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  * 3. 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|M_NULLOK);
 235         if (inp == NULL)
 236                 return (ENOMEM);
 237         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 238         inp->inp_pcbinfo = inp->inp_cpcbinfo = pcbinfo;
 239         inp->inp_socket = so;
 240 #ifdef IPSEC
 241         error = ipsec_init_policy(so, &inp->inp_sp);
 242         if (error != 0) {
 243                 kfree(inp, M_PCB);
 244                 return (error);
 245         }
 246 #endif
 247 #ifdef INET6
 248         if (INP_SOCKAF(so) == AF_INET6 && ip6_v6only)
 249                 inp->inp_flags |= IN6P_IPV6_V6ONLY;
 250         if (ip6_auto_flowlabel)
 251                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 252 #endif
 253         soreference(so);
 254         so->so_pcb = inp;
 255         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 256         pcbinfo->ipi_count++;
 257         return (0);
 258 }
 259
 260 /*
 261  * Unlink a pcb with the intention of moving it to another cpu with a
 262  * different pcbinfo.  While unlinked nothing should attempt to dereference
 263  * inp_pcbinfo, NULL it out so we assert if it does.
 264  */
 265 void
 266 in_pcbunlink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 267 {
 268         KASSERT(inp->inp_pcbinfo == pcbinfo, ("pcbinfo mismatch"));
 269         KASSERT(inp->inp_cpcbinfo == pcbinfo, ("cpcbinfo mismatch"));
 270         KASSERT((inp->inp_flags & (INP_WILDCARD | INP_CONNECTED)) == 0,
 271             ("already linked"));
 272
 273         LIST_REMOVE(inp, inp_list);
 274         pcbinfo->ipi_count--;
 275         inp->inp_pcbinfo = NULL;
 276         inp->inp_cpcbinfo = NULL;
 277 }
 278
 279 /*
 280  * Relink a pcb into a new pcbinfo.
 281  */
 282 void
 283 in_pcblink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 284 {
 285         KASSERT(inp->inp_pcbinfo == NULL, ("has pcbinfo"));
 286         KASSERT(inp->inp_cpcbinfo == NULL, ("has cpcbinfo"));
 287         KASSERT((inp->inp_flags & (INP_WILDCARD | INP_CONNECTED)) == 0,
 288             ("already linked"));
 289
 290         inp->inp_cpcbinfo = pcbinfo;
 291         inp->inp_pcbinfo = pcbinfo;
 292         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 293         pcbinfo->ipi_count++;
 294 }
 295
 296 int
 297 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 298 {
 299         struct socket *so = inp->inp_socket;
 300         unsigned short *lastport;
 301         struct sockaddr_in *sin;
 302         struct sockaddr_in jsin;
 303         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 304         struct ucred *cred = NULL;
 305         u_short lport = 0;
 306         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 307         int error;
 308
 309         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 310                 return (EADDRNOTAVAIL);
 311         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 312                 return (EINVAL);        /* already bound */
 313
 314         if (!(so->so_options & (SO_REUSEADDR|SO_REUSEPORT)))
 315                 wild = 1;    /* neither SO_REUSEADDR nor SO_REUSEPORT is set */
 316         if (td->td_proc)
 317                 cred = td->td_proc->p_ucred;
 318
 319         /*
 320          * This has to be atomic.  If the porthash is shared across multiple
 321          * protocol threads (aka tcp) then the token will be non-NULL.
 322          */
 323         if (pcbinfo->porttoken)
 324                 lwkt_gettoken(pcbinfo->porttoken);
 325
 326         if (nam != NULL) {
 327                 sin = (struct sockaddr_in *)nam;
 328                 if (nam->sa_len != sizeof *sin) {
 329                         error = EINVAL;
 330                         goto done;
 331                 }
 332 #ifdef notdef
 333                 /*
 334                  * We should check the family, but old programs
 335                  * incorrectly fail to initialize it.
 336                  */
 337                 if (sin->sin_family != AF_INET) {
 338                         error = EAFNOSUPPORT;
 339                         goto done;
 340                 }
 341 #endif
 342                 if (!prison_replace_wildcards(td, nam)) {
 343                         error = EINVAL;
 344                         goto done;
 345                 }
 346                 lport = sin->sin_port;
 347                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 348                         /*
 349                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 350                          * allow complete duplication of binding if
 351                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 352                          * and a multicast address is bound on both
 353                          * new and duplicated sockets.
 354                          */
 355                         if (so->so_options & SO_REUSEADDR)
 356                                 reuseport = SO_REUSEADDR | SO_REUSEPORT;
 357                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 358                         sin->sin_port = 0;              /* yech... */
 359                         bzero(&sin->sin_zero, sizeof sin->sin_zero);
 360                         if (ifa_ifwithaddr((struct sockaddr *)sin) == NULL) {
 361                                 error = EADDRNOTAVAIL;
 362                                 goto done;
 363                         }
 364                 }
 365                 if (lport != 0) {
 366                         struct inpcb *t;
 367
 368                         /* GROSS */
 369                         if (ntohs(lport) < IPPORT_RESERVED &&
 370                             cred &&
 371                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0)) {
 372                                 error = EACCES;
 373                                 goto done;
 374                         }
 375                         if (so->so_cred->cr_uid != 0 &&
 376                             !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 377                                 t = in_pcblookup_local(pcbinfo,
 378                                                        sin->sin_addr,
 379                                                        lport,
 380                                                        INPLOOKUP_WILDCARD,
 381                                                        cred);
 382                                 if (t &&
 383                                     (!in_nullhost(sin->sin_addr) ||
 384                                      !in_nullhost(t->inp_laddr) ||
 385                                      (t->inp_socket->so_options &
 386                                          SO_REUSEPORT) == 0) &&
 387                                     (so->so_cred->cr_uid !=
 388                                      t->inp_socket->so_cred->cr_uid)) {
 389 #ifdef INET6
 390                                         if (!in_nullhost(sin->sin_addr) ||
 391                                             !in_nullhost(t->inp_laddr) ||
 392                                             INP_SOCKAF(so) ==
 393                                             INP_SOCKAF(t->inp_socket))
 394 #endif
 395                                         {
 396                                                 error = EADDRINUSE;
 397                                                 goto done;
 398                                         }
 399                                 }
 400                         }
 401                         if (cred && !prison_replace_wildcards(td, nam)) {
 402                                 error = EADDRNOTAVAIL;
 403                                 goto done;
 404                         }
 405                         t = in_pcblookup_local(pcbinfo, sin->sin_addr, lport,
 406                                                wild, cred);
 407                         if (t && !(reuseport & t->inp_socket->so_options)) {
 408 #ifdef INET6
 409                                 if (!in_nullhost(sin->sin_addr) ||
 410                                     !in_nullhost(t->inp_laddr) ||
 411                                     INP_SOCKAF(so) == INP_SOCKAF(t->inp_socket))
 412 #endif
 413                                 {
 414                                         error = EADDRINUSE;
 415                                         goto done;
 416                                 }
 417                         }
 418                 }
 419                 inp->inp_laddr = sin->sin_addr;
 420         }
 421
 422         jsin.sin_family = AF_INET;
 423         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 424         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 425                 inp->inp_laddr.s_addr = INADDR_ANY;
 426                 error = EINVAL;
 427                 goto done;
 428         }
 429         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 430
 431         if (lport == 0) {
 432                 ushort first, last;
 433                 int count;
 434
 435                 inp->inp_flags |= INP_ANONPORT;
 436
 437                 if (inp->inp_flags & INP_HIGHPORT) {
 438                         first = ipport_hifirstauto;     /* sysctl */
 439                         last  = ipport_hilastauto;
 440                         lastport = &pcbinfo->lasthi;
 441                 } else if (inp->inp_flags & INP_LOWPORT) {
 442                         if (cred &&
 443                             (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 444                                 inp->inp_laddr.s_addr = INADDR_ANY;
 445                                 goto done;
 446                         }
 447                         first = ipport_lowfirstauto;    /* 1023 */
 448                         last  = ipport_lowlastauto;     /* 600 */
 449                         lastport = &pcbinfo->lastlow;
 450                 } else {
 451                         first = ipport_firstauto;       /* sysctl */
 452                         last  = ipport_lastauto;
 453                         lastport = &pcbinfo->lastport;
 454                 }
 455                 /*
 456                  * Simple check to ensure all ports are not used up causing
 457                  * a deadlock here.
