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 struct in_addr zeroin_addr;
 119
 120 /*
 121  * These configure the range of local port addresses assigned to
 122  * "unspecified" outgoing connections/packets/whatever.
 123  */
 124 int ipport_lowfirstauto = IPPORT_RESERVED - 1;  /* 1023 */
 125 int ipport_lowlastauto = IPPORT_RESERVEDSTART;  /* 600 */
 126
 127 int ipport_firstauto = IPPORT_RESERVED;         /* 1024 */
 128 int ipport_lastauto = IPPORT_USERRESERVED;      /* 5000 */
 129
 130 int ipport_hifirstauto = IPPORT_HIFIRSTAUTO;    /* 49152 */
 131 int ipport_hilastauto = IPPORT_HILASTAUTO;      /* 65535 */
 132
 133 #define RANGECHK(var, min, max) \
 134         if ((var) < (min)) { (var) = (min); } \
 135         else if ((var) > (max)) { (var) = (max); }
 136
 137 int udpencap_enable = 1;        /* enabled by default */
 138 int udpencap_port = 4500;       /* triggers decapsulation */
 139
 140 static int
 141 sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS)
 142 {
 143         int error;
 144
 145         error = sysctl_handle_int(oidp, oidp->oid_arg1, oidp->oid_arg2, req);
 146         if (!error) {
 147                 RANGECHK(ipport_lowfirstauto, 1, IPPORT_RESERVED - 1);
 148                 RANGECHK(ipport_lowlastauto, 1, IPPORT_RESERVED - 1);
 149
 150                 RANGECHK(ipport_firstauto, IPPORT_RESERVED, USHRT_MAX);
 151                 RANGECHK(ipport_lastauto, IPPORT_RESERVED, USHRT_MAX);
 152
 153                 RANGECHK(ipport_hifirstauto, IPPORT_RESERVED, USHRT_MAX);
 154                 RANGECHK(ipport_hilastauto, IPPORT_RESERVED, USHRT_MAX);
 155         }
 156         return (error);
 157 }
 158
 159 #undef RANGECHK
 160
 161 SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW, 0, "IP Ports");
 162
 163 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLTYPE_INT|CTLFLAG_RW,
 164            &ipport_lowfirstauto, 0, &sysctl_net_ipport_check, "I", "");
 165 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLTYPE_INT|CTLFLAG_RW,
 166            &ipport_lowlastauto, 0, &sysctl_net_ipport_check, "I", "");
 167 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLTYPE_INT|CTLFLAG_RW,
 168            &ipport_firstauto, 0, &sysctl_net_ipport_check, "I", "");
 169 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLTYPE_INT|CTLFLAG_RW,
 170            &ipport_lastauto, 0, &sysctl_net_ipport_check, "I", "");
 171 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLTYPE_INT|CTLFLAG_RW,
 172            &ipport_hifirstauto, 0, &sysctl_net_ipport_check, "I", "");
 173 SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLTYPE_INT|CTLFLAG_RW,
 174            &ipport_hilastauto, 0, &sysctl_net_ipport_check, "I", "");
 175
 176 /*
 177  * in_pcb.c: manage the Protocol Control Blocks.
 178  *
 179  * NOTE: It is assumed that most of these functions will be called from
 180  * a critical section.  XXX - There are, unfortunately, a few exceptions
 181  * to this rule that should be fixed.
 182  *
 183  * NOTE: The caller should initialize the cpu field to the cpu running the
 184  * protocol stack associated with this inpcbinfo.
 185  */
 186
 187 void
 188 in_pcbinfo_init(struct inpcbinfo *pcbinfo)
 189 {
 190         LIST_INIT(&pcbinfo->pcblisthead);
 191         pcbinfo->cpu = -1;
 192         pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave), M_PCB,
 193                                     M_WAITOK | M_ZERO);
 194 }
 195
 196 struct baddynamicports baddynamicports;
 197
 198 /*
 199  * Check if the specified port is invalid for dynamic allocation.
 200  */
 201 int
 202 in_baddynamic(u_int16_t port, u_int16_t proto)
 203 {
 204         switch (proto) {
 205         case IPPROTO_TCP:
 206                 return (DP_ISSET(baddynamicports.tcp, port));
 207         case IPPROTO_UDP:
 208 #ifdef IPSEC
 209                 /* Cannot preset this as it is a sysctl */
 210                 if (port == udpencap_port)
 211                         return (1);
 212 #endif
 213                 return (DP_ISSET(baddynamicports.udp, port));
 214         default:
 215                 return (0);
 216         }
 217 }
 218
 219
 220 /*
 221  * Allocate a PCB and associate it with the socket.
 222  */
 223 int
 224 in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo)
 225 {
 226         struct inpcb *inp;
 227 #ifdef IPSEC
 228         int error;
 229 #endif
 230
 231         inp = kmalloc(pcbinfo->ipi_size, M_PCB, M_WAITOK|M_ZERO);
 232         inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 233         inp->inp_pcbinfo = inp->inp_cpcbinfo = pcbinfo;
 234         inp->inp_socket = so;
 235 #ifdef IPSEC
 236         error = ipsec_init_policy(so, &inp->inp_sp);
 237         if (error != 0) {
 238                 kfree(inp, M_PCB);
 239                 return (error);
 240         }
 241 #endif
 242 #ifdef INET6
 243         if (INP_SOCKAF(so) == AF_INET6 && ip6_v6only)
 244                 inp->inp_flags |= IN6P_IPV6_V6ONLY;
 245         if (ip6_auto_flowlabel)
 246                 inp->inp_flags |= IN6P_AUTOFLOWLABEL;
 247 #endif
 248         soreference(so);
 249         so->so_pcb = inp;
 250         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 251         pcbinfo->ipi_count++;
 252         return (0);
 253 }
 254
 255 /*
 256  * Unlink a pcb with the intention of moving it to another cpu with a
 257  * different pcbinfo.  While unlinked nothing should attempt to dereference
 258  * inp_pcbinfo, NULL it out so we assert if it does.
