contrib/unbound/services/listen_dnsport.c

   1 /*
   2  * services/listen_dnsport.c - listen on port 53 for incoming DNS queries.
   3  *
   4  * Copyright (c) 2007, NLnet Labs. All rights reserved.
   5  *
   6  * This software is open source.
   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  *
  12  * Redistributions of source code must retain the above copyright notice,
  13  * this list of conditions and the following disclaimer.
  14  *
  15  * Redistributions in binary form must reproduce the above copyright notice,
  16  * this list of conditions and the following disclaimer in the documentation
  17  * and/or other materials provided with the distribution.
  18  *
  19  * Neither the name of the NLNET LABS nor the names of its contributors may
  20  * be used to endorse or promote products derived from this software without
  21  * specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  26  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  27  * HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  28  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
  29  * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  30  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  31  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  32  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  33  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  34  */
  35
  36 /**
  37  * \file
  38  *
  39  * This file has functions to get queries from clients.
  40  */
  41 #include "config.h"
  42 #ifdef HAVE_SYS_TYPES_H
  43 #  include <sys/types.h>
  44 #endif
  45 #include <sys/time.h>
  46 #include "services/listen_dnsport.h"
  47 #include "services/outside_network.h"
  48 #include "util/netevent.h"
  49 #include "util/log.h"
  50 #include "util/config_file.h"
  51 #include "util/net_help.h"
  52 #include "ldns/sbuffer.h"
  53
  54 #ifdef HAVE_NETDB_H
  55 #include <netdb.h>
  56 #endif
  57 #include <fcntl.h>
  58
  59 #ifdef HAVE_SYS_UN_H
  60 #include <sys/un.h>
  61 #endif
  62
  63 /** number of queued TCP connections for listen() */
  64 #define TCP_BACKLOG 256
  65
  66 /**
  67  * Debug print of the getaddrinfo returned address.
  68  * @param addr: the address returned.
  69  */
  70 static void
  71 verbose_print_addr(struct addrinfo *addr)
  72 {
  73         if(verbosity >= VERB_ALGO) {
  74                 char buf[100];
  75                 void* sinaddr = &((struct sockaddr_in*)addr->ai_addr)->sin_addr;
  76 #ifdef INET6
  77                 if(addr->ai_family == AF_INET6)
  78                         sinaddr = &((struct sockaddr_in6*)addr->ai_addr)->
  79                                 sin6_addr;
  80 #endif /* INET6 */
  81                 if(inet_ntop(addr->ai_family, sinaddr, buf,
  82                         (socklen_t)sizeof(buf)) == 0) {
  83                         (void)strlcpy(buf, "(null)", sizeof(buf));
  84                 }
  85                 buf[sizeof(buf)-1] = 0;
  86                 verbose(VERB_ALGO, "creating %s%s socket %s %d",
  87                         addr->ai_socktype==SOCK_DGRAM?"udp":
  88                         addr->ai_socktype==SOCK_STREAM?"tcp":"otherproto",
  89                         addr->ai_family==AF_INET?"4":
  90                         addr->ai_family==AF_INET6?"6":
  91                         "_otherfam", buf,
  92                         ntohs(((struct sockaddr_in*)addr->ai_addr)->sin_port));
  93         }
  94 }
  95
  96 int
  97 create_udp_sock(int family, int socktype, struct sockaddr* addr,
  98         socklen_t addrlen, int v6only, int* inuse, int* noproto,
  99         int rcv, int snd, int listen, int* reuseport)
 100 {
 101         int s;
 102 #if defined(SO_REUSEADDR) || defined(SO_REUSEPORT) || defined(IPV6_USE_MIN_MTU)
 103         int on=1;
 104 #endif
 105 #ifdef IPV6_MTU
 106         int mtu = IPV6_MIN_MTU;
 107 #endif
 108 #if !defined(SO_RCVBUFFORCE) && !defined(SO_RCVBUF)
 109         (void)rcv;
 110 #endif
 111 #if !defined(SO_SNDBUFFORCE) && !defined(SO_SNDBUF)
 112         (void)snd;
 113 #endif
 114 #ifndef IPV6_V6ONLY
 115         (void)v6only;
 116 #endif
 117         if((s = socket(family, socktype, 0)) == -1) {
 118                 *inuse = 0;
 119 #ifndef USE_WINSOCK
 120                 if(errno == EAFNOSUPPORT || errno == EPROTONOSUPPORT) {
 121                         *noproto = 1;
 122                         return -1;
 123                 }
 124                 log_err("can't create socket: %s", strerror(errno));
 125 #else
 126                 if(WSAGetLastError() == WSAEAFNOSUPPORT ||
 127                         WSAGetLastError() == WSAEPROTONOSUPPORT) {
 128                         *noproto = 1;
 129                         return -1;
 130                 }
 131                 log_err("can't create socket: %s",
 132                         wsa_strerror(WSAGetLastError()));
 133 #endif
 134                 *noproto = 0;
 135                 return -1;
 136         }
 137         if(listen) {
 138 #ifdef SO_REUSEADDR
 139                 if(setsockopt(s, SOL_SOCKET, SO_REUSEADDR, (void*)&on,
 140                         (socklen_t)sizeof(on)) < 0) {
 141 #ifndef USE_WINSOCK
 142                         log_err("setsockopt(.. SO_REUSEADDR ..) failed: %s",
 143                                 strerror(errno));
 144                         if(errno != ENOSYS) {
 145                                 close(s);
 146                                 *noproto = 0;
 147                                 *inuse = 0;
 148                                 return -1;
 149                         }
 150 #else
 151                         log_err("setsockopt(.. SO_REUSEADDR ..) failed: %s",
 152                                 wsa_strerror(WSAGetLastError()));
 153                         closesocket(s);
 154                         *noproto = 0;
 155                         *inuse = 0;
 156                         return -1;
 157 #endif
 158                 }
 159 #endif /* SO_REUSEADDR */
 160 #ifdef SO_REUSEPORT
 161                 /* try to set SO_REUSEPORT so that incoming
 162                  * queries are distributed evenly among the receiving threads.
