, int))
1049 {
1050         struct inpcb *inp, *ninp;
1051
1052         /*
1053          * note: if INP_PLACEMARKER is set we must ignore the rest of
1054          * the structure and skip it.
1055          */
1056         crit_enter();
1057         LIST_FOREACH_MUTABLE(inp, head, inp_list, ninp) {
1058                 if (inp->inp_flags & INP_PLACEMARKER)
1059                         continue;
1060 #ifdef INET6
1061                 if (!(inp->inp_vflag & INP_IPV4))
1062                         continue;
1063 #endif
1064                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
1065                     inp->inp_socket == NULL)
1066                         continue;
1067                 (*notify)(inp, err);            /* can remove inp from list! */
1068         }
1069         crit_exit();
1070 }
1071
1072 void
1073 in_pcbpurgeif0(struct inpcb *head, struct ifnet *ifp)
1074 {
1075         struct inpcb *inp;
1076         struct ip_moptions *imo;
1077         int i, gap;
1078
1079         for (inp = head; inp != NULL; inp = LIST_NEXT(inp, inp_list)) {
1080                 if (inp->inp_flags & INP_PLACEMARKER)
1081                         continue;
1082                 imo = inp->inp_moptions;
1083                 if ((inp->inp_vflag & INP_IPV4) && imo != NULL) {
1084                         /*
1085                          * Unselect the outgoing interface if it is being
1086                          * detached.
1087                          */
1088                         if (imo->imo_multicast_ifp == ifp)
1089                                 imo->imo_multicast_ifp = NULL;
1090
1091                         /*
1092                          * Drop multicast group membership if we joined
1093                          * through the interface being detached.
1094                          */
1095                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
1096                             i++) {
1097                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
1098                                         in_delmulti(imo->imo_membership[i]);
1099                                         gap++;
1100                                 } else if (gap != 0)
1101                                         imo->imo_membership[i - gap] =
1102                                             imo->imo_membership[i];
1103                         }
1104                         imo->imo_num_memberships -= gap;
1105                 }
1106         }
1107 }
1108
1109 /*
1110  * Check for alternatives when higher level complains
1111  * about service problems.  For now, invalidate cached
1112  * routing information.  If the route was created dynamically
1113  * (by a redirect), time to try a default gateway again.
1114  */
1115 void
1116 in_losing(struct inpcb *inp)
1117 {
1118         struct rtentry *rt;
1119         struct rt_addrinfo rtinfo;
1120
1121         if ((rt = inp->inp_route.ro_rt)) {
1122                 bzero(&rtinfo, sizeof(struct rt_addrinfo));
1123                 rtinfo.rti_info[RTAX_DST] = rt_key(rt);
1124                 rtinfo.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
1125                 rtinfo.rti_info[RTAX_NETMASK] = rt_mask(rt);
1126                 rtinfo.rti_flags = rt->rt_flags;
1127                 rt_missmsg(RTM_LOSING, &rtinfo, rt->rt_flags, 0);
1128                 if (rt->rt_flags & RTF_DYNAMIC)
1129                         rtrequest1_global(RTM_DELETE, &rtinfo, NULL, NULL);
1130                 inp->inp_route.ro_rt = NULL;
1131                 rtfree(rt);
1132                 /*
1133                  * A new route can be allocated
1134                  * the next time output is attempted.
1135                  */
1136         }
1137 }
1138
1139 /*
1140  * After a routing change, flush old routing
1141  * and allocate a (hopefully) better one.
1142  */
1143 void
1144 in_rtchange(struct inpcb *inp, int err)
1145 {
1146         if (inp->inp_route.ro_rt) {
1147                 rtfree(inp->inp_route.ro_rt);
1148                 inp->inp_route.ro_rt = NULL;
1149                 /*
1150                  * A new route can be allocated the next time
1151                  * output is attempted.
1152                  */
1153         }
1154 }
1155
1156 /*
1157  * Lookup a PCB based on the local address and port.
