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 we need
 526          * to lock.
 527          */
 528         if (pcbinfo->porttoken)
 529                 lwkt_gettoken(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         if (pcbinfo->porttoken)
 569                 lwkt_reltoken(pcbinfo->porttoken);
 570         return (match);
 571 }
 572
 573 int
 574 in_pcbconn_bind(struct inpcb *inp, const struct sockaddr *nam,
 575     struct thread *td)
 576 {
 577         struct proc *p = td->td_proc;
 578         unsigned short *lastport;
 579         const struct sockaddr_in *sin = (const struct sockaddr_in *)nam;
 580         struct sockaddr_in jsin;
 581         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 582         struct ucred *cred = NULL;
 583         u_short lport = 0;
 584         ushort first, last;
 585         int count, error, dup = 0;
 586
 587         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 588                 return (EADDRNOTAVAIL);
 589
 590         KKASSERT(inp->inp_laddr.s_addr != INADDR_ANY);
 591         if (inp->inp_lport != 0)
 592                 return (EINVAL);        /* already bound */
 593
 594         KKASSERT(p);
 595         cred = p->p_ucred;
 596
 597         jsin.sin_family = AF_INET;
 598         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 599         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 600                 inp->inp_laddr.s_addr = INADDR_ANY;
 601                 return (EINVAL);
 602         }
 603         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 604
 605         inp->inp_flags |= INP_ANONPORT;
 606
 607         if (inp->inp_flags & INP_HIGHPORT) {
 608                 first = ipport_hifirstauto;     /* sysctl */
 609                 last  = ipport_hilastauto;
 610                 lastport = &pcbinfo->lasthi;
 611         } else if (inp->inp_flags & INP_LOWPORT) {
 612                 if (cred &&
 613                     (error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 614                         inp->inp_laddr.s_addr = INADDR_ANY;
 615                         return (error);
 616                 }
 617                 first = ipport_lowfirstauto;    /* 1023 */
 618                 last  = ipport_lowlastauto;     /* 600 */
 619                 lastport = &pcbinfo->lastlow;
 620         } else {
 621                 first = ipport_firstauto;       /* sysctl */
 622                 last  = ipport_lastauto;
 623                 lastport = &pcbinfo->lastport;
 624         }
 625
 626         /*
 627          * This has to be atomic.  If the porthash is shared across multiple
 628          * protocol threads (aka tcp) then the token will be non-NULL.
 629          */
 630         if (pcbinfo->porttoken)
 631                 lwkt_gettoken(pcbinfo->porttoken);
 632
 633 again:
 634         /*
 635          * Simple check to ensure all ports are not used up causing
 636          * a deadlock here.
 637          *
 638          * We split the two cases (up and down) so that the direction
 639          * is not being tested on each round of the loop.
 640          */
 641         if (first > last) {
 642                 /*
 643                  * counting down
 644                  */
 645                 count = first - last;
 646
 647                 do {
 648                         if (count-- < 0) {      /* completely used? */
 649                                 inp->inp_laddr.s_addr = INADDR_ANY;
 650                                 error = EADDRNOTAVAIL;
 651                                 goto done;
 652                         }
 653                         --*lastport;
 654                         if (*lastport > first || *lastport < last)
 655                                 *lastport = first;
 656                         lport = htons(*lastport);
 657                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 658                                 sin->sin_addr, sin->sin_port, cred));
 659         } else {
 660                 /*
 661                  * counting up
 662                  */
 663                 count = last - first;
 664
 665                 do {
 666                         if (count-- < 0) {      /* completely used? */
 667                                 inp->inp_laddr.s_addr = INADDR_ANY;
 668                                 error = EADDRNOTAVAIL;
 669                                 goto done;
 670                         }
 671                         ++*lastport;
 672                         if (*lastport < first || *lastport > last)
 673                                 *lastport = first;
 674                         lport = htons(*lastport);
 675                 } while (in_pcblookup_addrport(pcbinfo, inp->inp_laddr, lport,
 676                                 sin->sin_addr, sin->sin_port, cred));
 677         }
 678
 679         /* This could happen on loopback interface */
 680         if (sin->sin_port == lport &&
 681             sin->sin_addr.s_addr == inp->inp_laddr.s_addr) {
 682                 if (dup) {
 683                         /*
 684                          * Duplicate again; give up
 685                          */
 686                         inp->inp_laddr.s_addr = INADDR_ANY;
 687                         error = EADDRNOTAVAIL;
 688                         goto done;
 689                 }
 690                 dup = 1;
 691                 goto again;
 692         }
 693         inp->inp_lport = lport;
 694
 695         if (in_pcbinsporthash(inp) != 0) {
 696                 inp->inp_laddr.s_addr = INADDR_ANY;
 697                 inp->inp_lport = 0;
 698                 error = EAGAIN;
 699                 goto done;
 700         }
 701         error = 0;
 702 done:
 703         if (pcbinfo->porttoken)
 704                 lwkt_reltoken(pcbinfo->porttoken);
 705         return error;
 706 }
 707
 708 /*
 709  *   Transform old in_pcbconnect() into an inner subroutine for new
 710  *   in_pcbconnect(): Do some validity-checking on the remote
 711  *   address (in mbuf 'nam') and then determine local host address
 712  *   (i.e., which interface) to use to access that remote host.
 713  *
 714  *   This preserves definition of in_pcbconnect(), while supporting a
 715  *   slightly different version for T/TCP.  (This is more than
 716  *   a bit of a kludge, but cleaning up the internal interfaces would
 717  *   have forced minor changes in every protocol).
 718  */
 719 int
 720 in_pcbladdr_find(struct inpcb *inp, struct sockaddr *nam,
 721     struct sockaddr_in **plocal_sin, struct thread *td, int find)
 722 {
 723         struct in_ifaddr *ia;
 724         struct ucred *cred = NULL;
 725         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 726         struct sockaddr *jsin;
 727         int jailed = 0, alloc_route = 0;
 728
 729         if (nam->sa_len != sizeof *sin)
 730                 return (EINVAL);
 731         if (sin->sin_family != AF_INET)
 732                 return (EAFNOSUPPORT);
 733         if (sin->sin_port == 0)
 734                 return (EADDRNOTAVAIL);
 735         if (td && td->td_proc && td->td_proc->p_ucred)
 736                 cred = td->td_proc->p_ucred;
 737         if (cred && cred->cr_prison)
 738                 jailed = 1;
 739         if (!TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) {
 740                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 741                 /*
 742                  * If the destination address is INADDR_ANY,
 743                  * use the primary local address.