 458                  *
 459                  * We split the two cases (up and down) so that the direction
 460                  * is not being tested on each round of the loop.
 461                  */
 462                 if (first > last) {
 463                         /*
 464                          * counting down
 465                          */
 466                         count = first - last;
 467
 468                         do {
 469                                 if (count-- < 0) {      /* completely used? */
 470                                         inp->inp_laddr.s_addr = INADDR_ANY;
 471                                         error = EADDRNOTAVAIL;
 472                                         goto done;
 473                                 }
 474                                 --*lastport;
 475                                 if (*lastport > first || *lastport < last)
 476                                         *lastport = first;
 477                                 lport = htons(*lastport);
 478                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 479                                                     lport, wild, cred));
 480                 } else {
 481                         /*
 482                          * counting up
 483                          */
 484                         count = last - first;
 485
 486                         do {
 487                                 if (count-- < 0) {      /* completely used? */
 488                                         inp->inp_laddr.s_addr = INADDR_ANY;
 489                                         error = EADDRNOTAVAIL;
 490                                         goto done;
 491                                 }
 492                                 ++*lastport;
 493                                 if (*lastport < first || *lastport > last)
 494                                         *lastport = first;
 495                                 lport = htons(*lastport);
 496                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 497                                                     lport, wild, cred));
 498                 }
 499         }
 500         inp->inp_lport = lport;
 501
 502         if (in_pcbinsporthash(inp) != 0) {
 503                 inp->inp_laddr.s_addr = INADDR_ANY;
 504                 inp->inp_lport = 0;
 505                 error = EAGAIN;
 506                 goto done;
 507         }
 508         error = 0;
 509 done:
 510         if (pcbinfo->porttoken)
 511                 lwkt_reltoken(pcbinfo->porttoken);
 512         return error;
 513 }
 514
 515 static struct inpcb *
 516 in_pcblookup_addrport(struct inpcbinfo *pcbinfo, struct in_addr laddr,
 517     u_short lport, struct in_addr faddr, u_short fport, struct ucred *cred)
 518 {
 519         struct inpcb *inp;
 520         struct inpcbporthead *porthash;
 521         struct inpcbport *phd;
 522         struct inpcb *match = NULL;
 523
 524         /*
 525          * If the porthashbase is shared across several cpus, it must
 526          * have been locked.
 527          */
 528         if (pcbinfo->porttoken)
 529                 ASSERT_LWKT_TOKEN_HELD(pcbinfo->porttoken);
 530
 531         /*
 532          * Best fit PCB lookup.
 533          *
 534          * First see if this local port is in use by looking on the
 535          * port hash list.
 536          */
 537         porthash = &pcbinfo->porthashbase[
 538                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
 539         LIST_FOREACH(phd, porthash, phd_hash) {
 540                 if (phd->phd_port == lport)
 541                         break;
 542         }
 543         if (phd != NULL) {
 544                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 545 #ifdef INET6
 546                         if ((inp->inp_vflag & INP_IPV4) == 0)
 547                                 continue;
 548 #endif
 549                         if (inp->inp_laddr.s_addr != INADDR_ANY &&
 550                             inp->inp_laddr.s_addr != laddr.s_addr)
 551                                 continue;
 552
 553                         if (inp->inp_faddr.s_addr != INADDR_ANY &&
 554                             inp->inp_faddr.s_addr != faddr.s_addr)
 555                                 continue;
 556
 557                         if (inp->inp_fport != 0 && inp->inp_fport != fport)
 558                                 continue;
 559
 560                         if (cred == NULL ||
 561                             cred->cr_prison ==
 562                             inp->inp_socket->so_cred->cr_prison) {
 563                                 match = inp;
 564                                 break;
 565                         }
 566                 }
 567         }
 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         jsin.sin_family = AF_INET;
 596         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 597         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 598                 inp->inp_laddr.s_addr = INADDR_ANY;
 599                 return (EINVAL);
 600         }
 601         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 602
 603         inp->inp_flags |= INP_ANONPORT;
 604
 605         if (inp->inp_flags & INP_HIGHPORT) {
 606                 first = ipport_hifirstauto;     /* sysctl */
 607                 last  = ipport_hilastauto;
 608                 lastport = &pcbinfo->lasthi;
 609         } else if (inp->inp_flags & INP_LOWPORT) {
 610                 if (cred &&
 611                     (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 612                         inp->inp_laddr.s_addr = INADDR_ANY;
 613                         return (error);
 614                 }
 615                 first = ipport_lowfirstauto;    /* 1023 */
 616                 last  = ipport_lowlastauto;     /* 600 */
 617                 lastport = &pcbinfo->lastlow;
 618         } else {
 619                 first = ipport_firstauto;       /* sysctl */
 620                 last  = ipport_lastauto;
 621                 lastport = &pcbinfo->lastport;
 622         }
 623
 624         /*
 625          * This has to be atomic.  If the porthash is shared across multiple
 626          * protocol threads (aka tcp) then the token will be non-NULL.
 627          */
 628         if (pcbinfo->porttoken)
 629                 lwkt_gettoken(pcbinfo->porttoken);
 630
 631 again:
 632         /*
 633          * Simple check to ensure all ports are not used up causing
 634          * a deadlock here.
 635          *
 636          * We split the two cases (up and down) so that the direction
 637          * is not being tested on each round of the loop.
 638          */
 639         if (first > last) {
 640                 /*
 641                  * counting down
 642                  */
 643                 count = first - last;
 644
 645                 do {
 646                         if (count-- < 0) {      /* completely used? */
 647                                 inp->inp_laddr.s_addr = INADDR_ANY;
 648                                 error = EADDRNOTAVAIL;
 649                                 goto done;
 650                         }
 651                         --*lastport;
 652                         if (*lastport > first || *lastport < last)
 653                                 *lastport = first;
 654                         lport = htons(*lastport);
 655                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 656                                 sin->sin_addr, sin->sin_port, cred));
 657         } else {
 658                 /*
 659                  * counting up
 660                  */
 661                 count = last - first;
 662
 663                 do {
 664                         if (count-- < 0) {      /* completely used? */
 665                                 inp->inp_laddr.s_addr = INADDR_ANY;
 666                                 error = EADDRNOTAVAIL;
 667                                 goto done;
 668                         }
 669                         ++*lastport;
 670                         if (*lastport < first || *lastport > last)
 671                                 *lastport = first;
 672                         lport = htons(*lastport);
 673                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 674                                 sin->sin_addr, sin->sin_port, cred));
 675         }
 676
 677         /* This could happen on loopback interface */
 678         if (sin->sin_port == lport &&
 679             sin->sin_addr.s_addr == inp->inp_laddr.s_addr) {
 680                 if (dup) {
 681                         /*
 682                          * Duplicate again; give up
 683                          */
 684                         inp->inp_laddr.s_addr = INADDR_ANY;
 685                         error = EADDRNOTAVAIL;
 686                         goto done;
 687                 }
 688                 dup = 1;
 689                 goto again;
 690         }
 691         inp->inp_lport = lport;
 692
 693         if (in_pcbinsporthash(inp) != 0) {
 694                 inp->inp_laddr.s_addr = INADDR_ANY;
 695                 inp->inp_lport = 0;
 696                 error = EAGAIN;
 697                 goto done;
 698         }
 699         error = 0;
 700 done:
 701         if (pcbinfo->porttoken)
 702                 lwkt_reltoken(pcbinfo->porttoken);
 703         return error;
 704 }
 705
 706 /*
 707  *   Transform old in_pcbconnect() into an inner subroutine for new
 708  *   in_pcbconnect(): Do some validity-checking on the remote
 709  *   address (in mbuf 'nam') and then determine local host address
 710  *   (i.e., which interface) to use to access that remote host.