1158  */
1159 struct inpcb *
1160 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
1161                    u_int lport_arg, int wild_okay, struct ucred *cred)
1162 {
1163         struct inpcb *inp;
1164         int matchwild = 3, wildcard;
1165         u_short lport = lport_arg;
1166         struct inpcbporthead *porthash;
1167         struct inpcbport *phd;
1168         struct inpcb *match = NULL;
1169
1170         /*
1171          * If the porthashbase is shared across several cpus we need
1172          * to lock.
1173          */
1174         if (pcbinfo->porttoken)
1175                 lwkt_gettoken(pcbinfo->porttoken);
1176
1177         /*
1178          * Best fit PCB lookup.
1179          *
1180          * First see if this local port is in use by looking on the
1181          * port hash list.
1182          */
1183         porthash = &pcbinfo->porthashbase[
1184                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
1185         LIST_FOREACH(phd, porthash, phd_hash) {
1186                 if (phd->phd_port == lport)
1187                         break;
1188         }
1189         if (phd != NULL) {
1190                 /*
1191                  * Port is in use by one or more PCBs. Look for best
1192                  * fit.
1193                  */
1194                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
1195                         wildcard = 0;
1196 #ifdef INET6
1197                         if ((inp->inp_vflag & INP_IPV4) == 0)
1198                                 continue;
1199 #endif
1200                         if (inp->inp_faddr.s_addr != INADDR_ANY)
1201                                 wildcard++;
1202                         if (inp->inp_laddr.s_addr != INADDR_ANY) {
1203                                 if (laddr.s_addr == INADDR_ANY)
1204                                         wildcard++;
1205                                 else if (inp->inp_laddr.s_addr != laddr.s_addr)
1206                                         continue;
1207                         } else {
1208                                 if (laddr.s_addr != INADDR_ANY)
1209                                         wildcard++;
1210                         }
1211                         if (wildcard && !wild_okay)
1212                                 continue;
1213                         if (wildcard < matchwild &&
1214                             (cred == NULL ||
1215                              cred->cr_prison ==
1216                                         inp->inp_socket->so_cred->cr_prison)) {
1217                                 match = inp;
1218                                 matchwild = wildcard;
1219                                 if (matchwild == 0) {
1220                                         break;
1221                                 }
1222                         }
1223                 }
1224         }
1225         if (pcbinfo->porttoken)
1226                 lwkt_reltoken(pcbinfo->porttoken);
1227         return (match);
1228 }
1229
1230 static struct inpcb *
1231 inp_localgroup_lookup(const struct inpcbinfo *pcbinfo,
1232     struct in_addr laddr, uint16_t lport, uint32_t pkt_hash)
1233 {
1234         struct inpcb *local_wild = NULL;
1235         const struct inp_localgrphead *hdr;
1236         const struct inp_localgroup *grp;
1237
1238         hdr = &pcbinfo->localgrphashbase[
1239             INP_PCBLOCALGRPHASH(lport, pcbinfo->localgrphashmask)];
1240         pkt_hash >>= ncpus2_shift;
1241
1242         /*
1243          * Order of socket selection:
1244          * 1. non-wild.
1245          * 2. wild.
1246          *
1247          * NOTE:
1248          * - Local group does not contain jailed sockets
1249          * - Local group does not contain IPv4 mapped INET6 wild sockets
1250          */
1251         LIST_FOREACH(grp, hdr, il_list) {
1252 #ifdef INET6
1253                 if (!(grp->il_vflag & INP_IPV4))
1254                         continue;
1255 #endif
1256                 if (grp->il_lport == lport) {
1257                         int idx;
1258
1259                         idx = pkt_hash / grp->il_factor;
1260                         KASSERT(idx < grp->il_inpcnt && idx >= 0,
1261                             ("invalid hash %04x, cnt %d or fact %d",
1262                              pkt_hash, grp->il_inpcnt, grp->il_factor));
1263
1264                         if (grp->il_laddr.s_addr == laddr.s_addr)
1265                                 return grp->il_inp[idx];
1266                         else if (grp->il_laddr.s_addr == INADDR_ANY)
1267                                 local_wild = grp->il_inp[idx];
1268                 }
1269         }
1270         if (local_wild != NULL)
1271                 return local_wild;
1272         return NULL;
1273 }
1274
1275 /*
1276  * Lookup PCB in hash list.
