max_parallel transport option. Bug 1698

[users/heiko/exim.git] / src / src / daemon.c

/*************************************************
*     Exim - an Internet mail transport agent    *
*************************************************/

/* Copyright (c) University of Cambridge 1995 - 2015 */
/* See the file NOTICE for conditions of use and distribution. */

/* Functions concerned with running Exim as a daemon */


#include "exim.h"


/* Structure for holding data for each SMTP connection */

typedef struct smtp_slot {
  pid_t pid;                       /* pid of the spawned reception process */
  uschar *host_address;            /* address of the client host */
} smtp_slot;

/* An empty slot for initializing (Standard C does not allow constructor
expressions in assigments except as initializers in declarations). */

static smtp_slot empty_smtp_slot = { 0, NULL };



/*************************************************
*               Local static variables           *
*************************************************/

static SIGNAL_BOOL sigchld_seen;
static SIGNAL_BOOL sighup_seen;

static int   accept_retry_count = 0;
static int   accept_retry_errno;
static BOOL  accept_retry_select_failed;

static int   queue_run_count = 0;
static pid_t *queue_pid_slots = NULL;
static smtp_slot *smtp_slots = NULL;

static BOOL  write_pid = TRUE;



/*************************************************
*             SIGHUP Handler                     *
*************************************************/

/* All this handler does is to set a flag and re-enable the signal.

Argument: the signal number
Returns:  nothing
*/

static void
sighup_handler(int sig)
{
sig = sig;    /* Keep picky compilers happy */
sighup_seen = TRUE;
signal(SIGHUP, sighup_handler);
}



/*************************************************
*     SIGCHLD handler for main daemon process    *
*************************************************/

/* Don't re-enable the handler here, since we aren't doing the
waiting here. If the signal is re-enabled, there will just be an
infinite sequence of calls to this handler. The SIGCHLD signal is
used just as a means of waking up the daemon so that it notices
terminated subprocesses as soon as possible.

Argument: the signal number
Returns:  nothing
*/

static void
main_sigchld_handler(int sig)
{
sig = sig;    /* Keep picky compilers happy */
os_non_restarting_signal(SIGCHLD, SIG_DFL);
sigchld_seen = TRUE;
}




/*************************************************
*          Unexpected errors in SMTP calls       *
*************************************************/

/* This function just saves a bit of repetitious coding.

Arguments:
  log_msg        Text of message to be logged
  smtp_msg       Text of SMTP error message
  was_errno      The failing errno

Returns:         nothing
*/

static void
never_error(uschar *log_msg, uschar *smtp_msg, int was_errno)
{
uschar *emsg = (was_errno <= 0)? US"" :
  string_sprintf(": %s", strerror(was_errno));
log_write(0, LOG_MAIN|LOG_PANIC, "%s%s", log_msg, emsg);
if (smtp_out != NULL) smtp_printf("421 %s\r\n", smtp_msg);
}




/*************************************************
*            Handle a connected SMTP call        *
*************************************************/

/* This function is called when an SMTP connection has been accepted.
If there are too many, give an error message and close down. Otherwise
spin off a sub-process to handle the call. The list of listening sockets
is required so that they can be closed in the sub-process. Take care not to
leak store in this process - reset the stacking pool at the end.

Arguments:
  listen_sockets        sockets which are listening for incoming calls
  listen_socket_count   count of listening sockets
  accept_socket         socket of the current accepted call
  accepted              socket information about the current call

Returns:            nothing
*/

static void
handle_smtp_call(int *listen_sockets, int listen_socket_count,
  int accept_socket, struct sockaddr *accepted)
{
pid_t pid;
union sockaddr_46 interface_sockaddr;
EXIM_SOCKLEN_T ifsize = sizeof(interface_sockaddr);
int dup_accept_socket = -1;
int max_for_this_host = 0;
int wfsize = 0;
int wfptr = 0;
int save_log_selector = *log_selector;
uschar *whofrom = NULL;

void *reset_point = store_get(0);

/* Make the address available in ASCII representation, and also fish out
the remote port. */

sender_host_address = host_ntoa(-1, accepted, NULL, &sender_host_port);
DEBUG(D_any) debug_printf("Connection request from %s port %d\n",
  sender_host_address, sender_host_port);

/* Set up the output stream, check the socket has duplicated, and set up the
input stream. These operations fail only the exceptional circumstances. Note
that never_error() won't use smtp_out if it is NULL. */

smtp_out = fdopen(accept_socket, "wb");
if (smtp_out == NULL)
  {
  never_error(US"daemon: fdopen() for smtp_out failed", US"", errno);
  goto ERROR_RETURN;
  }

dup_accept_socket = dup(accept_socket);
if (dup_accept_socket < 0)
  {
  never_error(US"daemon: couldn't dup socket descriptor",
    US"Connection setup failed", errno);
  goto ERROR_RETURN;
  }

smtp_in = fdopen(dup_accept_socket, "rb");
if (smtp_in == NULL)
  {
  never_error(US"daemon: fdopen() for smtp_in failed",
    US"Connection setup failed", errno);
  goto ERROR_RETURN;
  }

/* Get the data for the local interface address. Panic for most errors, but
"connection reset by peer" just means the connection went away. */

if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
     &ifsize) < 0)
  {
  log_write(0, LOG_MAIN | ((errno == ECONNRESET)? 0 : LOG_PANIC),
    "getsockname() failed: %s", strerror(errno));
  smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
  goto ERROR_RETURN;
  }

interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
  interface_address, interface_port);

/* Build a string identifying the remote host and, if requested, the port and
the local interface data. This is for logging; at the end of this function the
memory is reclaimed. */

whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");

if (LOGGING(incoming_port))
  whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
    sender_host_port));

if (LOGGING(incoming_interface))
  whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
    interface_address, "]:", string_sprintf("%d", interface_port));

whofrom[wfptr] = 0;    /* Terminate the newly-built string */

/* Check maximum number of connections. We do not check for reserved
connections or unacceptable hosts here. That is done in the subprocess because
it might take some time. */

if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
  {
  DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
    smtp_accept_count, smtp_accept_max);
  smtp_printf("421 Too many concurrent SMTP connections; "
    "please try again later.\r\n");
  log_write(L_connection_reject,
            LOG_MAIN, "Connection from %s refused: too many connections",
    whofrom);
  goto ERROR_RETURN;
  }

/* If a load limit above which only reserved hosts are acceptable is defined,
get the load average here, and if there are in fact no reserved hosts, do
the test right away (saves a fork). If there are hosts, do the check in the
subprocess because it might take time. */

if (smtp_load_reserve >= 0)
  {
  load_average = OS_GETLOADAVG();
  if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
    {
    DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
      (double)load_average/1000.0);
    smtp_printf("421 Too much load; please try again later.\r\n");
    log_write(L_connection_reject,
              LOG_MAIN, "Connection from %s refused: load average = %.2f",
      whofrom, (double)load_average/1000.0);
    goto ERROR_RETURN;
    }
  }

/* Check that one specific host (strictly, IP address) is not hogging
resources. This is done here to prevent a denial of service attack by someone
forcing you to fork lots of times before denying service. The value of
smtp_accept_max_per_host is a string which is expanded. This makes it possible
to provide host-specific limits according to $sender_host address, but because
this is in the daemon mainline, only fast expansions (such as inline address
checks) should be used. The documentation is full of warnings. */

if (smtp_accept_max_per_host != NULL)
  {
  uschar *expanded = expand_string(smtp_accept_max_per_host);
  if (expanded == NULL)
    {
    if (!expand_string_forcedfail)
      log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
        "failed for %s: %s", whofrom, expand_string_message);
    }
  /* For speed, interpret a decimal number inline here */
  else
    {
    uschar *s = expanded;
    while (isdigit(*s))
      max_for_this_host = max_for_this_host * 10 + *s++ - '0';
    if (*s != 0)
      log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
        "for %s contains non-digit: %s", whofrom, expanded);
    }
  }