 259  */
 260 void
 261 in_pcbunlink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 262 {
 263         KKASSERT(inp->inp_pcbinfo == pcbinfo);
 264
 265         LIST_REMOVE(inp, inp_list);
 266         pcbinfo->ipi_count--;
 267         inp->inp_pcbinfo = NULL;
 268 }
 269
 270 /*
 271  * Relink a pcb into a new pcbinfo.
 272  */
 273 void
 274 in_pcblink(struct inpcb *inp, struct inpcbinfo *pcbinfo)
 275 {
 276         KKASSERT(inp->inp_pcbinfo == NULL);
 277         inp->inp_pcbinfo = pcbinfo;
 278         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, inp, inp_list);
 279         pcbinfo->ipi_count++;
 280 }
 281
 282 int
 283 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 284 {
 285         struct socket *so = inp->inp_socket;
 286         unsigned short *lastport;
 287         struct sockaddr_in *sin;
 288         struct sockaddr_in jsin;
 289         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 290         struct ucred *cred = NULL;
 291         u_short lport = 0;
 292         int wild = 0, reuseport = (so->so_options & SO_REUSEPORT);
 293         int error;
 294
 295         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 296                 return (EADDRNOTAVAIL);
 297         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 298                 return (EINVAL);        /* already bound */
 299
 300         if (!(so->so_options & (SO_REUSEADDR|SO_REUSEPORT)))
 301                 wild = 1;    /* neither SO_REUSEADDR nor SO_REUSEPORT is set */
 302         if (td->td_proc)
 303                 cred = td->td_proc->p_ucred;
 304
 305         /*
 306          * This has to be atomic.  If the porthash is shared across multiple
 307          * protocol threads (aka tcp) then the token will be non-NULL.
 308          */
 309         if (pcbinfo->porttoken)
 310                 lwkt_gettoken(pcbinfo->porttoken);
 311
 312         if (nam != NULL) {
 313                 sin = (struct sockaddr_in *)nam;
 314                 if (nam->sa_len != sizeof *sin) {
 315                         error = EINVAL;
 316                         goto done;
 317                 }
 318 #ifdef notdef
 319                 /*
 320                  * We should check the family, but old programs
 321                  * incorrectly fail to initialize it.
 322                  */
 323                 if (sin->sin_family != AF_INET) {
 324                         error = EAFNOSUPPORT;
 325                         goto done;
 326                 }
 327 #endif
 328                 if (!prison_replace_wildcards(td, nam)) {
 329                         error = EINVAL;
 330                         goto done;
 331                 }
 332                 lport = sin->sin_port;
 333                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 334                         /*
 335                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 336                          * allow complete duplication of binding if
 337                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 338                          * and a multicast address is bound on both
 339                          * new and duplicated sockets.
 340                          */
 341                         if (so->so_options & SO_REUSEADDR)
 342                                 reuseport = SO_REUSEADDR | SO_REUSEPORT;
 343                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 344                         sin->sin_port = 0;              /* yech... */
 345                         bzero(&sin->sin_zero, sizeof sin->sin_zero);
 346                         if (ifa_ifwithaddr((struct sockaddr *)sin) == NULL) {
 347                                 error = EADDRNOTAVAIL;
 348                                 goto done;
 349                         }
 350                 }
 351                 if (lport != 0) {
 352                         struct inpcb *t;
 353
 354                         /* GROSS */
 355                         if (ntohs(lport) < IPPORT_RESERVED &&
 356                             cred &&
 357                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0)) {
 358                                 error = EACCES;
 359                                 goto done;
 360                         }
 361                         if (so->so_cred->cr_uid != 0 &&
 362                             !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 363                                 t = in_pcblookup_local(pcbinfo,
 364                                                        sin->sin_addr,
 365                                                        lport,
 366                                                        INPLOOKUP_WILDCARD,
 367                                                        cred);
 368                                 if (t &&
 369                                     (!in_nullhost(sin->sin_addr) ||
 370                                      !in_nullhost(t->inp_laddr) ||
 371                                      (t->inp_socket->so_options &
 372                                          SO_REUSEPORT) == 0) &&
 373                                     (so->so_cred->cr_uid !=
 374                                      t->inp_socket->so_cred->cr_uid)) {
 375 #ifdef INET6
 376                                         if (!in_nullhost(sin->sin_addr) ||
 377                                             !in_nullhost(t->inp_laddr) ||
 378                                             INP_SOCKAF(so) ==
 379                                             INP_SOCKAF(t->inp_socket))
 380 #endif
 381                                         {
 382                                                 error = EADDRINUSE;
 383                                                 goto done;
 384                                         }
 385                                 }
 386                         }
 387                         if (cred && !prison_replace_wildcards(td, nam)) {
 388                                 error = EADDRNOTAVAIL;
 389                                 goto done;
 390                         }
 391                         t = in_pcblookup_local(pcbinfo, sin->sin_addr, lport,
 392                                                wild, cred);
 393                         if (t && !(reuseport & t->inp_socket->so_options)) {
 394 #ifdef INET6
 395                                 if (!in_nullhost(sin->sin_addr) ||
 396                                     !in_nullhost(t->inp_laddr) ||
 397                                     INP_SOCKAF(so) == INP_SOCKAF(t->inp_socket))
 398 #endif
 399                                 {
 400                                         error = EADDRINUSE;
 401                                         goto done;
 402                                 }
 403                         }
 404                 }
 405                 inp->inp_laddr = sin->sin_addr;
 406         }
 407         if (lport == 0) {
 408                 ushort first, last;
 409                 int count;
 410
 411                 jsin.sin_family = AF_INET;
 412                 jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 413                 if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 414                         inp->inp_laddr.s_addr = INADDR_ANY;
 415                         error = EINVAL;
 416                         goto done;
 417                 }
 418                 inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 419
 420                 inp->inp_flags |= INP_ANONPORT;
 421
 422                 if (inp->inp_flags & INP_HIGHPORT) {
 423                         first = ipport_hifirstauto;     /* sysctl */
 424                         last  = ipport_hilastauto;
 425                         lastport = &pcbinfo->lasthi;
 426                 } else if (inp->inp_flags & INP_LOWPORT) {
 427                         if (cred &&
 428                             (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 429                                 inp->inp_laddr.s_addr = INADDR_ANY;
 430                                 goto done;
 431                         }
 432                         first = ipport_lowfirstauto;    /* 1023 */
 433                         last  = ipport_lowlastauto;     /* 600 */
 434                         lastport = &pcbinfo->lastlow;
 435                 } else {
 436                         first = ipport_firstauto;       /* sysctl */
 437                         last  = ipport_lastauto;
 438                         lastport = &pcbinfo->lastport;
 439                 }
 440                 /*
 441                  * Simple check to ensure all ports are not used up causing
 442                  * a deadlock here.