1277  */
1278 struct inpcb *
1279 in_pcblookup_pkthash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1280     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1281     boolean_t wildcard, struct ifnet *ifp, const struct mbuf *m)
1282 {
1283         struct inpcbhead *head;
1284         struct inpcb *inp, *jinp=NULL;
1285         u_short fport = fport_arg, lport = lport_arg;
1286
1287         /*
1288          * First look for an exact match.
1289          */
1290         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr.s_addr, fport,
1291             laddr.s_addr, lport, pcbinfo->hashmask)];
1292         LIST_FOREACH(inp, head, inp_hash) {
1293 #ifdef INET6
1294                 if (!(inp->inp_vflag & INP_IPV4))
1295                         continue;
1296 #endif
1297                 if (in_hosteq(inp->inp_faddr, faddr) &&
1298                     in_hosteq(inp->inp_laddr, laddr) &&
1299                     inp->inp_fport == fport && inp->inp_lport == lport) {
1300                         /* found */
1301                         if (inp->inp_socket == NULL ||
1302                             inp->inp_socket->so_cred->cr_prison == NULL) {
1303                                 return (inp);
1304                         } else {
1305                                 if  (jinp == NULL)
1306                                         jinp = inp;
1307                         }
1308                 }
1309         }
1310         if (jinp != NULL)
1311                 return (jinp);
1312         if (wildcard) {
1313                 struct inpcb *local_wild = NULL;
1314                 struct inpcb *jinp_wild = NULL;
1315 #ifdef INET6
1316                 struct inpcb *local_wild_mapped = NULL;
1317 #endif
1318                 struct inpcontainer *ic;
1319                 struct inpcontainerhead *chead;
1320                 struct sockaddr_in jsin;
1321                 struct ucred *cred;
1322
1323                 /*
1324                  * Check local group first
1325                  */
1326                 if (pcbinfo->localgrphashbase != NULL &&
1327                     m != NULL && (m->m_flags & M_HASH) &&
1328                     !(ifp && ifp->if_type == IFT_FAITH)) {
1329                         inp = inp_localgroup_lookup(pcbinfo,
1330                             laddr, lport, m->m_pkthdr.hash);
1331                         if (inp != NULL)
1332                                 return inp;
1333                 }
1334
1335                 /*
1336                  * Order of socket selection:
1337                  * 1. non-jailed, non-wild.
1338                  * 2. non-jailed, wild.
1339                  * 3. jailed, non-wild.
1340                  * 4. jailed, wild.
1341                  */
1342                 jsin.sin_family = AF_INET;
1343                 chead = &pcbinfo->wildcardhashbase[
1344                     INP_PCBWILDCARDHASH(lport, pcbinfo->wildcardhashmask)];
1345                 LIST_FOREACH(ic, chead, ic_list) {
1346                         inp = ic->ic_inp;
1347                         jsin.sin_addr.s_addr = laddr.s_addr;
1348 #ifdef INET6
1349                         if (!(inp->inp_vflag & INP_IPV4))
1350                                 continue;
1351 #endif
1352                         if (inp->inp_socket != NULL)
1353                                 cred = inp->inp_socket->so_cred;
1354                         else
1355                                 cred = NULL;
1356                         if (cred != NULL && jailed(cred)) {
1357                                 if (jinp != NULL)
1358                                         continue;
1359                                 else
1360                                         if (!jailed_ip(cred->cr_prison,
1361                                             (struct sockaddr *)&jsin))
1362                                                 continue;
1363                         }
1364                         if (inp->inp_lport == lport) {
1365                                 if (ifp && ifp->if_type == IFT_FAITH &&
1366                                     !(inp->inp_flags & INP_FAITH))
1367                                         continue;
1368                                 if (inp->inp_laddr.s_addr == laddr.s_addr) {
1369                                         if (cred != NULL && jailed(cred))
1370                                                 jinp = inp;
1371                                         else
1372                                                 return (inp);
1373                                 }
1374                                 if (inp->inp_laddr.s_addr == INADDR_ANY) {
1375 #ifdef INET6
1376                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1377                                                              AF_INET6))
1378                                                 local_wild_mapped = inp;
1379                                         else
1380 #endif
1381                                                 if (cred != NULL &&
1382                                                     jailed(cred))
1383                                                         jinp_wild = inp;
1384                                                 else
1385                                                         local_wild = inp;
1386                                 }
1387                         }
1388                 }
1389                 if (local_wild != NULL)
1390                         return (local_wild);
1391 #ifdef INET6
1392                 if (local_wild_mapped != NULL)
1393                         return (local_wild_mapped);
1394 #endif
1395                 if (jinp != NULL)
1396                         return (jinp);
1397                 return (jinp_wild);
1398         }
1399
1400         /*
1401          * Not found.
