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 static int
 297 in_pcbsetlport(struct inpcb *inp, int wild, struct ucred *cred)
 298 {
 299         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 300         struct inpcbportinfo *portinfo;
 301         u_short first, last, lport, step;
 302         u_short *lastport;
 303         int count, error;
 304         int portinfo_first, portinfo_idx;
 305
 306         inp->inp_flags |= INP_ANONPORT;
 307
 308         step = pcbinfo->portinfo_mask + 1;
 309         portinfo_first = mycpuid & pcbinfo->portinfo_mask;
 310         portinfo_idx = portinfo_first;
 311 loop:
 312         portinfo = &pcbinfo->portinfo[portinfo_idx];
 313
 314         if (inp->inp_flags & INP_HIGHPORT) {
 315                 first = ipport_hifirstauto;     /* sysctl */
 316                 last  = ipport_hilastauto;
 317                 lastport = &portinfo->lasthi;
 318         } else if (inp->inp_flags & INP_LOWPORT) {
 319                 if (cred &&
 320                     (error =
 321                      priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 322                         inp->inp_laddr.s_addr = INADDR_ANY;
 323                         return error;
 324                 }
 325                 first = ipport_lowfirstauto;    /* 1023 */
 326                 last  = ipport_lowlastauto;     /* 600 */
 327                 lastport = &portinfo->lastlow;
 328         } else {
 329                 first = ipport_firstauto;       /* sysctl */
 330                 last  = ipport_lastauto;
 331                 lastport = &portinfo->lastport;
 332         }
 333
 334         /*
 335          * This has to be atomic.  If the porthash is shared across multiple
 336          * protocol threads (aka tcp) then the token must be held.
 337          */
 338         GET_PORT_TOKEN(portinfo);
 339
 340         /*
 341          * Simple check to ensure all ports are not used up causing
 342          * a deadlock here.
 343          *
 344          * We split the two cases (up and down) so that the direction
 345          * is not being tested on each round of the loop.
 346          */
 347         if (first > last) {
 348                 /*
 349                  * counting down
 350                  */
 351                 in_pcbportrange(&first, &last, portinfo->offset, step);
 352                 count = (first - last) / step;
 353
 354                 do {
 355                         if (count-- < 0) {      /* completely used? */
 356                                 error = EADDRNOTAVAIL;
 357                                 goto done;
 358                         }
 359                         *lastport -= step;
 360                         if (*lastport > first || *lastport < last)
 361                                 *lastport = first;
 362                         KKASSERT((*lastport & pcbinfo->portinfo_mask) ==
 363                             portinfo->offset);
 364                         lport = htons(*lastport);
 365                 } while (in_pcblookup_local(portinfo, inp->inp_laddr, lport,
 366                     wild, cred));
 367         } else {
 368                 /*
 369                  * counting up
 370                  */
 371                 in_pcbportrange(&last, &first, portinfo->offset, step);
 372                 count = (last - first) / step;
 373
 374                 do {
 375                         if (count-- < 0) {      /* completely used? */
 376                                 error = EADDRNOTAVAIL;
 377                                 goto done;
 378                         }
 379                         *lastport += step;
 380                         if (*lastport < first || *lastport > last)
 381                                 *lastport = first;
 382                         KKASSERT((*lastport & pcbinfo->portinfo_mask) ==
 383                             portinfo->offset);
 384                         lport = htons(*lastport);
 385                 } while (in_pcblookup_local(portinfo, inp->inp_laddr, lport,
 386                     wild, cred));
 387         }
 388         inp->inp_lport = lport;
 389         in_pcbinsporthash(portinfo, inp);
 390         error = 0;
 391 done:
 392         REL_PORT_TOKEN(portinfo);
 393
 394         if (error) {
 395                 /* Try next portinfo */
 396                 portinfo_idx++;
 397                 portinfo_idx &= pcbinfo->portinfo_mask;
 398                 if (portinfo_idx != portinfo_first)
 399                         goto loop;
 400                 inp->inp_laddr.s_addr = INADDR_ANY;
 401         }
 402         return error;
 403 }
 404
 405 int
 406 in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 407 {
 408         struct socket *so = inp->inp_socket;
 409         struct sockaddr_in jsin;
 410         struct ucred *cred = NULL;
 411         int wild = 0;
 412
 413         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 414                 return (EADDRNOTAVAIL);
 415         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY)
 416                 return (EINVAL);        /* already bound */
 417
 418         if (!(so->so_options & (SO_REUSEADDR|SO_REUSEPORT)))
 419                 wild = 1;    /* neither SO_REUSEADDR nor SO_REUSEPORT is set */
 420         if (td->td_proc)
 421                 cred = td->td_proc->p_ucred;
 422
 423         if (nam != NULL) {
 424                 struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 425                 struct inpcbinfo *pcbinfo;
 426                 struct inpcbportinfo *portinfo;
 427                 struct inpcb *t;
 428                 u_short lport, lport_ho;
 429                 int reuseport = (so->so_options & SO_REUSEPORT);
 430                 int error;
 431
 432                 if (nam->sa_len != sizeof *sin)
 433                         return (EINVAL);
 434 #ifdef notdef
 435                 /*
 436                  * We should check the family, but old programs
 437                  * incorrectly fail to initialize it.
 438                  */
 439                 if (sin->sin_family != AF_INET)
 440                         return (EAFNOSUPPORT);
 441 #endif
 442                 if (!prison_replace_wildcards(td, nam))
 443                         return (EINVAL);
 444
 445                 lport = sin->sin_port;
 446                 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 447                         /*
 448                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
 449                          * allow complete duplication of binding if
 450                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
 451                          * and a multicast address is bound on both
 452                          * new and duplicated sockets.
 453                          */
 454                         if (so->so_options & SO_REUSEADDR)
 455                                 reuseport = SO_REUSEADDR | SO_REUSEPORT;
 456                 } else if (sin->sin_addr.s_addr != INADDR_ANY) {
 457                         sin->sin_port = 0;              /* yech... */
 458                         bzero(&sin->sin_zero, sizeof sin->sin_zero);
 459                         if (ifa_ifwithaddr((struct sockaddr *)sin) == NULL)
 460                                 return (EADDRNOTAVAIL);
 461                 }
 462
 463                 inp->inp_laddr = sin->sin_addr;
 464
 465                 jsin.sin_family = AF_INET;
 466                 jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 467                 if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 468                         inp->inp_laddr.s_addr = INADDR_ANY;
 469                         return (EINVAL);
 470                 }
 471                 inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 472
 473                 if (lport == 0) {
 474                         /* Auto-select local port */
 475                         return in_pcbsetlport(inp, wild, cred);
 476                 }
 477                 lport_ho = ntohs(lport);
 478
 479                 /* GROSS */
 480                 if (lport_ho < IPPORT_RESERVED && cred &&
 481                     (error =
 482                      priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 483                         inp->inp_laddr.s_addr = INADDR_ANY;
 484                         return (error);
 485                 }
 486
 487                 /*
 488                  * Locate the proper portinfo based on lport
 489                  */
 490                 pcbinfo = inp->inp_pcbinfo;
 491                 portinfo =
 492                     &pcbinfo->portinfo[lport_ho & pcbinfo->portinfo_mask];
 493                 KKASSERT((lport_ho & pcbinfo->portinfo_mask) ==
 494                     portinfo->offset);
 495
 496                 /*
 497                  * This has to be atomic.  If the porthash is shared across
 498                  * multiple protocol threads (aka tcp) then the token must
 499                  * be held.
