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