Fix ultimate retry timeouts for intermittently deliverable recipients.

[exim.git] / src / src / retry.c

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

/* Copyright (c) University of Cambridge 1995 - 2009 */
/* See the file NOTICE for conditions of use and distribution. */

/* Functions concerned with retrying unsuccessful deliveries. */


#include "exim.h"



/*************************************************
*         Check the ultimate address timeout     *
*************************************************/

/* This function tests whether a message has been on the queue longer than
the maximum retry time for a particular host or address.

Arguments:
  retry_key     the key to look up a retry rule
  domain        the domain to look up a domain retry rule
  retry_record  contains error information for finding rule
  now           the time

Returns:        TRUE if the ultimate timeout has been reached
*/

BOOL
retry_ultimate_address_timeout(uschar *retry_key, uschar *domain,
  dbdata_retry *retry_record, time_t now)
{
BOOL address_timeout;

DEBUG(D_retry)
  {
  debug_printf("retry time not reached: checking ultimate address timeout\n");
  debug_printf("  now=%d first_failed=%d next_try=%d expired=%d\n",
    (int)now, (int)retry_record->first_failed,
    (int)retry_record->next_try, retry_record->expired);
  }

retry_config *retry =
  retry_find_config(retry_key+2, domain,
    retry_record->basic_errno, retry_record->more_errno);

if (retry != NULL && retry->rules != NULL)
  {
  retry_rule *last_rule;
  for (last_rule = retry->rules;
       last_rule->next != NULL;
       last_rule = last_rule->next);
  DEBUG(D_retry)
    debug_printf("  received_time=%d diff=%d timeout=%d\n",
      received_time, (int)(now - received_time), last_rule->timeout);
  address_timeout = (now - received_time > last_rule->timeout);
  }
else
  {
  DEBUG(D_retry)
    debug_printf("no retry rule found: assume timed out\n");
  address_timeout = TRUE;
  }

DEBUG(D_retry)
  if (address_timeout)
    debug_printf("on queue longer than maximum retry for address - "
      "allowing delivery\n");

return address_timeout;
}



/*************************************************
*     Set status of a host+address item          *
*************************************************/

/* This function is passed a host_item which contains a host name and an
IP address string. Its job is to set the status of the address if it is not
already set (indicated by hstatus_unknown). The possible values are:

   hstatus_usable    the address is not listed in the unusable tree, and does
                     not have a retry record, OR the time is past the next
                     try time, OR the message has been on the queue for more
                     than the maximum retry time for a failing host

   hstatus_unusable  the address is listed in the unusable tree, or does have
                     a retry record, and the time is not yet at the next retry
                     time.

   hstatus_unusable_expired  as above, but also the retry time has expired
                     for this address.

The reason a delivery is permitted when a message has been around for a very
long time is to allow the ultimate address timeout to operate after a delivery
failure. Otherwise some messages may stick around without being tried for too
long.

If a host retry record is retrieved from the hints database, the time of last
trying is filled into the last_try field of the host block. If a host is
generally usable, a check is made to see if there is a retry delay on this
specific message at this host.

If a non-standard port is being used, it is added to the retry key.

Arguments:
  domain              the address domain
  host                pointer to a host item
  portstring          "" for standard port, ":xxxx" for a non-standard port
  include_ip_address  TRUE to include the address in the key - this is
                        usual, but sometimes is not wanted
  retry_host_key      where to put a pointer to the key for the host-specific
                        retry record, if one is read and the host is usable
  retry_message_key   where to put a pointer to the key for the message+host
                        retry record, if one is read and the host is usable

Returns:    TRUE if the host has expired but is usable because
             its retry time has come
*/

BOOL
retry_check_address(uschar *domain, host_item *host, uschar *portstring,
  BOOL include_ip_address, uschar **retry_host_key, uschar **retry_message_key)
{
BOOL yield = FALSE;
time_t now = time(NULL);
uschar *host_key, *message_key;
open_db dbblock;
open_db *dbm_file;
tree_node *node;
dbdata_retry *host_retry_record, *message_retry_record;