 500                  */
 501                 GET_PORT_TOKEN(portinfo);
 502
 503                 if (so->so_cred->cr_uid != 0 &&
 504                     !IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
 505                         t = in_pcblookup_local(portinfo, sin->sin_addr, lport,
 506                             INPLOOKUP_WILDCARD, cred);
 507                         if (t &&
 508                             (!in_nullhost(sin->sin_addr) ||
 509                              !in_nullhost(t->inp_laddr) ||
 510                              (t->inp_socket->so_options & SO_REUSEPORT) == 0) &&
 511                             (so->so_cred->cr_uid !=
 512                              t->inp_socket->so_cred->cr_uid)) {
 513 #ifdef INET6
 514                                 if (!in_nullhost(sin->sin_addr) ||
 515                                     !in_nullhost(t->inp_laddr) ||
 516                                     INP_SOCKAF(so) == INP_SOCKAF(t->inp_socket))
 517 #endif
 518                                 {
 519                                         inp->inp_laddr.s_addr = INADDR_ANY;
 520                                         error = EADDRINUSE;
 521                                         goto done;
 522                                 }
 523                         }
 524                 }
 525                 if (cred && !prison_replace_wildcards(td, nam)) {
 526                         inp->inp_laddr.s_addr = INADDR_ANY;
 527                         error = EADDRNOTAVAIL;
 528                         goto done;
 529                 }
 530                 t = in_pcblookup_local(portinfo, sin->sin_addr, lport,
 531                     wild, cred);
 532                 if (t && !(reuseport & t->inp_socket->so_options)) {
 533 #ifdef INET6
 534                         if (!in_nullhost(sin->sin_addr) ||
 535                             !in_nullhost(t->inp_laddr) ||
 536                             INP_SOCKAF(so) == INP_SOCKAF(t->inp_socket))
 537 #endif
 538                         {
 539                                 inp->inp_laddr.s_addr = INADDR_ANY;
 540                                 error = EADDRINUSE;
 541                                 goto done;
 542                         }
 543                 }
 544                 inp->inp_lport = lport;
 545                 in_pcbinsporthash(portinfo, inp);
 546                 error = 0;
 547 done:
 548                 REL_PORT_TOKEN(portinfo);
 549                 return (error);
 550         } else {
 551                 jsin.sin_family = AF_INET;
 552                 jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 553                 if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 554                         inp->inp_laddr.s_addr = INADDR_ANY;
 555                         return (EINVAL);
 556                 }
 557                 inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 558
 559                 return in_pcbsetlport(inp, wild, cred);
 560         }
 561 }
 562
 563 static struct inpcb *
 564 in_pcblookup_localremote(struct inpcbportinfo *portinfo, struct in_addr laddr,
 565     u_short lport, struct in_addr faddr, u_short fport, struct ucred *cred)
 566 {
 567         struct inpcb *inp;
 568         struct inpcbporthead *porthash;
 569         struct inpcbport *phd;
 570         struct inpcb *match = NULL;
 571
 572         /*
 573          * If the porthashbase is shared across several cpus, it must
 574          * have been locked.
 575          */
 576         ASSERT_PORT_TOKEN_HELD(portinfo);
 577
 578         /*
 579          * Best fit PCB lookup.
 580          *
 581          * First see if this local port is in use by looking on the
 582          * port hash list.
 583          */
 584         porthash = &portinfo->porthashbase[
 585                         INP_PCBPORTHASH(lport, portinfo->porthashmask)];
 586         LIST_FOREACH(phd, porthash, phd_hash) {
 587                 if (phd->phd_port == lport)
 588                         break;
 589         }
 590         if (phd != NULL) {
 591                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
 592 #ifdef INET6
 593                         if ((inp->inp_vflag & INP_IPV4) == 0)
 594                                 continue;
 595 #endif
 596                         if (inp->inp_laddr.s_addr != INADDR_ANY &&
 597                             inp->inp_laddr.s_addr != laddr.s_addr)
 598                                 continue;
 599
 600                         if (inp->inp_faddr.s_addr != INADDR_ANY &&
 601                             inp->inp_faddr.s_addr != faddr.s_addr)
 602                                 continue;
 603
 604                         if (inp->inp_fport != 0 && inp->inp_fport != fport)
 605                                 continue;
 606
 607                         if (cred == NULL ||
 608                             cred->cr_prison ==
 609                             inp->inp_socket->so_cred->cr_prison) {
 610                                 match = inp;
 611                                 break;
 612                         }
 613                 }
 614         }
 615         return (match);
 616 }
 617
 618 int
 619 in_pcbsetlport_remote(struct inpcb *inp, const struct sockaddr *remote,
 620     struct thread *td)
 621 {
 622         struct proc *p = td->td_proc;
 623         unsigned short *lastport;
 624         const struct sockaddr_in *sin = (const struct sockaddr_in *)remote;
 625         struct sockaddr_in jsin;
 626         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
 627         struct inpcbportinfo *portinfo;
 628         struct ucred *cred = NULL;
 629         u_short first, last, lport, step;
 630         int count, error, dup;
 631         int portinfo_first, portinfo_idx;
 632
 633         if (TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) /* XXX broken! */
 634                 return (EADDRNOTAVAIL);
 635
 636         KKASSERT(inp->inp_laddr.s_addr != INADDR_ANY);
 637         if (inp->inp_lport != 0)
 638                 return (EINVAL);        /* already bound */
 639
 640         KKASSERT(p);
 641         cred = p->p_ucred;
 642
 643         jsin.sin_family = AF_INET;
 644         jsin.sin_addr.s_addr = inp->inp_laddr.s_addr;
 645         if (!prison_replace_wildcards(td, (struct sockaddr *)&jsin)) {
 646                 inp->inp_laddr.s_addr = INADDR_ANY;
 647                 return (EINVAL);
 648         }
 649         inp->inp_laddr.s_addr = jsin.sin_addr.s_addr;
 650
 651         inp->inp_flags |= INP_ANONPORT;
 652
 653         step = pcbinfo->portinfo_mask + 1;
 654         portinfo_first = mycpuid & pcbinfo->portinfo_mask;
 655         portinfo_idx = portinfo_first;
 656 loop:
 657         portinfo = &pcbinfo->portinfo[portinfo_idx];
 658         dup = 0;
 659
 660         if (inp->inp_flags & INP_HIGHPORT) {
 661                 first = ipport_hifirstauto;     /* sysctl */
 662                 last  = ipport_hilastauto;
 663                 lastport = &portinfo->lasthi;
 664         } else if (inp->inp_flags & INP_LOWPORT) {
 665                 if (cred &&
 666                     (error =
 667                      priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT, 0))) {
 668                         inp->inp_laddr.s_addr = INADDR_ANY;
 669                         return (error);
 670                 }
 671                 first = ipport_lowfirstauto;    /* 1023 */
 672                 last  = ipport_lowlastauto;     /* 600 */
 673                 lastport = &portinfo->lastlow;
 674         } else {
 675                 first = ipport_firstauto;       /* sysctl */
 676                 last  = ipport_lastauto;
 677                 lastport = &portinfo->lastport;
 678         }
 679
 680         /*
 681          * This has to be atomic.  If the porthash is shared across multiple
 682          * protocol threads (aka tcp) then the token must be held.
 683          */
 684         GET_PORT_TOKEN(portinfo);
 685
 686 again:
 687         /*
 688          * Simple check to ensure all ports are not used up causing
 689          * a deadlock here.
 690          *
 691          * We split the two cases (up and down) so that the direction
 692          * is not being tested on each round of the loop.
 693          */
 694         if (first > last) {
 695                 /*
 696                  * counting down
 697                  */
 698                 in_pcbportrange(&first, &last, portinfo->offset, step);
 699                 count = (first - last) / step;
 700
 701                 do {
 702                         if (count-- < 0) {      /* completely used? */
 703                                 error = EADDRNOTAVAIL;
 704                                 goto done;
 705                         }
 706                         *lastport -= step;
 707                         if (*lastport > first || *lastport < last)
 708                                 *lastport = first;
 709                         KKASSERT((*lastport & pcbinfo->portinfo_mask) ==
 710                             portinfo->offset);
 711                         lport = htons(*lastport);
 712                 } while (in_pcblookup_localremote(portinfo, inp->inp_laddr,
 713                     lport, sin->sin_addr, sin->sin_port, cred));
 714         } else {
 715                 /*
 716                  * counting up
 717                  */
 718                 in_pcbportrange(&last, &first, portinfo->offset, step);
 719                 count = (last - first) / step;
 720
 721                 do {
 722                         if (count-- < 0) {      /* completely used? */
 723                                 error = EADDRNOTAVAIL;
 724                                 goto done;
 725                         }
 726                         *lastport += step;
 727                         if (*lastport < first || *lastport > last)
 728                                 *lastport = first;
 729                         KKASSERT((*lastport & pcbinfo->portinfo_mask) ==
 730                             portinfo->offset);
 731                         lport = htons(*lastport);
 732                 } while (in_pcblookup_localremote(portinfo, inp->inp_laddr,
 733                     lport, sin->sin_addr, sin->sin_port, cred));
 734         }
 735
 736         /* This could happen on loopback interface */
 737         if (sin->sin_port == lport &&
 738             sin->sin_addr.s_addr == inp->inp_laddr.s_addr) {
 739                 if (dup) {
 740                         /*
 741                          * Duplicate again; give up
 742                          */
 743                         error = EADDRNOTAVAIL;
 744                         goto done;
 745                 }
 746                 dup = 1;
 747                 goto again;
 748         }
 749         inp->inp_lport = lport;
 750         in_pcbinsporthash(portinfo, inp);
 751         error = 0;
 752 done:
 753         REL_PORT_TOKEN(portinfo);
 754
 755         if (error) {
 756                 /* Try next portinfo */
 757                 portinfo_idx++;
 758                 portinfo_idx &= pcbinfo->portinfo_mask;
 759                 if (portinfo_idx != portinfo_first)
 760                         goto loop;
 761                 inp->inp_laddr.s_addr = INADDR_ANY;
 762         }
 763         return error;
 764 }
 765
 766 /*
 767  *   Transform old in_pcbconnect() into an inner subroutine for new
 768  *   in_pcbconnect(): Do some validity-checking on the remote
 769  *   address (in mbuf 'nam') and then determine local host address
 770  *   (i.e., which interface) to use to access that remote host.