 711  *
 712  *   This preserves definition of in_pcbconnect(), while supporting a
 713  *   slightly different version for T/TCP.  (This is more than
 714  *   a bit of a kludge, but cleaning up the internal interfaces would
 715  *   have forced minor changes in every protocol).
 716  */
 717 int
 718 in_pcbladdr_find(struct inpcb *inp, struct sockaddr *nam,
 719     struct sockaddr_in **plocal_sin, struct thread *td, int find)
 720 {
 721         struct in_ifaddr *ia;
 722         struct ucred *cred = NULL;
 723         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 724         struct sockaddr *jsin;
 725         int jailed = 0, alloc_route = 0;
 726
 727         if (nam->sa_len != sizeof *sin)
 728                 return (EINVAL);
 729         if (sin->sin_family != AF_INET)
 730                 return (EAFNOSUPPORT);
 731         if (sin->sin_port == 0)
 732                 return (EADDRNOTAVAIL);
 733         if (td && td->td_proc && td->td_proc->p_ucred)
 734                 cred = td->td_proc->p_ucred;
 735         if (cred && cred->cr_prison)
 736                 jailed = 1;
 737         if (!TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) {
 738                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 739                 /*
 740                  * If the destination address is INADDR_ANY,
 741                  * use the primary local address.
 742                  * If the supplied address is INADDR_BROADCAST,
 743                  * and the primary interface supports broadcast,
 744                  * choose the broadcast address for that interface.
 745                  */
 746                 if (sin->sin_addr.s_addr == INADDR_ANY)
 747                         sin->sin_addr = IA_SIN(ia)->sin_addr;
 748                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 749                     (ia->ia_ifp->if_flags & IFF_BROADCAST))
 750                         sin->sin_addr = satosin(&ia->ia_broadaddr)->sin_addr;
 751         }
 752         if (find) {
 753                 struct route *ro;
 754
 755                 ia = NULL;
 756                 /*
 757                  * If route is known or can be allocated now,
 758                  * our src addr is taken from the i/f, else punt.
 759                  * Note that we should check the address family of the cached
 760                  * destination, in case of sharing the cache with IPv6.
 761                  */
 762                 ro = &inp->inp_route;
 763                 if (ro->ro_rt &&
 764                     (!(ro->ro_rt->rt_flags & RTF_UP) ||
 765                      ro->ro_dst.sa_family != AF_INET ||
 766                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 767                                       sin->sin_addr.s_addr ||
 768                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 769                         RTFREE(ro->ro_rt);
 770                         ro->ro_rt = NULL;
 771                 }
 772                 if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
 773                     (ro->ro_rt == NULL ||
 774                     ro->ro_rt->rt_ifp == NULL)) {
 775                         /* No route yet, so try to acquire one */
 776                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 777                         ro->ro_dst.sa_family = AF_INET;
 778                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 779                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 780                                 sin->sin_addr;
 781                         rtalloc(ro);
 782                         alloc_route = 1;
 783                 }
 784                 /*
 785                  * If we found a route, use the address
 786                  * corresponding to the outgoing interface
 787                  * unless it is the loopback (in case a route
 788                  * to our address on another net goes to loopback).
 789                  */
 790                 if (ro->ro_rt && !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK)) {
 791                         if (jailed) {
 792                                 if (jailed_ip(cred->cr_prison,
 793                                     ro->ro_rt->rt_ifa->ifa_addr)) {
 794                                         ia = ifatoia(ro->ro_rt->rt_ifa);
 795                                 }
 796                         } else {
 797                                 ia = ifatoia(ro->ro_rt->rt_ifa);
 798                         }
 799                 }
 800                 if (ia == NULL) {
 801                         u_short fport = sin->sin_port;
 802
 803                         sin->sin_port = 0;
 804                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 805                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 806                             sintosa(&ia->ia_addr)))
 807                                 ia = NULL;
 808                         if (ia == NULL)
 809                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 810                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 811                             sintosa(&ia->ia_addr)))
 812                                 ia = NULL;
 813                         sin->sin_port = fport;
 814                         if (ia == NULL &&
 815                             !TAILQ_EMPTY(&in_ifaddrheads[mycpuid]))
 816                                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 817                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 818                             sintosa(&ia->ia_addr)))
 819                                 ia = NULL;
 820
 821                         if (!jailed && ia == NULL)
 822                                 goto fail;
 823                 }
 824                 /*
 825                  * If the destination address is multicast and an outgoing
 826                  * interface has been set as a multicast option, use the
 827                  * address of that interface as our source address.
 828                  */
 829                 if (!jailed && IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 830                     inp->inp_moptions != NULL) {
 831                         struct ip_moptions *imo;
 832                         struct ifnet *ifp;
 833
 834                         imo = inp->inp_moptions;
 835                         if (imo->imo_multicast_ifp != NULL) {
 836                                 struct in_ifaddr_container *iac;
 837
 838                                 ifp = imo->imo_multicast_ifp;
 839                                 ia = NULL;
 840                                 TAILQ_FOREACH(iac,
 841                                 &in_ifaddrheads[mycpuid], ia_link) {
 842                                         if (iac->ia->ia_ifp == ifp) {
 843                                                 ia = iac->ia;
 844                                                 break;
 845                                         }
 846                                 }
 847                                 if (ia == NULL)
 848                                         goto fail;
 849                         }
 850                 }
 851                 /*
 852                  * Don't do pcblookup call here; return interface in plocal_sin
 853                  * and exit to caller, that will do the lookup.
 854                  */
 855                 if (ia == NULL && jailed) {
 856                         if ((jsin = prison_get_nonlocal(cred->cr_prison, AF_INET, NULL)) != NULL ||
 857                             (jsin = prison_get_local(cred->cr_prison, AF_INET, NULL)) != NULL) {
 858                                 *plocal_sin = satosin(jsin);
 859                         } else {
 860                                 /* IPv6 only Jail */
 861                                 goto fail;
 862                         }
 863                 } else {
 864                         *plocal_sin = &ia->ia_addr;
 865                 }
 866         }
 867         return (0);
 868 fail:
 869         if (alloc_route) {
 870                 struct route *ro = &inp->inp_route;
 871
 872                 if (ro->ro_rt != NULL)
 873                         RTFREE(ro->ro_rt);
 874                 bzero(ro, sizeof(*ro));
 875         }
 876         return (EADDRNOTAVAIL);
 877 }
 878
 879 int
 880 in_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 881     struct sockaddr_in **plocal_sin, struct thread *td)
 882 {
 883         return in_pcbladdr_find(inp, nam, plocal_sin, td,
 884             (inp->inp_laddr.s_addr == INADDR_ANY));
 885 }
 886
 887 /*
 888  * Outer subroutine:
 889  * Connect from a socket to a specified address.
 890  * Both address and port must be specified in argument sin.
 891  * If don't have a local address for this socket yet,
 892  * then pick one.
 893  */
 894 int
 895 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 896 {
 897         struct sockaddr_in *if_sin;
 898         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 899         int error;
 900
 901         /* Call inner routine to assign local interface address. */
 902         if ((error = in_pcbladdr(inp, nam, &if_sin, td)) != 0)
 903                 return (error);
 904
 905         if (in_pcblookup_hash(inp->inp_cpcbinfo, sin->sin_addr, sin->sin_port,
 906                               inp->inp_laddr.s_addr ?