1402          */
1403         return (NULL);
1404 }
1405
1406 struct inpcb *
1407 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1408     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1409     boolean_t wildcard, struct ifnet *ifp)
1410 {
1411         return in_pcblookup_pkthash(pcbinfo, faddr, fport_arg,
1412             laddr, lport_arg, wildcard, ifp, NULL);
1413 }
1414
1415 /*
1416  * Insert PCB into connection hash table.
1417  */
1418 void
1419 in_pcbinsconnhash(struct inpcb *inp)
1420 {
1421         struct inpcbinfo *pcbinfo = inp->inp_cpcbinfo;
1422         struct inpcbhead *bucket;
1423         u_int32_t hashkey_faddr, hashkey_laddr;
1424
1425 #ifdef INET6
1426         if (inp->inp_vflag & INP_IPV6) {
1427                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX JH */;
1428                 hashkey_laddr = inp->in6p_laddr.s6_addr32[3] /* XXX JH */;
1429         } else {
1430 #endif
1431                 hashkey_faddr = inp->inp_faddr.s_addr;
1432                 hashkey_laddr = inp->inp_laddr.s_addr;
1433 #ifdef INET6
1434         }
1435 #endif
1436
1437         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1438                 ("already on wildcardhash"));
1439         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1440                 ("already on connhash"));
1441         inp->inp_flags |= INP_CONNECTED;
1442
1443         /*
1444          * Insert into the connection hash table.
1445          */
1446         bucket = &pcbinfo->hashbase[INP_PCBCONNHASH(hashkey_faddr,
1447             inp->inp_fport, hashkey_laddr, inp->inp_lport, pcbinfo->hashmask)];
1448         LIST_INSERT_HEAD(bucket, inp, inp_hash);
1449 }
1450
1451 /*
1452  * Remove PCB from connection hash table.
1453  */
1454 void
1455 in_pcbremconnhash(struct inpcb *inp)
1456 {
1457         KASSERT(inp->inp_flags & INP_CONNECTED, ("inp not connected"));
1458         LIST_REMOVE(inp, inp_hash);
1459         inp->inp_flags &= ~INP_CONNECTED;
1460 }
1461
1462 /*
1463  * Insert PCB into port hash table.
1464  */
1465 int
1466 in_pcbinsporthash(struct inpcb *inp)
1467 {
1468         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1469         struct inpcbporthead *pcbporthash;
1470         struct inpcbport *phd;
1471
1472         /*
1473          * If the porthashbase is shared across several cpus we need
1474          * to lock.
1475          */
1476         if (pcbinfo->porttoken)
1477                 lwkt_gettoken(pcbinfo->porttoken);
1478
1479         /*
1480          * Insert into the port hash table.