 771  *
 772  *   This preserves definition of in_pcbconnect(), while supporting a
 773  *   slightly different version for T/TCP.  (This is more than
 774  *   a bit of a kludge, but cleaning up the internal interfaces would
 775  *   have forced minor changes in every protocol).
 776  */
 777 int
 778 in_pcbladdr_find(struct inpcb *inp, struct sockaddr *nam,
 779     struct sockaddr_in **plocal_sin, struct thread *td, int find)
 780 {
 781         struct in_ifaddr *ia;
 782         struct ucred *cred = NULL;
 783         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 784         struct sockaddr *jsin;
 785         int jailed = 0, alloc_route = 0;
 786
 787         if (nam->sa_len != sizeof *sin)
 788                 return (EINVAL);
 789         if (sin->sin_family != AF_INET)
 790                 return (EAFNOSUPPORT);
 791         if (sin->sin_port == 0)
 792                 return (EADDRNOTAVAIL);
 793         if (td && td->td_proc && td->td_proc->p_ucred)
 794                 cred = td->td_proc->p_ucred;
 795         if (cred && cred->cr_prison)
 796                 jailed = 1;
 797         if (!TAILQ_EMPTY(&in_ifaddrheads[mycpuid])) {
 798                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 799                 /*
 800                  * If the destination address is INADDR_ANY,
 801                  * use the primary local address.
 802                  * If the supplied address is INADDR_BROADCAST,
 803                  * and the primary interface supports broadcast,
 804                  * choose the broadcast address for that interface.
 805                  */
 806                 if (sin->sin_addr.s_addr == INADDR_ANY)
 807                         sin->sin_addr = IA_SIN(ia)->sin_addr;
 808                 else if (sin->sin_addr.s_addr == (u_long)INADDR_BROADCAST &&
 809                     (ia->ia_ifp->if_flags & IFF_BROADCAST))
 810                         sin->sin_addr = satosin(&ia->ia_broadaddr)->sin_addr;
 811         }
 812         if (find) {
 813                 struct route *ro;
 814
 815                 ia = NULL;
 816                 /*
 817                  * If route is known or can be allocated now,
 818                  * our src addr is taken from the i/f, else punt.
 819                  * Note that we should check the address family of the cached
 820                  * destination, in case of sharing the cache with IPv6.
 821                  */
 822                 ro = &inp->inp_route;
 823                 if (ro->ro_rt &&
 824                     (!(ro->ro_rt->rt_flags & RTF_UP) ||
 825                      ro->ro_dst.sa_family != AF_INET ||
 826                      satosin(&ro->ro_dst)->sin_addr.s_addr !=
 827                                       sin->sin_addr.s_addr ||
 828                      inp->inp_socket->so_options & SO_DONTROUTE)) {
 829                         RTFREE(ro->ro_rt);
 830                         ro->ro_rt = NULL;
 831                 }
 832                 if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
 833                     (ro->ro_rt == NULL ||
 834                     ro->ro_rt->rt_ifp == NULL)) {
 835                         /* No route yet, so try to acquire one */
 836                         bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
 837                         ro->ro_dst.sa_family = AF_INET;
 838                         ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
 839                         ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
 840                                 sin->sin_addr;
 841                         rtalloc(ro);
 842                         alloc_route = 1;
 843                 }
 844                 /*
 845                  * If we found a route, use the address
 846                  * corresponding to the outgoing interface
 847                  * unless it is the loopback (in case a route
 848                  * to our address on another net goes to loopback).
 849                  */
 850                 if (ro->ro_rt &&
 851                     !(ro->ro_rt->rt_ifp->if_flags & IFF_LOOPBACK)) {
 852                         if (jailed) {
 853                                 if (jailed_ip(cred->cr_prison,
 854                                     ro->ro_rt->rt_ifa->ifa_addr)) {
 855                                         ia = ifatoia(ro->ro_rt->rt_ifa);
 856                                 }
 857                         } else {
 858                                 ia = ifatoia(ro->ro_rt->rt_ifa);
 859                         }
 860                 }
 861                 if (ia == NULL) {
 862                         u_short fport = sin->sin_port;
 863
 864                         sin->sin_port = 0;
 865                         ia = ifatoia(ifa_ifwithdstaddr(sintosa(sin)));
 866                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 867                             sintosa(&ia->ia_addr)))
 868                                 ia = NULL;
 869                         if (ia == NULL)
 870                                 ia = ifatoia(ifa_ifwithnet(sintosa(sin)));
 871                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 872                             sintosa(&ia->ia_addr)))
 873                                 ia = NULL;
 874                         sin->sin_port = fport;
 875                         if (ia == NULL &&
 876                             !TAILQ_EMPTY(&in_ifaddrheads[mycpuid]))
 877                                 ia = TAILQ_FIRST(&in_ifaddrheads[mycpuid])->ia;
 878                         if (ia && jailed && !jailed_ip(cred->cr_prison,
 879                             sintosa(&ia->ia_addr)))
 880                                 ia = NULL;
 881
 882                         if (!jailed && ia == NULL)
 883                                 goto fail;
 884                 }
 885                 /*
 886                  * If the destination address is multicast and an outgoing
 887                  * interface has been set as a multicast option, use the
 888                  * address of that interface as our source address.
 889                  */
 890                 if (!jailed && IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) &&
 891                     inp->inp_moptions != NULL) {
 892                         struct ip_moptions *imo;
 893                         struct ifnet *ifp;
 894
 895                         imo = inp->inp_moptions;
 896                         if (imo->imo_multicast_ifp != NULL) {
 897                                 struct in_ifaddr_container *iac;
 898
 899                                 ifp = imo->imo_multicast_ifp;
 900                                 ia = NULL;
 901                                 TAILQ_FOREACH(iac,
 902                                 &in_ifaddrheads[mycpuid], ia_link) {
 903                                         if (iac->ia->ia_ifp == ifp) {
 904                                                 ia = iac->ia;
 905                                                 break;
 906                                         }
 907                                 }
 908                                 if (ia == NULL)
 909                                         goto fail;
 910                         }
 911                 }
 912                 /*
 913                  * Don't do pcblookup call here; return interface in plocal_sin
 914                  * and exit to caller, that will do the lookup.
 915                  */
 916                 if (ia == NULL && jailed) {
 917                         if ((jsin = prison_get_nonlocal(
 918                                 cred->cr_prison, AF_INET, NULL)) != NULL ||
 919                             (jsin = prison_get_local(
 920                                 cred->cr_prison, AF_INET, NULL)) != NULL) {
 921                                 *plocal_sin = satosin(jsin);
 922                         } else {
 923                                 /* IPv6 only Jail */
 924                                 goto fail;
 925                         }
 926                 } else {
 927                         *plocal_sin = &ia->ia_addr;
 928                 }
 929         }
 930         return (0);
 931 fail:
 932         if (alloc_route) {
 933                 struct route *ro = &inp->inp_route;
 934
 935                 if (ro->ro_rt != NULL)
 936                         RTFREE(ro->ro_rt);
 937                 bzero(ro, sizeof(*ro));
 938         }
 939         return (EADDRNOTAVAIL);
 940 }
 941
 942 int
 943 in_pcbladdr(struct inpcb *inp, struct sockaddr *nam,
 944     struct sockaddr_in **plocal_sin, struct thread *td)
 945 {
 946         return in_pcbladdr_find(inp, nam, plocal_sin, td,
 947             (inp->inp_laddr.s_addr == INADDR_ANY));
 948 }
 949
 950 /*
 951  * Outer subroutine:
 952  * Connect from a socket to a specified address.
 953  * Both address and port must be specified in argument sin.
 954  * If don't have a local address for this socket yet,
 955  * then pick one.
 956  */
 957 int
 958 in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct thread *td)
 959 {
 960         struct sockaddr_in *if_sin;
 961         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
 962         int error;
 963
 964         /* Call inner routine to assign local interface address. */
 965         if ((error = in_pcbladdr(inp, nam, &if_sin, td)) != 0)
 966                 return (error);
 967
 968         if (in_pcblookup_hash(inp->inp_cpcbinfo, sin->sin_addr, sin->sin_port,
 969                               inp->inp_laddr.s_addr ?