 163                  * Each thread must have its own socket bound to the same port,
 164                  * with SO_REUSEPORT set on each socket.
 165                  */
 166                 if (reuseport && *reuseport &&
 167                     setsockopt(s, SOL_SOCKET, SO_REUSEPORT, (void*)&on,
 168                         (socklen_t)sizeof(on)) < 0) {
 169 #ifdef ENOPROTOOPT
 170                         if(errno != ENOPROTOOPT || verbosity >= 3)
 171                                 log_warn("setsockopt(.. SO_REUSEPORT ..) failed: %s",
 172                                         strerror(errno));
 173 #endif
 174                         /* this option is not essential, we can continue */
 175                         *reuseport = 0;
 176                 }
 177 #else
 178                 (void)reuseport;
 179 #endif /* defined(SO_REUSEPORT) */
 180         }
 181         if(rcv) {
 182 #ifdef SO_RCVBUF
 183                 int got;
 184                 socklen_t slen = (socklen_t)sizeof(got);
 185 #  ifdef SO_RCVBUFFORCE
 186                 /* Linux specific: try to use root permission to override
 187                  * system limits on rcvbuf. The limit is stored in
 188                  * /proc/sys/net/core/rmem_max or sysctl net.core.rmem_max */
 189                 if(setsockopt(s, SOL_SOCKET, SO_RCVBUFFORCE, (void*)&rcv,
 190                         (socklen_t)sizeof(rcv)) < 0) {
 191                         if(errno != EPERM) {
 192 #    ifndef USE_WINSOCK
 193                                 log_err("setsockopt(..., SO_RCVBUFFORCE, "
 194                                         "...) failed: %s", strerror(errno));
 195                                 close(s);
 196 #    else
 197                                 log_err("setsockopt(..., SO_RCVBUFFORCE, "
 198                                         "...) failed: %s",
 199                                         wsa_strerror(WSAGetLastError()));
 200                                 closesocket(s);
 201 #    endif
 202                                 *noproto = 0;
 203                                 *inuse = 0;
 204                                 return -1;
 205                         }
 206 #  endif /* SO_RCVBUFFORCE */
 207                         if(setsockopt(s, SOL_SOCKET, SO_RCVBUF, (void*)&rcv,
 208                                 (socklen_t)sizeof(rcv)) < 0) {
 209 #  ifndef USE_WINSOCK
 210                                 log_err("setsockopt(..., SO_RCVBUF, "
 211                                         "...) failed: %s", strerror(errno));
 212                                 close(s);
 213 #  else
 214                                 log_err("setsockopt(..., SO_RCVBUF, "
 215                                         "...) failed: %s",
 216                                         wsa_strerror(WSAGetLastError()));
 217                                 closesocket(s);
 218 #  endif
 219                                 *noproto = 0;
 220                                 *inuse = 0;
 221                                 return -1;
 222                         }
 223                         /* check if we got the right thing or if system
 224                          * reduced to some system max.  Warn if so */
 225                         if(getsockopt(s, SOL_SOCKET, SO_RCVBUF, (void*)&got,
 226                                 &slen) >= 0 && got < rcv/2) {
 227                                 log_warn("so-rcvbuf %u was not granted. "
 228                                         "Got %u. To fix: start with "
 229                                         "root permissions(linux) or sysctl "
 230                                         "bigger net.core.rmem_max(linux) or "
 231                                         "kern.ipc.maxsockbuf(bsd) values.",
 232                                         (unsigned)rcv, (unsigned)got);
 233                         }
 234 #  ifdef SO_RCVBUFFORCE
 235                 }
 236 #  endif
 237 #endif /* SO_RCVBUF */
 238         }
 239         /* first do RCVBUF as the receive buffer is more important */
 240         if(snd) {
 241 #ifdef SO_SNDBUF
 242                 int got;
 243                 socklen_t slen = (socklen_t)sizeof(got);
 244 #  ifdef SO_SNDBUFFORCE
 245                 /* Linux specific: try to use root permission to override
 246                  * system limits on sndbuf. The limit is stored in
 247                  * /proc/sys/net/core/wmem_max or sysctl net.core.wmem_max */
 248                 if(setsockopt(s, SOL_SOCKET, SO_SNDBUFFORCE, (void*)&snd,
 249                         (socklen_t)sizeof(snd)) < 0) {
 250                         if(errno != EPERM) {
 251 #    ifndef USE_WINSOCK
 252                                 log_err("setsockopt(..., SO_SNDBUFFORCE, "
 253                                         "...) failed: %s", strerror(errno));
 254                                 close(s);
 255 #    else
 256                                 log_err("setsockopt(..., SO_SNDBUFFORCE, "
 257                                         "...) failed: %s",
 258                                         wsa_strerror(WSAGetLastError()));
 259                                 closesocket(s);
 260 #    endif
 261                                 *noproto = 0;
 262                                 *inuse = 0;
 263                                 return -1;
 264                         }
 265 #  endif /* SO_SNDBUFFORCE */
 266                         if(setsockopt(s, SOL_SOCKET, SO_SNDBUF, (void*)&snd,