 744                  * If the supplied address is INADDR_BROADCAST,
 745                  * and the primary interface supports broadcast,
 746                  * choose the broadcast address for that interface.
 747                  */
 748                 if (sin->sin_addr.s_addr == INADDR_ANY)
 749                         sin->sin_addr = IA_SIN(ia)->sin_addr;
 750                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 751                     (ia->ia_ifp->if_flags & IFF_BROADCAST))
 752                         sin->sin_addr = satosin(&ia->ia_broadaddr)->sin_addr;
 753         }
 754         if (find) {
 755                 struct route *ro;
 756
 757                 ia = NULL;
 758                 /*
 759                  * If route is known or can be allocated now,
 760                  * our src addr is taken from the i/f, else punt.
 761                  * Note that we should check the address family of the cached
 762                  * destination, in case of sharing the cache with IPv6.
 763                  */
 764                 ro = &inp->inp_route;
 765                 if (ro->ro_rt &&
 766                     (!(ro->ro_rt->rt_flags & RTF_UP) ||
 767                      ro->ro_dst.sa_family != AF_INET ||
 768                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 769                                       sin->sin_addr.s_addr ||
 770                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 771                         RTFREE(ro->ro_rt);
 772                         ro->ro_rt = NULL;
 773                 }
 774                 if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
 775                     (ro->ro_rt == NULL ||
 776                     ro->ro_rt->rt_ifp == NULL)) {
 777                         /* No route yet, so try to acquire one */
 778                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 779                         ro->ro_dst.sa_family = AF_INET;
 780                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 781                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 782                                 sin->sin_addr;
 783                         rtalloc(ro);
 784                         alloc_route = 1;
 785                 }
 786                 /*
 787                  * If we found a route, use the address
 788                  * corresponding to the outgoing interface
 789                  * unless it is the loopback (in case a route
 790                  * to our address on another net goes to loopback).
 791                  */
 792                 if (ro->ro_rt && !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK)) {
 793                         if (jailed) {
 794                                 if (jailed_ip(cred->cr_prison,
 795                                     ro->ro_rt->rt_ifa->ifa_addr)) {
 796                                         ia = ifatoia(ro->ro_rt->rt_ifa);
 797                                 }
 798                         } else {
 799                                 ia = ifatoia(ro->ro_rt->rt_ifa);
 800                         }
 801                 }
 802                 if (ia == NULL) {
 803                         u_short fport = sin->sin_port;
 804
 805                         sin->sin_port = 0;
 806                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 807                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 808                             sintosa(&ia->ia_addr)))
 809                                 ia = NULL;
 810                         if (ia == NULL)
 811                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 812                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 813                             sintosa(&ia->ia_addr)))
 814                                 ia = NULL;
 815                         sin->sin_port = fport;
 816                         if (ia == NULL &&
 817                             !TAILQ_EMPTY(&in_ifaddrheads[mycpuid]))
 818                                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 819                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 820                             sintosa(&ia->ia_addr)))
 821                                 ia = NULL;
 822
 823                         if (!jailed && ia == NULL)
 824                                 goto fail;
 825                 }
 826                 /*
 827                  * If the destination address is multicast and an outgoing
 828                  * interface has been set as a multicast option, use the
 829                  * address of that interface as our source address.
 830                  */
 831                 if (!jailed && IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 832                     inp->inp_moptions != NULL) {
 833                         struct ip_moptions *imo;
 834                         struct ifnet *ifp;
 835
 836                         imo = inp->inp_moptions;
 837                         if (imo->imo_multicast_ifp != NULL) {
 838                                 struct in_ifaddr_container *iac;
 839
 840                                 ifp = imo->imo_multicast_ifp;
 841                                 ia = NULL;
 842                                 TAILQ_FOREACH(iac,
 843                                 &in_ifaddrheads[mycpuid], ia_link) {
 844                                         if (iac->ia->ia_ifp == ifp) {
 845                                                 ia = iac->ia;
 846                                                 break;
 847                                         }
 848                                 }
 849                                 if (ia == NULL)
 850                                         goto fail;
 851                         }
 852                 }
 853                 /*
 854                  * Don't do pcblookup call here; return interface in plocal_sin
 855                  * and exit to caller, that will do the lookup.
 856                  */
 857                 if (ia == NULL && jailed) {
 858                         if ((jsin = prison_get_nonlocal(cred->cr_prison, AF_INET, NULL)) != NULL ||
 859                             (jsin = prison_get_local(cred->cr_prison, AF_INET, NULL)) != NULL) {
 860                                 *plocal_sin = satosin(jsin);
 861                         } else {
 862                                 /* IPv6 only Jail */
 863                                 goto fail;
 864                         }
 865                 } else {
 866                         *plocal_sin = &ia->ia_addr;
 867                 }
 868         }
 869         return (0);
 870 fail:
 871         if (alloc_route) {
 872                 struct route *ro = &inp->inp_route;
 873
 874                 if (ro->ro_rt != NULL)
 875                         RTFREE(ro->ro_rt);
 876                 bzero(ro, sizeof(*ro));
 877         }
 878         return (EADDRNOTAVAIL);
 879 }
 880
 881 int
 882 in_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 883     struct sockaddr_in **plocal_sin, struct thread *td)
 884 {
 885         return in_pcbladdr_find(inp, nam, plocal_sin, td,
 886             (inp->inp_laddr.s_addr == INADDR_ANY));
 887 }
 888
 889 /*
 890  * Outer subroutine:
 891  * Connect from a socket to a specified address.
 892  * Both address and port must be specified in argument sin.
 893  * If don't have a local address for this socket yet,
 894  * then pick one.
 895  */
 896 int
 897 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 898 {
 899         struct sockaddr_in *if_sin;
 900         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 901         int error;
 902
 903         /* Call inner routine to assign local interface address. */
 904         if ((error = in_pcbladdr(inp, nam, &if_sin, td)) != 0)
 905                 return (error);
 906
 907         if (in_pcblookup_hash(inp->inp_cpcbinfo, sin->sin_addr, sin->sin_port,
 908                               inp->inp_laddr.s_addr ?
