src/ip.c

   1 /*************************************************
   2 *     Exim - an Internet mail transport agent    *
   3 *************************************************/
   4
   5 /* Copyright (c) University of Cambridge 1995 - 2014 */
   6 /* See the file NOTICE for conditions of use and distribution. */
   7
   8 /* Functions for doing things with sockets. With the advent of IPv6 this has
   9 got messier, so that it's worth pulling out the code into separate functions
  10 that other parts of Exim can call, expecially as there are now several
  11 different places in the code where sockets are used. */
  12
  13
  14 #include "exim.h"
  15
  16
  17 /*************************************************
  18 *             Create a socket                    *
  19 *************************************************/
  20
  21 /* Socket creation happens in a number of places so it's packaged here for
  22 convenience.
  23
  24 Arguments:
  25   type       SOCK_DGRAM or SOCK_STREAM
  26   af         AF_INET or AF_INET6
  27
  28 Returns:     socket number or -1 on failure
  29 */
  30
  31 int
  32 ip_socket(int type, int af)
  33 {
  34 int sock = socket(af, type, 0);
  35 if (sock < 0)
  36   log_write(0, LOG_MAIN, "IPv%c socket creation failed: %s",
  37     (af == AF_INET6)? '6':'4', strerror(errno));
  38 return sock;
  39 }
  40
  41
  42
  43
  44 #if HAVE_IPV6
  45 /*************************************************
  46 *      Convert printing address to numeric       *
  47 *************************************************/
  48
  49 /* This function converts the textual form of an IP address into a numeric form
  50 in an appropriate structure in an IPv6 environment. The getaddrinfo() function
  51 can (apparently) handle more complicated addresses (e.g. those containing
  52 scopes) than inet_pton() in some environments. We use hints to tell it that the
  53 input must be a numeric address.
  54
  55 However, apparently some operating systems (or libraries) don't support
  56 getaddrinfo(), so there is a build-time option to revert to inet_pton() (which
  57 does not support scopes).
  58
  59 Arguments:
  60   address     textual form of the address
  61   addr        where to copy back the answer
  62
  63 Returns:      nothing - failure provokes a panic-die
  64 */
  65
  66 static void
  67 ip_addrinfo(uschar *address, struct sockaddr_in6 *saddr)
  68 {
  69 #ifdef IPV6_USE_INET_PTON
  70
  71   if (inet_pton(AF_INET6, CS address, &saddr->sin6_addr) != 1)
  72     log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
  73       "IP address", address);
  74   saddr->sin6_family = AF_INET6;
  75
  76 #else
  77
  78   int rc;
  79   struct addrinfo hints, *res;
  80   memset(&hints, 0, sizeof(hints));
  81   hints.ai_family = AF_INET6;
  82   hints.ai_socktype = SOCK_STREAM;
  83   hints.ai_flags = AI_NUMERICHOST;
  84   if ((rc = getaddrinfo(CS address, NULL, &hints, &res)) != 0 || res == NULL)
  85     log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
  86       "IP address: %s", address,
  87       (rc == 0)? "NULL result returned" : gai_strerror(rc));
  88   memcpy(saddr, res->ai_addr, res->ai_addrlen);
  89   freeaddrinfo(res);
  90
  91 #endif
  92 }
  93 #endif  /* HAVE_IPV6 */
  94
  95
  96 /*************************************************
  97 *         Bind socket to interface and port      *
  98 *************************************************/
  99
 100 /* This function binds a socket to a local interface address and port. For a
 101 wildcard IPv6 bind, the address is ":".