*retry_host_key = *retry_message_key = NULL;

DEBUG(D_transport|D_retry) debug_printf("checking status of %s\n", host->name);

/* Do nothing if status already set; otherwise initialize status as usable. */

if (host->status != hstatus_unknown) return FALSE;
host->status = hstatus_usable;

/* Generate the host key for the unusable tree and the retry database. Ensure
host names are lower cased (that's what %S does). */

host_key = include_ip_address?
  string_sprintf("T:%S:%s%s", host->name, host->address, portstring) :
  string_sprintf("T:%S%s", host->name, portstring);

/* Generate the message-specific key */

message_key = string_sprintf("%s:%s", host_key, message_id);

/* Search the tree of unusable IP addresses. This is filled in when deliveries
fail, because the retry database itself is not updated until the end of all
deliveries (so as to do it all in one go). The tree records addresses that have
become unusable during this delivery process (i.e. those that will get put into
the retry database when it is updated). */

node = tree_search(tree_unusable, host_key);
if (node != NULL)
  {
  DEBUG(D_transport|D_retry) debug_printf("found in tree of unusables\n");
  host->status = (node->data.val > 255)?
    hstatus_unusable_expired : hstatus_unusable;
  host->why = node->data.val & 255;
  return FALSE;
  }

/* Open the retry database, giving up if there isn't one. Otherwise, search for
the retry records, and then close the database again. */

if ((dbm_file = dbfn_open(US"retry", O_RDONLY, &dbblock, FALSE)) == NULL)
  {
  DEBUG(D_deliver|D_retry|D_hints_lookup)
    debug_printf("no retry data available\n");
  return FALSE;
  }
host_retry_record = dbfn_read(dbm_file, host_key);
message_retry_record = dbfn_read(dbm_file, message_key);
dbfn_close(dbm_file);

/* Ignore the data if it is too old - too long since it was written */

if (host_retry_record == NULL)
  {
  DEBUG(D_transport|D_retry) debug_printf("no host retry record\n");
  }
else if (now - host_retry_record->time_stamp > retry_data_expire)
  {
  host_retry_record = NULL;
  DEBUG(D_transport|D_retry) debug_printf("host retry record too old\n");
  }

if (message_retry_record == NULL)
  {
  DEBUG(D_transport|D_retry) debug_printf("no message retry record\n");
  }
else if (now - message_retry_record->time_stamp > retry_data_expire)
  {
  message_retry_record = NULL;
  DEBUG(D_transport|D_retry) debug_printf("message retry record too old\n");
  }

/* If there's a host-specific retry record, check for reaching the retry
time (or forcing). If not, and the host is not expired, check for the message
having been around for longer than the maximum retry time for this host or
address. Allow the delivery if it has. Otherwise set the appropriate unusable
flag and return FALSE. Otherwise arrange to return TRUE if this is an expired
host. */

if (host_retry_record != NULL)
  {
  *retry_host_key = host_key;

  /* We have not reached the next try time. Check for the ultimate address
  timeout if the host has not expired. */

  if (now < host_retry_record->next_try && !deliver_force)
    {
    if (!host_retry_record->expired &&
        retry_ultimate_address_timeout(host_key, domain,
          host_retry_record, now))
      return FALSE;

    /* We have not hit the ultimate address timeout; host is unusable. */

    host->status = (host_retry_record->expired)?
      hstatus_unusable_expired : hstatus_unusable;
    host->why = hwhy_retry;
    host->last_try = host_retry_record->last_try;
    return FALSE;
    }