 909                                 inp->inp_laddr : if_sin->sin_addr,
 910                               inp->inp_lport, FALSE, NULL) != NULL) {
 911                 return (EADDRINUSE);
 912         }
 913         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 914                 if (inp->inp_lport == 0) {
 915                         error = in_pcbbind(inp, NULL, td);
 916                         if (error)
 917                                 return (error);
 918                 }
 919                 inp->inp_laddr = if_sin->sin_addr;
 920         }
 921         inp->inp_faddr = sin->sin_addr;
 922         inp->inp_fport = sin->sin_port;
 923         in_pcbinsconnhash(inp);
 924         return (0);
 925 }
 926
 927 void
 928 in_pcbdisconnect(struct inpcb *inp)
 929 {
 930
 931         inp->inp_faddr.s_addr = INADDR_ANY;
 932         inp->inp_fport = 0;
 933         in_pcbremconnhash(inp);
 934         if (inp->inp_socket->so_state & SS_NOFDREF)
 935                 in_pcbdetach(inp);
 936 }
 937
 938 void
 939 in_pcbdetach(struct inpcb *inp)
 940 {
 941         struct socket *so = inp->inp_socket;
 942         struct inpcbinfo *ipi = inp->inp_pcbinfo;
 943
 944 #ifdef IPSEC
 945         ipsec4_delete_pcbpolicy(inp);
 946 #endif /*IPSEC*/
 947         inp->inp_gencnt = ++ipi->ipi_gencnt;
 948         KKASSERT((so->so_state & SS_ASSERTINPROG) == 0);
 949         in_pcbremlists(inp);
 950         so->so_pcb = NULL;
 951         sofree(so);                     /* remove pcb ref */
 952         if (inp->inp_options)
 953                 m_free(inp->inp_options);
 954         if (inp->inp_route.ro_rt)
 955                 rtfree(inp->inp_route.ro_rt);
 956         ip_freemoptions(inp->inp_moptions);
 957         inp->inp_vflag = 0;
 958         kfree(inp, M_PCB);
 959 }
 960
 961 /*
 962  * The calling convention of in_setsockaddr() and in_setpeeraddr() was
 963  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
 964  * in struct pr_usrreqs, so that protocols can just reference then directly
 965  * without the need for a wrapper function.  The socket must have a valid
 966  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
 967  * except through a kernel programming error, so it is acceptable to panic
 968  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
 969  * because there actually /is/ a programming error somewhere... XXX)
 970  */
 971 int
 972 in_setsockaddr(struct socket *so, struct sockaddr **nam)
 973 {
 974         struct inpcb *inp;
 975         struct sockaddr_in *sin;
 976
 977         /*
 978          * Do the malloc first in case it blocks.
 979          */
 980         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
 981         sin->sin_family = AF_INET;
 982         sin->sin_len = sizeof *sin;
 983
 984         crit_enter();
 985         inp = so->so_pcb;
 986         if (!inp) {
 987                 crit_exit();
 988                 kfree(sin, M_SONAME);
 989                 return (ECONNRESET);
 990         }
 991         sin->sin_port = inp->inp_lport;
 992         sin->sin_addr = inp->inp_laddr;
 993         crit_exit();
 994
 995         *nam = (struct sockaddr *)sin;
 996         return (0);
 997 }
 998
 999 void
1000 in_setsockaddr_dispatch(netmsg_t msg)
1001 {
1002         int error;
1003
1004         error = in_setsockaddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1005         lwkt_replymsg(&msg->lmsg, error);
1006 }
1007
1008 int
1009 in_setpeeraddr(struct socket *so, struct sockaddr **nam)
1010 {
1011         struct inpcb *inp;
1012         struct sockaddr_in *sin;
1013
1014         /*
1015          * Do the malloc first in case it blocks.
1016          */
1017         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
1018         sin->sin_family = AF_INET;
1019         sin->sin_len = sizeof *sin;
1020
1021         crit_enter();
1022         inp = so->so_pcb;
1023         if (!inp) {
1024                 crit_exit();
1025                 kfree(sin, M_SONAME);
1026                 return (ECONNRESET);
1027         }
1028         sin->sin_port = inp->inp_fport;
1029         sin->sin_addr = inp->inp_faddr;
1030         crit_exit();
1031
1032         *nam = (struct sockaddr *)sin;
1033         return (0);
1034 }
1035
1036 void
1037 in_setpeeraddr_dispatch(netmsg_t msg)
1038 {
1039         int error;
1040
1041         error = in_setpeeraddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1042         lwkt_replymsg(&msg->lmsg, error);
1043 }
1044
1045 void
1046 in_pcbnotifyall(struct inpcbhead *head, struct in_addr faddr, int err,
1047                 void (*notify)(struct inpcb *, int))
1048 {
1049         struct inpcb *inp, *ninp;
1050
1051         /*
1052          * note: if INP_PLACEMARKER is set we must ignore the rest of
1053          * the structure and skip it.
1054          */
1055         crit_enter();
1056         LIST_FOREACH_MUTABLE(inp, head, inp_list, ninp) {
1057                 if (inp->inp_flags & INP_PLACEMARKER)
1058                         continue;
1059 #ifdef INET6
1060                 if (!(inp->inp_vflag & INP_IPV4))
1061                         continue;
1062 #endif
1063                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
1064                     inp->inp_socket == NULL)
1065                         continue;
1066                 (*notify)(inp, err);            /* can remove inp from list! */
1067         }
1068         crit_exit();
1069 }
1070
1071 void
1072 in_pcbpurgeif0(struct inpcb *head, struct ifnet *ifp)
1073 {
1074         struct inpcb *inp;
1075         struct ip_moptions *imo;
1076         int i, gap;
1077
1078         for (inp = head; inp != NULL; inp = LIST_NEXT(inp, inp_list)) {
1079                 if (inp->inp_flags & INP_PLACEMARKER)
1080                         continue;
1081                 imo = inp->inp_moptions;
1082                 if ((inp->inp_vflag & INP_IPV4) && imo != NULL) {
1083                         /*
1084                          * Unselect the outgoing interface if it is being
1085                          * detached.
1086                          */
1087                         if (imo->imo_multicast_ifp == ifp)
1088                                 imo->imo_multicast_ifp = NULL;
1089
1090                         /*
1091                          * Drop multicast group membership if we joined
1092                          * through the interface being detached.
1093                          */
1094                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
1095                             i++) {
1096                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
1097                                         in_delmulti(imo->imo_membership[i]);
1098                                         gap++;
1099                                 } else if (gap != 0)
1100                                         imo->imo_membership[i - gap] =
1101                                             imo->imo_membership[i];
1102                         }
1103                         imo->imo_num_memberships -= gap;
1104                 }
1105         }
1106 }
1107
1108 /*
1109  * Check for alternatives when higher level complains
1110  * about service problems.  For now, invalidate cached
1111  * routing information.  If the route was created dynamically
1112  * (by a redirect), time to try a default gateway again.