 102
 103 Arguments:
 104   sock           the socket
 105   af             AF_INET or AF_INET6 - the socket type
 106   address        the IP address, in text form
 107   port           the IP port (host order)
 108
 109 Returns:         the result of bind()
 110 */
 111
 112 int
 113 ip_bind(int sock, int af, uschar *address, int port)
 114 {
 115 int s_len;
 116 union sockaddr_46 sin;
 117 memset(&sin, 0, sizeof(sin));
 118
 119 /* Setup code when using an IPv6 socket. The wildcard address is ":", to
 120 ensure an IPv6 socket is used. */
 121
 122 #if HAVE_IPV6
 123 if (af == AF_INET6)
 124   {
 125   if (address[0] == ':' && address[1] == 0)
 126     {
 127     sin.v6.sin6_family = AF_INET6;
 128     sin.v6.sin6_addr = in6addr_any;
 129     }
 130   else
 131     {
 132     ip_addrinfo(address, &sin.v6);  /* Panic-dies on error */
 133     }
 134   sin.v6.sin6_port = htons(port);
 135   s_len = sizeof(sin.v6);
 136   }
 137 else
 138 #else     /* HAVE_IPv6 */
 139 af = af;  /* Avoid compiler warning */
 140 #endif    /* HAVE_IPV6 */
 141
 142 /* Setup code when using IPv4 socket. The wildcard address is "". */
 143
 144   {
 145   sin.v4.sin_family = AF_INET;
 146   sin.v4.sin_port = htons(port);
 147   s_len = sizeof(sin.v4);
 148   if (address[0] == 0)
 149     sin.v4.sin_addr.s_addr = (S_ADDR_TYPE)INADDR_ANY;
 150   else
 151     sin.v4.sin_addr.s_addr = (S_ADDR_TYPE)inet_addr(CS address);
 152   }
 153
 154 /* Now we can call the bind() function */
 155
 156 return bind(sock, (struct sockaddr *)&sin, s_len);
 157 }
 158
 159
 160
 161 /*************************************************
 162 *        Connect socket to remote host           *
 163 *************************************************/
 164
 165 /* This function connects a socket to a remote address and port. The socket may
 166 or may not have previously been bound to a local interface. The socket is not
 167 closed, even in cases of error. It is expected that the calling function, which
 168 created the socket, will be the one that closes it.
 169
 170 Arguments:
 171   sock        the socket
 172   af          AF_INET6 or AF_INET for the socket type
 173   address     the remote address, in text form
 174   port        the remote port
 175   timeout     a timeout (zero for indefinite timeout)
 176
 177 Returns:      0 on success; -1 on failure, with errno set
 178 */
 179
 180 int
 181 ip_connect(int sock, int af, uschar *address, int port, int timeout)
 182 {
 183 struct sockaddr_in s_in4;
 184 struct sockaddr *s_ptr;
 185 int s_len, rc, save_errno;
 186
 187 /* For an IPv6 address, use an IPv6 sockaddr structure. */
 188
 189 #if HAVE_IPV6
 190 struct sockaddr_in6 s_in6;
 191 if (af == AF_INET6)
 192   {
 193   memset(&s_in6, 0, sizeof(s_in6));
 194   ip_addrinfo(address, &s_in6);   /* Panic-dies on error */
 195   s_in6.sin6_port = htons(port);
 196   s_ptr = (struct sockaddr *)&s_in6;
 197   s_len = sizeof(s_in6);
 198   }
 199 else
 200 #else     /* HAVE_IPV6 */
 201 af = af;  /* Avoid compiler warning */
 202 #endif    /* HAVE_IPV6 */
 203
 204 /* For an IPv4 address, use an IPv4 sockaddr structure, even on a system with
 205 IPv6 support. */
 206
 207   {
 208   memset(&s_in4, 0, sizeof(s_in4));
 209   s_in4.sin_family = AF_INET;
 210   s_in4.sin_port = htons(port);
 211   s_in4.sin_addr.s_addr = (S_ADDR_TYPE)inet_addr(CS address);
 212   s_ptr = (struct sockaddr *)&s_in4;
 213   s_len = sizeof(s_in4);
 214   }
 215
 216 /* If no connection timeout is set, just call connect() without setting a
 217 timer, thereby allowing the inbuilt OS timeout to operate. */
 218
 219 sigalrm_seen = FALSE;
 220 if (timeout > 0) alarm(timeout);
 221 rc = connect(sock, s_ptr, s_len);
 222 save_errno = errno;
 223 alarm(0);
 224
 225 /* There is a testing facility for simulating a connection timeout, as I
 226 can't think of any other way of doing this. It converts a connection refused
 227 into a timeout if the timeout is set to 999999. */
 228
 229 if (running_in_test_harness)