 267                                 (socklen_t)sizeof(snd)) < 0) {
 268 #  ifndef USE_WINSOCK
 269                                 log_err("setsockopt(..., SO_SNDBUF, "
 270                                         "...) failed: %s", strerror(errno));
 271                                 close(s);
 272 #  else
 273                                 log_err("setsockopt(..., SO_SNDBUF, "
 274                                         "...) failed: %s",
 275                                         wsa_strerror(WSAGetLastError()));
 276                                 closesocket(s);
 277 #  endif
 278                                 *noproto = 0;
 279                                 *inuse = 0;
 280                                 return -1;
 281                         }
 282                         /* check if we got the right thing or if system
 283                          * reduced to some system max.  Warn if so */
 284                         if(getsockopt(s, SOL_SOCKET, SO_SNDBUF, (void*)&got,
 285                                 &slen) >= 0 && got < snd/2) {
 286                                 log_warn("so-sndbuf %u was not granted. "
 287                                         "Got %u. To fix: start with "
 288                                         "root permissions(linux) or sysctl "
 289                                         "bigger net.core.wmem_max(linux) or "
 290                                         "kern.ipc.maxsockbuf(bsd) values.",
 291                                         (unsigned)snd, (unsigned)got);
 292                         }
 293 #  ifdef SO_SNDBUFFORCE
 294                 }
 295 #  endif
 296 #endif /* SO_SNDBUF */
 297         }
 298         if(family == AF_INET6) {
 299 # if defined(IPV6_V6ONLY)
 300                 if(v6only) {
 301                         int val=(v6only==2)?0:1;
 302                         if (setsockopt(s, IPPROTO_IPV6, IPV6_V6ONLY,
 303                                 (void*)&val, (socklen_t)sizeof(val)) < 0) {
 304 #ifndef USE_WINSOCK
 305                                 log_err("setsockopt(..., IPV6_V6ONLY"
 306                                         ", ...) failed: %s", strerror(errno));
 307                                 close(s);
 308 #else
 309                                 log_err("setsockopt(..., IPV6_V6ONLY"
 310                                         ", ...) failed: %s",
 311                                         wsa_strerror(WSAGetLastError()));
 312                                 closesocket(s);
 313 #endif
 314                                 *noproto = 0;
 315                                 *inuse = 0;
 316                                 return -1;
 317                         }
 318                 }
 319 # endif
 320 # if defined(IPV6_USE_MIN_MTU)
 321                 /*
 322                  * There is no fragmentation of IPv6 datagrams
 323                  * during forwarding in the network. Therefore
 324                  * we do not send UDP datagrams larger than
 325                  * the minimum IPv6 MTU of 1280 octets. The
 326                  * EDNS0 message length can be larger if the
 327                  * network stack supports IPV6_USE_MIN_MTU.
 328                  */
 329                 if (setsockopt(s, IPPROTO_IPV6, IPV6_USE_MIN_MTU,
 330                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 331 #  ifndef USE_WINSOCK
 332                         log_err("setsockopt(..., IPV6_USE_MIN_MTU, "
 333                                 "...) failed: %s", strerror(errno));
 334                         close(s);
 335 #  else
 336                         log_err("setsockopt(..., IPV6_USE_MIN_MTU, "
 337                                 "...) failed: %s",
 338                                 wsa_strerror(WSAGetLastError()));
 339                         closesocket(s);
 340 #  endif
 341                         *noproto = 0;
 342                         *inuse = 0;
 343                         return -1;
 344                 }
 345 # elif defined(IPV6_MTU)
 346                 /*
 347                  * On Linux, to send no larger than 1280, the PMTUD is
 348                  * disabled by default for datagrams anyway, so we set
 349                  * the MTU to use.
 350                  */
 351                 if (setsockopt(s, IPPROTO_IPV6, IPV6_MTU,
 352                         (void*)&mtu, (socklen_t)sizeof(mtu)) < 0) {
 353 #  ifndef USE_WINSOCK
 354                         log_err("setsockopt(..., IPV6_MTU, ...) failed: %s",
 355                                 strerror(errno));
 356                         close(s);
 357 #  else
 358                         log_err("setsockopt(..., IPV6_MTU, ...) failed: %s",
 359                                 wsa_strerror(WSAGetLastError()));
 360                         closesocket(s);
 361 #  endif
 362                         *noproto = 0;
 363                         *inuse = 0;
 364                         return -1;
 365                 }
 366 # endif /* IPv6 MTU */
 367         } else if(family == AF_INET) {
 368 #  if defined(IP_MTU_DISCOVER) && defined(IP_PMTUDISC_DONT)
 369 /* linux 3.15 has IP_PMTUDISC_OMIT, Hannes Frederic Sowa made it so that
 370  * PMTU information is not accepted, but fragmentation is allowed
 371  * if and only if the packet size exceeds the outgoing interface MTU
 372  * (and also uses the interface mtu to determine the size of the packets).
 373  * So there won't be any EMSGSIZE error.  Against DNS fragmentation attacks.
