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

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

/* Copyright (c) University of Cambridge 1995 - 2016 */
/* See the file NOTICE for conditions of use and distribution. */

/* The main code for delivering a message. */


#include "exim.h"
#include <assert.h>


/* Data block for keeping track of subprocesses for parallel remote
delivery. */

typedef struct pardata {
  address_item *addrlist;      /* chain of addresses */
  address_item *addr;          /* next address data expected for */
  pid_t pid;                   /* subprocess pid */
  int fd;                      /* pipe fd for getting result from subprocess */
  int transport_count;         /* returned transport count value */
  BOOL done;                   /* no more data needed */
  uschar *msg;                 /* error message */
  uschar *return_path;         /* return_path for these addresses */
} pardata;

/* Values for the process_recipients variable */

enum { RECIP_ACCEPT, RECIP_IGNORE, RECIP_DEFER,
       RECIP_FAIL, RECIP_FAIL_FILTER, RECIP_FAIL_TIMEOUT,
       RECIP_FAIL_LOOP};

/* Mutually recursive functions for marking addresses done. */

static void child_done(address_item *, uschar *);
static void address_done(address_item *, uschar *);

/* Table for turning base-62 numbers into binary */

static uschar tab62[] =
          {0,1,2,3,4,5,6,7,8,9,0,0,0,0,0,0,     /* 0-9 */
           0,10,11,12,13,14,15,16,17,18,19,20,  /* A-K */
          21,22,23,24,25,26,27,28,29,30,31,32,  /* L-W */
          33,34,35, 0, 0, 0, 0, 0,              /* X-Z */
           0,36,37,38,39,40,41,42,43,44,45,46,  /* a-k */
          47,48,49,50,51,52,53,54,55,56,57,58,  /* l-w */
          59,60,61};                            /* x-z */


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

/* addr_duplicate is global because it needs to be seen from the Envelope-To
writing code. */

static address_item *addr_defer = NULL;
static address_item *addr_failed = NULL;
static address_item *addr_fallback = NULL;
static address_item *addr_local = NULL;
static address_item *addr_new = NULL;
static address_item *addr_remote = NULL;
static address_item *addr_route = NULL;
static address_item *addr_succeed = NULL;
static address_item *addr_dsntmp = NULL;
static address_item *addr_senddsn = NULL;

static FILE *message_log = NULL;
static BOOL update_spool;
static BOOL remove_journal;
static int  parcount = 0;
static pardata *parlist = NULL;
static int  return_count;
static uschar *frozen_info = US"";
static uschar *used_return_path = NULL;



/*************************************************
*             Make a new address item            *
*************************************************/

/* This function gets the store and initializes with default values. The
transport_return value defaults to DEFER, so that any unexpected failure to
deliver does not wipe out the message. The default unique string is set to a
copy of the address, so that its domain can be lowercased.

Argument:
  address     the RFC822 address string
  copy        force a copy of the address

Returns:      a pointer to an initialized address_item
*/

address_item *
deliver_make_addr(uschar *address, BOOL copy)
{
address_item *addr = store_get(sizeof(address_item));
*addr = address_defaults;
if (copy) address = string_copy(address);
addr->address = address;
addr->unique = string_copy(address);
return addr;
}




/*************************************************
*     Set expansion values for an address        *
*************************************************/

/* Certain expansion variables are valid only when handling an address or
address list. This function sets them up or clears the values, according to its
argument.

Arguments:
  addr          the address in question, or NULL to clear values
Returns:        nothing
*/

void
deliver_set_expansions(address_item *addr)
{
if (!addr)
  {
  const uschar ***p = address_expansions;
  while (*p) **p++ = NULL;
  return;
  }

/* Exactly what gets set depends on whether there is one or more addresses, and
what they contain. These first ones are always set, taking their values from
the first address. */

if (!addr->host_list)
  {
  deliver_host = deliver_host_address = US"";
  deliver_host_port = 0;
  }
else
  {
  deliver_host = addr->host_list->name;
  deliver_host_address = addr->host_list->address;
  deliver_host_port = addr->host_list->port;
  }

deliver_recipients = addr;
deliver_address_data = addr->prop.address_data;
deliver_domain_data = addr->prop.domain_data;
deliver_localpart_data = addr->prop.localpart_data;

/* These may be unset for multiple addresses */

deliver_domain = addr->domain;
self_hostname = addr->self_hostname;