 230   {
 231   if (save_errno == ECONNREFUSED && timeout == 999999)
 232     {
 233     rc = -1;
 234     save_errno = EINTR;
 235     sigalrm_seen = TRUE;
 236     }
 237   }
 238
 239 /* Success */
 240
 241 if (rc >= 0) return 0;
 242
 243 /* A failure whose error code is "Interrupted system call" is in fact
 244 an externally applied timeout if the signal handler has been run. */
 245
 246 errno = (save_errno == EINTR && sigalrm_seen)? ETIMEDOUT : save_errno;
 247 return -1;
 248 }
 249
 250
 251
 252 /*************************************************
 253 *    Create connected socket to remote host      *
 254 *************************************************/
 255
 256 /* Create a socket and connect to host (name or number, ipv6 ok)
 257    at one of port-range.
 258
 259 Arguments:
 260   type          SOCK_DGRAM or SOCK_STREAM
 261   af            AF_INET6 or AF_INET for the socket type
 262   address       the remote address, in text form
 263   portlo,porthi the remote port range
 264   timeout       a timeout
 265   connhost      if not NULL, host_item filled in with connection details
 266   errstr        pointer for allocated string on error
 267
 268 Return:
 269   socket fd, or -1 on failure (having allocated an error string)
 270 */
 271 int
 272 ip_connectedsocket(int type, const uschar * hostname, int portlo, int porthi,
 273         int timeout, host_item * connhost, uschar ** errstr)
 274 {
 275 int namelen, port;
 276 host_item shost;
 277 host_item *h;
 278 int af = 0, fd, fd4 = -1, fd6 = -1;
 279
 280 shost.next = NULL;
 281 shost.address = NULL;
 282 shost.port = portlo;
 283 shost.mx = -1;
 284
 285 namelen = Ustrlen(hostname);
 286
 287 /* Anything enclosed in [] must be an IP address. */
 288
 289 if (hostname[0] == '[' &&
 290     hostname[namelen - 1] == ']')
 291   {
 292   uschar * host = string_copy(hostname);
 293   host[namelen - 1] = 0;
 294   host++;
 295   if (string_is_ip_address(host, NULL) == 0)
 296     {
 297     *errstr = string_sprintf("malformed IP address \"%s\"", hostname);
 298     return -1;
 299     }
 300   shost.name = shost.address = host;
 301   }
 302
 303 /* Otherwise check for an unadorned IP address */
 304
 305 else if (string_is_ip_address(hostname, NULL) != 0)
 306   shost.name = shost.address = string_copy(hostname);
 307
 308 /* Otherwise lookup IP address(es) from the name */
 309
 310 else
 311   {
 312   shost.name = string_copy(hostname);
 313   if (host_find_byname(&shost, NULL, HOST_FIND_QUALIFY_SINGLE, NULL,
 314       FALSE) != HOST_FOUND)
 315     {
 316     *errstr = string_sprintf("no IP address found for host %s", shost.name);
 317     return -1;
 318     }
 319   }
 320
 321 /* Try to connect to the server - test each IP till one works */
 322
 323 for (h = &shost; h != NULL; h = h->next)
 324   {
 325   fd = (Ustrchr(h->address, ':') != 0)
 326     ? (fd6 < 0) ? (fd6 = ip_socket(type, af = AF_INET6)) : fd6
 327     : (fd4 < 0) ? (fd4 = ip_socket(type, af = AF_INET )) : fd4;
 328
 329   if (fd < 0)
 330     {
 331     *errstr = string_sprintf("failed to create socket: %s", strerror(errno));
 332     goto bad;
 333     }
 334
 335   for(port = portlo; port <= porthi; port++)
 336     if (ip_connect(fd, af, h->address, port, timeout) == 0)
 337       {
 338       if (fd != fd6) close(fd6);
 339       if (fd != fd4) close(fd4);
 340       if (connhost) {
 341         h->port = port;
 342         *connhost = *h;
 343         connhost->next = NULL;
 344         }
 345       return fd;
 346       }
 347   }
 348
 349 *errstr = string_sprintf("failed to connect to %s: "
 350   "couldn't connect to any host: %s", hostname, strerror(errno));
 351
 352 bad:
 353   close(fd4); close(fd6); return -1;
 354 }
 355
 356
 357 /*************************************************
 358 *         Set keepalive on a socket              *
 359 *************************************************/
 360
 361 /* Can be called for both incoming and outgoing sockets.