/* If we have fewer connections than max_for_this_host, we can skip the tedious
per host_address checks. Note that at this stage smtp_accept_count contains the
count of *other* connections, not including this one. */

if ((max_for_this_host > 0) &&
    (smtp_accept_count >= max_for_this_host))
  {
  int i;
  int host_accept_count = 0;
  int other_host_count = 0;    /* keep a count of non matches to optimise */

  for (i = 0; i < smtp_accept_max; ++i)
    {
    if (smtp_slots[i].host_address != NULL)
      {
      if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
       host_accept_count++;
      else
       other_host_count++;

      /* Testing all these strings is expensive - see if we can drop out
      early, either by hitting the target, or finding there are not enough
      connections left to make the target. */

      if ((host_accept_count >= max_for_this_host) ||
         ((smtp_accept_count - other_host_count) < max_for_this_host))
       break;
      }
    }

  if (host_accept_count >= max_for_this_host)
    {
    DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
      "IP address: count=%d max=%d\n",
      host_accept_count, max_for_this_host);
    smtp_printf("421 Too many concurrent SMTP connections "
      "from this IP address; please try again later.\r\n");
    log_write(L_connection_reject,
              LOG_MAIN, "Connection from %s refused: too many connections "
      "from that IP address", whofrom);
    goto ERROR_RETURN;
    }
  }

/* OK, the connection count checks have been passed. Before we can fork the
accepting process, we must first log the connection if requested. This logging
used to happen in the subprocess, but doing that means that the value of
smtp_accept_count can be out of step by the time it is logged. So we have to do
the logging here and accept the performance cost. Note that smtp_accept_count
hasn't yet been incremented to take account of this connection.

In order to minimize the cost (because this is going to happen for every
connection), do a preliminary selector test here. This saves ploughing through
the generalized logging code each time when the selector is false. If the
selector is set, check whether the host is on the list for logging. If not,
arrange to unset the selector in the subprocess. */

if (LOGGING(smtp_connection))
  {
  uschar *list = hosts_connection_nolog;
  if (list != NULL && verify_check_host(&list) == OK)
    save_log_selector &= ~L_smtp_connection;
  else
    log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
      "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
  }

/* Now we can fork the accepting process; do a lookup tidy, just in case any
expansion above did a lookup. */

search_tidyup();
pid = fork();

/* Handle the child process */

if (pid == 0)
  {
  int i;
  int queue_only_reason = 0;
  int old_pool = store_pool;
  int save_debug_selector = debug_selector;
  BOOL local_queue_only;
  BOOL session_local_queue_only;
  #ifdef SA_NOCLDWAIT
  struct sigaction act;
  #endif

  smtp_accept_count++;    /* So that it includes this process */

  /* May have been modified for the subprocess */

  *log_selector = save_log_selector;

  /* Get the local interface address into permanent store */

  store_pool = POOL_PERM;
  interface_address = string_copy(interface_address);
  store_pool = old_pool;

  /* Check for a tls-on-connect port */

  if (host_is_tls_on_connect_port(interface_port)) tls_in.on_connect = TRUE;

  /* Expand smtp_active_hostname if required. We do not do this any earlier,
  because it may depend on the local interface address (indeed, that is most
  likely what it depends on.) */

  smtp_active_hostname = primary_hostname;
  if (raw_active_hostname != NULL)
    {
    uschar *nah = expand_string(raw_active_hostname);
    if (nah == NULL)
      {
      if (!expand_string_forcedfail)
        {
        log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
          "(smtp_active_hostname): %s", raw_active_hostname,
          expand_string_message);
        smtp_printf("421 Local configuration error; "
          "please try again later.\r\n");
        mac_smtp_fflush();
        search_tidyup();
        _exit(EXIT_FAILURE);
        }
      }
    else if (nah[0] != 0) smtp_active_hostname = nah;
    }

  /* Initialize the queueing flags */

  queue_check_only();
  session_local_queue_only = queue_only;

  /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
  We also attempt to set things up so that children are automatically reaped,
  but just in case this isn't available, there's a paranoid waitpid() in the
  loop too (except for systems where we are sure it isn't needed). See the more
  extensive comment before the reception loop in exim.c for a fuller
  explanation of this logic. */

  for (i = 0; i < listen_socket_count; i++) (void)close(listen_sockets[i]);

  /* Set FD_CLOEXEC on the SMTP socket. We don't want any rogue child processes
  to be able to communicate with them, under any circumstances. */
  (void)fcntl(accept_socket, F_SETFD,
              fcntl(accept_socket, F_GETFD) | FD_CLOEXEC);
  (void)fcntl(dup_accept_socket, F_SETFD,
              fcntl(dup_accept_socket, F_GETFD) | FD_CLOEXEC);

  #ifdef SA_NOCLDWAIT
  act.sa_handler = SIG_IGN;
  sigemptyset(&(act.sa_mask));
  act.sa_flags = SA_NOCLDWAIT;
  sigaction(SIGCHLD, &act, NULL);
  #else
  signal(SIGCHLD, SIG_IGN);
  #endif

  /* Attempt to get an id from the sending machine via the RFC 1413
  protocol. We do this in the sub-process in order not to hold up the
  main process if there is any delay. Then set up the fullhost information
  in case there is no HELO/EHLO.

  If debugging is enabled only for the daemon, we must turn if off while
  finding the id, but turn it on again afterwards so that information about the
  incoming connection is output. */

  if (debug_daemon) debug_selector = 0;
  verify_get_ident(IDENT_PORT);
  host_build_sender_fullhost();
  debug_selector = save_debug_selector;

  DEBUG(D_any)
    debug_printf("Process %d is handling incoming connection from %s\n",
      (int)getpid(), sender_fullhost);

  /* Now disable debugging permanently if it's required only for the daemon
  process. */

  if (debug_daemon) debug_selector = 0;

  /* If there are too many child processes for immediate delivery,
  set the session_local_queue_only flag, which is initialized from the
  configured value and may therefore already be TRUE. Leave logging
  till later so it will have a message id attached. Note that there is no
  possibility of re-calculating this per-message, because the value of
  smtp_accept_count does not change in this subprocess. */

  if (smtp_accept_queue > 0 && smtp_accept_count > smtp_accept_queue)
    {
    session_local_queue_only = TRUE;
    queue_only_reason = 1;
    }

  /* Handle the start of the SMTP session, then loop, accepting incoming
  messages from the SMTP connection. The end will come at the QUIT command,
  when smtp_setup_msg() returns 0. A break in the connection causes the
  process to die (see accept.c).