  /* Host is usable; set return TRUE if expired. */

  yield = host_retry_record->expired;
  }

/* It's OK to try the host. If there's a message-specific retry record, check
for reaching its retry time (or forcing). If not, mark the host unusable,
unless the ultimate address timeout has been reached. */

if (message_retry_record != NULL)
  {
  *retry_message_key = message_key;
  if (now < message_retry_record->next_try && !deliver_force)
    {
    if (!retry_ultimate_address_timeout(host_key, domain,
        message_retry_record, now))
      {
      host->status = hstatus_unusable;
      host->why = hwhy_retry;
      }
    return FALSE;
    }
  }

return yield;
}




/*************************************************
*           Add a retry item to an address       *
*************************************************/

/* Retry items are chained onto an address when it is deferred either by router
or by a transport, or if it succeeds or fails and there was a previous retry
item that now needs to be deleted. Sometimes there can be both kinds of item:
for example, if routing was deferred but then succeeded, and delivery then
deferred. In that case there is a delete item for the routing retry, and an
updating item for the delivery.

(But note that that is only visible at the outer level, because in remote
delivery subprocesses, the address starts "clean", with no retry items carried
in.)

These items are used at the end of a delivery attempt to update the retry
database. The keys start R: for routing delays and T: for transport delays.

Arguments:
  addr    the address block onto which to hang the item
  key     the retry key
  flags   delete, host, and message flags, copied into the block

Returns:  nothing
*/

void
retry_add_item(address_item *addr, uschar *key, int flags)
{
retry_item *rti = store_get(sizeof(retry_item));
rti->next = addr->retries;
addr->retries = rti;
rti->key = key;
rti->basic_errno = addr->basic_errno;
rti->more_errno = addr->more_errno;
rti->message = addr->message;
rti->flags = flags;

DEBUG(D_transport|D_retry)
  {
  int letter = rti->more_errno & 255;
  debug_printf("added retry item for %s: errno=%d more_errno=", rti->key,
    rti->basic_errno);
  if (letter == 'A' || letter == 'M')
    debug_printf("%d,%c", (rti->more_errno >> 8) & 255, letter);
  else
    debug_printf("%d", rti->more_errno);
  debug_printf(" flags=%d\n", flags);
  }
}



/*************************************************
*        Find retry configuration data           *
*************************************************/

/* Search the in-store retry information for the first retry item that applies
to a given destination. If the key contains an @ we are probably handling a
local delivery and have a complete address to search for; this happens when
retry_use_local_part is set on a router. Otherwise, the key is likely to be a
host name for a remote delivery, or a domain name for a local delivery. We
prepend *@ on the front of it so that it will match a retry item whose address
item pattern is independent of the local part. The alternate key, if set, is
always just a domain, so we treat it likewise.

Arguments:
  key          key for which retry info is wanted
  alternate    alternative key, always just a domain
  basic_errno  specific error predicate on the retry rule, or zero
  more_errno   additional data for errno predicate

Returns:       pointer to retry rule, or NULL
*/

retry_config *
retry_find_config(uschar *key, uschar *alternate, int basic_errno,
  int more_errno)
{
int replace = 0;
uschar *use_key, *use_alternate;
uschar *colon = Ustrchr(key, ':');
retry_config *yield;

/* If there's a colon in the key, there are two possibilities:

(1) This is a key for a host, ip address, and possibly port, in the format

      hostname:ip+port

    In this case, we temporarily replace the colon with a zero, to terminate
    the string after the host name.

(2) This is a key for a pipe, file, or autoreply delivery, in the format

      pipe-or-file-or-auto:x@y

    where x@y is the original address that provoked the delivery. The pipe or
    file or auto will start with | or / or >, whereas a host name will start
    with a letter or a digit. In this case we want to use the original address
    to search for a retry rule. */

if (colon != NULL)
  {
  if (isalnum(*key))
    replace = ':';
  else
    key = Ustrrchr(key, ':') + 1;   /* Take from the last colon */
  }

if (replace == 0) colon = key + Ustrlen(key);
*colon = 0;