1113  */
1114 void
1115 in_losing(struct inpcb *inp)
1116 {
1117         struct rtentry *rt;
1118         struct rt_addrinfo rtinfo;
1119
1120         if ((rt = inp->inp_route.ro_rt)) {
1121                 bzero(&rtinfo, sizeof(struct rt_addrinfo));
1122                 rtinfo.rti_info[RTAX_DST] = rt_key(rt);
1123                 rtinfo.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
1124                 rtinfo.rti_info[RTAX_NETMASK] = rt_mask(rt);
1125                 rtinfo.rti_flags = rt->rt_flags;
1126                 rt_missmsg(RTM_LOSING, &rtinfo, rt->rt_flags, 0);
1127                 if (rt->rt_flags & RTF_DYNAMIC) {
1128                         rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
1129                             rt_mask(rt), rt->rt_flags, NULL);
1130                 }
1131                 inp->inp_route.ro_rt = NULL;
1132                 rtfree(rt);
1133                 /*
1134                  * A new route can be allocated
1135                  * the next time output is attempted.
1136                  */
1137         }
1138 }
1139
1140 /*
1141  * After a routing change, flush old routing
1142  * and allocate a (hopefully) better one.
1143  */
1144 void
1145 in_rtchange(struct inpcb *inp, int err)
1146 {
1147         if (inp->inp_route.ro_rt) {
1148                 rtfree(inp->inp_route.ro_rt);
1149                 inp->inp_route.ro_rt = NULL;
1150                 /*
1151                  * A new route can be allocated the next time
1152                  * output is attempted.
1153                  */
1154         }
1155 }
1156
1157 /*
1158  * Lookup a PCB based on the local address and port.
1159  */
1160 struct inpcb *
1161 in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr,
1162                    u_int lport_arg, int wild_okay, struct ucred *cred)
1163 {
1164         struct inpcb *inp;
1165         int matchwild = 3, wildcard;
1166         u_short lport = lport_arg;
1167         struct inpcbporthead *porthash;
1168         struct inpcbport *phd;
1169         struct inpcb *match = NULL;
1170
1171         /*
1172          * If the porthashbase is shared across several cpus we need
1173          * to lock.
1174          */
1175         if (pcbinfo->porttoken)
1176                 lwkt_gettoken(pcbinfo->porttoken);
1177
1178         /*
1179          * Best fit PCB lookup.
1180          *
1181          * First see if this local port is in use by looking on the
1182          * port hash list.
1183          */
1184         porthash = &pcbinfo->porthashbase[
1185                         INP_PCBPORTHASH(lport, pcbinfo->porthashmask)];
1186         LIST_FOREACH(phd, porthash, phd_hash) {
1187                 if (phd->phd_port == lport)
1188                         break;
1189         }
1190         if (phd != NULL) {
1191                 /*
1192                  * Port is in use by one or more PCBs. Look for best
1193                  * fit.
1194                  */
1195                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
1196                         wildcard = 0;
1197 #ifdef INET6
1198                         if ((inp->inp_vflag & INP_IPV4) == 0)
1199                                 continue;
1200 #endif
1201                         if (inp->inp_faddr.s_addr != INADDR_ANY)
1202                                 wildcard++;
1203                         if (inp->inp_laddr.s_addr != INADDR_ANY) {
1204                                 if (laddr.s_addr == INADDR_ANY)
1205                                         wildcard++;
1206                                 else if (inp->inp_laddr.s_addr != laddr.s_addr)
1207                                         continue;
1208                         } else {
1209                                 if (laddr.s_addr != INADDR_ANY)
1210                                         wildcard++;
1211                         }
1212                         if (wildcard && !wild_okay)
1213                                 continue;
1214                         if (wildcard < matchwild &&
1215                             (cred == NULL ||
1216                              cred->cr_prison ==
1217                                         inp->inp_socket->so_cred->cr_prison)) {
1218                                 match = inp;
1219                                 matchwild = wildcard;
1220                                 if (matchwild == 0) {
1221                                         break;
1222                                 }
1223                         }
1224                 }
1225         }
1226         if (pcbinfo->porttoken)
1227                 lwkt_reltoken(pcbinfo->porttoken);
1228         return (match);
1229 }
1230
1231 struct inpcb *
1232 in_pcblocalgroup_last(const struct inpcbinfo *pcbinfo,
1233     const struct inpcb *inp)
1234 {
1235         const struct inp_localgrphead *hdr;
1236         const struct inp_localgroup *grp;
1237         int i;
1238
1239         if (pcbinfo->localgrphashbase == NULL)
1240                 return NULL;
1241
1242         hdr = &pcbinfo->localgrphashbase[
1243             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1244
1245         LIST_FOREACH(grp, hdr, il_list) {
1246                 if (grp->il_vflag == inp->inp_vflag &&
1247                     grp->il_lport == inp->inp_lport &&
1248                     memcmp(&grp->il_dependladdr,
1249                         &inp->inp_inc.inc_ie.ie_dependladdr,
1250                         sizeof(grp->il_dependladdr)) == 0) {
1251                         break;
1252                 }
1253         }
1254         if (grp == NULL || grp->il_inpcnt == 1)
1255                 return NULL;
1256
1257         KASSERT(grp->il_inpcnt >= 2,
1258             ("invalid localgroup inp count %d", grp->il_inpcnt));
1259         for (i = 0; i < grp->il_inpcnt; ++i) {
1260                 if (grp->il_inp[i] == inp) {
1261                         int last = grp->il_inpcnt - 1;
1262
1263                         if (i == last)
1264                                 last = grp->il_inpcnt - 2;
1265                         return grp->il_inp[last];
1266                 }
1267         }
1268         return NULL;
1269 }
1270
1271 static struct inpcb *
1272 inp_localgroup_lookup(const struct inpcbinfo *pcbinfo,
1273     struct in_addr laddr, uint16_t lport, uint32_t pkt_hash)
1274 {
1275         struct inpcb *local_wild = NULL;
1276         const struct inp_localgrphead *hdr;
1277         const struct inp_localgroup *grp;
1278
1279         hdr = &pcbinfo->localgrphashbase[
1280             INP_PCBLOCALGRPHASH(lport, pcbinfo->localgrphashmask)];
1281 #ifdef INP_LOCALGROUP_HASHTHR
1282         pkt_hash >>= ncpus2_shift;
1283 #endif
1284
1285         /*
1286          * Order of socket selection:
1287          * 1. non-wild.