  NOTE: We do *not* call smtp_log_no_mail() if smtp_start_session() fails,
  because a log line has already been written for all its failure exists
  (usually "connection refused: <reason>") and writing another one is
  unnecessary clutter. */

  if (!smtp_start_session())
    {
    mac_smtp_fflush();
    search_tidyup();
    _exit(EXIT_SUCCESS);
    }

  for (;;)
    {
    int rc;
    message_id[0] = 0;            /* Clear out any previous message_id */
    reset_point = store_get(0);   /* Save current store high water point */

    DEBUG(D_any)
      debug_printf("Process %d is ready for new message\n", (int)getpid());

    /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
    unacceptable host or if an ACL "drop" command was triggered, -1 on
    connection lost, and +1 on validly reaching DATA. Receive_msg() almost
    always returns TRUE when smtp_input is true; just retry if no message was
    accepted (can happen for invalid message parameters). However, it can yield
    FALSE if the connection was forcibly dropped by the DATA ACL. */

    if ((rc = smtp_setup_msg()) > 0)
      {
      BOOL ok = receive_msg(FALSE);
      search_tidyup();                    /* Close cached databases */
      if (!ok)                            /* Connection was dropped */
        {
        mac_smtp_fflush();
        smtp_log_no_mail();               /* Log no mail if configured */
        _exit(EXIT_SUCCESS);
        }
      if (message_id[0] == 0) continue;   /* No message was accepted */
      }
    else
      {
      mac_smtp_fflush();
      search_tidyup();
      smtp_log_no_mail();                 /* Log no mail if configured */
      _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
      }

    /* Show the recipients when debugging */

    DEBUG(D_receive)
      {
      int i;
      if (sender_address != NULL)
        debug_printf("Sender: %s\n", sender_address);
      if (recipients_list != NULL)
        {
        debug_printf("Recipients:\n");
        for (i = 0; i < recipients_count; i++)
          debug_printf("  %s\n", recipients_list[i].address);
        }
      }

    /* A message has been accepted. Clean up any previous delivery processes
    that have completed and are defunct, on systems where they don't go away
    by themselves (see comments when setting SIG_IGN above). On such systems
    (if any) these delivery processes hang around after termination until
    the next message is received. */

    #ifndef SIG_IGN_WORKS
    while (waitpid(-1, NULL, WNOHANG) > 0);
    #endif

    /* Reclaim up the store used in accepting this message */

    store_reset(reset_point);

    /* If queue_only is set or if there are too many incoming connections in
    existence, session_local_queue_only will be TRUE. If it is not, check
    whether we have received too many messages in this session for immediate
    delivery. */

    if (!session_local_queue_only &&
        smtp_accept_queue_per_connection > 0 &&
        receive_messagecount > smtp_accept_queue_per_connection)
      {
      session_local_queue_only = TRUE;
      queue_only_reason = 2;
      }

    /* Initialize local_queue_only from session_local_queue_only. If it is not
    true, and queue_only_load is set, check that the load average is below it.
    If local_queue_only is set by this means, we also set if for the session if
    queue_only_load_latch is true (the default). This means that, once set,
    local_queue_only remains set for any subsequent messages on the same SMTP
    connection. This is a deliberate choice; even though the load average may
    fall, it doesn't seem right to deliver later messages on the same call when
    not delivering earlier ones. However, the are special circumstances such as
    very long-lived connections from scanning appliances where this is not the
    best strategy. In such cases, queue_only_load_latch should be set false. */

    local_queue_only = session_local_queue_only;
    if (!local_queue_only && queue_only_load >= 0)
      {
      local_queue_only = (load_average = OS_GETLOADAVG()) > queue_only_load;
      if (local_queue_only)
        {
        queue_only_reason = 3;
        if (queue_only_load_latch) session_local_queue_only = TRUE;
        }
      }

    /* Log the queueing here, when it will get a message id attached, but
    not if queue_only is set (case 0). */

    if (local_queue_only) switch(queue_only_reason)
      {
      case 1:
      log_write(L_delay_delivery,
                LOG_MAIN, "no immediate delivery: too many connections "
                "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
      break;

      case 2:
      log_write(L_delay_delivery,
                LOG_MAIN, "no immediate delivery: more than %d messages "
                "received in one connection", smtp_accept_queue_per_connection);
      break;

      case 3:
      log_write(L_delay_delivery,
                LOG_MAIN, "no immediate delivery: load average %.2f",
                (double)load_average/1000.0);
      break;
      }

    /* If a delivery attempt is required, spin off a new process to handle it.
    If we are not root, we have to re-exec exim unless deliveries are being
    done unprivileged. */

    else if (!queue_only_policy && !deliver_freeze)
      {
      pid_t dpid;

      /* Before forking, ensure that the C output buffer is flushed. Otherwise
      anything that it in it will get duplicated, leading to duplicate copies
      of the pending output. */

      mac_smtp_fflush();

      if ((dpid = fork()) == 0)
        {
        (void)fclose(smtp_in);
        (void)fclose(smtp_out);

        /* Don't ever molest the parent's SSL connection, but do clean up
        the data structures if necessary. */

        #ifdef SUPPORT_TLS
        tls_close(TRUE, FALSE);
        #endif

        /* Reset SIGHUP and SIGCHLD in the child in both cases. */

        signal(SIGHUP,  SIG_DFL);
        signal(SIGCHLD, SIG_DFL);

        if (geteuid() != root_uid && !deliver_drop_privilege)
          {
          signal(SIGALRM, SIG_DFL);
          (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
            message_id);
          /* Control does not return here. */
          }

        /* No need to re-exec; SIGALRM remains set to the default handler */

        (void)deliver_message(message_id, FALSE, FALSE);
        search_tidyup();
        _exit(EXIT_SUCCESS);
        }

      if (dpid > 0)
        {
        DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
        }
      else
        {
        log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
          "failed: %s", strerror(errno));
        }
      }
    }
  }


/* Carrying on in the parent daemon process... Can't do much if the fork
failed. Otherwise, keep count of the number of accepting processes and
remember the pid for ticking off when the child completes. */

if (pid < 0)
  {
  never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
  }
else
  {
  int i;
  for (i = 0; i < smtp_accept_max; ++i)
    {
    if (smtp_slots[i].pid <= 0)
      {
      smtp_slots[i].pid = pid;
      if (smtp_accept_max_per_host != NULL)
        smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
      smtp_accept_count++;
      break;
      }
    }
  DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
    smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
  }

/* Get here via goto in error cases */

ERROR_RETURN:

/* Close the streams associated with the socket which will also close the
socket fds in this process. We can't do anything if fclose() fails, but
logging brings it to someone's attention. However, "connection reset by peer"
isn't really a problem, so skip that one. On Solaris, a dropped connection can
manifest itself as a broken pipe, so drop that one too. If the streams don't
exist, something went wrong while setting things up. Make sure the socket
descriptors are closed, in order to drop the connection. */

if (smtp_out != NULL)
  {
  if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
    log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
      strerror(errno));
  smtp_out = NULL;
  }
else (void)close(accept_socket);

if (smtp_in != NULL)
  {
  if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
    log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
      strerror(errno));
  smtp_in = NULL;
  }
else (void)close(dup_accept_socket);

/* Release any store used in this process, including the store used for holding
the incoming host address and an expanded active_hostname. */

log_close_all();
store_reset(reset_point);
sender_host_address = NULL;
}




/*************************************************
*       Check wildcard listen special cases      *
*************************************************/

/* This function is used when binding and listening on lists of addresses and
ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
sockets may interact in different ways in different operating systems. It is
passed an error number, the list of listening addresses, and the current
address. Two checks are available: for a previous wildcard IPv6 address, or for
a following wildcard IPv4 address, in both cases on the same port.

In practice, pairs of wildcard addresses should be adjacent in the address list
because they are sorted that way below.