/* Sort out the keys */

use_key = (Ustrchr(key, '@') != NULL)? key : string_sprintf("*@%s", key);
use_alternate = (alternate == NULL)? NULL : string_sprintf("*@%s", alternate);

/* Scan the configured retry items. */

for (yield = retries; yield != NULL; yield = yield->next)
  {
  uschar *plist = yield->pattern;
  uschar *slist = yield->senders;

  /* If a specific error is set for this item, check that we are handling that
  specific error, and if so, check any additional error information if
  required. */

  if (yield->basic_errno != 0)
    {
    /* Special code is required for quota errors, as these can either be system
    quota errors, or Exim's own quota imposition, which has a different error
    number. Full partitions are also treated in the same way as quota errors.
    */

    if (yield->basic_errno == ERRNO_EXIMQUOTA)
      {
      if ((basic_errno != ERRNO_EXIMQUOTA && basic_errno != errno_quota &&
           basic_errno != ENOSPC) ||
          (yield->more_errno != 0 && yield->more_errno > more_errno))
        continue;
      }

    /* The TLSREQUIRED error also covers TLSFAILURE. These are subtly different
    errors, but not worth separating at this level. */

    else if (yield->basic_errno == ERRNO_TLSREQUIRED)
      {
      if (basic_errno != ERRNO_TLSREQUIRED && basic_errno != ERRNO_TLSFAILURE)
        continue;
      }

    /* Handle 4xx responses to MAIL, RCPT, or DATA. The code that was received
    is in the 2nd least significant byte of more_errno (with 400 subtracted).
    The required value is coded in the 2nd least significant byte of the
    yield->more_errno field as follows:

      255     => any 4xx code
      >= 100  => the decade must match the value less 100
      < 100   => the exact value must match
    */

    else if (yield->basic_errno == ERRNO_MAIL4XX ||
             yield->basic_errno == ERRNO_RCPT4XX ||
             yield->basic_errno == ERRNO_DATA4XX)
      {
      int wanted;
      if (basic_errno != yield->basic_errno) continue;
      wanted = (yield->more_errno >> 8) & 255;
      if (wanted != 255)
        {
        int evalue = (more_errno >> 8) & 255;
        if (wanted >= 100)
          {
          if ((evalue/10)*10 != wanted - 100) continue;
          }
        else if (evalue != wanted) continue;
        }
      }

    /* There are some special cases for timeouts */

    else if (yield->basic_errno == ETIMEDOUT)
      {
      if (basic_errno != ETIMEDOUT) continue;

      /* Just RTEF_CTOUT in the rule => don't care about 'A'/'M' addresses */
      if (yield->more_errno == RTEF_CTOUT)
        {
        if ((more_errno & RTEF_CTOUT) == 0) continue;
        }

      else if (yield->more_errno != 0)
        {
        int cf_errno = more_errno;
        if ((yield->more_errno & RTEF_CTOUT) == 0) cf_errno &= ~RTEF_CTOUT;
        if (yield->more_errno != cf_errno) continue;
        }
      }

    /* Default checks for exact match */

    else
      {
      if (yield->basic_errno != basic_errno ||
         (yield->more_errno != 0 && yield->more_errno != more_errno))
       continue;
      }
    }

  /* If the "senders" condition is set, check it. Note that sender_address may
  be null during -brt checking, in which case we do not use this rule. */

  if (slist != NULL && (sender_address == NULL ||
      match_address_list(sender_address, TRUE, TRUE, &slist, NULL, -1, 0,
        NULL) != OK))
    continue;

  /* Check for a match between the address list item at the start of this retry
  rule and either the main or alternate keys. */

  if (match_address_list(use_key, TRUE, TRUE, &plist, NULL, -1, UCHAR_MAX+1,
        NULL) == OK ||
     (use_alternate != NULL &&
      match_address_list(use_alternate, TRUE, TRUE, &plist, NULL, -1,
        UCHAR_MAX+1, NULL) == OK))
    break;
  }

*colon = replace;
return yield;
}




/*************************************************
*              Update retry database             *
*************************************************/

/* Update the retry data for any directing/routing/transporting that was
deferred, or delete it for those that succeeded after a previous defer. This is
done all in one go to minimize opening/closing/locking of the database file.