1481          */
1482         pcbporthash = &pcbinfo->porthashbase[
1483             INP_PCBPORTHASH(inp->inp_lport, pcbinfo->porthashmask)];
1484
1485         /* Go through port list and look for a head for this lport. */
1486         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1487                 if (phd->phd_port == inp->inp_lport)
1488                         break;
1489         }
1490
1491         /* If none exists, malloc one and tack it on. */
1492         if (phd == NULL) {
1493                 KKASSERT(pcbinfo->portsave != NULL);
1494                 phd = pcbinfo->portsave;
1495                 pcbinfo->portsave = NULL;
1496                 phd->phd_port = inp->inp_lport;
1497                 LIST_INIT(&phd->phd_pcblist);
1498                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1499         }
1500
1501         inp->inp_phd = phd;
1502         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1503
1504         if (pcbinfo->porttoken)
1505                 lwkt_reltoken(pcbinfo->porttoken);
1506         if (pcbinfo->portsave == NULL) {
1507                 pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave),
1508                                             M_PCB, M_INTWAIT | M_ZERO);
1509         }
1510         return (0);
1511 }
1512
1513 static struct inp_localgroup *
1514 inp_localgroup_alloc(struct inp_localgrphead *hdr, u_char vflag,
1515     uint16_t port, const union in_dependaddr *addr, int size)
1516 {
1517         struct inp_localgroup *grp;
1518
1519         grp = kmalloc(__offsetof(struct inp_localgroup, il_inp[size]),
1520             M_TEMP, M_INTWAIT | M_ZERO);
1521         grp->il_vflag = vflag;
1522         grp->il_lport = port;
1523         grp->il_dependladdr = *addr;
1524         grp->il_inpsiz = size;
1525
1526         LIST_INSERT_HEAD(hdr, grp, il_list);
1527
1528         return grp;
1529 }
1530
1531 static void
1532 inp_localgroup_free(struct inp_localgroup *grp)
1533 {
1534         LIST_REMOVE(grp, il_list);
1535         kfree(grp, M_TEMP);
1536 }
1537
1538 static struct inp_localgroup *
1539 inp_localgroup_resize(struct inp_localgrphead *hdr,
1540     struct inp_localgroup *old_grp, int size)
1541 {
1542         struct inp_localgroup *grp;
1543         int i;
1544
1545         grp = inp_localgroup_alloc(hdr, old_grp->il_vflag,
1546             old_grp->il_lport, &old_grp->il_dependladdr, size);
1547
1548         KASSERT(old_grp->il_inpcnt < grp->il_inpsiz,
1549             ("invalid new local group size %d and old local group count %d",
1550              grp->il_inpsiz, old_grp->il_inpcnt));
1551         for (i = 0; i < old_grp->il_inpcnt; ++i)
1552                 grp->il_inp[i] = old_grp->il_inp[i];
1553         grp->il_inpcnt = old_grp->il_inpcnt;
1554         grp->il_factor = old_grp->il_factor;
1555
1556         inp_localgroup_free(old_grp);
1557
1558         return grp;
1559 }
1560
1561 static void
1562 inp_localgroup_factor(struct inp_localgroup *grp)
1563 {
1564         grp->il_factor =
1565             ((uint32_t)(0xffff >> ncpus2_shift) / grp->il_inpcnt) + 1;
1566         KASSERT(grp->il_factor != 0, ("invalid local group factor, "
1567             "ncpus2_shift %d, inpcnt %d", ncpus2_shift, grp->il_inpcnt));
1568 }
1569
1570 static void
1571 in_pcbinslocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1572 {
1573         struct inp_localgrphead *hdr;
1574         struct inp_localgroup *grp;
1575         struct ucred *cred;
1576
1577         if (pcbinfo->localgrphashbase == NULL)
1578                 return;
1579
1580         /*
1581          * XXX don't allow jailed socket to join local group
1582          */
1583         if (inp->inp_socket != NULL)
1584                 cred = inp->inp_socket->so_cred;
1585         else
1586                 cred = NULL;
1587         if (cred != NULL && jailed(cred))
1588                 return;
1589
1590 #ifdef INET6
1591         /*
1592          * XXX don't allow IPv4 mapped INET6 wild socket
1593          */
1594         if ((inp->inp_vflag & INP_IPV4) &&
1595             inp->inp_laddr.s_addr == INADDR_ANY &&
1596             INP_CHECK_SOCKAF(inp->inp_socket, AF_INET6))
1597                 return;
1598 #endif
1599
1600         hdr = &pcbinfo->localgrphashbase[
1601             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1602
1603         LIST_FOREACH(grp, hdr, il_list) {
1604                 if (grp->il_vflag == inp->inp_vflag &&
1605                     grp->il_lport == inp->inp_lport &&
1606                     memcmp(&grp->il_dependladdr,
1607                         &inp->inp_inc.inc_ie.ie_dependladdr,
1608                         sizeof(grp->il_dependladdr)) == 0) {
1609                         break;
1610                 }
1611         }
1612         if (grp == NULL) {
1613                 /* Create new local group */