 443                  *
 444                  * We split the two cases (up and down) so that the direction
 445                  * is not being tested on each round of the loop.
 446                  */
 447                 if (first > last) {
 448                         /*
 449                          * counting down
 450                          */
 451                         count = first - last;
 452
 453                         do {
 454                                 if (count-- < 0) {      /* completely used? */
 455                                         inp->inp_laddr.s_addr = INADDR_ANY;
 456                                         error = EADDRNOTAVAIL;
 457                                         goto done;
 458                                 }
 459                                 --*lastport;
 460                                 if (*lastport > first || *lastport < last)
 461                                         *lastport = first;
 462                                 lport = htons(*lastport);
 463                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 464                                                     lport, wild, cred));
 465                 } else {
 466                         /*
 467                          * counting up
 468                          */
 469                         count = last - first;
 470
 471                         do {
 472                                 if (count-- < 0) {      /* completely used? */
 473                                         inp->inp_laddr.s_addr = INADDR_ANY;
 474                                         error = EADDRNOTAVAIL;
 475                                         goto done;
 476                                 }
 477                                 ++*lastport;
 478                                 if (*lastport < first || *lastport > last)
 479                                         *lastport = first;
 480                                 lport = htons(*lastport);
 481                         } while (in_pcblookup_local(pcbinfo, inp->inp_laddr,
 482                                                     lport, wild, cred));
 483                 }
 484         }
 485         inp->inp_lport = lport;
 486
 487         jsin.sin_family = AF_INET;
 488         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 489         if (!prison_replace_wildcards(td, (struct sockaddr*)&jsin)) {
 490                 inp->inp_laddr.s_addr = INADDR_ANY;
 491                 inp->inp_lport = 0;
 492                 error = EINVAL;
 493                 goto done;
 494         }
 495         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 496
 497         if (in_pcbinsporthash(inp) != 0) {
 498                 inp->inp_laddr.s_addr = INADDR_ANY;
 499                 inp->inp_lport = 0;
 500                 error = EAGAIN;
 501                 goto done;
 502         }
 503         error = 0;
 504 done:
 505         if (pcbinfo->porttoken)
 506                 lwkt_reltoken(pcbinfo->porttoken);
 507         return error;
 508 }
 509
 510 static struct inpcb *
 511 in_pcblookup_addrport(struct inpcbinfo *pcbinfo, struct in_addr laddr,
 512     u_short lport, struct in_addr faddr, u_short fport, struct ucred *cred)
 513 {
 514         struct inpcb *inp;
 515         struct inpcbporthead *porthash;
 516         struct inpcbport *phd;
 517         struct inpcb *match = NULL;
 518
 519         /*
 520          * If the porthashbase is shared across several cpus we need
 521          * to lock.
 522          */
 523         if (pcbinfo->porttoken)
 524                 lwkt_gettoken(pcbinfo->porttoken);
 525
 526         /*
 527          * Best fit PCB lookup.
 528          *
 529          * First see if this local port is in use by looking on the
 530          * port hash list.
 531          */
 532         porthash = &pcbinfo->porthashbase[
 533                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
 534         LIST_FOREACH(phd, porthash, phd_hash) {
 535                 if (phd->phd_port == lport)
 536                         break;
 537         }
 538         if (phd != NULL) {
 539                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 540 #ifdef INET6
 541                         if ((inp->inp_vflag & INP_IPV4) == 0)
 542                                 continue;
 543 #endif
 544                         if (inp->inp_laddr.s_addr != INADDR_ANY &&
 545                             inp->inp_laddr.s_addr != laddr.s_addr)
 546                                 continue;
 547
 548                         if (inp->inp_faddr.s_addr != INADDR_ANY &&
 549                             inp->inp_faddr.s_addr != faddr.s_addr)
 550                                 continue;
 551
 552                         if (inp->inp_fport != 0 && inp->inp_fport != fport)
 553                                 continue;
 554
 555                         if (cred == NULL ||
 556                             cred->cr_prison ==
 557                             inp->inp_socket->so_cred->cr_prison) {
 558                                 match = inp;
 559                                 break;
 560                         }
 561                 }
 562         }
 563         if (pcbinfo->porttoken)
 564                 lwkt_reltoken(pcbinfo->porttoken);
 565         return (match);
 566 }
 567
 568 int
 569 in_pcbconn_bind(struct inpcb *inp, const struct sockaddr *nam,
 570     struct thread *td)
 571 {
 572         struct proc *p = td->td_proc;
 573         unsigned short *lastport;
 574         const struct sockaddr_in *sin = (const struct sockaddr_in *)nam;
 575         struct sockaddr_in jsin;
 576         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 577         struct ucred *cred = NULL;
 578         u_short lport = 0;
 579         ushort first, last;
 580         int count, error, dup = 0;
 581
 582         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 583                 return (EADDRNOTAVAIL);
 584
 585         KKASSERT(inp->inp_laddr.s_addr != INADDR_ANY);
 586         if (inp->inp_lport != 0)
 587                 return (EINVAL);        /* already bound */
 588
 589         KKASSERT(p);
 590         cred = p->p_ucred;
 591
 592         /*
 593          * This has to be atomic.  If the porthash is shared across multiple
 594          * protocol threads (aka tcp) then the token will be non-NULL.