Note that, because SMTP delivery involves a list of destinations to try, there
may be defer-type retry information for some of them even when the message was
successfully delivered. Likewise if it eventually failed.

This function may move addresses from the defer to the failed queue if the
ultimate retry time has expired.

Arguments:
  addr_defer    queue of deferred addresses
  addr_failed   queue of failed addresses
  addr_succeed  queue of successful addresses

Returns:        nothing
*/

void
retry_update(address_item **addr_defer, address_item **addr_failed,
  address_item **addr_succeed)
{
open_db dbblock;
open_db *dbm_file = NULL;
time_t now = time(NULL);
int i;

DEBUG(D_retry) debug_printf("Processing retry items\n");

/* Three-times loop to handle succeeded, failed, and deferred addresses.
Deferred addresses must be handled after failed ones, because some may be moved
to the failed chain if they have timed out. */

for (i = 0; i < 3; i++)
  {
  address_item *endaddr, *addr;
  address_item *last_first = NULL;
  address_item **paddr = (i==0)? addr_succeed :
    (i==1)? addr_failed : addr_defer;
  address_item **saved_paddr = NULL;

  DEBUG(D_retry) debug_printf("%s addresses:\n", (i == 0)? "Succeeded" :
    (i == 1)? "Failed" : "Deferred");

  /* Loop for each address on the chain. For deferred addresses, the whole
  address times out unless one of its retry addresses has a retry rule that
  hasn't yet timed out. Deferred addresses should not be requesting deletion
  of retry items, but just in case they do by accident, treat that case
  as "not timed out".

  As well as handling the addresses themselves, we must also process any
  retry items for any parent addresses - these are typically "delete" items,
  because the parent must have succeeded in order to generate the child. */

  while ((endaddr = *paddr) != NULL)
    {
    BOOL timed_out = FALSE;
    retry_item *rti;

    for (addr = endaddr; addr != NULL; addr = addr->parent)
      {
      int update_count = 0;
      int timedout_count = 0;

      DEBUG(D_retry) debug_printf("%s%s\n", addr->address, (addr->retries == NULL)?
        ": no retry items" : "");

      /* Loop for each retry item. */

      for (rti = addr->retries; rti != NULL; rti = rti->next)
        {
        uschar *message;
        int message_length, message_space, failing_interval, next_try;
        retry_rule *rule, *final_rule;
        retry_config *retry;
        dbdata_retry *retry_record;

        /* Open the retry database if it is not already open; failure to open
        the file is logged, but otherwise ignored - deferred addresses will
        get retried at the next opportunity. Not opening earlier than this saves
        opening if no addresses have retry items - common when none have yet
        reached their retry next try time. */

        if (dbm_file == NULL)
          dbm_file = dbfn_open(US"retry", O_RDWR, &dbblock, TRUE);

        if (dbm_file == NULL)
          {
          DEBUG(D_deliver|D_retry|D_hints_lookup)
            debug_printf("retry database not available for updating\n");
          return;
          }

        /* If there are no deferred addresses, that is, if this message is
        completing, and the retry item is for a message-specific SMTP error,
        force it to be deleted, because there's no point in keeping data for
        no-longer-existing messages. This situation can occur when a domain has
        two hosts and a message-specific error occurs for the first of them,
        but the address gets delivered to the second one. This optimization
        doesn't succeed in cleaning out all the dead entries, but it helps. */

        if (*addr_defer == NULL && (rti->flags & rf_message) != 0)
          rti->flags |= rf_delete;