 907                                 inp->inp_laddr : if_sin->sin_addr,
 908                               inp->inp_lport, FALSE, NULL) != NULL) {
 909                 return (EADDRINUSE);
 910         }
 911         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 912                 if (inp->inp_lport == 0) {
 913                         error = in_pcbbind(inp, NULL, td);
 914                         if (error)
 915                                 return (error);
 916                 }
 917                 inp->inp_laddr = if_sin->sin_addr;
 918         }
 919         inp->inp_faddr = sin->sin_addr;
 920         inp->inp_fport = sin->sin_port;
 921         in_pcbinsconnhash(inp);
 922         return (0);
 923 }
 924
 925 void
 926 in_pcbdisconnect(struct inpcb *inp)
 927 {
 928
 929         inp->inp_faddr.s_addr = INADDR_ANY;
 930         inp->inp_fport = 0;
 931         in_pcbremconnhash(inp);
 932         if (inp->inp_socket->so_state & SS_NOFDREF)
 933                 in_pcbdetach(inp);
 934 }
 935
 936 void
 937 in_pcbdetach(struct inpcb *inp)
 938 {
 939         struct socket *so = inp->inp_socket;
 940         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 941
 942 #ifdef IPSEC
 943         ipsec4_delete_pcbpolicy(inp);
 944 #endif /*IPSEC*/
 945         inp->inp_gencnt = ++ipi->ipi_gencnt;
 946         KKASSERT((so->so_state & SS_ASSERTINPROG) == 0);
 947         in_pcbremlists(inp);
 948         so->so_pcb = NULL;
 949         sofree(so);                     /* remove pcb ref */
 950         if (inp->inp_options)
 951                 m_free(inp->inp_options);
 952         if (inp->inp_route.ro_rt)
 953                 rtfree(inp->inp_route.ro_rt);
 954         ip_freemoptions(inp->inp_moptions);
 955         inp->inp_vflag = 0;
 956         kfree(inp, M_PCB);
 957 }
 958
 959 /*
 960  * The calling convention of in_setsockaddr() and in_setpeeraddr() was
 961  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 962  * in struct pr_usrreqs, so that protocols can just reference then directly
 963  * without the need for a wrapper function.  The socket must have a valid
 964  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 965  * except through a kernel programming error, so it is acceptable to panic
 966  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 967  * because there actually /is/ a programming error somewhere... XXX)
 968  */
 969 int
 970 in_setsockaddr(struct socket *so, struct sockaddr **nam)
 971 {
 972         struct inpcb *inp;
 973         struct sockaddr_in *sin;
 974
 975         /*
 976          * Do the malloc first in case it blocks.
 977          */
 978         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
 979         sin->sin_family = AF_INET;
 980         sin->sin_len = sizeof *sin;
 981
 982         crit_enter();
 983         inp = so->so_pcb;
 984         if (!inp) {
 985                 crit_exit();
 986                 kfree(sin, M_SONAME);
 987                 return (ECONNRESET);
 988         }
 989         sin->sin_port = inp->inp_lport;
 990         sin->sin_addr = inp->inp_laddr;
 991         crit_exit();
 992
 993         *nam = (struct sockaddr *)sin;
 994         return (0);
 995 }
 996
 997 void
 998 in_setsockaddr_dispatch(netmsg_t msg)
 999 {
1000         int error;
1001
1002         error = in_setsockaddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1003         lwkt_replymsg(&msg->lmsg, error);
1004 }
1005
1006 int
1007 in_setpeeraddr(struct socket *so, struct sockaddr **nam)
1008 {
1009         struct inpcb *inp;
1010         struct sockaddr_in *sin;
1011
1012         /*
1013          * Do the malloc first in case it blocks.
1014          */
1015         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
1016         sin->sin_family = AF_INET;
1017         sin->sin_len = sizeof *sin;
1018
1019         crit_enter();
1020         inp = so->so_pcb;
1021         if (!inp) {
1022                 crit_exit();
1023                 kfree(sin, M_SONAME);
1024                 return (ECONNRESET);
1025         }
1026         sin->sin_port = inp->inp_fport;
1027         sin->sin_addr = inp->inp_faddr;
1028         crit_exit();
1029
1030         *nam = (struct sockaddr *)sin;
1031         return (0);
1032 }
1033
1034 void
1035 in_setpeeraddr_dispatch(netmsg_t msg)
1036 {
1037         int error;
1038
1039         error = in_setpeeraddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1040         lwkt_replymsg(&msg->lmsg, error);
1041 }
1042
1043 void
1044 in_pcbnotifyall(struct inpcbhead *head, struct in_addr faddr, int err,
1045                 void (*notify)(struct inpcb *, int))
1046 {
1047         struct inpcb *inp, *ninp;
1048
1049         /*
1050          * note: if INP_PLACEMARKER is set we must ignore the rest of
1051          * the structure and skip it.
1052          */
1053         crit_enter();
1054         LIST_FOREACH_MUTABLE(inp, head, inp_list, ninp) {
1055                 if (inp->inp_flags & INP_PLACEMARKER)
1056                         continue;
1057 #ifdef INET6
1058                 if (!(inp->inp_vflag & INP_IPV4))
1059                         continue;
1060 #endif
1061                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
1062                     inp->inp_socket == NULL)
1063                         continue;
1064                 (*notify)(inp, err);            /* can remove inp from list! */
1065         }
1066         crit_exit();
1067 }
1068
1069 void
1070 in_pcbpurgeif0(struct inpcb *head, struct ifnet *ifp)
1071 {
1072         struct inpcb *inp;
1073         struct ip_moptions *imo;
1074         int i, gap;
1075
1076         for (inp = head; inp != NULL; inp = LIST_NEXT(inp, inp_list)) {
1077                 if (inp->inp_flags & INP_PLACEMARKER)
1078                         continue;
1079                 imo = inp->inp_moptions;
1080                 if ((inp->inp_vflag & INP_IPV4) && imo != NULL) {
1081                         /*
1082                          * Unselect the outgoing interface if it is being
1083                          * detached.
1084                          */
1085                         if (imo->imo_multicast_ifp == ifp)
1086                                 imo->imo_multicast_ifp = NULL;
1087
1088                         /*
1089                          * Drop multicast group membership if we joined
1090                          * through the interface being detached.
1091                          */
1092                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
1093                             i++) {
1094                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
1095                                         in_delmulti(imo->imo_membership[i]);
1096                                         gap++;
1097                                 } else if (gap != 0)
1098                                         imo->imo_membership[i - gap] =
1099                                             imo->imo_membership[i];
1100                         }
1101                         imo->imo_num_memberships -= gap;
1102                 }
1103         }
1104 }
1105
1106 /*
1107  * Check for alternatives when higher level complains
1108  * about service problems.  For now, invalidate cached
1109  * routing information.  If the route was created dynamically
1110  * (by a redirect), time to try a default gateway again.
1111  */
1112 void
1113 in_losing(struct inpcb *inp)
1114 {
1115         struct rtentry *rt;
1116         struct rt_addrinfo rtinfo;
1117
1118         if ((rt = inp->inp_route.ro_rt)) {
1119                 bzero(&rtinfo, sizeof(struct rt_addrinfo));
1120                 rtinfo.rti_info[RTAX_DST] = rt_key(rt);
1121                 rtinfo.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
1122                 rtinfo.rti_info[RTAX_NETMASK] = rt_mask(rt);
1123                 rtinfo.rti_flags = rt->rt_flags;
1124                 rt_missmsg(RTM_LOSING, &rtinfo, rt->rt_flags, 0);
1125                 if (rt->rt_flags & RTF_DYNAMIC) {
1126                         rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
1127                             rt_mask(rt), rt->rt_flags, NULL);
1128                 }
1129                 inp->inp_route.ro_rt = NULL;
1130                 rtfree(rt);
1131                 /*
1132                  * A new route can be allocated
1133                  * the next time output is attempted.