 970                                 inp->inp_laddr : if_sin->sin_addr,
 971                               inp->inp_lport, FALSE, NULL) != NULL) {
 972                 return (EADDRINUSE);
 973         }
 974         if (inp->inp_laddr.s_addr == INADDR_ANY) {
 975                 if (inp->inp_lport == 0) {
 976                         error = in_pcbbind(inp, NULL, td);
 977                         if (error)
 978                                 return (error);
 979                 }
 980                 inp->inp_laddr = if_sin->sin_addr;
 981         }
 982         inp->inp_faddr = sin->sin_addr;
 983         inp->inp_fport = sin->sin_port;
 984         in_pcbinsconnhash(inp);
 985         return (0);
 986 }
 987
 988 void
 989 in_pcbdisconnect(struct inpcb *inp)
 990 {
 991
 992         inp->inp_faddr.s_addr = INADDR_ANY;
 993         inp->inp_fport = 0;
 994         in_pcbremconnhash(inp);
 995 }
 996
 997 void
 998 in_pcbdetach(struct inpcb *inp)
 999 {
1000         struct socket *so = inp->inp_socket;
1001         struct inpcbinfo *ipi = inp->inp_pcbinfo;
1002
1003 #ifdef IPSEC
1004         ipsec4_delete_pcbpolicy(inp);
1005 #endif /*IPSEC*/
1006         inp->inp_gencnt = ++ipi->ipi_gencnt;
1007         KKASSERT((so->so_state & SS_ASSERTINPROG) == 0);
1008         in_pcbremlists(inp);
1009         so->so_pcb = NULL;
1010         sofree(so);                     /* remove pcb ref */
1011         if (inp->inp_options)
1012                 m_free(inp->inp_options);
1013         if (inp->inp_route.ro_rt)
1014                 rtfree(inp->inp_route.ro_rt);
1015         ip_freemoptions(inp->inp_moptions);
1016         inp->inp_vflag = 0;
1017         kfree(inp, M_PCB);
1018 }
1019
1020 /*
1021  * The calling convention of in_setsockaddr() and in_setpeeraddr() was
1022  * modified to match the pru_sockaddr() and pru_peeraddr() entry points
1023  * in struct pr_usrreqs, so that protocols can just reference then directly
1024  * without the need for a wrapper function.  The socket must have a valid
1025  * (i.e., non-nil) PCB, but it should be impossible to get an invalid one
1026  * except through a kernel programming error, so it is acceptable to panic
1027  * (or in this case trap) if the PCB is invalid.  (Actually, we don't trap
1028  * because there actually /is/ a programming error somewhere... XXX)
1029  */
1030 int
1031 in_setsockaddr(struct socket *so, struct sockaddr **nam)
1032 {
1033         struct inpcb *inp;
1034         struct sockaddr_in *sin;
1035
1036         /*
1037          * Do the malloc first in case it blocks.
1038          */
1039         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
1040         sin->sin_family = AF_INET;
1041         sin->sin_len = sizeof *sin;
1042
1043         crit_enter();
1044         inp = so->so_pcb;
1045         if (!inp) {
1046                 crit_exit();
1047                 kfree(sin, M_SONAME);
1048                 return (ECONNRESET);
1049         }
1050         sin->sin_port = inp->inp_lport;
1051         sin->sin_addr = inp->inp_laddr;
1052         crit_exit();
1053
1054         *nam = (struct sockaddr *)sin;
1055         return (0);
1056 }
1057
1058 void
1059 in_setsockaddr_dispatch(netmsg_t msg)
1060 {
1061         int error;
1062
1063         error = in_setsockaddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1064         lwkt_replymsg(&msg->lmsg, error);
1065 }
1066
1067 int
1068 in_setpeeraddr(struct socket *so, struct sockaddr **nam)
1069 {
1070         struct inpcb *inp;
1071         struct sockaddr_in *sin;
1072
1073         /*
1074          * Do the malloc first in case it blocks.
1075          */
1076         sin = kmalloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO);
1077         sin->sin_family = AF_INET;
1078         sin->sin_len = sizeof *sin;
1079
1080         crit_enter();
1081         inp = so->so_pcb;
1082         if (!inp) {
1083                 crit_exit();
1084                 kfree(sin, M_SONAME);
1085                 return (ECONNRESET);
1086         }
1087         sin->sin_port = inp->inp_fport;
1088         sin->sin_addr = inp->inp_faddr;
1089         crit_exit();
1090
1091         *nam = (struct sockaddr *)sin;
1092         return (0);
1093 }
1094
1095 void
1096 in_setpeeraddr_dispatch(netmsg_t msg)
1097 {
1098         int error;
1099
1100         error = in_setpeeraddr(msg->base.nm_so, msg->peeraddr.nm_nam);
1101         lwkt_replymsg(&msg->lmsg, error);
1102 }
1103
1104 void
1105 in_pcbnotifyall(struct inpcbhead *head, struct in_addr faddr, int err,
1106                 void (*notify)(struct inpcb *, int))
1107 {
1108         struct inpcb *inp, *ninp;
1109
1110         /*
1111          * note: if INP_PLACEMARKER is set we must ignore the rest of
1112          * the structure and skip it.
1113          */
1114         crit_enter();
1115         LIST_FOREACH_MUTABLE(inp, head, inp_list, ninp) {
1116                 if (inp->inp_flags & INP_PLACEMARKER)
1117                         continue;
1118 #ifdef INET6
1119                 if (!(inp->inp_vflag & INP_IPV4))
1120                         continue;
1121 #endif
1122                 if (inp->inp_faddr.s_addr != faddr.s_addr ||
1123                     inp->inp_socket == NULL)
1124                         continue;
1125                 (*notify)(inp, err);            /* can remove inp from list! */
1126         }
1127         crit_exit();
1128 }
1129
1130 void
1131 in_pcbpurgeif0(struct inpcb *head, struct ifnet *ifp)
1132 {
1133         struct inpcb *inp;
1134         struct ip_moptions *imo;
1135         int i, gap;
1136
1137         for (inp = head; inp != NULL; inp = LIST_NEXT(inp, inp_list)) {
1138                 if (inp->inp_flags & INP_PLACEMARKER)
1139                         continue;
1140                 imo = inp->inp_moptions;
1141                 if ((inp->inp_vflag & INP_IPV4) && imo != NULL) {
1142                         /*
1143                          * Unselect the outgoing interface if it is being
1144                          * detached.
1145                          */
1146                         if (imo->imo_multicast_ifp == ifp)
1147                                 imo->imo_multicast_ifp = NULL;
1148
1149                         /*
1150                          * Drop multicast group membership if we joined
1151                          * through the interface being detached.
1152                          */
1153                         for (i = 0, gap = 0; i < imo->imo_num_memberships;
1154                             i++) {
1155                                 if (imo->imo_membership[i]->inm_ifp == ifp) {
1156                                         in_delmulti(imo->imo_membership[i]);
1157                                         gap++;
1158                                 } else if (gap != 0)
1159                                         imo->imo_membership[i - gap] =
1160                                             imo->imo_membership[i];
1161                         }
1162                         imo->imo_num_memberships -= gap;
1163                 }
1164         }
1165 }
1166
1167 /*
1168  * Check for alternatives when higher level complains
1169  * about service problems.  For now, invalidate cached
1170  * routing information.  If the route was created dynamically
1171  * (by a redirect), time to try a default gateway again.
1172  */
1173 void
1174 in_losing(struct inpcb *inp)
1175 {
1176         struct rtentry *rt;
1177         struct rt_addrinfo rtinfo;
1178
1179         if ((rt = inp->inp_route.ro_rt)) {
1180                 bzero(&rtinfo, sizeof(struct rt_addrinfo));
1181                 rtinfo.rti_info[RTAX_DST] = rt_key(rt);
1182                 rtinfo.rti_info[RTAX_GATEWAY] = rt->rt_gateway;
1183                 rtinfo.rti_info[RTAX_NETMASK] = rt_mask(rt);
1184                 rtinfo.rti_flags = rt->rt_flags;
1185                 rt_missmsg(RTM_LOSING, &rtinfo, rt->rt_flags, 0);
1186                 if (rt->rt_flags & RTF_DYNAMIC) {
1187                         rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway,
1188                             rt_mask(rt), rt->rt_flags, NULL);
1189                 }
1190                 inp->inp_route.ro_rt = NULL;
1191                 rtfree(rt);
1192                 /*
1193                  * A new route can be allocated
1194                  * the next time output is attempted.
1195                  */
1196         }
1197 }
1198
1199 /*
1200  * After a routing change, flush old routing
1201  * and allocate a (hopefully) better one.