#ifdef EXPERIMENTAL_BRIGHTMAIL
bmi_deliver = 1;    /* deliver by default */
bmi_alt_location = NULL;
bmi_base64_verdict = NULL;
bmi_base64_tracker_verdict = NULL;
#endif

/* If there's only one address we can set everything. */

if (!addr->next)
  {
  address_item *addr_orig;

  deliver_localpart = addr->local_part;
  deliver_localpart_prefix = addr->prefix;
  deliver_localpart_suffix = addr->suffix;

  for (addr_orig = addr; addr_orig->parent; addr_orig = addr_orig->parent) ;
  deliver_domain_orig = addr_orig->domain;

  /* Re-instate any prefix and suffix in the original local part. In all
  normal cases, the address will have a router associated with it, and we can
  choose the caseful or caseless version accordingly. However, when a system
  filter sets up a pipe, file, or autoreply delivery, no router is involved.
  In this case, though, there won't be any prefix or suffix to worry about. */

  deliver_localpart_orig = !addr_orig->router
    ? addr_orig->local_part
    : addr_orig->router->caseful_local_part
    ? addr_orig->cc_local_part
    : addr_orig->lc_local_part;

  /* If there's a parent, make its domain and local part available, and if
  delivering to a pipe or file, or sending an autoreply, get the local
  part from the parent. For pipes and files, put the pipe or file string
  into address_pipe and address_file. */

  if (addr->parent)
    {
    deliver_domain_parent = addr->parent->domain;
    deliver_localpart_parent = !addr->parent->router
      ? addr->parent->local_part
      : addr->parent->router->caseful_local_part
      ? addr->parent->cc_local_part
      : addr->parent->lc_local_part;

    /* File deliveries have their own flag because they need to be picked out
    as special more often. */

    if (testflag(addr, af_pfr))
      {
      if (testflag(addr, af_file))          address_file = addr->local_part;
      else if (deliver_localpart[0] == '|') address_pipe = addr->local_part;
      deliver_localpart = addr->parent->local_part;
      deliver_localpart_prefix = addr->parent->prefix;
      deliver_localpart_suffix = addr->parent->suffix;
      }
    }

#ifdef EXPERIMENTAL_BRIGHTMAIL
    /* Set expansion variables related to Brightmail AntiSpam */
    bmi_base64_verdict = bmi_get_base64_verdict(deliver_localpart_orig, deliver_domain_orig);
    bmi_base64_tracker_verdict = bmi_get_base64_tracker_verdict(bmi_base64_verdict);
    /* get message delivery status (0 - don't deliver | 1 - deliver) */
    bmi_deliver = bmi_get_delivery_status(bmi_base64_verdict);
    /* if message is to be delivered, get eventual alternate location */
    if (bmi_deliver == 1)
      bmi_alt_location = bmi_get_alt_location(bmi_base64_verdict);
#endif

  }

/* For multiple addresses, don't set local part, and leave the domain and
self_hostname set only if it is the same for all of them. It is possible to
have multiple pipe and file addresses, but only when all addresses have routed
to the same pipe or file. */

else
  {
  address_item *addr2;
  if (testflag(addr, af_pfr))
    {
    if (testflag(addr, af_file))         address_file = addr->local_part;
    else if (addr->local_part[0] == '|') address_pipe = addr->local_part;
    }
  for (addr2 = addr->next; addr2; addr2 = addr2->next)
    {
    if (deliver_domain && Ustrcmp(deliver_domain, addr2->domain) != 0)
      deliver_domain = NULL;
    if (  self_hostname
       && (  !addr2->self_hostname
          || Ustrcmp(self_hostname, addr2->self_hostname) != 0
       )  )
      self_hostname = NULL;
    if (!deliver_domain && !self_hostname) break;
    }
  }
}




/*************************************************
*                Open a msglog file              *
*************************************************/

/* This function is used both for normal message logs, and for files in the
msglog directory that are used to catch output from pipes. Try to create the
directory if it does not exist. From release 4.21, normal message logs should
be created when the message is received.