 595          */
 596         if (pcbinfo->porttoken)
 597                 lwkt_gettoken(pcbinfo->porttoken);
 598
 599         jsin.sin_family = AF_INET;
 600         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 601         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 602                 inp->inp_laddr.s_addr = INADDR_ANY;
 603                 error = EINVAL;
 604                 goto done;
 605         }
 606         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 607
 608         inp->inp_flags |= INP_ANONPORT;
 609
 610         if (inp->inp_flags & INP_HIGHPORT) {
 611                 first = ipport_hifirstauto;     /* sysctl */
 612                 last  = ipport_hilastauto;
 613                 lastport = &pcbinfo->lasthi;
 614         } else if (inp->inp_flags & INP_LOWPORT) {
 615                 if (cred &&
 616                     (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 617                         inp->inp_laddr.s_addr = INADDR_ANY;
 618                         goto done;
 619                 }
 620                 first = ipport_lowfirstauto;    /* 1023 */
 621                 last  = ipport_lowlastauto;     /* 600 */
 622                 lastport = &pcbinfo->lastlow;
 623         } else {
 624                 first = ipport_firstauto;       /* sysctl */
 625                 last  = ipport_lastauto;
 626                 lastport = &pcbinfo->lastport;
 627         }
 628
 629 again:
 630         /*
 631          * Simple check to ensure all ports are not used up causing
 632          * a deadlock here.
 633          *
 634          * We split the two cases (up and down) so that the direction
 635          * is not being tested on each round of the loop.
 636          */
 637         if (first > last) {
 638                 /*
 639                  * counting down
 640                  */
 641                 count = first - last;
 642
 643                 do {
 644                         if (count-- < 0) {      /* completely used? */
 645                                 inp->inp_laddr.s_addr = INADDR_ANY;
 646                                 error = EADDRNOTAVAIL;
 647                                 goto done;
 648                         }
 649                         --*lastport;
 650                         if (*lastport > first || *lastport < last)
 651                                 *lastport = first;
 652                         lport = htons(*lastport);
 653                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 654                                 sin->sin_addr, sin->sin_port, cred));
 655         } else {
 656                 /*
 657                  * counting up
 658                  */
 659                 count = last - first;
 660
 661                 do {
 662                         if (count-- < 0) {      /* completely used? */
 663                                 inp->inp_laddr.s_addr = INADDR_ANY;
 664                                 error = EADDRNOTAVAIL;
 665                                 goto done;
 666                         }
 667                         ++*lastport;
 668                         if (*lastport < first || *lastport > last)
 669                                 *lastport = first;
 670                         lport = htons(*lastport);
 671                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 672                                 sin->sin_addr, sin->sin_port, cred));
 673         }
 674
 675         /* This could happen on loopback interface */
 676         if (sin->sin_port == lport &&
 677             sin->sin_addr.s_addr == inp->inp_laddr.s_addr) {
 678                 if (dup) {
 679                         /*
 680                          * Duplicate again; give up
 681                          */
 682                         inp->inp_laddr.s_addr = INADDR_ANY;
 683                         error = EADDRNOTAVAIL;
 684                         goto done;
 685                 }
 686                 dup = 1;
 687                 goto again;
 688         }
 689         inp->inp_lport = lport;
 690
 691         jsin.sin_family = AF_INET;
 692         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 693         if (!prison_replace_wildcards(td, (struct sockaddr*)&jsin)) {
 694                 inp->inp_laddr.s_addr = INADDR_ANY;
 695                 inp->inp_lport = 0;
 696                 error = EINVAL;
 697                 goto done;
 698         }
 699         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 700
 701         if (in_pcbinsporthash(inp) != 0) {
 702                 inp->inp_laddr.s_addr = INADDR_ANY;
 703                 inp->inp_lport = 0;
 704                 error = EAGAIN;
 705                 goto done;
 706         }
 707         error = 0;
 708 done:
 709         if (pcbinfo->porttoken)
 710                 lwkt_reltoken(pcbinfo->porttoken);
 711         return error;
 712 }
 713
 714 /*
 715  *   Transform old in_pcbconnect() into an inner subroutine for new
 716  *   in_pcbconnect(): Do some validity-checking on the remote
 717  *   address (in mbuf 'nam') and then determine local host address
 718  *   (i.e., which interface) to use to access that remote host.
 719  *
 720  *   This preserves definition of in_pcbconnect(), while supporting a
 721  *   slightly different version for T/TCP.  (This is more than
 722  *   a bit of a kludge, but cleaning up the internal interfaces would
 723  *   have forced minor changes in every protocol).
 724  */
 725 int
 726 in_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 727         struct sockaddr_in **plocal_sin, struct thread *td)
 728 {
 729         struct in_ifaddr *ia;
 730         struct ucred *cred = NULL;
 731         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 732         struct sockaddr *jsin;
 733         int jailed = 0, alloc_route = 0;
 734
 735         if (nam->sa_len != sizeof *sin)
 736                 return (EINVAL);
 737         if (sin->sin_family != AF_INET)
 738                 return (EAFNOSUPPORT);
 739         if (sin->sin_port == 0)
 740                 return (EADDRNOTAVAIL);
 741         if (td && td->td_proc && td->td_proc->p_ucred)
 742                 cred = td->td_proc->p_ucred;
 743         if (cred && cred->cr_prison)
 744                 jailed = 1;
 745         if (!TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) {
 746                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 747                 /*
 748                  * If the destination address is INADDR_ANY,
 749                  * use the primary local address.
 750                  * If the supplied address is INADDR_BROADCAST,
 751                  * and the primary interface supports broadcast,
 752                  * choose the broadcast address for that interface.
 753                  */
 754                 if (sin->sin_addr.s_addr == INADDR_ANY)
 755                         sin->sin_addr = IA_SIN(ia)->sin_addr;
 756                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 757                     (ia->ia_ifp->if_flags & IFF_BROADCAST))
 758                         sin->sin_addr = satosin(&ia->ia_broadaddr)->sin_addr;
 759         }
 760         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 761                 struct route *ro;
 762
 763                 ia = NULL;
 764                 /*
 765                  * If route is known or can be allocated now,
 766                  * our src addr is taken from the i/f, else punt.