1202  */
1203 void
1204 in_rtchange(struct inpcb *inp, int err)
1205 {
1206         if (inp->inp_route.ro_rt) {
1207                 rtfree(inp->inp_route.ro_rt);
1208                 inp->inp_route.ro_rt = NULL;
1209                 /*
1210                  * A new route can be allocated the next time
1211                  * output is attempted.
1212                  */
1213         }
1214 }
1215
1216 /*
1217  * Lookup a PCB based on the local address and port.
1218  */
1219 struct inpcb *
1220 in_pcblookup_local(struct inpcbportinfo *portinfo, struct in_addr laddr,
1221                    u_int lport_arg, int wild_okay, struct ucred *cred)
1222 {
1223         struct inpcb *inp;
1224         int matchwild = 3, wildcard;
1225         u_short lport = lport_arg;
1226         struct inpcbporthead *porthash;
1227         struct inpcbport *phd;
1228         struct inpcb *match = NULL;
1229
1230         /*
1231          * If the porthashbase is shared across several cpus, it must
1232          * have been locked.
1233          */
1234         ASSERT_PORT_TOKEN_HELD(portinfo);
1235
1236         /*
1237          * Best fit PCB lookup.
1238          *
1239          * First see if this local port is in use by looking on the
1240          * port hash list.
1241          */
1242         porthash = &portinfo->porthashbase[
1243                         INP_PCBPORTHASH(lport, portinfo->porthashmask)];
1244         LIST_FOREACH(phd, porthash, phd_hash) {
1245                 if (phd->phd_port == lport)
1246                         break;
1247         }
1248         if (phd != NULL) {
1249                 /*
1250                  * Port is in use by one or more PCBs. Look for best
1251                  * fit.
1252                  */
1253                 LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
1254                         wildcard = 0;
1255 #ifdef INET6
1256                         if ((inp->inp_vflag & INP_IPV4) == 0)
1257                                 continue;
1258 #endif
1259                         if (inp->inp_faddr.s_addr != INADDR_ANY)
1260                                 wildcard++;
1261                         if (inp->inp_laddr.s_addr != INADDR_ANY) {
1262                                 if (laddr.s_addr == INADDR_ANY)
1263                                         wildcard++;
1264                                 else if (inp->inp_laddr.s_addr != laddr.s_addr)
1265                                         continue;
1266                         } else {
1267                                 if (laddr.s_addr != INADDR_ANY)
1268                                         wildcard++;
1269                         }
1270                         if (wildcard && !wild_okay)
1271                                 continue;
1272                         if (wildcard < matchwild &&
1273                             (cred == NULL ||
1274                              cred->cr_prison ==
1275                                         inp->inp_socket->so_cred->cr_prison)) {
1276                                 match = inp;
1277                                 matchwild = wildcard;
1278                                 if (matchwild == 0) {
1279                                         break;
1280                                 }
1281                         }
1282                 }
1283         }
1284         return (match);
1285 }
1286
1287 struct inpcb *
1288 in_pcblocalgroup_last(const struct inpcbinfo *pcbinfo,
1289     const struct inpcb *inp)
1290 {
1291         const struct inp_localgrphead *hdr;
1292         const struct inp_localgroup *grp;
1293         int i;
1294
1295         if (pcbinfo->localgrphashbase == NULL)
1296                 return NULL;
1297
1298         hdr = &pcbinfo->localgrphashbase[
1299             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1300
1301         LIST_FOREACH(grp, hdr, il_list) {
1302                 if (grp->il_vflag == inp->inp_vflag &&
1303                     grp->il_lport == inp->inp_lport &&
1304                     memcmp(&grp->il_dependladdr,
1305                         &inp->inp_inc.inc_ie.ie_dependladdr,
1306                         sizeof(grp->il_dependladdr)) == 0) {
1307                         break;
1308                 }
1309         }
1310         if (grp == NULL || grp->il_inpcnt == 1)
1311                 return NULL;
1312
1313         KASSERT(grp->il_inpcnt >= 2,
1314             ("invalid localgroup inp count %d", grp->il_inpcnt));
1315         for (i = 0; i < grp->il_inpcnt; ++i) {
1316                 if (grp->il_inp[i] == inp) {
1317                         int last = grp->il_inpcnt - 1;
1318
1319                         if (i == last)
1320                                 last = grp->il_inpcnt - 2;
1321                         return grp->il_inp[last];
1322                 }
1323         }
1324         return NULL;
1325 }
1326
1327 static struct inpcb *
1328 inp_localgroup_lookup(const struct inpcbinfo *pcbinfo,
1329     struct in_addr laddr, uint16_t lport, uint32_t pkt_hash)
1330 {
1331         struct inpcb *local_wild = NULL;
1332         const struct inp_localgrphead *hdr;
1333         const struct inp_localgroup *grp;
1334
1335         hdr = &pcbinfo->localgrphashbase[
1336             INP_PCBLOCALGRPHASH(lport, pcbinfo->localgrphashmask)];
1337 #ifdef INP_LOCALGROUP_HASHTHR
1338         pkt_hash >>= ncpus2_shift;
1339 #endif
1340
1341         /*
1342          * Order of socket selection:
1343          * 1. non-wild.
1344          * 2. wild.
1345          *
1346          * NOTE:
1347          * - Local group does not contain jailed sockets
1348          * - Local group does not contain IPv4 mapped INET6 wild sockets
1349          */
1350         LIST_FOREACH(grp, hdr, il_list) {
1351 #ifdef INET6
1352                 if (!(grp->il_vflag & INP_IPV4))
1353                         continue;
1354 #endif
1355                 if (grp->il_lport == lport) {
1356                         int idx;
1357
1358 #ifdef INP_LOCALGROUP_HASHTHR
1359                         idx = pkt_hash / grp->il_factor;
1360                         KASSERT(idx < grp->il_inpcnt && idx >= 0,
1361                             ("invalid hash %04x, cnt %d or fact %d",
1362                              pkt_hash, grp->il_inpcnt, grp->il_factor));
1363 #else
1364                         /*
1365                          * Modulo-N is used here, which greatly reduces
1366                          * completion queue token contention, thus more
1367                          * cpu time is saved.
1368                          */
1369                         idx = pkt_hash % grp->il_inpcnt;
1370 #endif
1371
1372                         if (grp->il_laddr.s_addr == laddr.s_addr)
1373                                 return grp->il_inp[idx];
1374                         else if (grp->il_laddr.s_addr == INADDR_ANY)
1375                                 local_wild = grp->il_inp[idx];
1376                 }
1377         }
1378         if (local_wild != NULL)
1379                 return local_wild;
1380         return NULL;
1381 }
1382
1383 /*
1384  * Lookup PCB in hash list.
1385  */
1386 struct inpcb *
1387 in_pcblookup_pkthash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1388     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1389     boolean_t wildcard, struct ifnet *ifp, const struct mbuf *m)
1390 {
1391         struct inpcbhead *head;
1392         struct inpcb *inp, *jinp=NULL;
1393         u_short fport = fport_arg, lport = lport_arg;
1394
1395         /*
1396          * First look for an exact match.
1397          */
1398         head = &pcbinfo->hashbase[INP_PCBCONNHASH(faddr.s_addr, fport,
1399             laddr.s_addr, lport, pcbinfo->hashmask)];
1400         LIST_FOREACH(inp, head, inp_hash) {
1401 #ifdef INET6
1402                 if (!(inp->inp_vflag & INP_IPV4))
1403                         continue;
1404 #endif
1405                 if (in_hosteq(inp->inp_faddr, faddr) &&
1406                     in_hosteq(inp->inp_laddr, laddr) &&
1407                     inp->inp_fport == fport && inp->inp_lport == lport) {
1408                         /* found */
1409                         if (inp->inp_socket == NULL ||
1410                             inp->inp_socket->so_cred->cr_prison == NULL) {
1411                                 return (inp);
1412                         } else {
1413                                 if  (jinp == NULL)
1414                                         jinp = inp;
1415                         }
1416                 }
1417         }
1418         if (jinp != NULL)
1419                 return (jinp);
1420         if (wildcard) {
1421                 struct inpcb *local_wild = NULL;
1422                 struct inpcb *jinp_wild = NULL;
1423 #ifdef INET6
1424                 struct inpcb *local_wild_mapped = NULL;
1425 #endif
1426                 struct inpcontainer *ic;
1427                 struct inpcontainerhead *chead;
1428                 struct sockaddr_in jsin;
1429                 struct ucred *cred;
1430
1431                 /*
1432                  * Check local group first
1433                  */
1434                 if (pcbinfo->localgrphashbase != NULL &&
1435                     m != NULL && (m->m_flags & M_HASH) &&
1436                     !(ifp && ifp->if_type == IFT_FAITH)) {
1437                         inp = inp_localgroup_lookup(pcbinfo,
1438                             laddr, lport, m->m_pkthdr.hash);
1439                         if (inp != NULL)
1440                                 return inp;
1441                 }
1442
1443                 /*
1444                  * Order of socket selection:
1445                  * 1. non-jailed, non-wild.