1614                 grp = inp_localgroup_alloc(hdr, inp->inp_vflag,
1615                     inp->inp_lport, &inp->inp_inc.inc_ie.ie_dependladdr,
1616                     INP_LOCALGROUP_SIZMIN);
1617         } else if (grp->il_inpcnt == grp->il_inpsiz) {
1618                 if (grp->il_inpsiz >= INP_LOCALGROUP_SIZMAX) {
1619                         static int limit_logged = 0;
1620
1621                         if (!limit_logged) {
1622                                 limit_logged = 1;
1623                                 kprintf("local group port %d, "
1624                                     "limit reached\n", ntohs(grp->il_lport));
1625                         }
1626                         return;
1627                 }
1628
1629                 /* Expand this local group */
1630                 grp = inp_localgroup_resize(hdr, grp, grp->il_inpsiz * 2);
1631         }
1632
1633         KASSERT(grp->il_inpcnt < grp->il_inpsiz,
1634             ("invalid local group size %d and count %d",
1635              grp->il_inpsiz, grp->il_inpcnt));
1636         grp->il_inp[grp->il_inpcnt] = inp;
1637         grp->il_inpcnt++;
1638         inp_localgroup_factor(grp);
1639 }
1640
1641 void
1642 in_pcbinswildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1643 {
1644         struct inpcontainer *ic;
1645         struct inpcontainerhead *bucket;
1646
1647         in_pcbinslocalgrphash_oncpu(inp, pcbinfo);
1648
1649         bucket = &pcbinfo->wildcardhashbase[
1650             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1651
1652         ic = kmalloc(sizeof(struct inpcontainer), M_TEMP, M_INTWAIT);
1653         ic->ic_inp = inp;
1654         LIST_INSERT_HEAD(bucket, ic, ic_list);
1655 }
1656
1657 /*
1658  * Insert PCB into wildcard hash table.
1659  */
1660 void
1661 in_pcbinswildcardhash(struct inpcb *inp)
1662 {
1663         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1664
1665         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1666                 ("already on connhash"));
1667         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1668                 ("already on wildcardhash"));
1669         inp->inp_flags |= INP_WILDCARD;
1670
1671         in_pcbinswildcardhash_oncpu(inp, pcbinfo);
1672 }
1673
1674 static void
1675 in_pcbremlocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1676 {
1677         struct inp_localgrphead *hdr;
1678         struct inp_localgroup *grp;
1679
1680         if (pcbinfo->localgrphashbase == NULL)
1681                 return;
1682
1683         hdr = &pcbinfo->localgrphashbase[
1684             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1685
1686         LIST_FOREACH(grp, hdr, il_list) {
1687                 int i;
1688
1689                 for (i = 0; i < grp->il_inpcnt; ++i) {
1690                         if (grp->il_inp[i] != inp)
1691                                 continue;
1692
1693                         if (grp->il_inpcnt == 1) {
1694                                 /* Free this local group */
1695                                 inp_localgroup_free(grp);
1696                         } else {
1697                                 /* Pull up inpcbs */
1698                                 for (; i + 1 < grp->il_inpcnt; ++i)
1699                                         grp->il_inp[i] = grp->il_inp[i + 1];
1700                                 grp->il_inpcnt--;
1701                                 inp_localgroup_factor(grp);
1702
1703                                 if (grp->il_inpsiz > INP_LOCALGROUP_SIZMIN &&
1704                                     grp->il_inpcnt <= (grp->il_inpsiz / 4)) {
1705                                         /* Shrink this local group */
1706                                         grp = inp_localgroup_resize(hdr, grp,
1707                                             grp->il_inpsiz / 2);
1708                                 }
1709                         }
1710                         return;
1711                 }
1712         }
1713 }
1714
1715 void
1716 in_pcbremwildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1717 {
1718         struct inpcontainer *ic;
1719         struct inpcontainerhead *head;
1720
1721         in_pcbremlocalgrphash_oncpu(inp, pcbinfo);
1722
1723         /* find bucket */
1724         head = &pcbinfo->wildcardhashbase[
1725             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1726
1727         LIST_FOREACH(ic, head, ic_list) {
1728                 if (ic->ic_inp == inp)
1729                         goto found;
1730         }
1731         return;                 /* not found! */
1732
1733 found:
1734         LIST_REMOVE(ic, ic_list);       /* remove container from bucket chain */
1735         kfree(ic, M_TEMP);              /* deallocate container */
1736 }
1737
1738 /*
1739  * Remove PCB from wildcard hash table.