1134                  */
1135         }
1136 }
1137
1138 /*
1139  * After a routing change, flush old routing
1140  * and allocate a (hopefully) better one.
1141  */
1142 void
1143 in_rtchange(struct inpcb *inp, int err)
1144 {
1145         if (inp->inp_route.ro_rt) {
1146                 rtfree(inp->inp_route.ro_rt);
1147                 inp->inp_route.ro_rt = NULL;
1148                 /*
1149                  * A new route can be allocated the next time
1150                  * output is attempted.
1151                  */
1152         }
1153 }
1154
1155 /*
1156  * Lookup a PCB based on the local address and port.
1157  */
1158 struct inpcb *
1159 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
1160                    u_int lport_arg, int wild_okay, struct ucred *cred)
1161 {
1162         struct inpcb *inp;
1163         int matchwild = 3, wildcard;
1164         u_short lport = lport_arg;
1165         struct inpcbporthead *porthash;
1166         struct inpcbport *phd;
1167         struct inpcb *match = NULL;
1168
1169         /*
1170          * If the porthashbase is shared across several cpus we need
1171          * to lock.
1172          */
1173         if (pcbinfo->porttoken)
1174                 lwkt_gettoken(pcbinfo->porttoken);
1175
1176         /*
1177          * Best fit PCB lookup.
1178          *
1179          * First see if this local port is in use by looking on the
1180          * port hash list.
1181          */
1182         porthash = &pcbinfo->porthashbase[
1183                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
1184         LIST_FOREACH(phd, porthash, phd_hash) {
1185                 if (phd->phd_port == lport)
1186                         break;
1187         }
1188         if (phd != NULL) {
1189                 /*
1190                  * Port is in use by one or more PCBs. Look for best
1191                  * fit.
1192                  */
1193                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
1194                         wildcard = 0;
1195 #ifdef INET6
1196                         if ((inp->inp_vflag & INP_IPV4) == 0)
1197                                 continue;
1198 #endif
1199                         if (inp->inp_faddr.s_addr != INADDR_ANY)
1200                                 wildcard++;
1201                         if (inp->inp_laddr.s_addr != INADDR_ANY) {
1202                                 if (laddr.s_addr == INADDR_ANY)
1203                                         wildcard++;
1204                                 else if (inp->inp_laddr.s_addr != laddr.s_addr)
1205                                         continue;
1206                         } else {
1207                                 if (laddr.s_addr != INADDR_ANY)
1208                                         wildcard++;
1209                         }
1210                         if (wildcard && !wild_okay)
1211                                 continue;
1212                         if (wildcard < matchwild &&
1213                             (cred == NULL ||
1214                              cred->cr_prison ==
1215                                         inp->inp_socket->so_cred->cr_prison)) {
1216                                 match = inp;
1217                                 matchwild = wildcard;
1218                                 if (matchwild == 0) {
1219                                         break;
1220                                 }
1221                         }
1222                 }
1223         }
1224         if (pcbinfo->porttoken)
1225                 lwkt_reltoken(pcbinfo->porttoken);
1226         return (match);
1227 }
1228
1229 struct inpcb *
1230 in_pcblocalgroup_last(const struct inpcbinfo *pcbinfo,
1231     const struct inpcb *inp)
1232 {
1233         const struct inp_localgrphead *hdr;
1234         const struct inp_localgroup *grp;
1235         int i;
1236
1237         if (pcbinfo->localgrphashbase == NULL)
1238                 return NULL;
1239
1240         hdr = &pcbinfo->localgrphashbase[
1241             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1242
1243         LIST_FOREACH(grp, hdr, il_list) {
1244                 if (grp->il_vflag == inp->inp_vflag &&
1245                     grp->il_lport == inp->inp_lport &&
1246                     memcmp(&grp->il_dependladdr,
1247                         &inp->inp_inc.inc_ie.ie_dependladdr,
1248                         sizeof(grp->il_dependladdr)) == 0) {
1249                         break;
1250                 }
1251         }
1252         if (grp == NULL || grp->il_inpcnt == 1)
1253                 return NULL;
1254
1255         KASSERT(grp->il_inpcnt >= 2,
1256             ("invalid localgroup inp count %d", grp->il_inpcnt));
1257         for (i = 0; i < grp->il_inpcnt; ++i) {
1258                 if (grp->il_inp[i] == inp) {
1259                         int last = grp->il_inpcnt - 1;
1260
1261                         if (i == last)
1262                                 last = grp->il_inpcnt - 2;
1263                         return grp->il_inp[last];
1264                 }
1265         }
1266         return NULL;
1267 }
1268
1269 static struct inpcb *
1270 inp_localgroup_lookup(const struct inpcbinfo *pcbinfo,
1271     struct in_addr laddr, uint16_t lport, uint32_t pkt_hash)
1272 {
1273         struct inpcb *local_wild = NULL;
1274         const struct inp_localgrphead *hdr;
1275         const struct inp_localgroup *grp;
1276
1277         hdr = &pcbinfo->localgrphashbase[
1278             INP_PCBLOCALGRPHASH(lport, pcbinfo->localgrphashmask)];
1279 #ifdef INP_LOCALGROUP_HASHTHR
1280         pkt_hash >>= ncpus2_shift;
1281 #endif
1282
1283         /*
1284          * Order of socket selection:
1285          * 1. non-wild.
1286          * 2. wild.
1287          *
1288          * NOTE:
1289          * - Local group does not contain jailed sockets
1290          * - Local group does not contain IPv4 mapped INET6 wild sockets
1291          */
1292         LIST_FOREACH(grp, hdr, il_list) {
1293 #ifdef INET6
1294                 if (!(grp->il_vflag & INP_IPV4))
1295                         continue;
1296 #endif
1297                 if (grp->il_lport == lport) {
1298                         int idx;
1299
1300 #ifdef INP_LOCALGROUP_HASHTHR
1301                         idx = pkt_hash / grp->il_factor;
1302                         KASSERT(idx < grp->il_inpcnt && idx >= 0,
1303                             ("invalid hash %04x, cnt %d or fact %d",
1304                              pkt_hash, grp->il_inpcnt, grp->il_factor));
1305 #else
1306                         /*
1307                          * Modulo-N is used here, which greatly reduces
1308                          * completion queue token contention, thus more
1309                          * cpu time is saved.
1310                          */
1311                         idx = pkt_hash % grp->il_inpcnt;
1312 #endif
1313
1314                         if (grp->il_laddr.s_addr == laddr.s_addr)
1315                                 return grp->il_inp[idx];
1316                         else if (grp->il_laddr.s_addr == INADDR_ANY)
1317                                 local_wild = grp->il_inp[idx];
1318                 }
1319         }
1320         if (local_wild != NULL)
1321                 return local_wild;
1322         return NULL;
1323 }
1324
1325 /*
1326  * Lookup PCB in hash list.
1327  */
1328 struct inpcb *
1329 in_pcblookup_pkthash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1330     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1331     boolean_t wildcard, struct ifnet *ifp, const struct mbuf *m)
1332 {
1333         struct inpcbhead *head;
1334         struct inpcb *inp, *jinp=NULL;
1335         u_short fport = fport_arg, lport = lport_arg;
1336
1337         /*
1338          * First look for an exact match.