 362
 363 Arguments:
 364   sock       the socket
 365   address    the remote host address, for failure logging
 366   torf       true for outgoing connection, false for incoming
 367
 368 Returns:     nothing
 369 */
 370
 371 void
 372 ip_keepalive(int sock, uschar *address, BOOL torf)
 373 {
 374 int fodder = 1;
 375 if (setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 376     (uschar *)(&fodder), sizeof(fodder)) != 0)
 377   log_write(0, LOG_MAIN, "setsockopt(SO_KEEPALIVE) on connection %s %s "
 378     "failed: %s", torf? "to":"from", address, strerror(errno));
 379 }
 380
 381
 382
 383 /*************************************************
 384 *         Receive from a socket with timeout     *
 385 *************************************************/
 386
 387 /* The timeout is implemented using select(), and we loop to cover select()
 388 getting interrupted, and the possibility of select() returning with a positive
 389 result but no ready descriptor. Is this in fact possible?
 390
 391 Arguments:
 392   sock        the socket
 393   buffer      to read into
 394   bufsize     the buffer size
 395   timeout     the timeout
 396
 397 Returns:      > 0 => that much data read
 398               <= 0 on error or EOF; errno set - zero for EOF
 399 */
 400
 401 int
 402 ip_recv(int sock, uschar *buffer, int buffsize, int timeout)
 403 {
 404 fd_set select_inset;
 405 struct timeval tv;
 406 time_t start_recv = time(NULL);
 407 int rc;
 408
 409 /* Wait until the socket is ready */
 410
 411 for (;;)
 412   {
 413   FD_ZERO (&select_inset);
 414   FD_SET (sock, &select_inset);
 415   tv.tv_sec = timeout;
 416   tv.tv_usec = 0;
 417
 418   DEBUG(D_transport) debug_printf("waiting for data on socket\n");
 419   rc = select(sock + 1, (SELECT_ARG2_TYPE *)&select_inset, NULL, NULL, &tv);
 420
 421   /* If some interrupt arrived, just retry. We presume this to be rare,
 422   but it can happen (e.g. the SIGUSR1 signal sent by exiwhat causes
 423   select() to exit).
 424
 425   Aug 2004: Somebody set up a cron job that ran exiwhat every 2 minutes, making
 426   the interrupt not at all rare. Since the timeout is typically more than 2
 427   minutes, the effect was to block the timeout completely. To prevent this
 428   happening again, we do an explicit time test. */
 429
 430   if (rc < 0 && errno == EINTR)
 431     {
 432     DEBUG(D_transport) debug_printf("EINTR while waiting for socket data\n");
 433     if (time(NULL) - start_recv < timeout) continue;
 434     DEBUG(D_transport) debug_printf("total wait time exceeds timeout\n");
 435     }
 436
 437   /* Handle a timeout, and treat any other select error as a timeout, including
 438   an EINTR when we have been in this loop for longer than timeout. */
 439
 440   if (rc <= 0)
 441     {
 442     errno = ETIMEDOUT;
 443     return -1;
 444     }
 445
 446   /* If the socket is ready, break out of the loop. */
 447
 448   if (FD_ISSET(sock, &select_inset)) break;
 449   }
 450
 451 /* The socket is ready, read from it (via TLS if it's active). On EOF (i.e.