Arguments:
  eno            the error number
  addresses      the list of addresses
  ipa            the current IP address
  back           if TRUE, check for previous wildcard IPv6 address
                 if FALSE, check for a following wildcard IPv4 address

Returns:         TRUE or FALSE
*/

static BOOL
check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
  BOOL back)
{
ip_address_item *ipa2;

/* For the "back" case, if the failure was "address in use" for a wildcard IPv4
address, seek a previous IPv6 wildcard address on the same port. As it is
previous, it must have been successfully bound and be listening. Flag it as a
"6 including 4" listener. */

if (back)
  {
  if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
  for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
    {
    if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
      {
      ipa2->v6_include_v4 = TRUE;
      return TRUE;
      }
    }
  }

/* For the "forward" case, if the current address is a wildcard IPv6 address,
we seek a following wildcard IPv4 address on the same port. */

else
  {
  if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
  for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
    if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
  }

return FALSE;
}




/*************************************************
*         Handle terminating subprocesses        *
*************************************************/

/* Handle the termination of child processes. Theoretically, this need be done
only when sigchld_seen is TRUE, but rumour has it that some systems lose
SIGCHLD signals at busy times, so to be on the safe side, this function is
called each time round. It shouldn't be too expensive.

Arguments:  none
Returns:    nothing
*/

static void
handle_ending_processes(void)
{
int status;
pid_t pid;

while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
  {
  int i;
  DEBUG(D_any)
    {
    debug_printf("child %d ended: status=0x%x\n", (int)pid, status);
#ifdef WCOREDUMP
    if (WIFEXITED(status))
      debug_printf("  normal exit, %d\n", WEXITSTATUS(status));
    else if (WIFSIGNALED(status))
      debug_printf("  signal exit, signal %d%s\n", WTERMSIG(status),
          WCOREDUMP(status) ? " (core dumped)" : "");
#endif
    }

  /* If it's a listening daemon for which we are keeping track of individual
  subprocesses, deal with an accepting process that has terminated. */

  if (smtp_slots != NULL)
    {
    for (i = 0; i < smtp_accept_max; i++)
      {
      if (smtp_slots[i].pid == pid)
        {
        if (smtp_slots[i].host_address != NULL)
          store_free(smtp_slots[i].host_address);
        smtp_slots[i] = empty_smtp_slot;
        if (--smtp_accept_count < 0) smtp_accept_count = 0;
        DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
          smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
        break;
        }
      }
    if (i < smtp_accept_max) continue;  /* Found an accepting process */
    }

  /* If it wasn't an accepting process, see if it was a queue-runner
  process that we are tracking. */

  if (queue_pid_slots != NULL)
    {
    for (i = 0; i < queue_run_max; i++)
      {
      if (queue_pid_slots[i] == pid)
        {
        queue_pid_slots[i] = 0;
        if (--queue_run_count < 0) queue_run_count = 0;
        DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
          queue_run_count, (queue_run_count == 1)? "" : "es");
        break;
        }
      }
    }
  }
}



/*************************************************
*              Exim Daemon Mainline              *
*************************************************/

/* The daemon can do two jobs, either of which is optional:

(1) Listens for incoming SMTP calls and spawns off a sub-process to handle
each one. This is requested by the -bd option, with -oX specifying the SMTP
port on which to listen (for testing).

(2) Spawns a queue-running process every so often. This is controlled by the
-q option with a an interval time. (If no time is given, a single queue run
is done from the main function, and control doesn't get here.)

Root privilege is required in order to attach to port 25. Some systems require
it when calling socket() rather than bind(). To cope with all cases, we run as
root for both socket() and bind(). Some systems also require root in order to
write to the pid file directory. This function must therefore be called as root
if it is to work properly in all circumstances. Once the socket is bound and
the pid file written, root privilege is given up if there is an exim uid.

There are no arguments to this function, and it never returns. */

void
daemon_go(void)
{
struct passwd *pw;
int *listen_sockets = NULL;
int listen_socket_count = 0;
ip_address_item *addresses = NULL;
time_t last_connection_time = (time_t)0;

/* If any debugging options are set, turn on the D_pid bit so that all
debugging lines get the pid added. */

DEBUG(D_any|D_v) debug_selector |= D_pid;

if (inetd_wait_mode)
  {
  int on = 1;

  listen_socket_count = 1;
  listen_sockets = store_get(sizeof(int *));
  (void) close(3);
  if (dup2(0, 3) == -1)
    {
    log_write(0, LOG_MAIN|LOG_PANIC_DIE,
        "failed to dup inetd socket safely away: %s", strerror(errno));
    }
  listen_sockets[0] = 3;
  (void) close(0);
  (void) close(1);
  (void) close(2);
  exim_nullstd();

  if (debug_file == stderr)
    {
    /* need a call to log_write before call to open debug_file, so that
    log.c:file_path has been initialised.  This is unfortunate. */
    log_write(0, LOG_MAIN, "debugging Exim in inetd wait mode starting");

    fclose(debug_file);
    debug_file = NULL;
    exim_nullstd(); /* re-open fd2 after we just closed it again */
    debug_logging_activate(US"-wait", NULL);
    }

  DEBUG(D_any) debug_printf("running in inetd wait mode\n");

  /* As per below, when creating sockets ourselves, we handle tcp_nodelay for
  our own buffering; we assume though that inetd set the socket REUSEADDR. */

  if (tcp_nodelay) setsockopt(3, IPPROTO_TCP, TCP_NODELAY,
    (uschar *)(&on), sizeof(on));
  }


if (inetd_wait_mode || daemon_listen)
  {
  /* If any option requiring a load average to be available during the
  reception of a message is set, call os_getloadavg() while we are root
  for those OS for which this is necessary the first time it is called (in
  order to perform an "open" on the kernel memory file). */

  #ifdef LOAD_AVG_NEEDS_ROOT
  if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
       (deliver_queue_load_max >= 0 && deliver_drop_privilege))
    (void)os_getloadavg();
  #endif
  }


/* Do the preparation for setting up a listener on one or more interfaces, and
possible on various ports. This is controlled by the combination of
local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
(which is a list of default ports to use for those items in local_interfaces
that do not specify a port). The -oX command line option can be used to
override one or both of these options.

If local_interfaces is not set, the default is to listen on all interfaces.
When it is set, it can include "all IPvx interfaces" as an item. This is useful
when different ports are in use.

It turns out that listening on all interfaces is messy in an IPv6 world,
because several different implementation approaches have been taken. This code
is now supposed to work with all of them. The point of difference is whether an
IPv6 socket that is listening on all interfaces will receive incoming IPv4
calls or not. We also have to cope with the case when IPv6 libraries exist, but
there is no IPv6 support in the kernel.

. On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
  addresses. However, if an IPv4 socket is also listening on all interfaces,
  calls are directed to the appropriate socket.

. On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
  give them as mapped addresses, but an attempt also to listen on an IPv4
  socket on all interfaces causes an error.

. On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
  two sockets if you want to accept both kinds of call.

. FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
  can be turned off, to make it behave like the versions of Linux described
  above.

. I heard a report that the USAGI IPv6 stack for Linux has implemented
  IPV6_V6ONLY.

So, what we do when IPv6 is supported is as follows:

 (1) After it is set up, the list of interfaces is scanned for wildcard
     addresses. If an IPv6 and an IPv4 wildcard are both found for the same
     port, the list is re-arranged so that they are together, with the IPv6
     wildcard first.

 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
     carry on if an IPv4 wildcard socket for the same port follows later in the
     list. This allows Exim to carry on in the case when the kernel has no IPv6
     support.

 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
     option is defined. However, if setting fails, carry on regardless (but log
     the incident).

 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
     error.