1339          */
1340         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr.s_addr, fport,
1341             laddr.s_addr, lport, pcbinfo->hashmask)];
1342         LIST_FOREACH(inp, head, inp_hash) {
1343 #ifdef INET6
1344                 if (!(inp->inp_vflag & INP_IPV4))
1345                         continue;
1346 #endif
1347                 if (in_hosteq(inp->inp_faddr, faddr) &&
1348                     in_hosteq(inp->inp_laddr, laddr) &&
1349                     inp->inp_fport == fport && inp->inp_lport == lport) {
1350                         /* found */
1351                         if (inp->inp_socket == NULL ||
1352                             inp->inp_socket->so_cred->cr_prison == NULL) {
1353                                 return (inp);
1354                         } else {
1355                                 if  (jinp == NULL)
1356                                         jinp = inp;
1357                         }
1358                 }
1359         }
1360         if (jinp != NULL)
1361                 return (jinp);
1362         if (wildcard) {
1363                 struct inpcb *local_wild = NULL;
1364                 struct inpcb *jinp_wild = NULL;
1365 #ifdef INET6
1366                 struct inpcb *local_wild_mapped = NULL;
1367 #endif
1368                 struct inpcontainer *ic;
1369                 struct inpcontainerhead *chead;
1370                 struct sockaddr_in jsin;
1371                 struct ucred *cred;
1372
1373                 /*
1374                  * Check local group first
1375                  */
1376                 if (pcbinfo->localgrphashbase != NULL &&
1377                     m != NULL && (m->m_flags & M_HASH) &&
1378                     !(ifp && ifp->if_type == IFT_FAITH)) {
1379                         inp = inp_localgroup_lookup(pcbinfo,
1380                             laddr, lport, m->m_pkthdr.hash);
1381                         if (inp != NULL)
1382                                 return inp;
1383                 }
1384
1385                 /*
1386                  * Order of socket selection:
1387                  * 1. non-jailed, non-wild.
1388                  * 2. non-jailed, wild.
1389                  * 3. jailed, non-wild.
1390                  * 4. jailed, wild.
1391                  */
1392                 jsin.sin_family = AF_INET;
1393                 chead = &pcbinfo->wildcardhashbase[
1394                     INP_PCBWILDCARDHASH(lport, pcbinfo->wildcardhashmask)];
1395                 LIST_FOREACH(ic, chead, ic_list) {
1396                         inp = ic->ic_inp;
1397                         jsin.sin_addr.s_addr = laddr.s_addr;
1398 #ifdef INET6
1399                         if (!(inp->inp_vflag & INP_IPV4))
1400                                 continue;
1401 #endif
1402                         if (inp->inp_socket != NULL)
1403                                 cred = inp->inp_socket->so_cred;
1404                         else
1405                                 cred = NULL;
1406                         if (cred != NULL && jailed(cred)) {
1407                                 if (jinp != NULL)
1408                                         continue;
1409                                 else
1410                                         if (!jailed_ip(cred->cr_prison,
1411                                             (struct sockaddr *)&jsin))
1412                                                 continue;
1413                         }
1414                         if (inp->inp_lport == lport) {
1415                                 if (ifp && ifp->if_type == IFT_FAITH &&
1416                                     !(inp->inp_flags & INP_FAITH))
1417                                         continue;
1418                                 if (inp->inp_laddr.s_addr == laddr.s_addr) {
1419                                         if (cred != NULL && jailed(cred))
1420                                                 jinp = inp;
1421                                         else
1422                                                 return (inp);
1423                                 }
1424                                 if (inp->inp_laddr.s_addr == INADDR_ANY) {
1425 #ifdef INET6
1426                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1427                                                              AF_INET6))
1428                                                 local_wild_mapped = inp;
1429                                         else
1430 #endif
1431                                                 if (cred != NULL &&
1432                                                     jailed(cred))
1433                                                         jinp_wild = inp;
1434                                                 else
1435                                                         local_wild = inp;
1436                                 }
1437                         }
1438                 }
1439                 if (local_wild != NULL)
1440                         return (local_wild);
1441 #ifdef INET6
1442                 if (local_wild_mapped != NULL)
1443                         return (local_wild_mapped);
1444 #endif
1445                 if (jinp != NULL)
1446                         return (jinp);
1447                 return (jinp_wild);
1448         }
1449
1450         /*
1451          * Not found.
1452          */
1453         return (NULL);
1454 }
1455
1456 struct inpcb *
1457 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1458     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1459     boolean_t wildcard, struct ifnet *ifp)
1460 {
1461         return in_pcblookup_pkthash(pcbinfo, faddr, fport_arg,
1462             laddr, lport_arg, wildcard, ifp, NULL);
1463 }
1464
1465 /*
1466  * Insert PCB into connection hash table.
1467  */
1468 void
1469 in_pcbinsconnhash(struct inpcb *inp)
1470 {
1471         struct inpcbinfo *pcbinfo = inp->inp_cpcbinfo;
1472         struct inpcbhead *bucket;
1473         u_int32_t hashkey_faddr, hashkey_laddr;
1474
1475 #ifdef INET6
1476         if (inp->inp_vflag & INP_IPV6) {
1477                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX JH */;
1478                 hashkey_laddr = inp->in6p_laddr.s6_addr32[3] /* XXX JH */;
1479         } else {
1480 #endif
1481                 hashkey_faddr = inp->inp_faddr.s_addr;
1482                 hashkey_laddr = inp->inp_laddr.s_addr;
1483 #ifdef INET6
1484         }
1485 #endif
1486
1487         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1488                 ("already on wildcardhash"));
1489         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1490                 ("already on connhash"));
1491         inp->inp_flags |= INP_CONNECTED;
1492
1493         /*
1494          * Insert into the connection hash table.
1495          */
1496         bucket = &pcbinfo->hashbase[INP_PCBCONNHASH(hashkey_faddr,
1497             inp->inp_fport, hashkey_laddr, inp->inp_lport, pcbinfo->hashmask)];
1498         LIST_INSERT_HEAD(bucket, inp, inp_hash);
1499 }
1500
1501 /*
1502  * Remove PCB from connection hash table.
1503  */
1504 void
1505 in_pcbremconnhash(struct inpcb *inp)
1506 {
1507         KASSERT(inp->inp_flags & INP_CONNECTED, ("inp not connected"));
1508         LIST_REMOVE(inp, inp_hash);
1509         inp->inp_flags &= ~INP_CONNECTED;
1510 }
1511
1512 /*
1513  * Insert PCB into port hash table.
1514  */
1515 int
1516 in_pcbinsporthash(struct inpcb *inp)
1517 {
1518         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1519         struct inpcbporthead *pcbporthash;
1520         struct inpcbport *phd;
1521
1522         /*
1523          * If the porthashbase is shared across several cpus we need
1524          * to lock.
1525          */
1526         if (pcbinfo->porttoken)
1527                 lwkt_gettoken(pcbinfo->porttoken);
1528
1529         /*
1530          * Insert into the port hash table.