1446                  * 2. non-jailed, wild.
1447                  * 3. jailed, non-wild.
1448                  * 4. jailed, wild.
1449                  */
1450                 jsin.sin_family = AF_INET;
1451                 chead = &pcbinfo->wildcardhashbase[
1452                     INP_PCBWILDCARDHASH(lport, pcbinfo->wildcardhashmask)];
1453                 LIST_FOREACH(ic, chead, ic_list) {
1454                         inp = ic->ic_inp;
1455                         jsin.sin_addr.s_addr = laddr.s_addr;
1456 #ifdef INET6
1457                         if (!(inp->inp_vflag & INP_IPV4))
1458                                 continue;
1459 #endif
1460                         if (inp->inp_socket != NULL)
1461                                 cred = inp->inp_socket->so_cred;
1462                         else
1463                                 cred = NULL;
1464                         if (cred != NULL && jailed(cred)) {
1465                                 if (jinp != NULL)
1466                                         continue;
1467                                 else
1468                                         if (!jailed_ip(cred->cr_prison,
1469                                             (struct sockaddr *)&jsin))
1470                                                 continue;
1471                         }
1472                         if (inp->inp_lport == lport) {
1473                                 if (ifp && ifp->if_type == IFT_FAITH &&
1474                                     !(inp->inp_flags & INP_FAITH))
1475                                         continue;
1476                                 if (inp->inp_laddr.s_addr == laddr.s_addr) {
1477                                         if (cred != NULL && jailed(cred))
1478                                                 jinp = inp;
1479                                         else
1480                                                 return (inp);
1481                                 }
1482                                 if (inp->inp_laddr.s_addr == INADDR_ANY) {
1483 #ifdef INET6
1484                                         if (INP_CHECK_SOCKAF(inp->inp_socket,
1485                                                              AF_INET6))
1486                                                 local_wild_mapped = inp;
1487                                         else
1488 #endif
1489                                                 if (cred != NULL &&
1490                                                     jailed(cred))
1491                                                         jinp_wild = inp;
1492                                                 else
1493                                                         local_wild = inp;
1494                                 }
1495                         }
1496                 }
1497                 if (local_wild != NULL)
1498                         return (local_wild);
1499 #ifdef INET6
1500                 if (local_wild_mapped != NULL)
1501                         return (local_wild_mapped);
1502 #endif
1503                 if (jinp != NULL)
1504                         return (jinp);
1505                 return (jinp_wild);
1506         }
1507
1508         /*
1509          * Not found.
1510          */
1511         return (NULL);
1512 }
1513
1514 struct inpcb *
1515 in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,
1516     u_int fport_arg, struct in_addr laddr, u_int lport_arg,
1517     boolean_t wildcard, struct ifnet *ifp)
1518 {
1519         return in_pcblookup_pkthash(pcbinfo, faddr, fport_arg,
1520             laddr, lport_arg, wildcard, ifp, NULL);
1521 }
1522
1523 /*
1524  * Insert PCB into connection hash table.
1525  */
1526 void
1527 in_pcbinsconnhash(struct inpcb *inp)
1528 {
1529         struct inpcbinfo *pcbinfo = inp->inp_cpcbinfo;
1530         struct inpcbhead *bucket;
1531         u_int32_t hashkey_faddr, hashkey_laddr;
1532
1533 #ifdef INET6
1534         if (inp->inp_vflag & INP_IPV6) {
1535                 hashkey_faddr = inp->in6p_faddr.s6_addr32[3] /* XXX JH */;
1536                 hashkey_laddr = inp->in6p_laddr.s6_addr32[3] /* XXX JH */;
1537         } else {
1538 #endif
1539                 hashkey_faddr = inp->inp_faddr.s_addr;
1540                 hashkey_laddr = inp->inp_laddr.s_addr;
1541 #ifdef INET6
1542         }
1543 #endif
1544
1545         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1546                 ("already on wildcardhash"));
1547         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1548                 ("already on connhash"));
1549         inp->inp_flags |= INP_CONNECTED;
1550
1551         /*
1552          * Insert into the connection hash table.
1553          */
1554         bucket = &pcbinfo->hashbase[INP_PCBCONNHASH(hashkey_faddr,
1555             inp->inp_fport, hashkey_laddr, inp->inp_lport, pcbinfo->hashmask)];
1556         LIST_INSERT_HEAD(bucket, inp, inp_hash);
1557 }
1558
1559 /*
1560  * Remove PCB from connection hash table.
1561  */
1562 void
1563 in_pcbremconnhash(struct inpcb *inp)
1564 {
1565         KASSERT(inp->inp_flags & INP_CONNECTED, ("inp not connected"));
1566         LIST_REMOVE(inp, inp_hash);
1567         inp->inp_flags &= ~INP_CONNECTED;
1568 }
1569
1570 /*
1571  * Insert PCB into port hash table.
1572  */
1573 void
1574 in_pcbinsporthash(struct inpcbportinfo *portinfo, struct inpcb *inp)
1575 {
1576         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1577         struct inpcbporthead *pcbporthash;
1578         struct inpcbport *phd;
1579
1580         /*
1581          * If the porthashbase is shared across several cpus, it must
1582          * have been locked.
1583          */
1584         ASSERT_PORT_TOKEN_HELD(portinfo);
1585
1586         /*
1587          * Insert into the port hash table.
1588          */
1589         pcbporthash = &portinfo->porthashbase[
1590             INP_PCBPORTHASH(inp->inp_lport, portinfo->porthashmask)];
1591
1592         /* Go through port list and look for a head for this lport. */
1593         LIST_FOREACH(phd, pcbporthash, phd_hash) {
1594                 if (phd->phd_port == inp->inp_lport)
1595                         break;
1596         }
1597
1598         /* If none exists, use saved one and tack it on. */
1599         if (phd == NULL) {
1600                 KKASSERT(pcbinfo->portsave != NULL);
1601                 phd = pcbinfo->portsave;
1602                 pcbinfo->portsave = NULL;
1603                 phd->phd_port = inp->inp_lport;
1604                 LIST_INIT(&phd->phd_pcblist);
1605                 LIST_INSERT_HEAD(pcbporthash, phd, phd_hash);
1606         }
1607
1608         inp->inp_portinfo = portinfo;
1609         inp->inp_phd = phd;
1610         LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist);
1611
1612         /*
1613          * Malloc one inpcbport for later use.  It is safe to use
1614          * "wait" malloc here (port token would be released, if
1615          * malloc ever blocked), since all changes to the porthash
1616          * are done.