 374  * FreeBSD already has same semantics without setting the option. */
 375 #    if defined(IP_PMTUDISC_OMIT)
 376                 int action = IP_PMTUDISC_OMIT;
 377 #    else
 378                 int action = IP_PMTUDISC_DONT;
 379 #    endif
 380                 if (setsockopt(s, IPPROTO_IP, IP_MTU_DISCOVER,
 381                         &action, (socklen_t)sizeof(action)) < 0) {
 382                         log_err("setsockopt(..., IP_MTU_DISCOVER, "
 383 #    if defined(IP_PMTUDISC_OMIT)
 384                                 "IP_PMTUDISC_OMIT"
 385 #    else
 386                                 "IP_PMTUDISC_DONT"
 387 #    endif
 388                                 "...) failed: %s",
 389                                 strerror(errno));
 390 #    ifndef USE_WINSOCK
 391                         close(s);
 392 #    else
 393                         closesocket(s);
 394 #    endif
 395                         *noproto = 0;
 396                         *inuse = 0;
 397                         return -1;
 398                 }
 399 #  elif defined(IP_DONTFRAG)
 400                 int off = 0;
 401                 if (setsockopt(s, IPPROTO_IP, IP_DONTFRAG,
 402                         &off, (socklen_t)sizeof(off)) < 0) {
 403                         log_err("setsockopt(..., IP_DONTFRAG, ...) failed: %s",
 404                                 strerror(errno));
 405 #    ifndef USE_WINSOCK
 406                         close(s);
 407 #    else
 408                         closesocket(s);
 409 #    endif
 410                         *noproto = 0;
 411                         *inuse = 0;
 412                         return -1;
 413                 }
 414 #  endif /* IPv4 MTU */
 415         }
 416         if(bind(s, (struct sockaddr*)addr, addrlen) != 0) {
 417                 *noproto = 0;
 418                 *inuse = 0;
 419 #ifndef USE_WINSOCK
 420 #ifdef EADDRINUSE
 421                 *inuse = (errno == EADDRINUSE);
 422                 /* detect freebsd jail with no ipv6 permission */
 423                 if(family==AF_INET6 && errno==EINVAL)
 424                         *noproto = 1;
 425                 else if(errno != EADDRINUSE) {
 426                         log_err_addr("can't bind socket", strerror(errno),
 427                                 (struct sockaddr_storage*)addr, addrlen);
 428                 }
 429 #endif /* EADDRINUSE */
 430                 close(s);
 431 #else /* USE_WINSOCK */
 432                 if(WSAGetLastError() != WSAEADDRINUSE &&
 433                         WSAGetLastError() != WSAEADDRNOTAVAIL) {
 434                         log_err_addr("can't bind socket",
 435                                 wsa_strerror(WSAGetLastError()),
 436                                 (struct sockaddr_storage*)addr, addrlen);
 437                 }
 438                 closesocket(s);
 439 #endif
 440                 return -1;
 441         }
 442         if(!fd_set_nonblock(s)) {
 443                 *noproto = 0;
 444                 *inuse = 0;
 445 #ifndef USE_WINSOCK
 446                 close(s);
 447 #else
 448                 closesocket(s);
 449 #endif
 450                 return -1;
 451         }
 452         return s;
 453 }
 454
 455 int
 456 create_tcp_accept_sock(struct addrinfo *addr, int v6only, int* noproto,
 457         int* reuseport)
 458 {
 459         int s;
 460 #if defined(SO_REUSEADDR) || defined(SO_REUSEPORT) || defined(IPV6_V6ONLY)
 461         int on = 1;
 462 #endif /* SO_REUSEADDR || IPV6_V6ONLY */
 463         verbose_print_addr(addr);
 464         *noproto = 0;
 465         if((s = socket(addr->ai_family, addr->ai_socktype, 0)) == -1) {
 466 #ifndef USE_WINSOCK
 467                 if(errno == EAFNOSUPPORT || errno == EPROTONOSUPPORT) {
 468                         *noproto = 1;
 469                         return -1;
 470                 }
 471                 log_err("can't create socket: %s", strerror(errno));
 472 #else
 473                 if(WSAGetLastError() == WSAEAFNOSUPPORT ||
 474                         WSAGetLastError() == WSAEPROTONOSUPPORT) {
 475                         *noproto = 1;
 476                         return -1;
 477                 }
 478                 log_err("can't create socket: %s",
 479                         wsa_strerror(WSAGetLastError()));
 480 #endif
 481                 return -1;
 482         }
 483 #ifdef SO_REUSEADDR
 484         if(setsockopt(s, SOL_SOCKET, SO_REUSEADDR, (void*)&on,
 485                 (socklen_t)sizeof(on)) < 0) {
 486 #ifndef USE_WINSOCK
 487                 log_err("setsockopt(.. SO_REUSEADDR ..) failed: %s",
 488                         strerror(errno));
 489                 close(s);
 490 #else
 491                 log_err("setsockopt(.. SO_REUSEADDR ..) failed: %s",
 492                         wsa_strerror(WSAGetLastError()));
 493                 closesocket(s);
 494 #endif
 495                 return -1;
 496         }
 497 #endif /* SO_REUSEADDR */
 498 #ifdef SO_REUSEPORT
 499         /* try to set SO_REUSEPORT so that incoming
 500          * connections are distributed evenly among the receiving threads.
 501          * Each thread must have its own socket bound to the same port,
 502          * with SO_REUSEPORT set on each socket.