 (5) If binding or listening on an IPv4 wildcard socket fails with the error
     EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
     port (which must have succeeded or we wouldn't have got this far), we
     assume we are in the situation where just a single socket is permitted,
     and ignore the error.

Phew!

The preparation code decodes options and sets up the relevant data. We do this
first, so that we can return non-zero if there are any syntax errors, and also
write to stderr. */

if (daemon_listen && !inetd_wait_mode)
  {
  int *default_smtp_port;
  int sep;
  int pct = 0;
  uschar *s;
  const uschar * list;
  uschar *local_iface_source = US"local_interfaces";
  ip_address_item *ipa;
  ip_address_item **pipa;

  /* If -oX was used, disable the writing of a pid file unless -oP was
  explicitly used to force it. Then scan the string given to -oX. Any items
  that contain neither a dot nor a colon are used to override daemon_smtp_port.
  Any other items are used to override local_interfaces. */

  if (override_local_interfaces != NULL)
    {
    uschar *new_smtp_port = NULL;
    uschar *new_local_interfaces = NULL;
    int portsize = 0;
    int portptr = 0;
    int ifacesize = 0;
    int ifaceptr = 0;

    if (override_pid_file_path == NULL) write_pid = FALSE;

    list = override_local_interfaces;
    sep = 0;
    while ((s = string_nextinlist(&list, &sep, big_buffer, big_buffer_size)))
      {
      uschar joinstr[4];
      uschar **ptr;
      int *sizeptr;
      int *ptrptr;

      if (Ustrpbrk(s, ".:") == NULL)
        {
        ptr = &new_smtp_port;
        sizeptr = &portsize;
        ptrptr = &portptr;
        }
      else
        {
        ptr = &new_local_interfaces;
        sizeptr = &ifacesize;
        ptrptr = &ifaceptr;
        }

      if (*ptr == NULL)
        {
        joinstr[0] = sep;
        joinstr[1] = ' ';
        *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
        }

      *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
      *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
      }

    if (new_smtp_port != NULL)
      {
      new_smtp_port[portptr] = 0;
      daemon_smtp_port = new_smtp_port;
      DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n  %s\n",
        daemon_smtp_port);
      }

    if (new_local_interfaces != NULL)
      {
      new_local_interfaces[ifaceptr] = 0;
      local_interfaces = new_local_interfaces;
      local_iface_source = US"-oX data";
      DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n  %s\n",
        local_interfaces);
      }
    }

  /* Create a list of default SMTP ports, to be used if local_interfaces
  contains entries without explict ports. First count the number of ports, then
  build a translated list in a vector. */

  list = daemon_smtp_port;
  sep = 0;
  while ((s = string_nextinlist(&list, &sep, big_buffer, big_buffer_size)))
    pct++;
  default_smtp_port = store_get((pct+1) * sizeof(int));
  list = daemon_smtp_port;
  sep = 0;
  for (pct = 0;
       (s = string_nextinlist(&list, &sep, big_buffer, big_buffer_size));
       pct++)
    {
    if (isdigit(*s))
      {
      uschar *end;
      default_smtp_port[pct] = Ustrtol(s, &end, 0);
      if (end != s + Ustrlen(s))
        log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
      }
    else
      {
      struct servent *smtp_service = getservbyname(CS s, "tcp");
      if (!smtp_service)
        log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
      default_smtp_port[pct] = ntohs(smtp_service->s_port);
      }
    }
  default_smtp_port[pct] = 0;

  /* Check the list of TLS-on-connect ports and do name lookups if needed */

  list = tls_in.on_connect_ports;
  sep = 0;
  while ((s = string_nextinlist(&list, &sep, big_buffer, big_buffer_size)))
    if (!isdigit(*s))
      {
      list = tls_in.on_connect_ports;
      tls_in.on_connect_ports = NULL;
      sep = 0;
      while ((s = string_nextinlist(&list, &sep, big_buffer, big_buffer_size)))
        {
        if (!isdigit(*s))
          {
          struct servent *smtp_service = getservbyname(CS s, "tcp");
          if (!smtp_service)
            log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
          s= string_sprintf("%d", (int)ntohs(smtp_service->s_port));
          }
        tls_in.on_connect_ports = string_append_listele(tls_in.on_connect_ports,
            ':', s);
        }
      break;
      }

  /* Create the list of local interfaces, possibly with ports included. This
  list may contain references to 0.0.0.0 and ::0 as wildcards. These special
  values are converted below. */

  addresses = host_build_ifacelist(local_interfaces, local_iface_source);

  /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
  into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
  fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
  anyway, but we need to know which are the wildcard addresses, and the shorter
  strings are neater.

  In the same scan, fill in missing port numbers from the default list. When
  there is more than one item in the list, extra items are created. */

  for (ipa = addresses; ipa != NULL; ipa = ipa->next)
    {
    int i;

    if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
    else if (Ustrcmp(ipa->address, "::0") == 0)
      {
      ipa->address[0] = ':';
      ipa->address[1] = 0;
      }

    if (ipa->port > 0) continue;

    if (daemon_smtp_port[0] <= 0)
      log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
        "%s and daemon_smtp_port is unset; cannot start daemon",
        (ipa->address[0] == 0)? US"\"all IPv4\"" :
        (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
    ipa->port = default_smtp_port[0];
    for (i = 1; default_smtp_port[i] > 0; i++)
      {
      ip_address_item *new = store_get(sizeof(ip_address_item));
      memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
      new->port = default_smtp_port[i];
      new->next = ipa->next;
      ipa->next = new;
      ipa = new;
      }
    }

  /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
  wildcard for the same port, ensure that (a) they are together and (b) the
  IPv6 address comes first. This makes handling the messy features easier, and
  also simplifies the construction of the "daemon started" log line. */

  pipa = &addresses;
  for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
    {
    ip_address_item *ipa2;

    /* Handle an IPv4 wildcard */

    if (ipa->address[0] == 0)
      {
      for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
        {
        ip_address_item *ipa3 = ipa2->next;
        if (ipa3->address[0] == ':' &&
            ipa3->address[1] == 0 &&
            ipa3->port == ipa->port)
          {
          ipa2->next = ipa3->next;
          ipa3->next = ipa;
          *pipa = ipa3;
          break;
          }
        }
      }

    /* Handle an IPv6 wildcard. */

    else if (ipa->address[0] == ':' && ipa->address[1] == 0)
      {
      for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
        {
        ip_address_item *ipa3 = ipa2->next;
        if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
          {
          ipa2->next = ipa3->next;
          ipa3->next = ipa->next;
          ipa->next = ipa3;
          ipa = ipa3;
          break;
          }
        }
      }
    }

  /* Get a vector to remember all the sockets in */

  for (ipa = addresses; ipa != NULL; ipa = ipa->next)
    listen_socket_count++;
  listen_sockets = store_get(sizeof(int *) * listen_socket_count);

  } /* daemon_listen but not inetd_wait_mode */

if (daemon_listen)
  {

  /* Do a sanity check on the max connects value just to save us from getting
  a huge amount of store. */

  if (smtp_accept_max > 4095) smtp_accept_max = 4096;

  /* There's no point setting smtp_accept_queue unless it is less than the max
  connects limit. The configuration reader ensures that the max is set if the
  queue-only option is set. */

  if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;

  /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
  track of them for total number and queue/host limits. */

  if (smtp_accept_max > 0)
    {
    int i;
    smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
    for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
    }
  }

/* The variable background_daemon is always false when debugging, but
can also be forced false in order to keep a non-debugging daemon in the
foreground. If background_daemon is true, close all open file descriptors that
we know about, but then re-open stdin, stdout, and stderr to /dev/null.  Also
do this for inetd_wait mode.