 767                  * Note that we should check the address family of the cached
 768                  * destination, in case of sharing the cache with IPv6.
 769                  */
 770                 ro = &inp->inp_route;
 771                 if (ro->ro_rt &&
 772                     (!(ro->ro_rt->rt_flags & RTF_UP) ||
 773                      ro->ro_dst.sa_family != AF_INET ||
 774                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 775                                       sin->sin_addr.s_addr ||
 776                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 777                         RTFREE(ro->ro_rt);
 778                         ro->ro_rt = NULL;
 779                 }
 780                 if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
 781                     (ro->ro_rt == NULL ||
 782                     ro->ro_rt->rt_ifp == NULL)) {
 783                         /* No route yet, so try to acquire one */
 784                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 785                         ro->ro_dst.sa_family = AF_INET;
 786                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 787                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 788                                 sin->sin_addr;
 789                         rtalloc(ro);
 790                         alloc_route = 1;
 791                 }
 792                 /*
 793                  * If we found a route, use the address
 794                  * corresponding to the outgoing interface
 795                  * unless it is the loopback (in case a route
 796                  * to our address on another net goes to loopback).
 797                  */
 798                 if (ro->ro_rt && !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK)) {
 799                         if (jailed) {
 800                                 if (jailed_ip(cred->cr_prison,
 801                                     ro->ro_rt->rt_ifa->ifa_addr)) {
 802                                         ia = ifatoia(ro->ro_rt->rt_ifa);
 803                                 }
 804                         } else {
 805                                 ia = ifatoia(ro->ro_rt->rt_ifa);
 806                         }
 807                 }
 808                 if (ia == NULL) {
 809                         u_short fport = sin->sin_port;
 810
 811                         sin->sin_port = 0;
 812                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 813                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 814                             sintosa(&ia->ia_addr)))
 815                                 ia = NULL;
 816                         if (ia == NULL)
 817                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 818                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 819                             sintosa(&ia->ia_addr)))
 820                                 ia = NULL;
 821                         sin->sin_port = fport;
 822                         if (ia == NULL &&
 823                             !TAILQ_EMPTY(&in_ifaddrheads[mycpuid]))
 824                                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 825                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 826                             sintosa(&ia->ia_addr)))
 827                                 ia = NULL;
 828
 829                         if (!jailed && ia == NULL)
 830                                 goto fail;
 831                 }
 832                 /*
 833                  * If the destination address is multicast and an outgoing
 834                  * interface has been set as a multicast option, use the
 835                  * address of that interface as our source address.
 836                  */
 837                 if (!jailed && IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 838                     inp->inp_moptions != NULL) {
 839                         struct ip_moptions *imo;
 840                         struct ifnet *ifp;
 841
 842                         imo = inp->inp_moptions;
 843                         if (imo->imo_multicast_ifp != NULL) {
 844                                 struct in_ifaddr_container *iac;
 845
 846                                 ifp = imo->imo_multicast_ifp;
 847                                 ia = NULL;
 848                                 TAILQ_FOREACH(iac,
 849                                 &in_ifaddrheads[mycpuid], ia_link) {
 850                                         if (iac->ia->ia_ifp == ifp) {
 851                                                 ia = iac->ia;
 852                                                 break;
 853                                         }
 854                                 }
 855                                 if (ia == NULL)
 856                                         goto fail;
 857                         }
 858                 }
 859                 /*
 860                  * Don't do pcblookup call here; return interface in plocal_sin
 861                  * and exit to caller, that will do the lookup.
 862                  */
 863                 if (ia == NULL && jailed) {
 864                         if ((jsin = prison_get_nonlocal(cred->cr_prison, AF_INET, NULL)) != NULL ||
 865                             (jsin = prison_get_local(cred->cr_prison, AF_INET, NULL)) != NULL) {
 866                                 *plocal_sin = satosin(jsin);
 867                         } else {
 868                                 /* IPv6 only Jail */
 869                                 goto fail;
 870                         }
 871                 } else {
 872                         *plocal_sin = &ia->ia_addr;
 873                 }
 874         }
 875         return (0);
 876 fail:
 877         if (alloc_route) {
 878                 struct route *ro = &inp->inp_route;
 879
 880                 if (ro->ro_rt != NULL)
 881                         RTFREE(ro->ro_rt);
 882                 bzero(ro, sizeof(*ro));
 883         }
 884         return (EADDRNOTAVAIL);
 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                         rtrequest1_global(RTM_DELETE, &rtinfo, NULL, NULL);
1127                 inp->inp_route.ro_rt = NULL;
1128                 rtfree(rt);
1129                 /*
1130                  * A new route can be allocated
1131                  * the next time output is attempted.
1132                  */
1133         }
1134 }
1135
1136 /*
1137  * After a routing change, flush old routing
1138  * and allocate a (hopefully) better one.
1139  */
1140 void
1141 in_rtchange(struct inpcb *inp, int err)
1142 {
1143         if (inp->inp_route.ro_rt) {
1144                 rtfree(inp->inp_route.ro_rt);
1145                 inp->inp_route.ro_rt = NULL;
1146                 /*
1147                  * A new route can be allocated the next time
1148                  * output is attempted.
1149                  */
1150         }
1151 }
1152
1153 /*
1154  * Lookup a PCB based on the local address and port.
1155  */
1156 struct inpcb *
1157 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
1158                    u_int lport_arg, int wild_okay, struct ucred *cred)
1159 {
1160         struct inpcb *inp;
1161         int matchwild = 3, wildcard;
1162         u_short lport = lport_arg;
1163         struct inpcbporthead *porthash;
1164         struct inpcbport *phd;
1165         struct inpcb *match = NULL;
1166
1167         /*
1168          * If the porthashbase is shared across several cpus we need
1169          * to lock.
1170          */
1171         if (pcbinfo->porttoken)
1172                 lwkt_gettoken(pcbinfo->porttoken);
1173
1174         /*
1175          * Best fit PCB lookup.
1176          *
1177          * First see if this local port is in use by looking on the
1178          * port hash list.