Argument:
  filename  the file name
  mode      the mode required
  error     used for saying what failed

Returns:    a file descriptor, or -1 (with errno set)
*/

static int
open_msglog_file(uschar *filename, int mode, uschar **error)
{
int fd = Uopen(filename, O_WRONLY|O_APPEND|O_CREAT, mode);

if (fd < 0 && errno == ENOENT)
  {
  (void)directory_make(spool_directory,
                        spool_sname(US"msglog", message_subdir),
                        MSGLOG_DIRECTORY_MODE, TRUE);
  fd = Uopen(filename, O_WRONLY|O_APPEND|O_CREAT, mode);
  }

/* Set the close-on-exec flag and change the owner to the exim uid/gid (this
function is called as root). Double check the mode, because the group setting
doesn't always get set automatically. */

if (fd >= 0)
  {
  (void)fcntl(fd, F_SETFD, fcntl(fd, F_GETFD) | FD_CLOEXEC);
  if (fchown(fd, exim_uid, exim_gid) < 0)
    {
    *error = US"chown";
    return -1;
    }
  if (fchmod(fd, mode) < 0)
    {
    *error = US"chmod";
    return -1;
    }
  }
else *error = US"create";

return fd;
}




/*************************************************
*           Write to msglog if required          *
*************************************************/

/* Write to the message log, if configured. This function may also be called
from transports.

Arguments:
  format       a string format

Returns:       nothing
*/

void
deliver_msglog(const char *format, ...)
{
va_list ap;
if (!message_logs) return;
va_start(ap, format);
vfprintf(message_log, format, ap);
fflush(message_log);
va_end(ap);
}




/*************************************************
*            Replicate status for batch          *
*************************************************/

/* When a transport handles a batch of addresses, it may treat them
individually, or it may just put the status in the first one, and return FALSE,
requesting that the status be copied to all the others externally. This is the
replication function. As well as the status, it copies the transport pointer,
which may have changed if appendfile passed the addresses on to a different
transport.

Argument:    pointer to the first address in a chain
Returns:     nothing
*/

static void
replicate_status(address_item *addr)
{
address_item *addr2;
for (addr2 = addr->next; addr2; addr2 = addr2->next)
  {
  addr2->transport =        addr->transport;
  addr2->transport_return = addr->transport_return;
  addr2->basic_errno =      addr->basic_errno;
  addr2->more_errno =       addr->more_errno;
  addr2->special_action =   addr->special_action;
  addr2->message =          addr->message;
  addr2->user_message =     addr->user_message;
  }
}



/*************************************************
*              Compare lists of hosts            *
*************************************************/

/* This function is given two pointers to chains of host items, and it yields
TRUE if the lists refer to the same hosts in the same order, except that

(1) Multiple hosts with the same non-negative MX values are permitted to appear
    in different orders. Round-robinning nameservers can cause this to happen.

(2) Multiple hosts with the same negative MX values less than MX_NONE are also
    permitted to appear in different orders. This is caused by randomizing
    hosts lists.

This enables Exim to use a single SMTP transaction for sending to two entirely
different domains that happen to end up pointing at the same hosts.

Arguments:
  one       points to the first host list
  two       points to the second host list

Returns:    TRUE if the lists refer to the same host set
*/

static BOOL
same_hosts(host_item *one, host_item *two)
{
while (one && two)
  {
  if (Ustrcmp(one->name, two->name) != 0)
    {
    int mx = one->mx;
    host_item *end_one = one;
    host_item *end_two = two;

    /* Batch up only if there was no MX and the list was not randomized */

    if (mx == MX_NONE) return FALSE;

    /* Find the ends of the shortest sequence of identical MX values */

    while (  end_one->next && end_one->next->mx == mx
          && end_two->next && end_two->next->mx == mx)
      {
      end_one = end_one->next;
      end_two = end_two->next;
      }

    /* If there aren't any duplicates, there's no match. */

    if (end_one == one) return FALSE;

    /* For each host in the 'one' sequence, check that it appears in the 'two'
    sequence, returning FALSE if not. */

    for (;;)
      {
      host_item *hi;
      for (hi = two; hi != end_two->next; hi = hi->next)
        if (Ustrcmp(one->name, hi->name) == 0) break;
      if (hi == end_two->next) return FALSE;
      if (one == end_one) break;
      one = one->next;
      }

    /* All the hosts in the 'one' sequence were found in the 'two' sequence.
    Ensure both are pointing at the last host, and carry on as for equality. */

    two = end_two;
    }

  /* Hosts matched */

  one = one->next;
  two = two->next;
  }

/* True if both are NULL */

return (one == two);
}



/*************************************************
*              Compare header lines              *
*************************************************/

/* This function is given two pointers to chains of header items, and it yields
TRUE if they are the same header texts in the same order.