1288          * 2. wild.
1289          *
1290          * NOTE:
1291          * - Local group does not contain jailed sockets
1292          * - Local group does not contain IPv4 mapped INET6 wild sockets
1293          */
1294         LIST_FOREACH(grp, hdr, il_list) {
1295 #ifdef INET6
1296                 if (!(grp->il_vflag & INP_IPV4))
1297                         continue;
1298 #endif
1299                 if (grp->il_lport == lport) {
1300                         int idx;
1301
1302 #ifdef INP_LOCALGROUP_HASHTHR
1303                         idx = pkt_hash / grp->il_factor;
1304                         KASSERT(idx < grp->il_inpcnt && idx >= 0,
1305                             ("invalid hash %04x, cnt %d or fact %d",
1306                              pkt_hash, grp->il_inpcnt, grp->il_factor));
1307 #else
1308                         /*
1309                          * Modulo-N is used here, which greatly reduces
1310                          * completion queue token contention, thus more
1311                          * cpu time is saved.
1312                          */
1313                         idx = pkt_hash % grp->il_inpcnt;
1314 #endif
1315
1316                         if (grp->il_laddr.s_addr == laddr.s_addr)
1317                                 return grp->il_inp[idx];
1318                         else if (grp->il_laddr.s_addr == INADDR_ANY)
1319                                 local_wild = grp->il_inp[idx];
1320                 }
1321         }
1322         if (local_wild != NULL)
1323                 return local_wild;
1324         return NULL;
1325 }
1326
1327 /*
1328  * Lookup PCB in hash list.
1329  */
1330 struct inpcb *
1331 in_pcblookup_pkthash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1332     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1333     boolean_t wildcard, struct ifnet *ifp, const struct mbuf *m)
1334 {
1335         struct inpcbhead *head;
1336         struct inpcb *inp, *jinp=NULL;
1337         u_short fport = fport_arg, lport = lport_arg;
1338
1339         /*
1340          * First look for an exact match.
1341          */
1342         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr.s_addr, fport,
1343             laddr.s_addr, lport, pcbinfo->hashmask)];
1344         LIST_FOREACH(inp, head, inp_hash) {
1345 #ifdef INET6
1346                 if (!(inp->inp_vflag & INP_IPV4))
1347                         continue;
1348 #endif
1349                 if (in_hosteq(inp->inp_faddr, faddr) &&
1350                     in_hosteq(inp->inp_laddr, laddr) &&
1351                     inp->inp_fport == fport && inp->inp_lport == lport) {
1352                         /* found */
1353                         if (inp->inp_socket == NULL ||
1354                             inp->inp_socket->so_cred->cr_prison == NULL) {
1355                                 return (inp);
1356                         } else {
1357                                 if  (jinp == NULL)
1358                                         jinp = inp;
1359                         }
1360                 }
1361         }
1362         if (jinp != NULL)
1363                 return (jinp);
1364         if (wildcard) {
1365                 struct inpcb *local_wild = NULL;
1366                 struct inpcb *jinp_wild = NULL;
1367 #ifdef INET6
1368                 struct inpcb *local_wild_mapped = NULL;
1369 #endif
1370                 struct inpcontainer *ic;
1371                 struct inpcontainerhead *chead;
1372                 struct sockaddr_in jsin;
1373                 struct ucred *cred;
1374
1375                 /*
1376                  * Check local group first
1377                  */
1378                 if (pcbinfo->localgrphashbase != NULL &&
1379                     m != NULL && (m->m_flags & M_HASH) &&
1380                     !(ifp && ifp->if_type == IFT_FAITH)) {
1381                         inp = inp_localgroup_lookup(pcbinfo,
1382                             laddr, lport, m->m_pkthdr.hash);
1383                         if (inp != NULL)
1384                                 return inp;
1385                 }
1386
1387                 /*
1388                  * Order of socket selection:
1389                  * 1. non-jailed, non-wild.
1390                  * 2. non-jailed, wild.
1391                  * 3. jailed, non-wild.
1392                  * 4. jailed, wild.
1393                  */
1394                 jsin.sin_family = AF_INET;
1395                 chead = &pcbinfo->wildcardhashbase[
1396                     INP_PCBWILDCARDHASH(lport, pcbinfo->wildcardhashmask)];
1397                 LIST_FOREACH(ic, chead, ic_list) {
1398                         inp = ic->ic_inp;
1399                         jsin.sin_addr.s_addr = laddr.s_addr;
1400 #ifdef INET6
1401                         if (!(inp->inp_vflag & INP_IPV4))
1402                                 continue;
1403 #endif
1404                         if (inp->inp_socket != NULL)
1405                                 cred = inp->inp_socket->so_cred;
1406                         else
1407                                 cred = NULL;
1408                         if (cred != NULL && jailed(cred)) {
1409                                 if (jinp != NULL)
1410                                         continue;
1411                                 else
1412                                         if (!jailed_ip(cred->cr_prison,
1413                                             (struct sockaddr *)&jsin))
1414                                                 continue;
1415                         }
1416                         if (inp->inp_lport == lport) {
1417                                 if (ifp && ifp->if_type == IFT_FAITH &&
1418                                     !(inp->inp_flags & INP_FAITH))
1419                                         continue;
1420                                 if (inp->inp_laddr.s_addr == laddr.s_addr) {
1421                                         if (cred != NULL && jailed(cred))
1422                                                 jinp = inp;
1423                                         else
1424                                                 return (inp);
1425                                 }
1426                                 if (inp->inp_laddr.s_addr == INADDR_ANY) {
1427 #ifdef INET6
1428                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1429                                                              AF_INET6))
1430                                                 local_wild_mapped = inp;
1431                                         else
1432 #endif
1433                                                 if (cred != NULL &&
1434                                                     jailed(cred))
1435                                                         jinp_wild = inp;
1436                                                 else
1437                                                         local_wild = inp;
1438                                 }
1439                         }
1440                 }
1441                 if (local_wild != NULL)
1442                         return (local_wild);
1443 #ifdef INET6
1444                 if (local_wild_mapped != NULL)
1445                         return (local_wild_mapped);
1446 #endif
1447                 if (jinp != NULL)
1448                         return (jinp);
1449                 return (jinp_wild);
1450         }
1451
1452         /*
1453          * Not found.