1740  */
1741 void
1742 in_pcbremwildcardhash(struct inpcb *inp)
1743 {
1744         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1745
1746         KASSERT(inp->inp_flags & INP_WILDCARD, ("inp not wildcard"));
1747         in_pcbremwildcardhash_oncpu(inp, pcbinfo);
1748         inp->inp_flags &= ~INP_WILDCARD;
1749 }
1750
1751 /*
1752  * Remove PCB from various lists.
1753  */
1754 void
1755 in_pcbremlists(struct inpcb *inp)
1756 {
1757         struct inpcbinfo *pcbinfo;
1758
1759         if (inp->inp_lport) {
1760                 struct inpcbport *phd;
1761
1762                 pcbinfo = inp->inp_pcbinfo;
1763                 if (pcbinfo->porttoken)
1764                         lwkt_gettoken(pcbinfo->porttoken);
1765
1766                 phd = inp->inp_phd;
1767                 LIST_REMOVE(inp, inp_portlist);
1768                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1769                         LIST_REMOVE(phd, phd_hash);
1770                         kfree(phd, M_PCB);
1771                 }
1772                 if (pcbinfo->porttoken)
1773                         lwkt_reltoken(pcbinfo->porttoken);
1774         }
1775         if (inp->inp_flags & INP_WILDCARD) {
1776                 in_pcbremwildcardhash(inp);
1777         } else if (inp->inp_flags & INP_CONNECTED) {
1778                 in_pcbremconnhash(inp);
1779         }
1780         LIST_REMOVE(inp, inp_list);
1781         inp->inp_pcbinfo->ipi_count--;
1782 }
1783
1784 int
1785 prison_xinpcb(struct thread *td, struct inpcb *inp)
1786 {
1787         struct ucred *cr;
1788
1789         if (td->td_proc == NULL)
1790                 return (0);
1791         cr = td->td_proc->p_ucred;
1792         if (cr->cr_prison == NULL)
1793                 return (0);
1794         if (inp->inp_socket && inp->inp_socket->so_cred &&
1795             inp->inp_socket->so_cred->cr_prison &&
1796             cr->cr_prison == inp->inp_socket->so_cred->cr_prison)
1797                 return (0);
1798         return (1);
1799 }
1800
1801 int
1802 in_pcblist_global(SYSCTL_HANDLER_ARGS)
1803 {
1804         struct inpcbinfo *pcbinfo = arg1;
1805         struct inpcb *inp, *marker;
1806         struct xinpcb xi;
1807         int error, i, n;
1808
1809         /*
1810          * The process of preparing the TCB list is too time-consuming and
1811          * resource-intensive to repeat twice on every request.
1812          */
1813         if (req->oldptr == NULL) {
1814                 n = pcbinfo->ipi_count;
1815                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1816                 return 0;
1817         }
1818
1819         if (req->newptr != NULL)
1820                 return EPERM;
1821
1822         /*
1823          * OK, now we're committed to doing something.  Re-fetch ipi_count
1824          * after obtaining the generation count.