Arguments:
  one       points to the first header list
  two       points to the second header list

Returns:    TRUE if the lists refer to the same header set
*/

static BOOL
same_headers(header_line *one, header_line *two)
{
for (;; one = one->next, two = two->next)
  {
  if (one == two) return TRUE;   /* Includes the case where both NULL */
  if (!one || !two) return FALSE;
  if (Ustrcmp(one->text, two->text) != 0) return FALSE;
  }
}



/*************************************************
*            Compare string settings             *
*************************************************/

/* This function is given two pointers to strings, and it returns
TRUE if they are the same pointer, or if the two strings are the same.

Arguments:
  one       points to the first string
  two       points to the second string

Returns:    TRUE or FALSE
*/

static BOOL
same_strings(uschar *one, uschar *two)
{
if (one == two) return TRUE;   /* Includes the case where both NULL */
if (!one || !two) return FALSE;
return (Ustrcmp(one, two) == 0);
}



/*************************************************
*        Compare uid/gid for addresses           *
*************************************************/

/* This function is given a transport and two addresses. It yields TRUE if the
uid/gid/initgroups settings for the two addresses are going to be the same when
they are delivered.

Arguments:
  tp            the transort
  addr1         the first address
  addr2         the second address

Returns:        TRUE or FALSE
*/

static BOOL
same_ugid(transport_instance *tp, address_item *addr1, address_item *addr2)
{
if (  !tp->uid_set && !tp->expand_uid
   && !tp->deliver_as_creator
   && (  testflag(addr1, af_uid_set) != testflag(addr2, af_gid_set)
      || (  testflag(addr1, af_uid_set)
         && (  addr1->uid != addr2->uid
            || testflag(addr1, af_initgroups) != testflag(addr2, af_initgroups)
   )  )  )  )
  return FALSE;

if (  !tp->gid_set && !tp->expand_gid
   && (  testflag(addr1, af_gid_set) != testflag(addr2, af_gid_set)
      || (  testflag(addr1, af_gid_set)
         && addr1->gid != addr2->gid
   )  )  )
  return FALSE;

return TRUE;
}




/*************************************************
*      Record that an address is complete        *
*************************************************/

/* This function records that an address is complete. This is straightforward
for most addresses, where the unique address is just the full address with the
domain lower cased. For homonyms (addresses that are the same as one of their
ancestors) their are complications. Their unique addresses have \x\ prepended
(where x = 0, 1, 2...), so that de-duplication works correctly for siblings and
cousins.

Exim used to record the unique addresses of homonyms as "complete". This,
however, fails when the pattern of redirection varies over time (e.g. if taking
unseen copies at only some times of day) because the prepended numbers may vary
from one delivery run to the next. This problem is solved by never recording
prepended unique addresses as complete. Instead, when a homonymic address has
actually been delivered via a transport, we record its basic unique address
followed by the name of the transport. This is checked in subsequent delivery
runs whenever an address is routed to a transport.

If the completed address is a top-level one (has no parent, which means it
cannot be homonymic) we also add the original address to the non-recipients
tree, so that it gets recorded in the spool file and therefore appears as
"done" in any spool listings. The original address may differ from the unique
address in the case of the domain.

Finally, this function scans the list of duplicates, marks as done any that
match this address, and calls child_done() for their ancestors.

Arguments:
  addr        address item that has been completed
  now         current time as a string

Returns:      nothing
*/

static void
address_done(address_item *addr, uschar *now)
{
address_item *dup;

update_spool = TRUE;        /* Ensure spool gets updated */

/* Top-level address */

if (!addr->parent)
  {
  tree_add_nonrecipient(addr->unique);
  tree_add_nonrecipient(addr->address);
  }

/* Homonymous child address */

else if (testflag(addr, af_homonym))
  {
  if (addr->transport)
    tree_add_nonrecipient(
      string_sprintf("%s/%s", addr->unique + 3, addr->transport->name));
  }

/* Non-homonymous child address */

else tree_add_nonrecipient(addr->unique);

/* Check the list of duplicate addresses and ensure they are now marked
done as well. */

for (dup = addr_duplicate; dup; dup = dup->next)
  if (Ustrcmp(addr->unique, dup->unique) == 0)
    {
    tree_add_nonrecipient(dup->unique);
    child_done(dup, now);
    }
}




/*************************************************
*      Decrease counts in parents and mark done  *
*************************************************/

/* This function is called when an address is complete. If there is a parent
address, its count of children is decremented. If there are still other
children outstanding, the function exits. Otherwise, if the count has become
zero, address_done() is called to mark the parent and its duplicates complete.
Then loop for any earlier ancestors.