 503          */
 504         if (reuseport && *reuseport &&
 505                 setsockopt(s, SOL_SOCKET, SO_REUSEPORT, (void*)&on,
 506                 (socklen_t)sizeof(on)) < 0) {
 507 #ifdef ENOPROTOOPT
 508                 if(errno != ENOPROTOOPT || verbosity >= 3)
 509                         log_warn("setsockopt(.. SO_REUSEPORT ..) failed: %s",
 510                                 strerror(errno));
 511 #endif
 512                 /* this option is not essential, we can continue */
 513                 *reuseport = 0;
 514         }
 515 #else
 516         (void)reuseport;
 517 #endif /* defined(SO_REUSEPORT) */
 518 #if defined(IPV6_V6ONLY)
 519         if(addr->ai_family == AF_INET6 && v6only) {
 520                 if(setsockopt(s, IPPROTO_IPV6, IPV6_V6ONLY,
 521                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 522 #ifndef USE_WINSOCK
 523                         log_err("setsockopt(..., IPV6_V6ONLY, ...) failed: %s",
 524                                 strerror(errno));
 525                         close(s);
 526 #else
 527                         log_err("setsockopt(..., IPV6_V6ONLY, ...) failed: %s",
 528                                 wsa_strerror(WSAGetLastError()));
 529                         closesocket(s);
 530 #endif
 531                         return -1;
 532                 }
 533         }
 534 #else
 535         (void)v6only;
 536 #endif /* IPV6_V6ONLY */
 537         if(bind(s, addr->ai_addr, addr->ai_addrlen) != 0) {
 538 #ifndef USE_WINSOCK
 539                 /* detect freebsd jail with no ipv6 permission */
 540                 if(addr->ai_family==AF_INET6 && errno==EINVAL)
 541                         *noproto = 1;
 542                 else {
 543                         log_err_addr("can't bind socket", strerror(errno),
 544                                 (struct sockaddr_storage*)addr->ai_addr,
 545                                 addr->ai_addrlen);
 546                 }
 547                 close(s);
 548 #else
 549                 log_err_addr("can't bind socket",
 550                         wsa_strerror(WSAGetLastError()),
 551                         (struct sockaddr_storage*)addr->ai_addr,
 552                         addr->ai_addrlen);
 553                 closesocket(s);
 554 #endif
 555                 return -1;
 556         }
 557         if(!fd_set_nonblock(s)) {
 558 #ifndef USE_WINSOCK
 559                 close(s);
 560 #else
 561                 closesocket(s);
 562 #endif
 563                 return -1;
 564         }
 565         if(listen(s, TCP_BACKLOG) == -1) {
 566 #ifndef USE_WINSOCK
 567                 log_err("can't listen: %s", strerror(errno));
 568                 close(s);
 569 #else
 570                 log_err("can't listen: %s", wsa_strerror(WSAGetLastError()));
 571                 closesocket(s);
 572 #endif
 573                 return -1;
 574         }
 575         return s;
 576 }
 577
 578 int
 579 create_local_accept_sock(const char *path, int* noproto)
 580 {
 581 #ifdef HAVE_SYS_UN_H
 582         int s;
 583         struct sockaddr_un sun;
 584
 585 #ifdef HAVE_STRUCT_SOCKADDR_UN_SUN_LEN
 586         /* this member exists on BSDs, not Linux */
 587         sun.sun_len = (sa_family_t)sizeof(sun);
 588 #endif
 589         sun.sun_family = AF_LOCAL;
 590         /* length is 92-108, 104 on FreeBSD */
 591         (void)strlcpy(sun.sun_path, path, sizeof(sun.sun_path));
 592
 593         if ((s = socket(PF_LOCAL, SOCK_STREAM, 0)) == -1) {
 594                 log_err("Cannot create local socket %s (%s)",
 595                         path, strerror(errno));
 596                 return -1;
 597         }
 598
 599         if (unlink(path) && errno != ENOENT) {
 600                 /* The socket already exists and cannot be removed */
 601                 log_err("Cannot remove old local socket %s (%s)",
 602                         path, strerror(errno));
 603                 return -1;
 604         }
 605
 606         if (bind(s, (struct sockaddr *)&sun,
 607                 (socklen_t)sizeof(struct sockaddr_un)) == -1) {
 608                 log_err("Cannot bind local socket %s (%s)",
 609                         path, strerror(errno));
 610                 return -1;
 611         }
 612
 613         if (!fd_set_nonblock(s)) {
 614                 log_err("Cannot set non-blocking mode");
 615                 return -1;
 616         }
 617
 618         if (listen(s, TCP_BACKLOG) == -1) {
 619                 log_err("can't listen: %s", strerror(errno));
 620                 return -1;
 621         }
 622
 623         (void)noproto; /*unused*/
 624         return s;
 625 #else
 626         log_err("Local sockets are not supported");
 627         *noproto = 1;
 628         return -1;
 629 #endif
 630 }
 631
 632
 633 /**
 634  * Create socket from getaddrinfo results
 635  */
 636 static int
 637 make_sock(int stype, const char* ifname, const char* port,
 638         struct addrinfo *hints, int v6only, int* noip6, size_t rcv, size_t snd,
 639         int* reuseport)
 640 {
 641         struct addrinfo *res = NULL;
 642         int r, s, inuse, noproto;
 643         hints->ai_socktype = stype;
 644         *noip6 = 0;
 645         if((r=getaddrinfo(ifname, port, hints, &res)) != 0 || !res) {
 646 #ifdef USE_WINSOCK
 647                 if(r == EAI_NONAME && hints->ai_family == AF_INET6){
 648                         *noip6 = 1; /* 'Host not found' for IP6 on winXP */
 649                         return -1;
 650                 }
 651 #endif
 652                 log_err("node %s:%s getaddrinfo: %s %s",
 653                         ifname?ifname:"default", port, gai_strerror(r),
 654 #ifdef EAI_SYSTEM
 655                         r==EAI_SYSTEM?(char*)strerror(errno):""
 656 #else
 657                         ""
 658 #endif
 659                 );
 660                 return -1;
 661         }
 662         if(stype == SOCK_DGRAM) {
 663                 verbose_print_addr(res);
 664                 s = create_udp_sock(res->ai_family, res->ai_socktype,
 665                         (struct sockaddr*)res->ai_addr, res->ai_addrlen,
 666                         v6only, &inuse, &noproto, (int)rcv, (int)snd, 1,
 667                         reuseport);
 668                 if(s == -1 && inuse) {
 669                         log_err("bind: address already in use");
 670                 } else if(s == -1 && noproto && hints->ai_family == AF_INET6){
 671                         *noip6 = 1;
 672                 }
 673         } else  {
 674                 s = create_tcp_accept_sock(res, v6only, &noproto, reuseport);
 675                 if(s == -1 && noproto && hints->ai_family == AF_INET6){
 676                         *noip6 = 1;
 677                 }
 678         }
 679         freeaddrinfo(res);
 680         return s;
 681 }
 682
 683 /** make socket and first see if ifname contains port override info */
 684 static int
 685 make_sock_port(int stype, const char* ifname, const char* port,
 686         struct addrinfo *hints, int v6only, int* noip6, size_t rcv, size_t snd,
 687         int* reuseport)
 688 {
 689         char* s = strchr(ifname, '@');
 690         if(s) {
 691                 /* override port with ifspec@port */
 692                 char p[16];
 693                 char newif[128];
 694                 if((size_t)(s-ifname) >= sizeof(newif)) {
 695                         log_err("ifname too long: %s", ifname);
 696                         *noip6 = 0;
 697                         return -1;
 698                 }
 699                 if(strlen(s+1) >= sizeof(p)) {
 700                         log_err("portnumber too long: %s", ifname);
 701                         *noip6 = 0;
 702                         return -1;
 703                 }
 704                 (void)strlcpy(newif, ifname, sizeof(newif));
 705                 newif[s-ifname] = 0;
 706                 (void)strlcpy(p, s+1, sizeof(p));
 707                 p[strlen(s+1)]=0;
 708                 return make_sock(stype, newif, p, hints, v6only, noip6,
 709                         rcv, snd, reuseport);
 710         }
 711         return make_sock(stype, ifname, port, hints, v6only, noip6, rcv, snd,
 712                 reuseport);
 713 }
 714
 715 /**
 716  * Add port to open ports list.