1531          */
1532         pcbporthash = &pcbinfo->porthashbase[
1533             INP_PCBPORTHASH(inp->inp_lport, pcbinfo->porthashmask)];
1534
1535         /* Go through port list and look for a head for this lport. */
1536         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1537                 if (phd->phd_port == inp->inp_lport)
1538                         break;
1539         }
1540
1541         /* If none exists, malloc one and tack it on. */
1542         if (phd == NULL) {
1543                 KKASSERT(pcbinfo->portsave != NULL);
1544                 phd = pcbinfo->portsave;
1545                 pcbinfo->portsave = NULL;
1546                 phd->phd_port = inp->inp_lport;
1547                 LIST_INIT(&phd->phd_pcblist);
1548                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1549         }
1550
1551         inp->inp_phd = phd;
1552         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1553
1554         if (pcbinfo->porttoken)
1555                 lwkt_reltoken(pcbinfo->porttoken);
1556         if (pcbinfo->portsave == NULL) {
1557                 pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave),
1558                                             M_PCB, M_INTWAIT | M_ZERO);
1559         }
1560         return (0);
1561 }
1562
1563 static struct inp_localgroup *
1564 inp_localgroup_alloc(struct inp_localgrphead *hdr, u_char vflag,
1565     uint16_t port, const union in_dependaddr *addr, int size)
1566 {
1567         struct inp_localgroup *grp;
1568
1569         grp = kmalloc(__offsetof(struct inp_localgroup, il_inp[size]),
1570             M_TEMP, M_INTWAIT | M_ZERO);
1571         grp->il_vflag = vflag;
1572         grp->il_lport = port;
1573         grp->il_dependladdr = *addr;
1574         grp->il_inpsiz = size;
1575
1576         LIST_INSERT_HEAD(hdr, grp, il_list);
1577
1578         return grp;
1579 }
1580
1581 static void
1582 inp_localgroup_free(struct inp_localgroup *grp)
1583 {
1584         LIST_REMOVE(grp, il_list);
1585         kfree(grp, M_TEMP);
1586 }
1587
1588 static struct inp_localgroup *
1589 inp_localgroup_resize(struct inp_localgrphead *hdr,
1590     struct inp_localgroup *old_grp, int size)
1591 {
1592         struct inp_localgroup *grp;
1593         int i;
1594
1595         grp = inp_localgroup_alloc(hdr, old_grp->il_vflag,
1596             old_grp->il_lport, &old_grp->il_dependladdr, size);
1597
1598         KASSERT(old_grp->il_inpcnt < grp->il_inpsiz,
1599             ("invalid new local group size %d and old local group count %d",
1600              grp->il_inpsiz, old_grp->il_inpcnt));
1601         for (i = 0; i < old_grp->il_inpcnt; ++i)
1602                 grp->il_inp[i] = old_grp->il_inp[i];
1603         grp->il_inpcnt = old_grp->il_inpcnt;
1604         grp->il_factor = old_grp->il_factor;
1605
1606         inp_localgroup_free(old_grp);
1607
1608         return grp;
1609 }
1610
1611 static void
1612 inp_localgroup_factor(struct inp_localgroup *grp)
1613 {
1614         grp->il_factor =
1615             ((uint32_t)(0xffff >> ncpus2_shift) / grp->il_inpcnt) + 1;
1616         KASSERT(grp->il_factor != 0, ("invalid local group factor, "
1617             "ncpus2_shift %d, inpcnt %d", ncpus2_shift, grp->il_inpcnt));
1618 }
1619
1620 static void
1621 in_pcbinslocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1622 {
1623         struct inp_localgrphead *hdr;
1624         struct inp_localgroup *grp;
1625         struct ucred *cred;
1626
1627         if (pcbinfo->localgrphashbase == NULL)
1628                 return;
1629
1630         /*
1631          * XXX don't allow jailed socket to join local group
1632          */
1633         if (inp->inp_socket != NULL)
1634                 cred = inp->inp_socket->so_cred;
1635         else
1636                 cred = NULL;
1637         if (cred != NULL && jailed(cred))
1638                 return;
1639
1640 #ifdef INET6
1641         /*
1642          * XXX don't allow IPv4 mapped INET6 wild socket
1643          */
1644         if ((inp->inp_vflag & INP_IPV4) &&
1645             inp->inp_laddr.s_addr == INADDR_ANY &&
1646             INP_CHECK_SOCKAF(inp->inp_socket, AF_INET6))
1647                 return;
1648 #endif
1649
1650         hdr = &pcbinfo->localgrphashbase[
1651             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1652
1653         LIST_FOREACH(grp, hdr, il_list) {
1654                 if (grp->il_vflag == inp->inp_vflag &&
1655                     grp->il_lport == inp->inp_lport &&
1656                     memcmp(&grp->il_dependladdr,
1657                         &inp->inp_inc.inc_ie.ie_dependladdr,
1658                         sizeof(grp->il_dependladdr)) == 0) {
1659                         break;
1660                 }
1661         }
1662         if (grp == NULL) {
1663                 /* Create new local group */
1664                 grp = inp_localgroup_alloc(hdr, inp->inp_vflag,
1665                     inp->inp_lport, &inp->inp_inc.inc_ie.ie_dependladdr,
1666                     INP_LOCALGROUP_SIZMIN);
1667         } else if (grp->il_inpcnt == grp->il_inpsiz) {
1668                 if (grp->il_inpsiz >= INP_LOCALGROUP_SIZMAX) {
1669                         static int limit_logged = 0;
1670
1671                         if (!limit_logged) {
1672                                 limit_logged = 1;
1673                                 kprintf("local group port %d, "
1674                                     "limit reached\n", ntohs(grp->il_lport));
1675                         }
1676                         return;
1677                 }
1678
1679                 /* Expand this local group */
1680                 grp = inp_localgroup_resize(hdr, grp, grp->il_inpsiz * 2);
1681         }
1682
1683         KASSERT(grp->il_inpcnt < grp->il_inpsiz,
1684             ("invalid local group size %d and count %d",
1685              grp->il_inpsiz, grp->il_inpcnt));
1686         grp->il_inp[grp->il_inpcnt] = inp;
1687         grp->il_inpcnt++;
1688         inp_localgroup_factor(grp);
1689 }
1690
1691 void
1692 in_pcbinswildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1693 {
1694         struct inpcontainer *ic;
1695         struct inpcontainerhead *bucket;
1696
1697         in_pcbinslocalgrphash_oncpu(inp, pcbinfo);
1698
1699         bucket = &pcbinfo->wildcardhashbase[
1700             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1701
1702         ic = kmalloc(sizeof(struct inpcontainer), M_TEMP, M_INTWAIT);
1703         ic->ic_inp = inp;
1704         LIST_INSERT_HEAD(bucket, ic, ic_list);
1705 }
1706
1707 /*
1708  * Insert PCB into wildcard hash table.
1709  */
1710 void
1711 in_pcbinswildcardhash(struct inpcb *inp)
1712 {
1713         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1714
1715         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1716                 ("already on connhash"));
1717         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1718                 ("already on wildcardhash"));
1719         inp->inp_flags |= INP_WILDCARD;
1720
1721         in_pcbinswildcardhash_oncpu(inp, pcbinfo);
1722 }
1723
1724 static void
1725 in_pcbremlocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1726 {
1727         struct inp_localgrphead *hdr;
1728         struct inp_localgroup *grp;
1729
1730         if (pcbinfo->localgrphashbase == NULL)
1731                 return;
1732
1733         hdr = &pcbinfo->localgrphashbase[
1734             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1735
1736         LIST_FOREACH(grp, hdr, il_list) {
1737                 int i;
1738
1739                 for (i = 0; i < grp->il_inpcnt; ++i) {
1740                         if (grp->il_inp[i] != inp)
1741                                 continue;
1742
1743                         if (grp->il_inpcnt == 1) {
1744                                 /* Free this local group */
1745                                 inp_localgroup_free(grp);
1746                         } else {
1747                                 /* Pull up inpcbs */
1748                                 for (; i + 1 < grp->il_inpcnt; ++i)
1749                                         grp->il_inp[i] = grp->il_inp[i + 1];
1750                                 grp->il_inpcnt--;
1751                                 inp_localgroup_factor(grp);
1752
1753                                 if (grp->il_inpsiz > INP_LOCALGROUP_SIZMIN &&
1754                                     grp->il_inpcnt <= (grp->il_inpsiz / 4)) {
1755                                         /* Shrink this local group */
1756                                         grp = inp_localgroup_resize(hdr, grp,
1757                                             grp->il_inpsiz / 2);
1758                                 }
1759                         }
1760                         return;
1761                 }
1762         }
1763 }
1764
1765 void
1766 in_pcbremwildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1767 {
1768         struct inpcontainer *ic;
1769         struct inpcontainerhead *head;
1770
1771         in_pcbremlocalgrphash_oncpu(inp, pcbinfo);
1772
1773         /* find bucket */
1774         head = &pcbinfo->wildcardhashbase[
1775             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1776
1777         LIST_FOREACH(ic, head, ic_list) {
1778                 if (ic->ic_inp == inp)
1779                         goto found;
1780         }
1781         return;                 /* not found! */
1782
1783 found:
1784         LIST_REMOVE(ic, ic_list);       /* remove container from bucket chain */
1785         kfree(ic, M_TEMP);              /* deallocate container */
1786 }
1787
1788 /*
1789  * Remove PCB from wildcard hash table.