        /* Handle the case of a request to delete the retry info for this
        destination. */

        if ((rti->flags & rf_delete) != 0)
          {
          (void)dbfn_delete(dbm_file, rti->key);
          DEBUG(D_retry)
            debug_printf("deleted retry information for %s\n", rti->key);
          continue;
          }

        /* Count the number of non-delete retry items. This is so that we
        can compare it to the count of timed_out ones, to check whether
        all are timed out. */

        update_count++;

        /* Get the retry information for this destination and error code, if
        any. If this item is for a remote host with ip address, then pass
        the domain name as an alternative to search for. If no retry
        information is found, we can't generate a retry time, so there is
        no point updating the database. This retry item is timed out. */

        if ((retry = retry_find_config(rti->key + 2,
             ((rti->flags & rf_host) != 0)? addr->domain : NULL,
             rti->basic_errno, rti->more_errno)) == NULL)
          {
          DEBUG(D_retry) debug_printf("No configured retry item for %s%s%s\n",
            rti->key,
            ((rti->flags & rf_host) != 0)? US" or " : US"",
            ((rti->flags & rf_host) != 0)? addr->domain : US"");
          if (addr == endaddr) timedout_count++;
          continue;
          }

        DEBUG(D_retry)
          {
          if ((rti->flags & rf_host) != 0)
            debug_printf("retry for %s (%s) = %s %d %d\n", rti->key,
              addr->domain, retry->pattern, retry->basic_errno,
              retry->more_errno);
          else
            debug_printf("retry for %s = %s %d %d\n", rti->key, retry->pattern,
              retry->basic_errno, retry->more_errno);
          }

        /* Set up the message for the database retry record. Because DBM
        records have a maximum data length, we enforce a limit. There isn't
        much point in keeping a huge message here, anyway. */

        message = (rti->basic_errno > 0)? US strerror(rti->basic_errno) :
          (rti->message == NULL)?
          US"unknown error" : string_printing(rti->message);
        message_length = Ustrlen(message);
        if (message_length > 150) message_length = 150;

        /* Read a retry record from the database or construct a new one.
        Ignore an old one if it is too old since it was last updated. */

        retry_record = dbfn_read(dbm_file, rti->key);
        if (retry_record != NULL &&
            now - retry_record->time_stamp > retry_data_expire)
          retry_record = NULL;

        if (retry_record == NULL)
          {
          retry_record = store_get(sizeof(dbdata_retry) + message_length);
          message_space = message_length;
          retry_record->first_failed = now;
          retry_record->last_try = now;
          retry_record->next_try = now;
          retry_record->expired = FALSE;
          retry_record->text[0] = 0;      /* just in case */
          }
        else message_space = Ustrlen(retry_record->text);

        /* Compute how long this destination has been failing */

        failing_interval = now - retry_record->first_failed;
        DEBUG(D_retry) debug_printf("failing_interval=%d message_age=%d\n",
          failing_interval, message_age);

        /* For a non-host error, if the message has been on the queue longer
        than the recorded time of failure, use the message's age instead. This
        can happen when some messages can be delivered and others cannot; a
        successful delivery will reset the first_failed time, and this can lead
        to a failing message being retried too often. */

        if ((rti->flags & rf_host) == 0 && message_age > failing_interval)
          failing_interval = message_age;

        /* Search for the current retry rule. The cutoff time of the
        last rule is handled differently to the others. The rule continues
        to operate for ever (the global maximum interval will eventually
        limit the gaps) but its cutoff time determines when an individual
        destination times out. If there are no retry rules, the destination
        always times out, but we can't compute a retry time. */

        final_rule = NULL;
        for (rule = retry->rules; rule != NULL; rule = rule->next)
          {
          if (failing_interval <= rule->timeout) break;
          final_rule = rule;
          }