 717  * @param list: list head. changed.
 718  * @param s: fd.
 719  * @param ftype: if fd is UDP.
 720  * @return false on failure. list in unchanged then.
 721  */
 722 static int
 723 port_insert(struct listen_port** list, int s, enum listen_type ftype)
 724 {
 725         struct listen_port* item = (struct listen_port*)malloc(
 726                 sizeof(struct listen_port));
 727         if(!item)
 728                 return 0;
 729         item->next = *list;
 730         item->fd = s;
 731         item->ftype = ftype;
 732         *list = item;
 733         return 1;
 734 }
 735
 736 /** set fd to receive source address packet info */
 737 static int
 738 set_recvpktinfo(int s, int family)
 739 {
 740 #if defined(IPV6_RECVPKTINFO) || defined(IPV6_PKTINFO) || (defined(IP_RECVDSTADDR) && defined(IP_SENDSRCADDR)) || defined(IP_PKTINFO)
 741         int on = 1;
 742 #else
 743         (void)s;
 744 #endif
 745         if(family == AF_INET6) {
 746 #           ifdef IPV6_RECVPKTINFO
 747                 if(setsockopt(s, IPPROTO_IPV6, IPV6_RECVPKTINFO,
 748                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 749                         log_err("setsockopt(..., IPV6_RECVPKTINFO, ...) failed: %s",
 750                                 strerror(errno));
 751                         return 0;
 752                 }
 753 #           elif defined(IPV6_PKTINFO)
 754                 if(setsockopt(s, IPPROTO_IPV6, IPV6_PKTINFO,
 755                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 756                         log_err("setsockopt(..., IPV6_PKTINFO, ...) failed: %s",
 757                                 strerror(errno));
 758                         return 0;
 759                 }
 760 #           else
 761                 log_err("no IPV6_RECVPKTINFO and no IPV6_PKTINFO option, please "
 762                         "disable interface-automatic in config");
 763                 return 0;
 764 #           endif /* defined IPV6_RECVPKTINFO */
 765
 766         } else if(family == AF_INET) {
 767 #           ifdef IP_PKTINFO
 768                 if(setsockopt(s, IPPROTO_IP, IP_PKTINFO,
 769                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 770                         log_err("setsockopt(..., IP_PKTINFO, ...) failed: %s",
 771                                 strerror(errno));
 772                         return 0;
 773                 }
 774 #           elif defined(IP_RECVDSTADDR) && defined(IP_SENDSRCADDR)
 775                 if(setsockopt(s, IPPROTO_IP, IP_RECVDSTADDR,
 776                         (void*)&on, (socklen_t)sizeof(on)) < 0) {
 777                         log_err("setsockopt(..., IP_RECVDSTADDR, ...) failed: %s",
 778                                 strerror(errno));
 779                         return 0;
 780                 }
 781 #           else
 782                 log_err("no IP_SENDSRCADDR or IP_PKTINFO option, please disable "
 783                         "interface-automatic in config");
 784                 return 0;
 785 #           endif /* IP_PKTINFO */
 786
 787         }
 788         return 1;
 789 }
 790
 791 /**
 792  * Helper for ports_open. Creates one interface (or NULL for default).
 793  * @param ifname: The interface ip address.
 794  * @param do_auto: use automatic interface detection.
 795  *      If enabled, then ifname must be the wildcard name.
 796  * @param do_udp: if udp should be used.
 797  * @param do_tcp: if udp should be used.
 798  * @param hints: for getaddrinfo. family and flags have to be set by caller.
 799  * @param port: Port number to use (as string).
 800  * @param list: list of open ports, appended to, changed to point to list head.
 801  * @param rcv: receive buffer size for UDP
 802  * @param snd: send buffer size for UDP
 803  * @param ssl_port: ssl service port number
 804  * @param reuseport: try to set SO_REUSEPORT if nonNULL and true.
 805  *      set to false on exit if reuseport failed due to no kernel support.
 806  * @return: returns false on error.