1790  */
1791 void
1792 in_pcbremwildcardhash(struct inpcb *inp)
1793 {
1794         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1795
1796         KASSERT(inp->inp_flags & INP_WILDCARD, ("inp not wildcard"));
1797         in_pcbremwildcardhash_oncpu(inp, pcbinfo);
1798         inp->inp_flags &= ~INP_WILDCARD;
1799 }
1800
1801 /*
1802  * Remove PCB from various lists.
1803  */
1804 void
1805 in_pcbremlists(struct inpcb *inp)
1806 {
1807         struct inpcbinfo *pcbinfo;
1808
1809         if (inp->inp_lport) {
1810                 struct inpcbport *phd;
1811
1812                 pcbinfo = inp->inp_pcbinfo;
1813                 if (pcbinfo->porttoken)
1814                         lwkt_gettoken(pcbinfo->porttoken);
1815
1816                 phd = inp->inp_phd;
1817                 LIST_REMOVE(inp, inp_portlist);
1818                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1819                         LIST_REMOVE(phd, phd_hash);
1820                         kfree(phd, M_PCB);
1821                 }
1822                 if (pcbinfo->porttoken)
1823                         lwkt_reltoken(pcbinfo->porttoken);
1824         }
1825         if (inp->inp_flags & INP_WILDCARD) {
1826                 in_pcbremwildcardhash(inp);
1827         } else if (inp->inp_flags & INP_CONNECTED) {
1828                 in_pcbremconnhash(inp);
1829         }
1830         LIST_REMOVE(inp, inp_list);
1831         inp->inp_pcbinfo->ipi_count--;
1832 }
1833
1834 int
1835 prison_xinpcb(struct thread *td, struct inpcb *inp)
1836 {
1837         struct ucred *cr;
1838
1839         if (td->td_proc == NULL)
1840                 return (0);
1841         cr = td->td_proc->p_ucred;
1842         if (cr->cr_prison == NULL)
1843                 return (0);
1844         if (inp->inp_socket && inp->inp_socket->so_cred &&
1845             inp->inp_socket->so_cred->cr_prison &&
1846             cr->cr_prison == inp->inp_socket->so_cred->cr_prison)
1847                 return (0);
1848         return (1);
1849 }
1850
1851 int
1852 in_pcblist_global(SYSCTL_HANDLER_ARGS)
1853 {
1854         struct inpcbinfo *pcbinfo = arg1;
1855         struct inpcb *inp, *marker;
1856         struct xinpcb xi;
1857         int error, i, n;
1858
1859         /*
1860          * The process of preparing the TCB list is too time-consuming and
1861          * resource-intensive to repeat twice on every request.
1862          */
1863         if (req->oldptr == NULL) {
1864                 n = pcbinfo->ipi_count;
1865                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1866                 return 0;
1867         }
1868
1869         if (req->newptr != NULL)
1870                 return EPERM;
1871
1872         /*
1873          * OK, now we're committed to doing something.  Re-fetch ipi_count
1874          * after obtaining the generation count.
1875          */
1876         n = pcbinfo->ipi_count;
1877
1878         marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1879         marker->inp_flags |= INP_PLACEMARKER;
1880         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, marker, inp_list);
1881
1882         i = 0;
1883         error = 0;
1884
1885         while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1886                 LIST_REMOVE(marker, inp_list);
1887                 LIST_INSERT_AFTER(inp, marker, inp_list);
1888
1889                 if (inp->inp_flags & INP_PLACEMARKER)
1890                         continue;
1891                 if (prison_xinpcb(req->td, inp))
1892                         continue;
1893                 bzero(&xi, sizeof xi);
1894                 xi.xi_len = sizeof xi;
1895                 bcopy(inp, &xi.xi_inp, sizeof *inp);
1896                 if (inp->inp_socket)
1897                         sotoxsocket(inp->inp_socket, &xi.xi_socket);
1898                 if ((error = SYSCTL_OUT(req, &xi, sizeof xi)) != 0)
1899                         break;
1900                 ++i;
1901         }
1902         LIST_REMOVE(marker, inp_list);
1903         if (error == 0 && i < n) {
1904                 bzero(&xi, sizeof xi);
1905                 xi.xi_len = sizeof xi;
1906                 while (i < n) {
1907                         error = SYSCTL_OUT(req, &xi, sizeof xi);
1908                         ++i;
1909                 }
1910         }
1911         kfree(marker, M_TEMP);
1912         return(error);
1913 }
1914
1915 int
1916 in_pcblist_global_nomarker(SYSCTL_HANDLER_ARGS, struct xinpcb **xi0, int *nxi0)
1917 {
1918         struct inpcbinfo *pcbinfo = arg1;
1919         struct inpcb *inp;
1920         struct xinpcb *xi;
1921         int nxi;
1922
1923         *nxi0 = 0;
1924         *xi0 = NULL;
1925
1926         /*
1927          * The process of preparing the PCB list is too time-consuming and
1928          * resource-intensive to repeat twice on every request.
1929          */
1930         if (req->oldptr == NULL) {
1931                 int n = pcbinfo->ipi_count;
1932
1933                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1934                 return 0;
1935         }
1936
1937         if (req->newptr != NULL)
1938                 return EPERM;
1939
1940         if (pcbinfo->ipi_count == 0)
1941                 return 0;
1942
1943         nxi = 0;
1944         xi = kmalloc(pcbinfo->ipi_count * sizeof(*xi), M_TEMP,
1945                      M_WAITOK | M_ZERO | M_NULLOK);
1946         if (xi == NULL)
1947                 return ENOMEM;
1948
1949         LIST_FOREACH(inp, &pcbinfo->pcblisthead, inp_list) {
1950                 struct xinpcb *xi_ptr = &xi[nxi];
1951
1952                 if (prison_xinpcb(req->td, inp))
1953                         continue;
1954
1955                 xi_ptr->xi_len = sizeof(*xi_ptr);
1956                 bcopy(inp, &xi_ptr->xi_inp, sizeof(*inp));
1957                 if (inp->inp_socket)
1958                         sotoxsocket(inp->inp_socket, &xi_ptr->xi_socket);
1959                 ++nxi;
1960         }
1961
1962         if (nxi == 0) {
1963                 kfree(xi, M_TEMP);
1964                 return 0;
1965         }
1966
1967         *nxi0 = nxi;
1968         *xi0 = xi;
1969
1970         return 0;
1971 }
1972
1973 void
1974 in_savefaddr(struct socket *so, const struct sockaddr *faddr)
1975 {
1976         struct sockaddr_in *sin;
1977
1978         KASSERT(faddr->sa_family == AF_INET,
1979             ("not AF_INET faddr %d", faddr->sa_family));
1980
1981         sin = kmalloc(sizeof(*sin), M_SONAME, M_WAITOK | M_ZERO);
1982         sin->sin_family = AF_INET;
1983         sin->sin_len = sizeof(*sin);
1984         sin->sin_port = ((const struct sockaddr_in *)faddr)->sin_port;
1985         sin->sin_addr = ((const struct sockaddr_in *)faddr)->sin_addr;
1986
1987         so->so_faddr = (struct sockaddr *)sin;
1988 }