1454          */
1455         return (NULL);
1456 }
1457
1458 struct inpcb *
1459 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1460     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1461     boolean_t wildcard, struct ifnet *ifp)
1462 {
1463         return in_pcblookup_pkthash(pcbinfo, faddr, fport_arg,
1464             laddr, lport_arg, wildcard, ifp, NULL);
1465 }
1466
1467 /*
1468  * Insert PCB into connection hash table.
1469  */
1470 void
1471 in_pcbinsconnhash(struct inpcb *inp)
1472 {
1473         struct inpcbinfo *pcbinfo = inp->inp_cpcbinfo;
1474         struct inpcbhead *bucket;
1475         u_int32_t hashkey_faddr, hashkey_laddr;
1476
1477 #ifdef INET6
1478         if (inp->inp_vflag & INP_IPV6) {
1479                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX JH */;
1480                 hashkey_laddr = inp->in6p_laddr.s6_addr32[3] /* XXX JH */;
1481         } else {
1482 #endif
1483                 hashkey_faddr = inp->inp_faddr.s_addr;
1484                 hashkey_laddr = inp->inp_laddr.s_addr;
1485 #ifdef INET6
1486         }
1487 #endif
1488
1489         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1490                 ("already on wildcardhash"));
1491         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1492                 ("already on connhash"));
1493         inp->inp_flags |= INP_CONNECTED;
1494
1495         /*
1496          * Insert into the connection hash table.
1497          */
1498         bucket = &pcbinfo->hashbase[INP_PCBCONNHASH(hashkey_faddr,
1499             inp->inp_fport, hashkey_laddr, inp->inp_lport, pcbinfo->hashmask)];
1500         LIST_INSERT_HEAD(bucket, inp, inp_hash);
1501 }
1502
1503 /*
1504  * Remove PCB from connection hash table.
1505  */
1506 void
1507 in_pcbremconnhash(struct inpcb *inp)
1508 {
1509         KASSERT(inp->inp_flags & INP_CONNECTED, ("inp not connected"));
1510         LIST_REMOVE(inp, inp_hash);
1511         inp->inp_flags &= ~INP_CONNECTED;
1512 }
1513
1514 /*
1515  * Insert PCB into port hash table.
1516  */
1517 int
1518 in_pcbinsporthash(struct inpcb *inp)
1519 {
1520         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1521         struct inpcbporthead *pcbporthash;
1522         struct inpcbport *phd;
1523
1524         /*
1525          * If the porthashbase is shared across several cpus we need
1526          * to lock.
1527          */
1528         if (pcbinfo->porttoken)
1529                 lwkt_gettoken(pcbinfo->porttoken);
1530
1531         /*
1532          * Insert into the port hash table.
1533          */
1534         pcbporthash = &pcbinfo->porthashbase[
1535             INP_PCBPORTHASH(inp->inp_lport, pcbinfo->porthashmask)];
1536
1537         /* Go through port list and look for a head for this lport. */
1538         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1539                 if (phd->phd_port == inp->inp_lport)
1540                         break;
1541         }
1542
1543         /* If none exists, malloc one and tack it on. */
1544         if (phd == NULL) {
1545                 KKASSERT(pcbinfo->portsave != NULL);
1546                 phd = pcbinfo->portsave;
1547                 pcbinfo->portsave = NULL;
1548                 phd->phd_port = inp->inp_lport;
1549                 LIST_INIT(&phd->phd_pcblist);
1550                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1551         }
1552
1553         inp->inp_phd = phd;
1554         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1555
1556         if (pcbinfo->porttoken)
1557                 lwkt_reltoken(pcbinfo->porttoken);
1558         if (pcbinfo->portsave == NULL) {
1559                 pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave),
1560                                             M_PCB, M_INTWAIT | M_ZERO);
1561         }
1562         return (0);
1563 }
1564
1565 static struct inp_localgroup *
1566 inp_localgroup_alloc(struct inp_localgrphead *hdr, u_char vflag,
1567     uint16_t port, const union in_dependaddr *addr, int size)
1568 {
1569         struct inp_localgroup *grp;
1570
1571         grp = kmalloc(__offsetof(struct inp_localgroup, il_inp[size]),
1572             M_TEMP, M_INTWAIT | M_ZERO);
1573         grp->il_vflag = vflag;
1574         grp->il_lport = port;
1575         grp->il_dependladdr = *addr;
1576         grp->il_inpsiz = size;
1577
1578         LIST_INSERT_HEAD(hdr, grp, il_list);
1579
1580         return grp;
1581 }
1582
1583 static void
1584 inp_localgroup_free(struct inp_localgroup *grp)
1585 {
1586         LIST_REMOVE(grp, il_list);
1587         kfree(grp, M_TEMP);
1588 }
1589
1590 static struct inp_localgroup *
1591 inp_localgroup_resize(struct inp_localgrphead *hdr,
1592     struct inp_localgroup *old_grp, int size)
1593 {
1594         struct inp_localgroup *grp;
1595         int i;
1596
1597         grp = inp_localgroup_alloc(hdr, old_grp->il_vflag,
1598             old_grp->il_lport, &old_grp->il_dependladdr, size);
1599
1600         KASSERT(old_grp->il_inpcnt < grp->il_inpsiz,
1601             ("invalid new local group size %d and old local group count %d",
1602              grp->il_inpsiz, old_grp->il_inpcnt));
1603         for (i = 0; i < old_grp->il_inpcnt; ++i)
1604                 grp->il_inp[i] = old_grp->il_inp[i];
1605         grp->il_inpcnt = old_grp->il_inpcnt;
1606         grp->il_factor = old_grp->il_factor;
1607
1608         inp_localgroup_free(old_grp);
1609
1610         return grp;
1611 }
1612
1613 static void
1614 inp_localgroup_factor(struct inp_localgroup *grp)
1615 {
1616         grp->il_factor =
1617             ((uint32_t)(0xffff >> ncpus2_shift) / grp->il_inpcnt) + 1;
1618         KASSERT(grp->il_factor != 0, ("invalid local group factor, "
1619             "ncpus2_shift %d, inpcnt %d", ncpus2_shift, grp->il_inpcnt));
1620 }
1621
1622 static void
1623 in_pcbinslocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1624 {
1625         struct inp_localgrphead *hdr;
1626         struct inp_localgroup *grp;
1627         struct ucred *cred;
1628
1629         if (pcbinfo->localgrphashbase == NULL)
1630                 return;
1631
1632         /*
1633          * XXX don't allow jailed socket to join local group
1634          */
1635         if (inp->inp_socket != NULL)
1636                 cred = inp->inp_socket->so_cred;