 807  */
 808 static int
 809 ports_create_if(const char* ifname, int do_auto, int do_udp, int do_tcp,
 810         struct addrinfo *hints, const char* port, struct listen_port** list,
 811         size_t rcv, size_t snd, int ssl_port, int* reuseport)
 812 {
 813         int s, noip6=0;
 814         if(!do_udp && !do_tcp)
 815                 return 0;
 816         if(do_auto) {
 817                 if((s = make_sock_port(SOCK_DGRAM, ifname, port, hints, 1,
 818                         &noip6, rcv, snd, reuseport)) == -1) {
 819                         if(noip6) {
 820                                 log_warn("IPv6 protocol not available");
 821                                 return 1;
 822                         }
 823                         return 0;
 824                 }
 825                 /* getting source addr packet info is highly non-portable */
 826                 if(!set_recvpktinfo(s, hints->ai_family)) {
 827 #ifndef USE_WINSOCK
 828                         close(s);
 829 #else
 830                         closesocket(s);
 831 #endif
 832                         return 0;
 833                 }
 834                 if(!port_insert(list, s, listen_type_udpancil)) {
 835 #ifndef USE_WINSOCK
 836                         close(s);
 837 #else
 838                         closesocket(s);
 839 #endif
 840                         return 0;
 841                 }
 842         } else if(do_udp) {
 843                 /* regular udp socket */
 844                 if((s = make_sock_port(SOCK_DGRAM, ifname, port, hints, 1,
 845                         &noip6, rcv, snd, reuseport)) == -1) {
 846                         if(noip6) {
 847                                 log_warn("IPv6 protocol not available");
 848                                 return 1;
 849                         }
 850                         return 0;
 851                 }
 852                 if(!port_insert(list, s, listen_type_udp)) {
 853 #ifndef USE_WINSOCK
 854                         close(s);
 855 #else
 856                         closesocket(s);
 857 #endif
 858                         return 0;
 859                 }
 860         }
 861         if(do_tcp) {
 862                 int is_ssl = ((strchr(ifname, '@') &&
 863                         atoi(strchr(ifname, '@')+1) == ssl_port) ||
 864                         (!strchr(ifname, '@') && atoi(port) == ssl_port));
 865                 if((s = make_sock_port(SOCK_STREAM, ifname, port, hints, 1,
 866                         &noip6, 0, 0, reuseport)) == -1) {
 867                         if(noip6) {
 868                                 /*log_warn("IPv6 protocol not available");*/
 869                                 return 1;
 870                         }
 871                         return 0;
 872                 }
 873                 if(is_ssl)
 874                         verbose(VERB_ALGO, "setup TCP for SSL service");
 875                 if(!port_insert(list, s, is_ssl?listen_type_ssl:
 876                         listen_type_tcp)) {
 877 #ifndef USE_WINSOCK
 878                         close(s);
 879 #else
 880                         closesocket(s);
 881 #endif
 882                         return 0;
 883                 }
 884         }
 885         return 1;
 886 }
 887
 888 /**
 889  * Add items to commpoint list in front.
 890  * @param c: commpoint to add.
 891  * @param front: listen struct.
 892  * @return: false on failure.
 893  */
 894 static int
 895 listen_cp_insert(struct comm_point* c, struct listen_dnsport* front)
 896 {
 897         struct listen_list* item = (struct listen_list*)malloc(
 898                 sizeof(struct listen_list));
 899         if(!item)
 900                 return 0;
 901         item->com = c;
 902         item->next = front->cps;
 903         front->cps = item;
 904         return 1;
 905 }
 906
 907 struct listen_dnsport*
 908 listen_create(struct comm_base* base, struct listen_port* ports,
 909         size_t bufsize, int tcp_accept_count, void* sslctx,
 910         struct dt_env* dtenv, comm_point_callback_t* cb, void *cb_arg)
 911 {
 912         struct listen_dnsport* front = (struct listen_dnsport*)
 913                 malloc(sizeof(struct listen_dnsport));
 914         if(!front)
 915                 return NULL;
 916         front->cps = NULL;
 917         front->udp_buff = sldns_buffer_new(bufsize);
 918         if(!front->udp_buff) {
 919                 free(front);
 920                 return NULL;
 921         }
 922
 923         /* create comm points as needed */
 924         while(ports) {
 925                 struct comm_point* cp = NULL;
 926                 if(ports->ftype == listen_type_udp)
 927                         cp = comm_point_create_udp(base, ports->fd,
 928                                 front->udp_buff, cb, cb_arg);
 929                 else if(ports->ftype == listen_type_tcp)
 930                         cp = comm_point_create_tcp(base, ports->fd,
 931                                 tcp_accept_count, bufsize, cb, cb_arg);
 932                 else if(ports->ftype == listen_type_ssl) {
 933                         cp = comm_point_create_tcp(base, ports->fd,
 934                                 tcp_accept_count, bufsize, cb, cb_arg);
 935                         cp->ssl = sslctx;
 936                 } else if(ports->ftype == listen_type_udpancil)
 937                         cp = comm_point_create_udp_ancil(base, ports->fd,
 938                                 front->udp_buff, cb, cb_arg);
 939                 if(!cp) {
 940                         log_err("can't create commpoint");
 941                         listen_delete(front);
 942                         return NULL;
 943                 }
 944                 cp->dtenv = dtenv;
 945                 cp->do_not_close = 1;
 946                 if(!listen_cp_insert(cp, front)) {
 947                         log_err("malloc failed");
 948                         comm_point_delete(cp);
 949                         listen_delete(front);
 950                         return NULL;
 951                 }
 952                 ports = ports->next;
 953         }
 954         if(!front->cps) {
 955                 log_err("Could not open sockets to accept queries.");
 956                 listen_delete(front);
 957                 return NULL;