1179          */
1180         porthash = &pcbinfo->porthashbase[
1181                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
1182         LIST_FOREACH(phd, porthash, phd_hash) {
1183                 if (phd->phd_port == lport)
1184                         break;
1185         }
1186         if (phd != NULL) {
1187                 /*
1188                  * Port is in use by one or more PCBs. Look for best
1189                  * fit.
1190                  */
1191                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
1192                         wildcard = 0;
1193 #ifdef INET6
1194                         if ((inp->inp_vflag & INP_IPV4) == 0)
1195                                 continue;
1196 #endif
1197                         if (inp->inp_faddr.s_addr != INADDR_ANY)
1198                                 wildcard++;
1199                         if (inp->inp_laddr.s_addr != INADDR_ANY) {
1200                                 if (laddr.s_addr == INADDR_ANY)
1201                                         wildcard++;
1202                                 else if (inp->inp_laddr.s_addr != laddr.s_addr)
1203                                         continue;
1204                         } else {
1205                                 if (laddr.s_addr != INADDR_ANY)
1206                                         wildcard++;
1207                         }
1208                         if (wildcard && !wild_okay)
1209                                 continue;
1210                         if (wildcard < matchwild &&
1211                             (cred == NULL ||
1212                              cred->cr_prison ==
1213                                         inp->inp_socket->so_cred->cr_prison)) {
1214                                 match = inp;
1215                                 matchwild = wildcard;
1216                                 if (matchwild == 0) {
1217                                         break;
1218                                 }
1219                         }
1220                 }
1221         }
1222         if (pcbinfo->porttoken)
1223                 lwkt_reltoken(pcbinfo->porttoken);
1224         return (match);
1225 }
1226
1227 /*
1228  * Lookup PCB in hash list.
1229  */
1230 struct inpcb *
1231 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1232                   u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1233                   boolean_t wildcard, struct ifnet *ifp)
1234 {
1235         struct inpcbhead *head;
1236         struct inpcb *inp, *jinp=NULL;
1237         u_short fport = fport_arg, lport = lport_arg;
1238
1239         /*
1240          * First look for an exact match.
1241          */
1242         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr.s_addr, fport,
1243             laddr.s_addr, lport, pcbinfo->hashmask)];
1244         LIST_FOREACH(inp, head, inp_hash) {
1245 #ifdef INET6
1246                 if (!(inp->inp_vflag & INP_IPV4))
1247                         continue;
1248 #endif
1249                 if (in_hosteq(inp->inp_faddr, faddr) &&
1250                     in_hosteq(inp->inp_laddr, laddr) &&
1251                     inp->inp_fport == fport && inp->inp_lport == lport) {
1252                         /* found */
1253                         if (inp->inp_socket == NULL ||
1254                             inp->inp_socket->so_cred->cr_prison == NULL) {
1255                                 return (inp);
1256                         } else {
1257                                 if  (jinp == NULL)
1258                                         jinp = inp;
1259                         }
1260                 }
1261         }
1262         if (jinp != NULL)
1263                 return (jinp);
1264         if (wildcard) {
1265                 struct inpcb *local_wild = NULL;
1266                 struct inpcb *jinp_wild = NULL;
1267 #ifdef INET6
1268                 struct inpcb *local_wild_mapped = NULL;
1269 #endif
1270                 struct inpcontainer *ic;
1271                 struct inpcontainerhead *chead;
1272                 struct sockaddr_in jsin;
1273                 struct ucred *cred;
1274
1275                 /*
1276                  * Order of socket selection:
1277                  * 1. non-jailed, non-wild.
1278                  * 2. non-jailed, wild.
1279                  * 3. jailed, non-wild.
1280                  * 4. jailed, wild.
1281                  */
1282                 jsin.sin_family = AF_INET;
1283                 chead = &pcbinfo->wildcardhashbase[
1284                     INP_PCBWILDCARDHASH(lport, pcbinfo->wildcardhashmask)];
1285                 LIST_FOREACH(ic, chead, ic_list) {
1286                         inp = ic->ic_inp;
1287                         jsin.sin_addr.s_addr = laddr.s_addr;
1288 #ifdef INET6
1289                         if (!(inp->inp_vflag & INP_IPV4))
1290                                 continue;
1291 #endif
1292                         if (inp->inp_socket != NULL)
1293                                 cred = inp->inp_socket->so_cred;
1294                         else
1295                                 cred = NULL;
1296                         if (cred != NULL && jailed(cred)) {
1297                                 if (jinp != NULL)
1298                                         continue;
1299                                 else
1300                                         if (!jailed_ip(cred->cr_prison,
1301                                             (struct sockaddr *)&jsin))
1302                                                 continue;
1303                         }
1304                         if (inp->inp_lport == lport) {
1305                                 if (ifp && ifp->if_type == IFT_FAITH &&
1306                                     !(inp->inp_flags & INP_FAITH))
1307                                         continue;
1308                                 if (inp->inp_laddr.s_addr == laddr.s_addr) {
1309                                         if (cred != NULL && jailed(cred))
1310                                                 jinp = inp;
1311                                         else
1312                                                 return (inp);
1313                                 }
1314                                 if (inp->inp_laddr.s_addr == INADDR_ANY) {
1315 #ifdef INET6
1316                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1317                                                              AF_INET6))
1318                                                 local_wild_mapped = inp;
1319                                         else
1320 #endif
1321                                                 if (cred != NULL &&
1322                                                     jailed(cred))
1323                                                         jinp_wild = inp;
1324                                                 else
1325                                                         local_wild = inp;
1326                                 }
1327                         }
1328                 }
1329                 if (local_wild != NULL)
1330                         return (local_wild);
1331 #ifdef INET6
1332                 if (local_wild_mapped != NULL)
1333                         return (local_wild_mapped);
1334 #endif
1335                 if (jinp != NULL)
1336                         return (jinp);
1337                 return (jinp_wild);
1338         }
1339
1340         /*
1341          * Not found.
1342          */
1343         return (NULL);
1344 }
1345
1346 /*
1347  * Insert PCB into connection hash table.