This is protection against any called functions (in libraries, or in
Perl, or whatever) that think they can write to stderr (or stdout). Before this
was added, it was quite likely that an SMTP connection would use one of these
file descriptors, in which case writing random stuff to it caused chaos.

Then disconnect from the controlling terminal, Most modern Unixes seem to have
setsid() for getting rid of the controlling terminal. For any OS that doesn't,
setsid() can be #defined as a no-op, or as something else. */

if (background_daemon || inetd_wait_mode)
  {
  log_close_all();    /* Just in case anything was logged earlier */
  search_tidyup();    /* Just in case any were used in reading the config. */
  (void)close(0);           /* Get rid of stdin/stdout/stderr */
  (void)close(1);
  (void)close(2);
  exim_nullstd();     /* Connect stdin/stdout/stderr to /dev/null */
  log_stderr = NULL;  /* So no attempt to copy paniclog output */
  }

if (background_daemon)
  {
  /* If the parent process of this one has pid == 1, we are re-initializing the
  daemon as the result of a SIGHUP. In this case, there is no need to do
  anything, because the controlling terminal has long gone. Otherwise, fork, in
  case current process is a process group leader (see 'man setsid' for an
  explanation) before calling setsid(). */

  if (getppid() != 1)
    {
    pid_t pid = fork();
    if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
      "fork() failed when starting daemon: %s", strerror(errno));
    if (pid > 0) exit(EXIT_SUCCESS);      /* in parent process, just exit */
    (void)setsid();                       /* release controlling terminal */
    }
  }

/* We are now in the disconnected, daemon process (unless debugging). Set up
the listening sockets if required. */

if (daemon_listen && !inetd_wait_mode)
  {
  int sk;
  int on = 1;
  ip_address_item *ipa;

  /* For each IP address, create a socket, bind it to the appropriate port, and
  start listening. See comments above about IPv6 sockets that may or may not
  accept IPv4 calls when listening on all interfaces. We also have to cope with
  the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
  listening, provided a wildcard IPv4 socket for the same port follows. */

  for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
    {
    BOOL wildcard;
    ip_address_item *ipa2;
    int af;

    if (Ustrchr(ipa->address, ':') != NULL)
      {
      af = AF_INET6;
      wildcard = ipa->address[1] == 0;
      }
    else
      {
      af = AF_INET;
      wildcard = ipa->address[0] == 0;
      }

    listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
    if (listen_sockets[sk] < 0)
      {
      if (check_special_case(0, addresses, ipa, FALSE))
        {
        log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
          "listening (%s): will use IPv4", strerror(errno));
        goto SKIP_SOCKET;
        }
      log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
        (af == AF_INET6)? '6' : '4', strerror(errno));
      }

    /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
    available. Just log failure (can get protocol not available, just like
    socket creation can). */

    #ifdef IPV6_V6ONLY
    if (af == AF_INET6 && wildcard &&
        setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
          sizeof(on)) < 0)
      log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
        "socket failed (%s): carrying on without it", strerror(errno));
    #endif  /* IPV6_V6ONLY */

    /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
    is being handled.  Without this, a connection will prevent reuse of the
    smtp port for listening. */

    if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
                   (uschar *)(&on), sizeof(on)) < 0)
      log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
        "failed when starting daemon: %s", strerror(errno));

    /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
    disable this because it breaks some broken clients. */

    if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
      (uschar *)(&on), sizeof(on));

    /* Now bind the socket to the required port; if Exim is being restarted
    it may not always be possible to bind immediately, even with SO_REUSEADDR
    set, so try 10 times, waiting between each try. After 10 failures, we give
    up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
    we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
    address for the same port, ignore the error on the grounds that we must be
    in a system where the IPv6 socket accepts both kinds of call. This is
    necessary for (some release of) USAGI Linux; other IP stacks fail at the
    listen() stage instead. */

    for(;;)
      {
      uschar *msg, *addr;
      if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
      if (check_special_case(errno, addresses, ipa, TRUE))
        {
        DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
          "listen() success; EADDRINUSE ignored\n");
        (void)close(listen_sockets[sk]);
        goto SKIP_SOCKET;
        }
      msg = US strerror(errno);
      addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
        ipa->address;
      if (daemon_startup_retries <= 0)
        log_write(0, LOG_MAIN|LOG_PANIC_DIE,
          "socket bind() to port %d for address %s failed: %s: "
          "daemon abandoned", ipa->port, addr, msg);
      log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
        "failed: %s: waiting %s before trying again (%d more %s)",
        ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
        daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
      daemon_startup_retries--;
      sleep(daemon_startup_sleep);
      }

    DEBUG(D_any)
      {
      if (wildcard)
        debug_printf("listening on all interfaces (IPv%c) port %d\n",
          (af == AF_INET6)? '6' : '4', ipa->port);
      else
        debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
      }

    /* Start listening on the bound socket, establishing the maximum backlog of
    connections that is allowed. On success, continue to the next address. */

    if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;

    /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
    fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
    and there was a previous successful IPv6 wildcard listen on the same port,
    we want to ignore the error on the grounds that we must be in a system
    where the IPv6 socket accepts both kinds of call. */

    if (!check_special_case(errno, addresses, ipa, TRUE))
      log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
        wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
        ipa->address,
        strerror(errno));

    DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
      "listen() success; EADDRINUSE ignored\n");
    (void)close(listen_sockets[sk]);

    /* Come here if there has been a problem with the socket which we
    are going to ignore. We remove the address from the chain, and back up the
    counts. */

    SKIP_SOCKET:
    sk--;                          /* Back up the count */
    listen_socket_count--;         /* Reduce the total */
    if (ipa == addresses) addresses = ipa->next; else
      {
      for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
      ipa2->next = ipa->next;
      ipa = ipa2;
      }
    }          /* End of bind/listen loop for each address */
  }            /* End of setup for listening */


/* If we are not listening, we want to write a pid file only if -oP was
explicitly given. */

else if (override_pid_file_path == NULL) write_pid = FALSE;

/* Write the pid to a known file for assistance in identification, if required.
We do this before giving up root privilege, because on some systems it is
necessary to be root in order to write into the pid file directory. There's
nothing to stop multiple daemons running, as long as no more than one listens
on a given TCP/IP port on the same interface(s). However, in these
circumstances it gets far too complicated to mess with pid file names
automatically. Consequently, Exim 4 writes a pid file only

  (a) When running in the test harness, or
  (b) When -bd is used and -oX is not used, or
  (c) When -oP is used to supply a path.