 452 close down of the connection), set errno to zero; otherwise leave it alone. */
 453
 454 #ifdef SUPPORT_TLS
 455 if (tls_out.active == sock)
 456   rc = tls_read(FALSE, buffer, buffsize);
 457 else if (tls_in.active == sock)
 458   rc = tls_read(TRUE, buffer, buffsize);
 459 else
 460 #endif
 461   rc = recv(sock, buffer, buffsize, 0);
 462
 463 if (rc > 0) return rc;
 464 if (rc == 0) errno = 0;
 465 return -1;
 466 }
 467
 468
 469
 470
 471 /*************************************************
 472 *    Lookup address family of potential socket   *
 473 *************************************************/
 474
 475 /* Given a file-descriptor, check to see if it's a socket and, if so,
 476 return the address family; detects IPv4 vs IPv6.  If not a socket then
 477 return -1.
 478
 479 The value 0 is typically AF_UNSPEC, which should not be seen on a connected
 480 fd.  If the return is -1, the errno will be from getsockname(); probably
 481 ENOTSOCK or ECONNRESET.
 482
 483 Arguments:     socket-or-not fd
 484 Returns:       address family or -1
 485 */
 486
 487 int
 488 ip_get_address_family(int fd)
 489 {
 490 struct sockaddr_storage ss;
 491 socklen_t sslen = sizeof(ss);
 492
 493 if (getsockname(fd, (struct sockaddr *) &ss, &sslen) < 0)
 494   return -1;
 495
 496 return (int) ss.ss_family;
 497 }
 498
 499
 500
 501
 502 /*************************************************
 503 *       Lookup DSCP settings for a socket        *
 504 *************************************************/
 505
 506 struct dscp_name_tableentry {
 507   const uschar *name;
 508   int value;
 509 };
 510 /* Keep both of these tables sorted! */
 511 static struct dscp_name_tableentry dscp_table[] = {
 512 #ifdef IPTOS_DSCP_AF11
 513     { CUS"af11", IPTOS_DSCP_AF11 },
 514     { CUS"af12", IPTOS_DSCP_AF12 },
 515     { CUS"af13", IPTOS_DSCP_AF13 },
 516     { CUS"af21", IPTOS_DSCP_AF21 },
 517     { CUS"af22", IPTOS_DSCP_AF22 },
 518     { CUS"af23", IPTOS_DSCP_AF23 },
 519     { CUS"af31", IPTOS_DSCP_AF31 },
 520     { CUS"af32", IPTOS_DSCP_AF32 },
 521     { CUS"af33", IPTOS_DSCP_AF33 },
 522     { CUS"af41", IPTOS_DSCP_AF41 },
 523     { CUS"af42", IPTOS_DSCP_AF42 },
 524     { CUS"af43", IPTOS_DSCP_AF43 },
 525     { CUS"ef", IPTOS_DSCP_EF },
 526 #endif
 527 #ifdef IPTOS_LOWCOST
 528     { CUS"lowcost", IPTOS_LOWCOST },
 529 #endif
 530     { CUS"lowdelay", IPTOS_LOWDELAY },
 531 #ifdef IPTOS_MINCOST
 532     { CUS"mincost", IPTOS_MINCOST },
 533 #endif
 534     { CUS"reliability", IPTOS_RELIABILITY },
 535     { CUS"throughput", IPTOS_THROUGHPUT }
 536 };
 537 static int dscp_table_size =
 538   sizeof(dscp_table) / sizeof(struct dscp_name_tableentry);
 539
 540 /* DSCP values change by protocol family, and so do the options used for
 541 setsockopt(); this utility does all the lookups.  It takes an unexpanded
 542 option string, expands it, strips off affix whitespace, then checks if it's
 543 a number.  If all of what's left is a number, then that's how the option will
 544 be parsed and success/failure is a range check.  If it's not all a number,
 545 then it must be a supported keyword.