1348  */
1349 void
1350 in_pcbinsconnhash(struct inpcb *inp)
1351 {
1352         struct inpcbinfo *pcbinfo = inp->inp_cpcbinfo;
1353         struct inpcbhead *bucket;
1354         u_int32_t hashkey_faddr, hashkey_laddr;
1355
1356 #ifdef INET6
1357         if (inp->inp_vflag & INP_IPV6) {
1358                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX JH */;
1359                 hashkey_laddr = inp->in6p_laddr.s6_addr32[3] /* XXX JH */;
1360         } else {
1361 #endif
1362                 hashkey_faddr = inp->inp_faddr.s_addr;
1363                 hashkey_laddr = inp->inp_laddr.s_addr;
1364 #ifdef INET6
1365         }
1366 #endif
1367
1368         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1369                 ("already on wildcardhash"));
1370         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1371                 ("already on connhash"));
1372         inp->inp_flags |= INP_CONNECTED;
1373
1374         /*
1375          * Insert into the connection hash table.
1376          */
1377         bucket = &pcbinfo->hashbase[INP_PCBCONNHASH(hashkey_faddr,
1378             inp->inp_fport, hashkey_laddr, inp->inp_lport, pcbinfo->hashmask)];
1379         LIST_INSERT_HEAD(bucket, inp, inp_hash);
1380 }
1381
1382 /*
1383  * Remove PCB from connection hash table.
1384  */
1385 void
1386 in_pcbremconnhash(struct inpcb *inp)
1387 {
1388         KASSERT(inp->inp_flags & INP_CONNECTED, ("inp not connected"));
1389         LIST_REMOVE(inp, inp_hash);
1390         inp->inp_flags &= ~INP_CONNECTED;
1391 }
1392
1393 /*
1394  * Insert PCB into port hash table.
1395  */
1396 int
1397 in_pcbinsporthash(struct inpcb *inp)
1398 {
1399         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1400         struct inpcbporthead *pcbporthash;
1401         struct inpcbport *phd;
1402
1403         /*
1404          * If the porthashbase is shared across several cpus we need
1405          * to lock.
1406          */
1407         if (pcbinfo->porttoken)
1408                 lwkt_gettoken(pcbinfo->porttoken);
1409
1410         /*
1411          * Insert into the port hash table.
1412          */
1413         pcbporthash = &pcbinfo->porthashbase[
1414             INP_PCBPORTHASH(inp->inp_lport, pcbinfo->porthashmask)];
1415
1416         /* Go through port list and look for a head for this lport. */
1417         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1418                 if (phd->phd_port == inp->inp_lport)
1419                         break;
1420         }
1421
1422         /* If none exists, malloc one and tack it on. */
1423         if (phd == NULL) {
1424                 KKASSERT(pcbinfo->portsave != NULL);
1425                 phd = pcbinfo->portsave;
1426                 pcbinfo->portsave = NULL;
1427                 phd->phd_port = inp->inp_lport;
1428                 LIST_INIT(&phd->phd_pcblist);
1429                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1430         }
1431
1432         inp->inp_phd = phd;
1433         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1434
1435         if (pcbinfo->porttoken)
1436                 lwkt_reltoken(pcbinfo->porttoken);
1437         if (pcbinfo->portsave == NULL) {
1438                 pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave),
1439                                             M_PCB, M_INTWAIT | M_ZERO);
1440         }
1441         return (0);
1442 }
1443
1444 void
1445 in_pcbinswildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1446 {
1447         struct inpcontainer *ic;
1448         struct inpcontainerhead *bucket;
1449
1450         bucket = &pcbinfo->wildcardhashbase[
1451             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1452
1453         ic = kmalloc(sizeof(struct inpcontainer), M_TEMP, M_INTWAIT);
1454         ic->ic_inp = inp;
1455         LIST_INSERT_HEAD(bucket, ic, ic_list);
1456 }
1457
1458 /*
1459  * Insert PCB into wildcard hash table.
1460  */
1461 void
1462 in_pcbinswildcardhash(struct inpcb *inp)
1463 {
1464         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1465
1466         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1467                 ("already on connhash"));
1468         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1469                 ("already on wildcardhash"));
1470         inp->inp_flags |= INP_WILDCARD;
1471
1472         in_pcbinswildcardhash_oncpu(inp, pcbinfo);
1473 }
1474
1475 void
1476 in_pcbremwildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1477 {
1478         struct inpcontainer *ic;
1479         struct inpcontainerhead *head;
1480
1481         /* find bucket */
1482         head = &pcbinfo->wildcardhashbase[
1483             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1484
1485         LIST_FOREACH(ic, head, ic_list) {
1486                 if (ic->ic_inp == inp)
1487                         goto found;
1488         }
1489         return;                 /* not found! */
1490
1491 found:
1492         LIST_REMOVE(ic, ic_list);       /* remove container from bucket chain */
1493         kfree(ic, M_TEMP);              /* deallocate container */
1494 }
1495
1496 /*
1497  * Remove PCB from wildcard hash table.
1498  */
1499 void
1500 in_pcbremwildcardhash(struct inpcb *inp)
1501 {
1502         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1503
1504         KASSERT(inp->inp_flags & INP_WILDCARD, ("inp not wildcard"));
1505         in_pcbremwildcardhash_oncpu(inp, pcbinfo);
1506         inp->inp_flags &= ~INP_WILDCARD;
1507 }
1508
1509 /*
1510  * Remove PCB from various lists.