        /* If there's an un-timed out rule, the destination has not
        yet timed out, so the address as a whole has not timed out (but we are
        interested in this only for the end address). Make sure the expired
        flag is false (can be forced via fixdb from outside, but ensure it is
        consistent with the rules whenever we go through here). */

        if (rule != NULL)
          {
          retry_record->expired = FALSE;
          }

        /* Otherwise, set the retry timeout expired, and set the final rule
        as the one from which to compute the next retry time. Subsequent
        messages will fail immediately until the retry time is reached (unless
        there are other, still active, retries). */

        else
          {
          rule = final_rule;
          retry_record->expired = TRUE;
          if (addr == endaddr) timedout_count++;
          }

        /* There is a special case to consider when some messages get through
        to a destination and others don't. This can happen locally when a
        large message pushes a user over quota, and it can happen remotely
        when a machine is on a dodgy Internet connection. The messages that
        get through wipe the retry information, causing those that don't to
        stay on the queue longer than the final retry time. In order to
        avoid this, we check, using the time of arrival of the message, to
        see if it has been on the queue for more than the final cutoff time,
        and if so, cause this retry item to time out, and the retry time to
        be set to "now" so that any subsequent messages in the same condition
        also get tried. We search for the last rule onwards from the one that
        is in use. If there are no retry rules for the item, rule will be null
        and timedout_count will already have been updated.

        This implements "timeout this rule if EITHER the host (or routing or
        directing) has been failing for more than the maximum time, OR if the
        message has been on the queue for more than the maximum time."

        February 2006: It is possible that this code is no longer needed
        following the change to the retry calculation to use the message age if
        it is larger than the time since first failure. It may be that the
        expired flag is always set when the other conditions are met. However,
        this is a small bit of code, and it does no harm to leave it in place,
        just in case. */

        if (received_time <= retry_record->first_failed &&
            addr == endaddr && !retry_record->expired && rule != NULL)
          {
          retry_rule *last_rule;
          for (last_rule = rule;
               last_rule->next != NULL;
               last_rule = last_rule->next);
          if (now - received_time > last_rule->timeout)
            {
            DEBUG(D_retry) debug_printf("on queue longer than maximum retry\n");
            timedout_count++;
            rule = NULL;
            }
          }

        /* Compute the next try time from the rule, subject to the global
        maximum, and update the retry database. If rule == NULL it means
        there were no rules at all (and the timeout will be set expired),
        or we have a message that is older than the final timeout. In this
        case set the next retry time to now, so that one delivery attempt
        happens for subsequent messages. */

        if (rule == NULL) next_try = now; else
          {
          if (rule->rule == 'F') next_try = now + rule->p1;
          else  /* rule = 'G' or 'H' */
            {
            int last_predicted_gap =
              retry_record->next_try - retry_record->last_try;
            int last_actual_gap = now - retry_record->last_try;
            int lastgap = (last_predicted_gap < last_actual_gap)?
              last_predicted_gap : last_actual_gap;
            int next_gap = (lastgap * rule->p2)/1000;
            if (rule->rule == 'G')
              {
              next_try = now + ((lastgap < rule->p1)? rule->p1 : next_gap);
              }
            else  /* The 'H' rule */
              {
              next_try = now + rule->p1;
              if (next_gap > rule->p1)
                next_try += random_number(next_gap - rule->p1)/2 +
                  (next_gap - rule->p1)/2;
              }
            }
          }

        /* Impose a global retry max */

        if (next_try - now > retry_interval_max)
          next_try = now + retry_interval_max;

        /* If the new message length is greater than the previous one, we
        have to copy the record first. */

        if (message_length > message_space)
          {
          dbdata_retry *newr = store_get(sizeof(dbdata_retry) + message_length);
          memcpy(newr, retry_record, sizeof(dbdata_retry));
          retry_record = newr;
          }