 546
 547 Arguments:
 548   dscp_name   a string, so far unvalidated
 549   af          address_family in use
 550   level       setsockopt level to use
 551   optname     setsockopt name to use
 552   dscp_value  value for dscp_name
 553
 554 Returns: TRUE if okay to setsockopt(), else FALSE
 555
 556 *level and *optname may be set even if FALSE is returned
 557 */
 558
 559 BOOL
 560 dscp_lookup(const uschar *dscp_name, int af,
 561     int *level, int *optname, int *dscp_value)
 562 {
 563 uschar *dscp_lookup, *p;
 564 int first, last;
 565 long rawlong;
 566
 567 if (af == AF_INET)
 568   {
 569   *level = IPPROTO_IP;
 570   *optname = IP_TOS;
 571   }
 572 #if HAVE_IPV6 && defined(IPV6_TCLASS)
 573 else if (af == AF_INET6)
 574   {
 575   *level = IPPROTO_IPV6;
 576   *optname = IPV6_TCLASS;
 577   }
 578 #endif
 579 else
 580   {
 581   DEBUG(D_transport)
 582     debug_printf("Unhandled address family %d in dscp_lookup()\n", af);
 583   return FALSE;
 584   }
 585 if (!dscp_name)
 586   {
 587   DEBUG(D_transport)
 588     debug_printf("[empty DSCP]\n");
 589   return FALSE;
 590   }
 591 dscp_lookup = expand_string(US dscp_name);
 592 if (dscp_lookup == NULL || *dscp_lookup == '\0')
 593   return FALSE;
 594
 595 p = dscp_lookup + Ustrlen(dscp_lookup) - 1;
 596 while (isspace(*p)) *p-- = '\0';
 597 while (isspace(*dscp_lookup) && dscp_lookup < p) dscp_lookup++;
 598 if (*dscp_lookup == '\0')
 599   return FALSE;
 600
 601 rawlong = Ustrtol(dscp_lookup, &p, 0);
 602 if (p != dscp_lookup && *p == '\0')
 603   {
 604   /* We have six bits available, which will end up shifted to fit in 0xFC mask.
 605   RFC 2597 defines the values unshifted. */
 606   if (rawlong < 0 || rawlong > 0x3F)
 607     {
 608     DEBUG(D_transport)
 609       debug_printf("DSCP value %ld out of range, ignored.\n", rawlong);
 610     return FALSE;
 611     }
 612   *dscp_value = rawlong << 2;
 613   return TRUE;
 614   }
 615
 616 first = 0;
 617 last = dscp_table_size;
 618 while (last > first)
 619   {
 620   int middle = (first + last)/2;
 621   int c = Ustrcmp(dscp_lookup, dscp_table[middle].name);
 622   if (c == 0)
 623     {
 624     *dscp_value = dscp_table[middle].value;
 625     return TRUE;
 626     }
 627   else if (c > 0)
 628     {
 629     first = middle + 1;
 630     }
 631   else
 632     {
 633     last = middle;
 634     }
 635   }
 636 return FALSE;
 637 }
 638
 639 void
 640 dscp_list_to_stream(FILE *stream)
 641 {
 642 int i;
 643 for (i=0; i < dscp_table_size; ++i)
 644   fprintf(stream, "%s\n", dscp_table[i].name);
 645 }
 646
 647
 648 /* End of ip.c */