1637         else
1638                 cred = NULL;
1639         if (cred != NULL && jailed(cred))
1640                 return;
1641
1642 #ifdef INET6
1643         /*
1644          * XXX don't allow IPv4 mapped INET6 wild socket
1645          */
1646         if ((inp->inp_vflag & INP_IPV4) &&
1647             inp->inp_laddr.s_addr == INADDR_ANY &&
1648             INP_CHECK_SOCKAF(inp->inp_socket, AF_INET6))
1649                 return;
1650 #endif
1651
1652         hdr = &pcbinfo->localgrphashbase[
1653             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1654
1655         LIST_FOREACH(grp, hdr, il_list) {
1656                 if (grp->il_vflag == inp->inp_vflag &&
1657                     grp->il_lport == inp->inp_lport &&
1658                     memcmp(&grp->il_dependladdr,
1659                         &inp->inp_inc.inc_ie.ie_dependladdr,
1660                         sizeof(grp->il_dependladdr)) == 0) {
1661                         break;
1662                 }
1663         }
1664         if (grp == NULL) {
1665                 /* Create new local group */
1666                 grp = inp_localgroup_alloc(hdr, inp->inp_vflag,
1667                     inp->inp_lport, &inp->inp_inc.inc_ie.ie_dependladdr,
1668                     INP_LOCALGROUP_SIZMIN);
1669         } else if (grp->il_inpcnt == grp->il_inpsiz) {
1670                 if (grp->il_inpsiz >= INP_LOCALGROUP_SIZMAX) {
1671                         static int limit_logged = 0;
1672
1673                         if (!limit_logged) {
1674                                 limit_logged = 1;
1675                                 kprintf("local group port %d, "
1676                                     "limit reached\n", ntohs(grp->il_lport));
1677                         }
1678                         return;
1679                 }
1680
1681                 /* Expand this local group */
1682                 grp = inp_localgroup_resize(hdr, grp, grp->il_inpsiz * 2);
1683         }
1684
1685         KASSERT(grp->il_inpcnt < grp->il_inpsiz,
1686             ("invalid local group size %d and count %d",
1687              grp->il_inpsiz, grp->il_inpcnt));
1688         grp->il_inp[grp->il_inpcnt] = inp;
1689         grp->il_inpcnt++;
1690         inp_localgroup_factor(grp);
1691 }
1692
1693 void
1694 in_pcbinswildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1695 {
1696         struct inpcontainer *ic;
1697         struct inpcontainerhead *bucket;
1698
1699         in_pcbinslocalgrphash_oncpu(inp, pcbinfo);
1700
1701         bucket = &pcbinfo->wildcardhashbase[
1702             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1703
1704         ic = kmalloc(sizeof(struct inpcontainer), M_TEMP, M_INTWAIT);
1705         ic->ic_inp = inp;
1706         LIST_INSERT_HEAD(bucket, ic, ic_list);
1707 }
1708
1709 /*
1710  * Insert PCB into wildcard hash table.
1711  */
1712 void
1713 in_pcbinswildcardhash(struct inpcb *inp)
1714 {
1715         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1716
1717         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1718                 ("already on connhash"));
1719         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1720                 ("already on wildcardhash"));
1721         inp->inp_flags |= INP_WILDCARD;
1722
1723         in_pcbinswildcardhash_oncpu(inp, pcbinfo);
1724 }
1725
1726 static void
1727 in_pcbremlocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1728 {
1729         struct inp_localgrphead *hdr;
1730         struct inp_localgroup *grp;
1731
1732         if (pcbinfo->localgrphashbase == NULL)
1733                 return;
1734
1735         hdr = &pcbinfo->localgrphashbase[
1736             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1737
1738         LIST_FOREACH(grp, hdr, il_list) {
1739                 int i;
1740
1741                 for (i = 0; i < grp->il_inpcnt; ++i) {
1742                         if (grp->il_inp[i] != inp)
1743                                 continue;
1744
1745                         if (grp->il_inpcnt == 1) {
1746                                 /* Free this local group */
1747                                 inp_localgroup_free(grp);
1748                         } else {
1749                                 /* Pull up inpcbs */
1750                                 for (; i + 1 < grp->il_inpcnt; ++i)
1751                                         grp->il_inp[i] = grp->il_inp[i + 1];
1752                                 grp->il_inpcnt--;
1753                                 inp_localgroup_factor(grp);
1754
1755                                 if (grp->il_inpsiz > INP_LOCALGROUP_SIZMIN &&
1756                                     grp->il_inpcnt <= (grp->il_inpsiz / 4)) {
1757                                         /* Shrink this local group */
1758                                         grp = inp_localgroup_resize(hdr, grp,
1759                                             grp->il_inpsiz / 2);
1760                                 }
1761                         }
1762                         return;
1763                 }
1764         }
1765 }
1766
1767 void
1768 in_pcbremwildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1769 {
1770         struct inpcontainer *ic;
1771         struct inpcontainerhead *head;
1772
1773         in_pcbremlocalgrphash_oncpu(inp, pcbinfo);
1774
1775         /* find bucket */
1776         head = &pcbinfo->wildcardhashbase[
1777             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1778
1779         LIST_FOREACH(ic, head, ic_list) {
1780                 if (ic->ic_inp == inp)
1781                         goto found;
1782         }
1783         return;                 /* not found! */
1784
1785 found:
1786         LIST_REMOVE(ic, ic_list);       /* remove container from bucket chain */
1787         kfree(ic, M_TEMP);              /* deallocate container */
1788 }
1789
1790 /*
1791  * Remove PCB from wildcard hash table.
1792  */
1793 void
1794 in_pcbremwildcardhash(struct inpcb *inp)
1795 {
1796         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1797
1798         KASSERT(inp->inp_flags & INP_WILDCARD, ("inp not wildcard"));
1799         in_pcbremwildcardhash_oncpu(inp, pcbinfo);
1800         inp->inp_flags &= ~INP_WILDCARD;
1801 }
1802
1803 /*
1804  * Remove PCB from various lists.