Arguments:
  addr      points to the completed address item
  now       the current time as a string, for writing to the message log

Returns:    nothing
*/

static void
child_done(address_item *addr, uschar *now)
{
address_item *aa;
while (addr->parent)
  {
  addr = addr->parent;
  if ((addr->child_count -= 1) > 0) return;   /* Incomplete parent */
  address_done(addr, now);

  /* Log the completion of all descendents only when there is no ancestor with
  the same original address. */

  for (aa = addr->parent; aa; aa = aa->parent)
    if (Ustrcmp(aa->address, addr->address) == 0) break;
  if (aa) continue;

  deliver_msglog("%s %s: children all complete\n", now, addr->address);
  DEBUG(D_deliver) debug_printf("%s: children all complete\n", addr->address);
  }
}



/*************************************************
*      Delivery logging support functions        *
*************************************************/

/* The LOGGING() checks in d_log_interface() are complicated for backwards
compatibility. When outgoing interface logging was originally added, it was
conditional on just incoming_interface (which is off by default). The
outgoing_interface option is on by default to preserve this behaviour, but
you can enable incoming_interface and disable outgoing_interface to get I=
fields on incoming lines only.

Arguments:
  s         The log line buffer
  sizep     Pointer to the buffer size
  ptrp      Pointer to current index into buffer
  addr      The address to be logged

Returns:    New value for s
*/

static uschar *
d_log_interface(uschar *s, int *sizep, int *ptrp)
{
if (LOGGING(incoming_interface) && LOGGING(outgoing_interface)
    && sending_ip_address)
  {
  s = string_append(s, sizep, ptrp, 2, US" I=[", sending_ip_address);
  s = LOGGING(outgoing_port)
    ? string_append(s, sizep, ptrp, 2, US"]:",
        string_sprintf("%d", sending_port))
    : string_catn(s, sizep, ptrp, US"]", 1);
  }
return s;
}



static uschar *
d_hostlog(uschar * s, int * sp, int * pp, address_item * addr)
{
host_item * h = addr->host_used;

s = string_append(s, sp, pp, 2, US" H=", h->name);

if (LOGGING(dnssec) && h->dnssec == DS_YES)
  s = string_cat(s, sp, pp, US" DS");

s = string_append(s, sp, pp, 3, US" [", h->address, US"]");

if (LOGGING(outgoing_port))
  s = string_append(s, sp, pp, 2, US":", string_sprintf("%d", h->port));

#ifdef SUPPORT_SOCKS
if (LOGGING(proxy) && proxy_local_address)
  {
  s = string_append(s, sp, pp, 3, US" PRX=[", proxy_local_address, US"]");
  if (LOGGING(outgoing_port))
    s = string_append(s, sp, pp, 2, US":", string_sprintf("%d",
      proxy_local_port));
  }
#endif

return d_log_interface(s, sp, pp);
}





#ifdef SUPPORT_TLS
static uschar *
d_tlslog(uschar * s, int * sizep, int * ptrp, address_item * addr)
{
if (LOGGING(tls_cipher) && addr->cipher)
  s = string_append(s, sizep, ptrp, 2, US" X=", addr->cipher);
if (LOGGING(tls_certificate_verified) && addr->cipher)
  s = string_append(s, sizep, ptrp, 2, US" CV=",
    testflag(addr, af_cert_verified)
    ?
#ifdef EXPERIMENTAL_DANE
      testflag(addr, af_dane_verified)
    ? "dane"
    :
#endif
      "yes"
    : "no");
if (LOGGING(tls_peerdn) && addr->peerdn)
  s = string_append(s, sizep, ptrp, 3, US" DN=\"",
    string_printing(addr->peerdn), US"\"");
return s;
}
#endif




#ifndef DISABLE_EVENT
uschar *
event_raise(uschar * action, const uschar * event, uschar * ev_data)
{
uschar * s;
if (action)
  {
  DEBUG(D_deliver)
    debug_printf("Event(%s): event_action=|%s| delivery_IP=%s\n",
      event,
      action, deliver_host_address);

  event_name = event;
  event_data = ev_data;

  if (!(s = expand_string(action)) && *expand_string_message)
    log_write(0, LOG_MAIN|LOG_PANIC,
      "failed to expand event_action %s in %s: %s\n",
      event, transport_name, expand_string_message);

  event_name = event_data = NULL;