1617          */
1618         if (pcbinfo->portsave == NULL) {
1619                 pcbinfo->portsave = kmalloc(sizeof(*pcbinfo->portsave),
1620                                             M_PCB, M_INTWAIT | M_ZERO);
1621         }
1622 }
1623
1624 void
1625 in_pcbinsporthash_lport(struct inpcb *inp)
1626 {
1627         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1628         struct inpcbportinfo *portinfo;
1629         u_short lport_ho;
1630
1631         /* Locate the proper portinfo based on lport */
1632         lport_ho = ntohs(inp->inp_lport);
1633         portinfo = &pcbinfo->portinfo[lport_ho & pcbinfo->portinfo_mask];
1634         KKASSERT((lport_ho & pcbinfo->portinfo_mask) == portinfo->offset);
1635
1636         GET_PORT_TOKEN(portinfo);
1637         in_pcbinsporthash(portinfo, inp);
1638         REL_PORT_TOKEN(portinfo);
1639 }
1640
1641 static struct inp_localgroup *
1642 inp_localgroup_alloc(struct inp_localgrphead *hdr, u_char vflag,
1643     uint16_t port, const union in_dependaddr *addr, int size)
1644 {
1645         struct inp_localgroup *grp;
1646
1647         grp = kmalloc(__offsetof(struct inp_localgroup, il_inp[size]),
1648             M_TEMP, M_INTWAIT | M_ZERO);
1649         grp->il_vflag = vflag;
1650         grp->il_lport = port;
1651         grp->il_dependladdr = *addr;
1652         grp->il_inpsiz = size;
1653
1654         LIST_INSERT_HEAD(hdr, grp, il_list);
1655
1656         return grp;
1657 }
1658
1659 static void
1660 inp_localgroup_free(struct inp_localgroup *grp)
1661 {
1662         LIST_REMOVE(grp, il_list);
1663         kfree(grp, M_TEMP);
1664 }
1665
1666 static struct inp_localgroup *
1667 inp_localgroup_resize(struct inp_localgrphead *hdr,
1668     struct inp_localgroup *old_grp, int size)
1669 {
1670         struct inp_localgroup *grp;
1671         int i;
1672
1673         grp = inp_localgroup_alloc(hdr, old_grp->il_vflag,
1674             old_grp->il_lport, &old_grp->il_dependladdr, size);
1675
1676         KASSERT(old_grp->il_inpcnt < grp->il_inpsiz,
1677             ("invalid new local group size %d and old local group count %d",
1678              grp->il_inpsiz, old_grp->il_inpcnt));
1679         for (i = 0; i < old_grp->il_inpcnt; ++i)
1680                 grp->il_inp[i] = old_grp->il_inp[i];
1681         grp->il_inpcnt = old_grp->il_inpcnt;
1682         grp->il_factor = old_grp->il_factor;
1683
1684         inp_localgroup_free(old_grp);
1685
1686         return grp;
1687 }
1688
1689 static void
1690 inp_localgroup_factor(struct inp_localgroup *grp)
1691 {
1692         grp->il_factor =
1693             ((uint32_t)(0xffff >> ncpus2_shift) / grp->il_inpcnt) + 1;
1694         KASSERT(grp->il_factor != 0, ("invalid local group factor, "
1695             "ncpus2_shift %d, inpcnt %d", ncpus2_shift, grp->il_inpcnt));
1696 }
1697
1698 static void
1699 in_pcbinslocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1700 {
1701         struct inp_localgrphead *hdr;
1702         struct inp_localgroup *grp;
1703         struct ucred *cred;
1704
1705         if (pcbinfo->localgrphashbase == NULL)
1706                 return;
1707
1708         /*
1709          * XXX don't allow jailed socket to join local group
1710          */
1711         if (inp->inp_socket != NULL)
1712                 cred = inp->inp_socket->so_cred;
1713         else
1714                 cred = NULL;
1715         if (cred != NULL && jailed(cred))
1716                 return;
1717
1718 #ifdef INET6
1719         /*
1720          * XXX don't allow IPv4 mapped INET6 wild socket
1721          */
1722         if ((inp->inp_vflag & INP_IPV4) &&
1723             inp->inp_laddr.s_addr == INADDR_ANY &&
1724             INP_CHECK_SOCKAF(inp->inp_socket, AF_INET6))
1725                 return;
1726 #endif
1727
1728         hdr = &pcbinfo->localgrphashbase[
1729             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1730
1731         LIST_FOREACH(grp, hdr, il_list) {
1732                 if (grp->il_vflag == inp->inp_vflag &&
1733                     grp->il_lport == inp->inp_lport &&
1734                     memcmp(&grp->il_dependladdr,
1735                         &inp->inp_inc.inc_ie.ie_dependladdr,
1736                         sizeof(grp->il_dependladdr)) == 0) {
1737                         break;
1738                 }
1739         }
1740         if (grp == NULL) {
1741                 /* Create new local group */
1742                 grp = inp_localgroup_alloc(hdr, inp->inp_vflag,
1743                     inp->inp_lport, &inp->inp_inc.inc_ie.ie_dependladdr,
1744                     INP_LOCALGROUP_SIZMIN);
1745         } else if (grp->il_inpcnt == grp->il_inpsiz) {
1746                 if (grp->il_inpsiz >= INP_LOCALGROUP_SIZMAX) {
1747                         static int limit_logged = 0;
1748
1749                         if (!limit_logged) {
1750                                 limit_logged = 1;
1751                                 kprintf("local group port %d, "
1752                                     "limit reached\n", ntohs(grp->il_lport));
1753                         }
1754                         return;
1755                 }
1756
1757                 /* Expand this local group */
1758                 grp = inp_localgroup_resize(hdr, grp, grp->il_inpsiz * 2);
1759         }
1760
1761         KASSERT(grp->il_inpcnt < grp->il_inpsiz,
1762             ("invalid local group size %d and count %d",
1763              grp->il_inpsiz, grp->il_inpcnt));
1764         grp->il_inp[grp->il_inpcnt] = inp;
1765         grp->il_inpcnt++;
1766         inp_localgroup_factor(grp);
1767 }
1768
1769 void
1770 in_pcbinswildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1771 {
1772         struct inpcontainer *ic;
1773         struct inpcontainerhead *bucket;
1774
1775         in_pcbinslocalgrphash_oncpu(inp, pcbinfo);
1776
1777         bucket = &pcbinfo->wildcardhashbase[
1778             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1779
1780         ic = kmalloc(sizeof(struct inpcontainer), M_TEMP, M_INTWAIT);
1781         ic->ic_inp = inp;
1782         LIST_INSERT_HEAD(bucket, ic, ic_list);
1783 }
1784
1785 /*
1786  * Insert PCB into wildcard hash table.
1787  */
1788 void
1789 in_pcbinswildcardhash(struct inpcb *inp)
1790 {
1791         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1792
1793         KASSERT(!(inp->inp_flags & INP_CONNECTED),
1794                 ("already on connhash"));
1795         KASSERT(!(inp->inp_flags & INP_WILDCARD),
1796                 ("already on wildcardhash"));
1797         inp->inp_flags |= INP_WILDCARD;
1798
1799         in_pcbinswildcardhash_oncpu(inp, pcbinfo);
1800 }
1801
1802 static void
1803 in_pcbremlocalgrphash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1804 {
1805         struct inp_localgrphead *hdr;
1806         struct inp_localgroup *grp;
1807
1808         if (pcbinfo->localgrphashbase == NULL)
1809                 return;
1810
1811         hdr = &pcbinfo->localgrphashbase[
1812             INP_PCBLOCALGRPHASH(inp->inp_lport, pcbinfo->localgrphashmask)];
1813
1814         LIST_FOREACH(grp, hdr, il_list) {
1815                 int i;
1816
1817                 for (i = 0; i < grp->il_inpcnt; ++i) {
1818                         if (grp->il_inp[i] != inp)
1819                                 continue;
1820
1821                         if (grp->il_inpcnt == 1) {
1822                                 /* Free this local group */
1823                                 inp_localgroup_free(grp);
1824                         } else {
1825                                 /* Pull up inpcbs */
1826                                 for (; i + 1 < grp->il_inpcnt; ++i)
1827                                         grp->il_inp[i] = grp->il_inp[i + 1];
1828                                 grp->il_inpcnt--;
1829                                 inp_localgroup_factor(grp);
1830
1831                                 if (grp->il_inpsiz > INP_LOCALGROUP_SIZMIN &&
1832                                     grp->il_inpcnt <= (grp->il_inpsiz / 4)) {
1833                                         /* Shrink this local group */
1834                                         grp = inp_localgroup_resize(hdr, grp,
1835                                             grp->il_inpsiz / 2);
1836                                 }
1837                         }
1838                         return;
1839                 }
1840         }
1841 }
1842
1843 void
1844 in_pcbremwildcardhash_oncpu(struct inpcb *inp, struct inpcbinfo *pcbinfo)
1845 {
1846         struct inpcontainer *ic;
1847         struct inpcontainerhead *head;
1848
1849         in_pcbremlocalgrphash_oncpu(inp, pcbinfo);
1850
1851         /* find bucket */
1852         head = &pcbinfo->wildcardhashbase[
1853             INP_PCBWILDCARDHASH(inp->inp_lport, pcbinfo->wildcardhashmask)];
1854
1855         LIST_FOREACH(ic, head, ic_list) {
1856                 if (ic->ic_inp == inp)
1857                         goto found;
1858         }
1859         return;                 /* not found! */
1860
1861 found:
1862         LIST_REMOVE(ic, ic_list);       /* remove container from bucket chain */
1863         kfree(ic, M_TEMP);              /* deallocate container */
1864 }
1865
1866 /*
1867  * Remove PCB from wildcard hash table.
1868  */
1869 void
1870 in_pcbremwildcardhash(struct inpcb *inp)
1871 {
1872         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
1873
1874         KASSERT(inp->inp_flags & INP_WILDCARD, ("inp not wildcard"));
1875         in_pcbremwildcardhash_oncpu(inp, pcbinfo);
1876         inp->inp_flags &= ~INP_WILDCARD;
1877 }
1878
1879 /*
1880  * Remove PCB from various lists.
1881  */
1882 void
1883 in_pcbremlists(struct inpcb *inp)
1884 {
1885         if (inp->inp_lport) {
1886                 struct inpcbportinfo *portinfo;
1887                 struct inpcbport *phd;
1888
1889                 /*
1890                  * NOTE:
1891                  * inp->inp_portinfo is _not_ necessary same as
1892                  * inp->inp_pcbinfo->portinfo.