1825          */
1826         n = pcbinfo->ipi_count;
1827
1828         marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1829         marker->inp_flags |= INP_PLACEMARKER;
1830         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, marker, inp_list);
1831
1832         i = 0;
1833         error = 0;
1834
1835         while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1836                 LIST_REMOVE(marker, inp_list);
1837                 LIST_INSERT_AFTER(inp, marker, inp_list);
1838
1839                 if (inp->inp_flags & INP_PLACEMARKER)
1840                         continue;
1841                 if (prison_xinpcb(req->td, inp))
1842                         continue;
1843                 bzero(&xi, sizeof xi);
1844                 xi.xi_len = sizeof xi;
1845                 bcopy(inp, &xi.xi_inp, sizeof *inp);
1846                 if (inp->inp_socket)
1847                         sotoxsocket(inp->inp_socket, &xi.xi_socket);
1848                 if ((error = SYSCTL_OUT(req, &xi, sizeof xi)) != 0)
1849                         break;
1850                 ++i;
1851         }
1852         LIST_REMOVE(marker, inp_list);
1853         if (error == 0 && i < n) {
1854                 bzero(&xi, sizeof xi);
1855                 xi.xi_len = sizeof xi;
1856                 while (i < n) {
1857                         error = SYSCTL_OUT(req, &xi, sizeof xi);
1858                         ++i;
1859                 }
1860         }
1861         kfree(marker, M_TEMP);
1862         return(error);
1863 }
1864
1865 int
1866 in_pcblist_global_nomarker(SYSCTL_HANDLER_ARGS, struct xinpcb **xi0, int *nxi0)
1867 {
1868         struct inpcbinfo *pcbinfo = arg1;
1869         struct inpcb *inp;
1870         struct xinpcb *xi;
1871         int nxi;
1872
1873         *nxi0 = 0;
1874         *xi0 = NULL;
1875
1876         /*
1877          * The process of preparing the PCB list is too time-consuming and
1878          * resource-intensive to repeat twice on every request.
1879          */
1880         if (req->oldptr == NULL) {
1881                 int n = pcbinfo->ipi_count;
1882
1883                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1884                 return 0;
1885         }
1886
1887         if (req->newptr != NULL)
1888                 return EPERM;
1889
1890         if (pcbinfo->ipi_count == 0)
1891                 return 0;
1892
1893         nxi = 0;
1894         xi = kmalloc(pcbinfo->ipi_count * sizeof(*xi), M_TEMP,
1895                      M_WAITOK | M_ZERO | M_NULLOK);
1896         if (xi == NULL)
1897                 return ENOMEM;
1898
1899         LIST_FOREACH(inp, &pcbinfo->pcblisthead, inp_list) {
1900                 struct xinpcb *xi_ptr = &xi[nxi];
1901
1902                 if (prison_xinpcb(req->td, inp))
1903                         continue;
1904
1905                 xi_ptr->xi_len = sizeof(*xi_ptr);
1906                 bcopy(inp, &xi_ptr->xi_inp, sizeof(*inp));
1907                 if (inp->inp_socket)
1908                         sotoxsocket(inp->inp_socket, &xi_ptr->xi_socket);
1909                 ++nxi;
1910         }
1911
1912         if (nxi == 0) {
1913                 kfree(xi, M_TEMP);
1914                 return 0;
1915         }
1916
1917         *nxi0 = nxi;
1918         *xi0 = xi;
1919
1920         return 0;
1921 }
1922
1923 void
1924 in_savefaddr(struct socket *so, const struct sockaddr *faddr)
1925 {
1926         struct sockaddr_in *sin;
1927
1928         KASSERT(faddr->sa_family == AF_INET,
1929             ("not AF_INET faddr %d", faddr->sa_family));
1930
1931         sin = kmalloc(sizeof(*sin), M_SONAME, M_WAITOK | M_ZERO);
1932         sin->sin_family = AF_INET;
1933         sin->sin_len = sizeof(*sin);
1934         sin->sin_port = ((const struct sockaddr_in *)faddr)->sin_port;
1935         sin->sin_addr = ((const struct sockaddr_in *)faddr)->sin_addr;
1936
1937         so->so_faddr = (struct sockaddr *)sin;
1938 }