        /* Set up the retry record; message_length may be less than the string
        length for very long error strings. */

        retry_record->last_try = now;
        retry_record->next_try = next_try;
        retry_record->basic_errno = rti->basic_errno;
        retry_record->more_errno = rti->more_errno;
        Ustrncpy(retry_record->text, message, message_length);
        retry_record->text[message_length] = 0;

        DEBUG(D_retry)
          {
          int letter = retry_record->more_errno & 255;
          debug_printf("Writing retry data for %s\n", rti->key);
          debug_printf("  first failed=%d last try=%d next try=%d expired=%d\n",
            (int)retry_record->first_failed, (int)retry_record->last_try,
            (int)retry_record->next_try, retry_record->expired);
          debug_printf("  errno=%d more_errno=", retry_record->basic_errno);
          if (letter == 'A' || letter == 'M')
            debug_printf("%d,%c", (retry_record->more_errno >> 8) & 255,
              letter);
          else
            debug_printf("%d", retry_record->more_errno);
          debug_printf(" %s\n", retry_record->text);
          }

        (void)dbfn_write(dbm_file, rti->key, retry_record,
          sizeof(dbdata_retry) + message_length);
        }                            /* Loop for each retry item */

      /* If all the non-delete retry items are timed out, the address is
      timed out, provided that we didn't skip any hosts because their retry
      time was not reached (or because of hosts_max_try). */

      if (update_count > 0 && update_count == timedout_count)
        {
        if (!testflag(endaddr, af_retry_skipped))
          {
          DEBUG(D_retry) debug_printf("timed out: all retries expired\n");
          timed_out = TRUE;
          }
        else
          {
          DEBUG(D_retry)
            debug_printf("timed out but some hosts were skipped\n");
          }
        }
      }     /* Loop for an address and its parents */

    /* If this is a deferred address, and retry processing was requested by
    means of one or more retry items, and they all timed out, move the address
    to the failed queue, and restart this loop without updating paddr.

    If there were several addresses batched in the same remote delivery, only
    the original top one will have host retry items attached to it, but we want
    to handle all the same. Each will have a pointer back to its "top" address,
    and they will now precede the item with the retries because addresses are
    inverted when added to these final queues. We have saved information about
    them in passing (below) so they can all be cut out at once. */

    if (i == 2)   /* Handling defers */
      {
      if (endaddr->retries != NULL && timed_out)
        {
        if (last_first == endaddr) paddr = saved_paddr;
        addr = *paddr;
        *paddr = endaddr->next;

        endaddr->next = *addr_failed;
        *addr_failed = addr;

        for (;; addr = addr->next)
          {
          setflag(addr, af_retry_timedout);
          addr->message = (addr->message == NULL)? US"retry timeout exceeded" :
            string_sprintf("%s: retry timeout exceeded", addr->message);
          addr->user_message = (addr->user_message == NULL)?
            US"retry timeout exceeded" :
            string_sprintf("%s: retry timeout exceeded", addr->user_message);
          log_write(0, LOG_MAIN, "** %s%s%s%s: retry timeout exceeded",
            addr->address,
           (addr->parent == NULL)? US"" : US" <",
           (addr->parent == NULL)? US"" : addr->parent->address,
           (addr->parent == NULL)? US"" : US">");

          if (addr == endaddr) break;
          }

        continue;                       /* Restart from changed *paddr */
        }

      /* This address is to remain on the defer chain. If it has a "first"
      pointer, save the pointer to it in case we want to fail the set of
      addresses when we get to the first one. */

      if (endaddr->first != last_first)
        {
        last_first = endaddr->first;
        saved_paddr = paddr;
        }
      }

    /* All cases (succeed, fail, defer left on queue) */

    paddr = &(endaddr->next);         /* Advance to next address */
    }                                 /* Loop for all addresses  */
  }                                   /* Loop for succeed, fail, defer */

/* Close and unlock the database */

if (dbm_file != NULL) dbfn_close(dbm_file);

DEBUG(D_retry) debug_printf("end of retry processing\n");
}

/* End of retry.c */