 958         }
 959
 960         return front;
 961 }
 962
 963 void
 964 listen_list_delete(struct listen_list* list)
 965 {
 966         struct listen_list *p = list, *pn;
 967         while(p) {
 968                 pn = p->next;
 969                 comm_point_delete(p->com);
 970                 free(p);
 971                 p = pn;
 972         }
 973 }
 974
 975 void
 976 listen_delete(struct listen_dnsport* front)
 977 {
 978         if(!front)
 979                 return;
 980         listen_list_delete(front->cps);
 981         sldns_buffer_free(front->udp_buff);
 982         free(front);
 983 }
 984
 985 struct listen_port*
 986 listening_ports_open(struct config_file* cfg, int* reuseport)
 987 {
 988         struct listen_port* list = NULL;
 989         struct addrinfo hints;
 990         int i, do_ip4, do_ip6;
 991         int do_tcp, do_auto;
 992         char portbuf[32];
 993         snprintf(portbuf, sizeof(portbuf), "%d", cfg->port);
 994         do_ip4 = cfg->do_ip4;
 995         do_ip6 = cfg->do_ip6;
 996         do_tcp = cfg->do_tcp;
 997         do_auto = cfg->if_automatic && cfg->do_udp;
 998         if(cfg->incoming_num_tcp == 0)
 999                 do_tcp = 0;
1000
1001         /* getaddrinfo */
1002         memset(&hints, 0, sizeof(hints));
1003         hints.ai_flags = AI_PASSIVE;
1004         /* no name lookups on our listening ports */
1005         if(cfg->num_ifs > 0)
1006                 hints.ai_flags |= AI_NUMERICHOST;
1007         hints.ai_family = AF_UNSPEC;
1008 #ifndef INET6
1009         do_ip6 = 0;
1010 #endif
1011         if(!do_ip4 && !do_ip6) {
1012                 return NULL;
1013         }
1014         /* create ip4 and ip6 ports so that return addresses are nice. */
1015         if(do_auto || cfg->num_ifs == 0) {
1016                 if(do_ip6) {
1017                         hints.ai_family = AF_INET6;
1018                         if(!ports_create_if(do_auto?"::0":"::1",
1019                                 do_auto, cfg->do_udp, do_tcp,
1020                                 &hints, portbuf, &list,
1021                                 cfg->so_rcvbuf, cfg->so_sndbuf,
1022                                 cfg->ssl_port, reuseport)) {
1023                                 listening_ports_free(list);
1024                                 return NULL;
1025                         }
1026                 }
1027                 if(do_ip4) {
1028                         hints.ai_family = AF_INET;
1029                         if(!ports_create_if(do_auto?"0.0.0.0":"127.0.0.1",
1030                                 do_auto, cfg->do_udp, do_tcp,
1031                                 &hints, portbuf, &list,
1032                                 cfg->so_rcvbuf, cfg->so_sndbuf,
1033                                 cfg->ssl_port, reuseport)) {
1034                                 listening_ports_free(list);
1035                                 return NULL;
1036                         }
1037                 }
1038         } else for(i = 0; i<cfg->num_ifs; i++) {
1039                 if(str_is_ip6(cfg->ifs[i])) {
1040                         if(!do_ip6)
1041                                 continue;
1042                         hints.ai_family = AF_INET6;
1043                         if(!ports_create_if(cfg->ifs[i], 0, cfg->do_udp,
1044                                 do_tcp, &hints, portbuf, &list,
1045                                 cfg->so_rcvbuf, cfg->so_sndbuf,
1046                                 cfg->ssl_port, reuseport)) {
1047                                 listening_ports_free(list);
1048                                 return NULL;
1049                         }
1050                 } else {
1051                         if(!do_ip4)
1052                                 continue;
1053                         hints.ai_family = AF_INET;
1054                         if(!ports_create_if(cfg->ifs[i], 0, cfg->do_udp,
1055                                 do_tcp, &hints, portbuf, &list,
1056                                 cfg->so_rcvbuf, cfg->so_sndbuf,
1057                                 cfg->ssl_port, reuseport)) {
1058                                 listening_ports_free(list);
1059                                 return NULL;
1060                         }
1061                 }
1062         }
1063         return list;
1064 }
1065
1066 void listening_ports_free(struct listen_port* list)
1067 {
1068         struct listen_port* nx;
1069         while(list) {
1070                 nx = list->next;
1071                 if(list->fd != -1) {
1072 #ifndef USE_WINSOCK
1073                         close(list->fd);
1074 #else
1075                         closesocket(list->fd);
1076 #endif
1077                 }
1078                 free(list);
1079                 list = nx;
1080         }
1081 }
1082
1083 size_t listen_get_mem(struct listen_dnsport* listen)
1084 {
1085         size_t s = sizeof(*listen) + sizeof(*listen->base) +
1086                 sizeof(*listen->udp_buff) +
1087                 sldns_buffer_capacity(listen->udp_buff);
1088         struct listen_list* p;
1089         for(p = listen->cps; p; p = p->next) {
1090                 s += sizeof(*p);
1091                 s += comm_point_get_mem(p->com);
1092         }
1093         return s;
1094 }
1095
1096 void listen_stop_accept(struct listen_dnsport* listen)
1097 {
1098         /* do not stop the ones that have no tcp_free list
1099          * (they have already stopped listening) */
1100         struct listen_list* p;
1101         for(p=listen->cps; p; p=p->next) {
1102                 if(p->com->type == comm_tcp_accept &&
1103                         p->com->tcp_free != NULL) {
1104                         comm_point_stop_listening(p->com);
1105                 }
1106         }
1107 }
1108
1109 void listen_start_accept(struct listen_dnsport* listen)
1110 {
1111         /* do not start the ones that have no tcp_free list, it is no
1112          * use to listen to them because they have no free tcp handlers */
1113         struct listen_list* p;
1114         for(p=listen->cps; p; p=p->next) {
1115                 if(p->com->type == comm_tcp_accept &&
1116                         p->com->tcp_free != NULL) {
1117                         comm_point_start_listening(p->com, -1, -1);
1118                 }
1119         }
1120 }
1121