1805  */
1806 void
1807 in_pcbremlists(struct inpcb *inp)
1808 {
1809         struct inpcbinfo *pcbinfo;
1810
1811         if (inp->inp_lport) {
1812                 struct inpcbport *phd;
1813
1814                 pcbinfo = inp->inp_pcbinfo;
1815                 if (pcbinfo->porttoken)
1816                         lwkt_gettoken(pcbinfo->porttoken);
1817
1818                 phd = inp->inp_phd;
1819                 LIST_REMOVE(inp, inp_portlist);
1820                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1821                         LIST_REMOVE(phd, phd_hash);
1822                         kfree(phd, M_PCB);
1823                 }
1824                 if (pcbinfo->porttoken)
1825                         lwkt_reltoken(pcbinfo->porttoken);
1826         }
1827         if (inp->inp_flags & INP_WILDCARD) {
1828                 in_pcbremwildcardhash(inp);
1829         } else if (inp->inp_flags & INP_CONNECTED) {
1830                 in_pcbremconnhash(inp);
1831         }
1832         LIST_REMOVE(inp, inp_list);
1833         inp->inp_pcbinfo->ipi_count--;
1834 }
1835
1836 int
1837 prison_xinpcb(struct thread *td, struct inpcb *inp)
1838 {
1839         struct ucred *cr;
1840
1841         if (td->td_proc == NULL)
1842                 return (0);
1843         cr = td->td_proc->p_ucred;
1844         if (cr->cr_prison == NULL)
1845                 return (0);
1846         if (inp->inp_socket && inp->inp_socket->so_cred &&
1847             inp->inp_socket->so_cred->cr_prison &&
1848             cr->cr_prison == inp->inp_socket->so_cred->cr_prison)
1849                 return (0);
1850         return (1);
1851 }
1852
1853 int
1854 in_pcblist_global(SYSCTL_HANDLER_ARGS)
1855 {
1856         struct inpcbinfo *pcbinfo = arg1;
1857         struct inpcb *inp, *marker;
1858         struct xinpcb xi;
1859         int error, i, n;
1860
1861         /*
1862          * The process of preparing the TCB list is too time-consuming and
1863          * resource-intensive to repeat twice on every request.
1864          */
1865         if (req->oldptr == NULL) {
1866                 n = pcbinfo->ipi_count;
1867                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1868                 return 0;
1869         }
1870
1871         if (req->newptr != NULL)
1872                 return EPERM;
1873
1874         /*
1875          * OK, now we're committed to doing something.  Re-fetch ipi_count
1876          * after obtaining the generation count.
1877          */
1878         n = pcbinfo->ipi_count;
1879
1880         marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1881         marker->inp_flags |= INP_PLACEMARKER;
1882         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, marker, inp_list);
1883
1884         i = 0;
1885         error = 0;
1886
1887         while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1888                 LIST_REMOVE(marker, inp_list);
1889                 LIST_INSERT_AFTER(inp, marker, inp_list);
1890
1891                 if (inp->inp_flags & INP_PLACEMARKER)
1892                         continue;
1893                 if (prison_xinpcb(req->td, inp))
1894                         continue;
1895                 bzero(&xi, sizeof xi);
1896                 xi.xi_len = sizeof xi;
1897                 bcopy(inp, &xi.xi_inp, sizeof *inp);
1898                 if (inp->inp_socket)
1899                         sotoxsocket(inp->inp_socket, &xi.xi_socket);
1900                 if ((error = SYSCTL_OUT(req, &xi, sizeof xi)) != 0)
1901                         break;
1902                 ++i;
1903         }
1904         LIST_REMOVE(marker, inp_list);
1905         if (error == 0 && i < n) {
1906                 bzero(&xi, sizeof xi);
1907                 xi.xi_len = sizeof xi;
1908                 while (i < n) {
1909                         error = SYSCTL_OUT(req, &xi, sizeof xi);
1910                         ++i;
1911                 }
1912         }
1913         kfree(marker, M_TEMP);
1914         return(error);
1915 }
1916
1917 int
1918 in_pcblist_global_nomarker(SYSCTL_HANDLER_ARGS, struct xinpcb **xi0, int *nxi0)
1919 {
1920         struct inpcbinfo *pcbinfo = arg1;
1921         struct inpcb *inp;
1922         struct xinpcb *xi;
1923         int nxi;
1924
1925         *nxi0 = 0;
1926         *xi0 = NULL;
1927
1928         /*
1929          * The process of preparing the PCB list is too time-consuming and
1930          * resource-intensive to repeat twice on every request.
1931          */
1932         if (req->oldptr == NULL) {
1933                 int n = pcbinfo->ipi_count;
1934
1935                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1936                 return 0;
1937         }
1938
1939         if (req->newptr != NULL)
1940                 return EPERM;
1941
1942         if (pcbinfo->ipi_count == 0)
1943                 return 0;
1944
1945         nxi = 0;
1946         xi = kmalloc(pcbinfo->ipi_count * sizeof(*xi), M_TEMP,
1947                      M_WAITOK | M_ZERO | M_NULLOK);
1948         if (xi == NULL)
1949                 return ENOMEM;
1950
1951         LIST_FOREACH(inp, &pcbinfo->pcblisthead, inp_list) {
1952                 struct xinpcb *xi_ptr = &xi[nxi];
1953
1954                 if (prison_xinpcb(req->td, inp))
1955                         continue;
1956
1957                 xi_ptr->xi_len = sizeof(*xi_ptr);
1958                 bcopy(inp, &xi_ptr->xi_inp, sizeof(*inp));
1959                 if (inp->inp_socket)
1960                         sotoxsocket(inp->inp_socket, &xi_ptr->xi_socket);
1961                 ++nxi;
1962         }
1963
1964         if (nxi == 0) {
1965                 kfree(xi, M_TEMP);
1966                 return 0;
1967         }
1968
1969         *nxi0 = nxi;
1970         *xi0 = xi;
1971
1972         return 0;
1973 }
1974
1975 void
1976 in_savefaddr(struct socket *so, const struct sockaddr *faddr)
1977 {
1978         struct sockaddr_in *sin;
1979
1980         KASSERT(faddr->sa_family == AF_INET,
1981             ("not AF_INET faddr %d", faddr->sa_family));
1982
1983         sin = kmalloc(sizeof(*sin), M_SONAME, M_WAITOK | M_ZERO);
1984         sin->sin_family = AF_INET;
1985         sin->sin_len = sizeof(*sin);
1986         sin->sin_port = ((const struct sockaddr_in *)faddr)->sin_port;
1987         sin->sin_addr = ((const struct sockaddr_in *)faddr)->sin_addr;
1988
1989         so->so_faddr = (struct sockaddr *)sin;
1990 }