The variable daemon_write_pid is used to control this. */

if (running_in_test_harness || write_pid)
  {
  FILE *f;

  if (override_pid_file_path != NULL)
    pid_file_path = override_pid_file_path;

  if (pid_file_path[0] == 0)
    pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);

  f = modefopen(pid_file_path, "wb", 0644);
  if (f != NULL)
    {
    (void)fprintf(f, "%d\n", (int)getpid());
    (void)fclose(f);
    DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
    }
  else
    {
    DEBUG(D_any)
      debug_printf("%s\n", string_open_failed(errno, "pid file %s",
        pid_file_path));
    }
  }

/* Set up the handler for SIGHUP, which causes a restart of the daemon. */

sighup_seen = FALSE;
signal(SIGHUP, sighup_handler);

/* Give up root privilege at this point (assuming that exim_uid and exim_gid
are not root). The third argument controls the running of initgroups().
Normally we do this, in order to set up the groups for the Exim user. However,
if we are not root at this time - some odd installations run that way - we
cannot do this. */

exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");

/* Update the originator_xxx fields so that received messages as listed as
coming from Exim, not whoever started the daemon. */

originator_uid = exim_uid;
originator_gid = exim_gid;
originator_login = ((pw = getpwuid(exim_uid)) != NULL)?
  string_copy_malloc(US pw->pw_name) : US"exim";

/* Get somewhere to keep the list of queue-runner pids if we are keeping track
of them (and also if we are doing queue runs). */

if (queue_interval > 0 && queue_run_max > 0)
  {
  int i;
  queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
  for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
  }

/* Set up the handler for termination of child processes. */

sigchld_seen = FALSE;
os_non_restarting_signal(SIGCHLD, main_sigchld_handler);

/* If we are to run the queue periodically, pretend the alarm has just gone
off. This will cause the first queue-runner to get kicked off straight away. */

sigalrm_seen = (queue_interval > 0);

/* Log the start up of a daemon - at least one of listening or queue running
must be set up. */

if (inetd_wait_mode)
  {
  uschar *p = big_buffer;

  if (inetd_wait_timeout >= 0)
    sprintf(CS p, "terminating after %d seconds", inetd_wait_timeout);
  else
    sprintf(CS p, "with no wait timeout");

  log_write(0, LOG_MAIN,
    "exim %s daemon started: pid=%d, launched with listening socket, %s",
    version_string, getpid(), big_buffer);
  set_process_info("daemon(%s): pre-listening socket", version_string);

  /* set up the timeout logic */
  sigalrm_seen = 1;
  }

else if (daemon_listen)
  {
  int i, j;
  int smtp_ports = 0;
  int smtps_ports = 0;
  ip_address_item *ipa;
  uschar *p = big_buffer;
  uschar *qinfo = (queue_interval > 0)?
    string_sprintf("-q%s", readconf_printtime(queue_interval))
    :
    US"no queue runs";

  /* Build a list of listening addresses in big_buffer, but limit it to 10
  items. The style is for backwards compatibility.

  It is now possible to have some ports listening for SMTPS (the old,
  deprecated protocol that starts TLS without using STARTTLS), and others
  listening for standard SMTP. Keep their listings separate. */

  for (j = 0; j < 2; j++)
    {
    for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
       {
       /* First time round, look for SMTP ports; second time round, look for
       SMTPS ports. For the first one of each, insert leading text. */

       if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
         {
         if (j == 0)
           {
           if (smtp_ports++ == 0)
             {
             memcpy(p, "SMTP on", 8);
             p += 7;
             }
           }
         else
           {
           if (smtps_ports++ == 0)
             {
             (void)sprintf(CS p, "%sSMTPS on",
               (smtp_ports == 0)? "":" and for ");
             while (*p != 0) p++;
             }
           }

         /* Now the information about the port (and sometimes interface) */

         if (ipa->address[0] == ':' && ipa->address[1] == 0)
           {
           if (ipa->next != NULL && ipa->next->address[0] == 0 &&
               ipa->next->port == ipa->port)
             {
             (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
             ipa = ipa->next;
             }
           else if (ipa->v6_include_v4)
             (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
           else
             (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
           }
         else if (ipa->address[0] == 0)
           (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
         else
           (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
         while (*p != 0) p++;
         }
       }

    if (ipa != NULL)
      {
      memcpy(p, " ...", 5);
      p += 4;
      }
    }

  log_write(0, LOG_MAIN,
    "exim %s daemon started: pid=%d, %s, listening for %s",
    version_string, getpid(), qinfo, big_buffer);
  set_process_info("daemon(%s): %s, listening for %s", version_string, qinfo, big_buffer);
  }

else
  {
  log_write(0, LOG_MAIN,
    "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
    version_string, getpid(), readconf_printtime(queue_interval));
  set_process_info("daemon(%s): -q%s, not listening",
    version_string,
    readconf_printtime(queue_interval));
  }

/* Do any work it might be useful to amortize over our children
(eg: compile regex) */

deliver_init();
dns_pattern_init();

#ifdef WITH_CONTENT_SCAN
malware_init();
#endif

/* Close the log so it can be renamed and moved. In the few cases below where
this long-running process writes to the log (always exceptional conditions), it
closes the log afterwards, for the same reason. */

log_close_all();

DEBUG(D_any) debug_print_ids(US"daemon running with");

/* Any messages accepted via this route are going to be SMTP. */

smtp_input = TRUE;

/* Enter the never-ending loop... */

for (;;)
  {
  #if HAVE_IPV6
  struct sockaddr_in6 accepted;
  #else
  struct sockaddr_in accepted;
  #endif

  EXIM_SOCKLEN_T len;
  pid_t pid;

  /* This code is placed first in the loop, so that it gets obeyed at the
  start, before the first wait, for the queue-runner case, so that the first
  one can be started immediately.

  The other option is that we have an inetd wait timeout specified to -bw. */

  if (sigalrm_seen)
    {
    if (inetd_wait_timeout > 0)
      {
      time_t resignal_interval = inetd_wait_timeout;

      if (last_connection_time == (time_t)0)
        {
        DEBUG(D_any)
          debug_printf("inetd wait timeout expired, but still not seen first message, ignoring\n");
        }
      else
        {
        time_t now = time(NULL);
        if (now == (time_t)-1)
          {
          DEBUG(D_any) debug_printf("failed to get time: %s\n", strerror(errno));
          }
        else
          {
          if ((now - last_connection_time) >= inetd_wait_timeout)
            {
            DEBUG(D_any)
              debug_printf("inetd wait timeout %d expired, ending daemon\n",
                  inetd_wait_timeout);
            log_write(0, LOG_MAIN, "exim %s daemon terminating, inetd wait timeout reached.\n",
                version_string);
            exit(EXIT_SUCCESS);
            }
          else
            {
            resignal_interval -= (now - last_connection_time);
            }
          }
        }

      sigalrm_seen = FALSE;
      alarm(resignal_interval);
      }

    else
      {
      DEBUG(D_any) debug_printf("SIGALRM received\n");

      /* Do a full queue run in a child process, if required, unless we already
      have enough queue runners on the go. If we are not running as root, a
      re-exec is required. */

      if (queue_interval > 0 &&
         (queue_run_max <= 0 || queue_run_count < queue_run_max))
        {
        if ((pid = fork()) == 0)
          {
          int sk;

          DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
            (int)getpid());

          /* Disable debugging if it's required only for the daemon process. We
          leave the above message, because it ties up with the "child ended"
          debugging messages. */

          if (debug_daemon) debug_selector = 0;

          /* Close any open listening sockets in the child */

          for (sk = 0; sk < listen_socket_count; sk++)
            (void)close(listen_sockets[sk]);

          /* Reset SIGHUP and SIGCHLD in the child in both cases. */

          signal(SIGHUP,  SIG_DFL);
          signal(SIGCHLD, SIG_DFL);

          /* Re-exec if privilege has been given up, unless deliver_drop_
          privilege is set. Reset SIGALRM before exec(). */

          if (geteuid() != root_uid && !deliver_drop_privilege)
            {
            uschar opt[8];
            uschar *p = opt;
            uschar *extra[5];
            int extracount = 1;

            signal(SIGALRM, SIG_DFL);
            *p++ = '-';
            *p++ = 'q';
            if (queue_2stage) *p++ = 'q';
            if (queue_run_first_delivery) *p++ = 'i';
            if (queue_run_force) *p++ = 'f';
            if (deliver_force_thaw) *p++ = 'f';
            if (queue_run_local) *p++ = 'l';
            *p = 0;
            extra[0] = opt;