1893                  */
1894                 portinfo = inp->inp_portinfo;
1895                 GET_PORT_TOKEN(portinfo);
1896
1897                 phd = inp->inp_phd;
1898                 LIST_REMOVE(inp, inp_portlist);
1899                 if (LIST_FIRST(&phd->phd_pcblist) == NULL) {
1900                         LIST_REMOVE(phd, phd_hash);
1901                         kfree(phd, M_PCB);
1902                 }
1903
1904                 REL_PORT_TOKEN(portinfo);
1905         }
1906         if (inp->inp_flags & INP_WILDCARD) {
1907                 in_pcbremwildcardhash(inp);
1908         } else if (inp->inp_flags & INP_CONNECTED) {
1909                 in_pcbremconnhash(inp);
1910         }
1911         LIST_REMOVE(inp, inp_list);
1912         inp->inp_pcbinfo->ipi_count--;
1913 }
1914
1915 int
1916 prison_xinpcb(struct thread *td, struct inpcb *inp)
1917 {
1918         struct ucred *cr;
1919
1920         if (td->td_proc == NULL)
1921                 return (0);
1922         cr = td->td_proc->p_ucred;
1923         if (cr->cr_prison == NULL)
1924                 return (0);
1925         if (inp->inp_socket && inp->inp_socket->so_cred &&
1926             inp->inp_socket->so_cred->cr_prison &&
1927             cr->cr_prison == inp->inp_socket->so_cred->cr_prison)
1928                 return (0);
1929         return (1);
1930 }
1931
1932 int
1933 in_pcblist_global(SYSCTL_HANDLER_ARGS)
1934 {
1935         struct inpcbinfo *pcbinfo = arg1;
1936         struct inpcb *inp, *marker;
1937         struct xinpcb xi;
1938         int error, i, n;
1939
1940         /*
1941          * The process of preparing the TCB list is too time-consuming and
1942          * resource-intensive to repeat twice on every request.
1943          */
1944         if (req->oldptr == NULL) {
1945                 n = pcbinfo->ipi_count;
1946                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
1947                 return 0;
1948         }
1949
1950         if (req->newptr != NULL)
1951                 return EPERM;
1952
1953         /*
1954          * OK, now we're committed to doing something.  Re-fetch ipi_count
1955          * after obtaining the generation count.
1956          */
1957         n = pcbinfo->ipi_count;
1958
1959         marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1960         marker->inp_flags |= INP_PLACEMARKER;
1961         LIST_INSERT_HEAD(&pcbinfo->pcblisthead, marker, inp_list);
1962
1963         i = 0;
1964         error = 0;
1965
1966         while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1967                 LIST_REMOVE(marker, inp_list);
1968                 LIST_INSERT_AFTER(inp, marker, inp_list);
1969
1970                 if (inp->inp_flags & INP_PLACEMARKER)
1971                         continue;
1972                 if (prison_xinpcb(req->td, inp))
1973                         continue;
1974                 bzero(&xi, sizeof xi);
1975                 xi.xi_len = sizeof xi;
1976                 bcopy(inp, &xi.xi_inp, sizeof *inp);
1977                 if (inp->inp_socket)
1978                         sotoxsocket(inp->inp_socket, &xi.xi_socket);
1979                 if ((error = SYSCTL_OUT(req, &xi, sizeof xi)) != 0)
1980                         break;
1981                 ++i;
1982         }
1983         LIST_REMOVE(marker, inp_list);
1984         if (error == 0 && i < n) {
1985                 bzero(&xi, sizeof xi);
1986                 xi.xi_len = sizeof xi;
1987                 while (i < n) {
1988                         error = SYSCTL_OUT(req, &xi, sizeof xi);
1989                         ++i;
1990                 }
1991         }
1992         kfree(marker, M_TEMP);
1993         return(error);
1994 }
1995
1996 int
1997 in_pcblist_global_cpu0(SYSCTL_HANDLER_ARGS)
1998 {
1999         boolean_t migrate = FALSE;
2000         int origcpu = mycpuid;
2001         int error;
2002
2003         if (origcpu != 0) {
2004                 migrate = TRUE;
2005                 lwkt_migratecpu(0);
2006         }
2007
2008         error = in_pcblist_global(oidp, arg1, arg2, req);
2009
2010         if (migrate)
2011                 lwkt_migratecpu(origcpu);
2012         return error;
2013 }
2014
2015 int
2016 in_pcblist_global_nomarker(SYSCTL_HANDLER_ARGS, struct xinpcb **xi0, int *nxi0)
2017 {
2018         struct inpcbinfo *pcbinfo = arg1;
2019         struct inpcb *inp;
2020         struct xinpcb *xi;
2021         int nxi;
2022
2023         *nxi0 = 0;
2024         *xi0 = NULL;
2025
2026         /*
2027          * The process of preparing the PCB list is too time-consuming and
2028          * resource-intensive to repeat twice on every request.
2029          */
2030         if (req->oldptr == NULL) {
2031                 int n = pcbinfo->ipi_count;
2032
2033                 req->oldidx = (n + n/8 + 10) * sizeof(struct xinpcb);
2034                 return 0;
2035         }
2036
2037         if (req->newptr != NULL)
2038                 return EPERM;
2039
2040         if (pcbinfo->ipi_count == 0)
2041                 return 0;
2042
2043         nxi = 0;
2044         xi = kmalloc(pcbinfo->ipi_count * sizeof(*xi), M_TEMP,
2045                      M_WAITOK | M_ZERO | M_NULLOK);
2046         if (xi == NULL)
2047                 return ENOMEM;
2048
2049         LIST_FOREACH(inp, &pcbinfo->pcblisthead, inp_list) {
2050                 struct xinpcb *xi_ptr = &xi[nxi];
2051
2052                 if (prison_xinpcb(req->td, inp))
2053                         continue;
2054
2055                 xi_ptr->xi_len = sizeof(*xi_ptr);
2056                 bcopy(inp, &xi_ptr->xi_inp, sizeof(*inp));
2057                 if (inp->inp_socket)
2058                         sotoxsocket(inp->inp_socket, &xi_ptr->xi_socket);
2059                 ++nxi;
2060         }
2061
2062         if (nxi == 0) {
2063                 kfree(xi, M_TEMP);
2064                 return 0;
2065         }
2066
2067         *nxi0 = nxi;
2068         *xi0 = xi;
2069
2070         return 0;
2071 }
2072
2073 void
2074 in_savefaddr(struct socket *so, const struct sockaddr *faddr)
2075 {
2076         struct sockaddr_in *sin;
2077
2078         KASSERT(faddr->sa_family == AF_INET,
2079             ("not AF_INET faddr %d", faddr->sa_family));
2080
2081         sin = kmalloc(sizeof(*sin), M_SONAME, M_WAITOK | M_ZERO);
2082         sin->sin_family = AF_INET;
2083         sin->sin_len = sizeof(*sin);
2084         sin->sin_port = ((const struct sockaddr_in *)faddr)->sin_port;
2085         sin->sin_addr = ((const struct sockaddr_in *)faddr)->sin_addr;
2086
2087         so->so_faddr = (struct sockaddr *)sin;
2088 }
2089
2090 void
2091 in_pcbportinfo_init(struct inpcbportinfo *portinfo, int hashsize,
2092     boolean_t shared, u_short offset)
2093 {
2094         memset(portinfo, 0, sizeof(*portinfo));
2095
2096         portinfo->offset = offset;
2097         portinfo->lastport = offset;
2098         portinfo->lastlow = offset;
2099         portinfo->lasthi = offset;
2100
2101         portinfo->porthashbase = hashinit(hashsize, M_PCB,
2102             &portinfo->porthashmask);
2103
2104         if (shared) {
2105                 portinfo->porttoken = kmalloc(sizeof(struct lwkt_token),
2106                     M_PCB, M_WAITOK);
2107                 lwkt_token_init(portinfo->porttoken, "porttoken");
2108         }
2109 }
2110
2111 void
2112 in_pcbportrange(u_short *hi0, u_short *lo0, u_short ofs, u_short step)
2113 {
2114         int hi, lo;
2115
2116         if (step == 1)
2117                 return;
2118
2119         hi = *hi0;
2120         lo = *lo0;
2121
2122         hi = rounddown2(hi, step);
2123         hi += ofs;
2124         if (hi > (int)*hi0)
2125                 hi -= step;
2126
2127         lo = roundup2(lo, step);
2128         lo -= (step - ofs);
2129         if (lo < (int)*lo0)
2130                 lo += step;
2131
2132         *hi0 = hi;
2133         *lo0 = lo;
2134 }