1511  */
1512 void
1513 in_pcbremlists(struct inpcb *inp)
1514 {
1515         struct inpcbinfo *pcbinfo;
1516
1517         if (inp->inp_lport) {
1518                 struct inpcbport *phd;
1519
1520                 pcbinfo = inp->inp_pcbinfo;
1521                 if (pcbinfo->porttoken)
1522                         lwkt_gettoken(pcbinfo->porttoken);
1523
1524                 phd = inp->inp_phd;
1525                 LIST_REMOVE(inp, inp_portlist);
1526                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1527                         LIST_REMOVE(phd, phd_hash);
1528                         kfree(phd, M_PCB);
1529                 }
1530                 if (pcbinfo->porttoken)
1531                         lwkt_reltoken(pcbinfo->porttoken);
1532         }
1533         if (inp->inp_flags & INP_WILDCARD) {
1534                 in_pcbremwildcardhash(inp);
1535         } else if (inp->inp_flags & INP_CONNECTED) {
1536                 in_pcbremconnhash(inp);
1537         }
1538         LIST_REMOVE(inp, inp_list);
1539         inp->inp_pcbinfo->ipi_count--;
1540 }
1541
1542 int
1543 prison_xinpcb(struct thread *td, struct inpcb *inp)
1544 {
1545         struct ucred *cr;
1546
1547         if (td->td_proc == NULL)
1548                 return (0);
1549         cr = td->td_proc->p_ucred;
1550         if (cr->cr_prison == NULL)
1551                 return (0);
1552         if (inp->inp_socket && inp->inp_socket->so_cred &&
1553             inp->inp_socket->so_cred->cr_prison &&
1554             cr->cr_prison == inp->inp_socket->so_cred->cr_prison)
1555                 return (0);
1556         return (1);
1557 }
1558
1559 int
1560 in_pcblist_global(SYSCTL_HANDLER_ARGS)
1561 {
1562         struct inpcbinfo *pcbinfo = arg1;
1563         struct inpcb *inp, *marker;
1564         struct xinpcb xi;
1565         int error, i, n;
1566
1567         /*
1568          * The process of preparing the TCB list is too time-consuming and
1569          * resource-intensive to repeat twice on every request.
1570          */
1571         if (req->oldptr == NULL) {
1572                 n = pcbinfo->ipi_count;
1573                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1574                 return 0;
1575         }
1576
1577         if (req->newptr != NULL)
1578                 return EPERM;
1579
1580         /*
1581          * OK, now we're committed to doing something.  Re-fetch ipi_count
1582          * after obtaining the generation count.
1583          */
1584         n = pcbinfo->ipi_count;
1585
1586         marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1587         marker->inp_flags |= INP_PLACEMARKER;
1588         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, marker, inp_list);
1589
1590         i = 0;
1591         error = 0;
1592
1593         while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1594                 LIST_REMOVE(marker, inp_list);
1595                 LIST_INSERT_AFTER(inp, marker, inp_list);
1596
1597                 if (inp->inp_flags & INP_PLACEMARKER)
1598                         continue;
1599                 if (prison_xinpcb(req->td, inp))
1600                         continue;
1601                 bzero(&xi, sizeof xi);
1602                 xi.xi_len = sizeof xi;
1603                 bcopy(inp, &xi.xi_inp, sizeof *inp);
1604                 if (inp->inp_socket)
1605                         sotoxsocket(inp->inp_socket, &xi.xi_socket);
1606                 if ((error = SYSCTL_OUT(req, &xi, sizeof xi)) != 0)
1607                         break;
1608                 ++i;
1609         }
1610         LIST_REMOVE(marker, inp_list);
1611         if (error == 0 && i < n) {
1612                 bzero(&xi, sizeof xi);
1613                 xi.xi_len = sizeof xi;
1614                 while (i < n) {
1615                         error = SYSCTL_OUT(req, &xi, sizeof xi);
1616                         ++i;
1617                 }
1618         }
1619         kfree(marker, M_TEMP);
1620         return(error);
1621 }
1622
1623 int
1624 in_pcblist_global_nomarker(SYSCTL_HANDLER_ARGS, struct xinpcb **xi0, int *nxi0)
1625 {
1626         struct inpcbinfo *pcbinfo = arg1;
1627         struct inpcb *inp;
1628         struct xinpcb *xi;
1629         int nxi;
1630
1631         *nxi0 = 0;
1632         *xi0 = NULL;
1633
1634         /*
1635          * The process of preparing the PCB list is too time-consuming and
1636          * resource-intensive to repeat twice on every request.
1637          */
1638         if (req->oldptr == NULL) {
1639                 int n = pcbinfo->ipi_count;
1640
1641                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1642                 return 0;
1643         }
1644
1645         if (req->newptr != NULL)
1646                 return EPERM;
1647
1648         if (pcbinfo->ipi_count == 0)
1649                 return 0;
1650
1651         nxi = 0;
1652         xi = kmalloc(pcbinfo->ipi_count * sizeof(*xi), M_TEMP,
1653                      M_WAITOK | M_ZERO | M_NULLOK);
1654         if (xi == NULL)
1655                 return ENOMEM;
1656
1657         LIST_FOREACH(inp, &pcbinfo->pcblisthead, inp_list) {
1658                 struct xinpcb *xi_ptr = &xi[nxi];
1659
1660                 if (prison_xinpcb(req->td, inp))
1661                         continue;
1662
1663                 xi_ptr->xi_len = sizeof(*xi_ptr);
1664                 bcopy(inp, &xi_ptr->xi_inp, sizeof(*inp));
1665                 if (inp->inp_socket)
1666                         sotoxsocket(inp->inp_socket, &xi_ptr->xi_socket);
1667                 ++nxi;
1668         }
1669
1670         if (nxi == 0) {
1671                 kfree(xi, M_TEMP);
1672                 return 0;
1673         }
1674
1675         *nxi0 = nxi;
1676         *xi0 = xi;
1677
1678         return 0;
1679 }
1680
1681 void
1682 in_savefaddr(struct socket *so, const struct sockaddr *faddr)
1683 {
1684         struct sockaddr_in *sin;
1685
1686         KASSERT(faddr->sa_family == AF_INET,
1687             ("not AF_INET faddr %d", faddr->sa_family));
1688
1689         sin = kmalloc(sizeof(*sin), M_SONAME, M_WAITOK | M_ZERO);
1690         sin->sin_family = AF_INET;
1691         sin->sin_len = sizeof(*sin);
1692         sin->sin_port = ((const struct sockaddr_in *)faddr)->sin_port;
1693         sin->sin_addr = ((const struct sockaddr_in *)faddr)->sin_addr;
1694
1695         so->so_faddr = (struct sockaddr *)sin;
1696 }