            /* If -R or -S were on the original command line, ensure they get
            passed on. */

            if (deliver_selectstring != NULL)
              {
              extra[extracount++] = deliver_selectstring_regex? US"-Rr" : US"-R";
              extra[extracount++] = deliver_selectstring;
              }

            if (deliver_selectstring_sender != NULL)
              {
              extra[extracount++] = deliver_selectstring_sender_regex?
                US"-Sr" : US"-S";
              extra[extracount++] = deliver_selectstring_sender;
              }

            /* Overlay this process with a new execution. */

            (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, extracount,
              extra[0], extra[1], extra[2], extra[3], extra[4]);

            /* Control never returns here. */
            }

          /* No need to re-exec; SIGALRM remains set to the default handler */

          queue_run(NULL, NULL, FALSE);
          _exit(EXIT_SUCCESS);
          }

        if (pid < 0)
          {
          log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
            "process failed: %s", strerror(errno));
          log_close_all();
          }
        else
          {
          int i;
          for (i = 0; i < queue_run_max; ++i)
            {
            if (queue_pid_slots[i] <= 0)
              {
              queue_pid_slots[i] = pid;
              queue_run_count++;
              break;
              }
            }
          DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
            queue_run_count, (queue_run_count == 1)? "" : "es");
          }
        }

      /* Reset the alarm clock */

      sigalrm_seen = FALSE;
      alarm(queue_interval);
      }

    } /* sigalrm_seen */


  /* Sleep till a connection happens if listening, and handle the connection if
  that is why we woke up. The FreeBSD operating system requires the use of
  select() before accept() because the latter function is not interrupted by
  a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
  OS do notice signals in accept() but it does no harm to have the select()
  in for all of them - and it won't then be a lurking problem for ports to
  new OS. In fact, the later addition of listening on specific interfaces only
  requires this way of working anyway. */

  if (daemon_listen)
    {
    int sk, lcount, select_errno;
    int max_socket = 0;
    BOOL select_failed = FALSE;
    fd_set select_listen;

    FD_ZERO(&select_listen);
    for (sk = 0; sk < listen_socket_count; sk++)
      {
      FD_SET(listen_sockets[sk], &select_listen);
      if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
      }

    DEBUG(D_any) debug_printf("Listening...\n");

    /* In rare cases we may have had a SIGCHLD signal in the time between
    setting the handler (below) and getting back here. If so, pretend that the
    select() was interrupted so that we reap the child. This might still leave
    a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
    only to do the reaping more quickly, it shouldn't result in anything other
    than a delay until something else causes a wake-up. */

    if (sigchld_seen)
      {
      lcount = -1;
      errno = EINTR;
      }
    else
      {
      lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
        NULL, NULL, NULL);
      }

    if (lcount < 0)
      {
      select_failed = TRUE;
      lcount = 1;
      }

    /* Clean up any subprocesses that may have terminated. We need to do this
    here so that smtp_accept_max_per_host works when a connection to that host
    has completed, and we are about to accept a new one. When this code was
    later in the sequence, a new connection could be rejected, even though an
    old one had just finished. Preserve the errno from any select() failure for
    the use of the common select/accept error processing below. */

    select_errno = errno;
    handle_ending_processes();
    errno = select_errno;

    /* Loop for all the sockets that are currently ready to go. If select
    actually failed, we have set the count to 1 and select_failed=TRUE, so as
    to use the common error code for select/accept below. */

    while (lcount-- > 0)
      {
      int accept_socket = -1;
      if (!select_failed)
        {
        for (sk = 0; sk < listen_socket_count; sk++)
          {
          if (FD_ISSET(listen_sockets[sk], &select_listen))
            {
            len = sizeof(accepted);
            accept_socket = accept(listen_sockets[sk],
              (struct sockaddr *)&accepted, &len);
            FD_CLR(listen_sockets[sk], &select_listen);
            break;
            }
          }
        }

      /* If select or accept has failed and this was not caused by an
      interruption, log the incident and try again. With asymmetric TCP/IP
      routing errors such as "No route to network" have been seen here. Also
      "connection reset by peer" has been seen. These cannot be classed as
      disastrous errors, but they could fill up a lot of log. The code in smail
      crashes the daemon after 10 successive failures of accept, on the grounds
      that some OS fail continuously. Exim originally followed suit, but this
      appears to have caused problems. Now it just keeps going, but instead of
      logging each error, it batches them up when they are continuous. */

      if (accept_socket < 0 && errno != EINTR)
        {
        if (accept_retry_count == 0)
          {
          accept_retry_errno = errno;
          accept_retry_select_failed = select_failed;
          }
        else
          {
          if (errno != accept_retry_errno ||
              select_failed != accept_retry_select_failed ||
              accept_retry_count >= 50)
            {
            log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
              "%d %s() failure%s: %s",
              accept_retry_count,
              accept_retry_select_failed? "select" : "accept",
              (accept_retry_count == 1)? "" : "s",
              strerror(accept_retry_errno));
            log_close_all();
            accept_retry_count = 0;
            accept_retry_errno = errno;
            accept_retry_select_failed = select_failed;
            }
          }
        accept_retry_count++;
        }

      else
        {
        if (accept_retry_count > 0)
          {
          log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
            accept_retry_count,
            accept_retry_select_failed? "select" : "accept",
            (accept_retry_count == 1)? "" : "s",
            strerror(accept_retry_errno));
          log_close_all();
          accept_retry_count = 0;
          }
        }

      /* If select/accept succeeded, deal with the connection. */

      if (accept_socket >= 0)
        {
        if (inetd_wait_timeout)
          last_connection_time = time(NULL);
        handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
          (struct sockaddr *)&accepted);
        }
      }
    }

  /* If not listening, then just sleep for the queue interval. If we woke
  up early the last time for some other signal, it won't matter because
  the alarm signal will wake at the right time. This code originally used
  sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
  rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
  select() can be used as an interruptible sleep() on all versions of Unix. */

  else
    {
    struct timeval tv;
    tv.tv_sec = queue_interval;
    tv.tv_usec = 0;
    select(0, NULL, NULL, NULL, &tv);
    handle_ending_processes();
    }

  /* Re-enable the SIGCHLD handler if it has been run. It can't do it
  for itself, because it isn't doing the waiting itself. */

  if (sigchld_seen)
    {
    sigchld_seen = FALSE;
    os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
    }

  /* Handle being woken by SIGHUP. We know at this point that the result
  of accept() has been dealt with, so we can re-exec exim safely, first
  closing the listening sockets so that they can be reused. Cancel any pending
  alarm in case it is just about to go off, and set SIGHUP to be ignored so
  that another HUP in quick succession doesn't clobber the new daemon before it
  gets going. All log files get closed by the close-on-exec flag; however, if
  the exec fails, we need to close the logs. */

  if (sighup_seen)
    {
    int sk;
    log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
      getpid());
    for (sk = 0; sk < listen_socket_count; sk++)
      (void)close(listen_sockets[sk]);
    alarm(0);
    signal(SIGHUP, SIG_IGN);
    sighup_argv[0] = exim_path;
    exim_nullstd();
    execv(CS exim_path, (char *const *)sighup_argv);
    log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
      getpid(), exim_path, strerror(errno));
    log_close_all();
    }

  }   /* End of main loop */

/* Control never reaches here */
}

/* vi: aw ai sw=2
*/
/* End of exim_daemon.c */