  /* If the expansion returns anything but an empty string, flag for
  the caller to modify his normal processing
  */
  if (s && *s)
    {
    DEBUG(D_deliver)
      debug_printf("Event(%s): event_action returned \"%s\"\n", event, s);
    return s;
    }
  }
return NULL;
}

void
msg_event_raise(const uschar * event, const address_item * addr)
{
const uschar * save_domain = deliver_domain;
uschar * save_local =  deliver_localpart;
const uschar * save_host = deliver_host;
const uschar * save_address = deliver_host_address;
const int      save_port =   deliver_host_port;

if (!addr->transport)
  return;

router_name =    addr->router ? addr->router->name : NULL;
transport_name = addr->transport->name;
deliver_domain = addr->domain;
deliver_localpart = addr->local_part;
deliver_host =   addr->host_used ? addr->host_used->name : NULL;

(void) event_raise(addr->transport->event_action, event,
          addr->host_used
          || Ustrcmp(addr->transport->driver_name, "smtp") == 0
          || Ustrcmp(addr->transport->driver_name, "lmtp") == 0
         ? addr->message : NULL); 

deliver_host_port =    save_port;
deliver_host_address = save_address;
deliver_host =      save_host;
deliver_localpart = save_local;
deliver_domain =    save_domain;
router_name = transport_name = NULL;
}
#endif  /*DISABLE_EVENT*/



/* If msg is NULL this is a delivery log and logchar is used. Otherwise
this is a nonstandard call; no two-character delivery flag is written
but sender-host and sender are prefixed and "msg" is inserted in the log line.

Arguments:
  flags         passed to log_write()
*/
void
delivery_log(int flags, address_item * addr, int logchar, uschar * msg)
{
uschar *log_address;
int size = 256;         /* Used for a temporary, */
int ptr = 0;            /* expanding buffer, for */
uschar *s;              /* building log lines;   */
void *reset_point;      /* released afterwards.  */

/* Log the delivery on the main log. We use an extensible string to build up
the log line, and reset the store afterwards. Remote deliveries should always
have a pointer to the host item that succeeded; local deliveries can have a
pointer to a single host item in their host list, for use by the transport. */

#ifndef DISABLE_EVENT
  /* presume no successful remote delivery */
  lookup_dnssec_authenticated = NULL;
#endif

s = reset_point = store_get(size);

log_address = string_log_address(addr, LOGGING(all_parents), TRUE);
if (msg)
  s = string_append(s, &size, &ptr, 3, host_and_ident(TRUE), US" ", log_address);
else
  {
  s[ptr++] = logchar;
  s = string_append(s, &size, &ptr, 2, US"> ", log_address);
  }

if (LOGGING(sender_on_delivery) || msg)
  s = string_append(s, &size, &ptr, 3, US" F=<",
#ifdef SUPPORT_I18N
    testflag(addr, af_utf8_downcvt)
    ? string_address_utf8_to_alabel(sender_address, NULL)
    :
#endif
      sender_address,
  US">");

if (*queue_name)
  s = string_append(s, &size, &ptr, 2, US" Q=", queue_name);

#ifdef EXPERIMENTAL_SRS
if(addr->prop.srs_sender)
  s = string_append(s, &size, &ptr, 3, US" SRS=<", addr->prop.srs_sender, US">");
#endif

/* You might think that the return path must always be set for a successful
delivery; indeed, I did for some time, until this statement crashed. The case
when it is not set is for a delivery to /dev/null which is optimised by not
being run at all. */

if (used_return_path && LOGGING(return_path_on_delivery))
  s = string_append(s, &size, &ptr, 3, US" P=<", used_return_path, US">");

if (msg)
  s = string_append(s, &size, &ptr, 2, US" ", msg);

/* For a delivery from a system filter, there may not be a router */
if (addr->router)
  s = string_append(s, &size, &ptr, 2, US" R=", addr->router->name);

s = string_append(s, &size, &ptr, 2, US" T=", addr->transport->name);

if (LOGGING(delivery_size))
  s = string_append(s, &size, &ptr, 2, US" S=",
    string_sprintf("%d", transport_count));

/* Local delivery */

if (addr->transport->info->local)
  {
  if (addr->host_list)
    s = string_append(s, &size, &ptr, 2, US" H=", addr->host_list->name);
  s = d_log_interface(s, &size, &ptr);
  if (addr->shadow_message)
    s = string_cat(s, &size, &ptr, addr->shadow_message);
  }

/* Remote delivery */

else
  {
  if (addr->host_used)
    {
    s = d_hostlog(s, &size, &ptr, addr);
    if (continue_sequence > 1)
      s = string_catn(s, &size, &ptr, US"*", 1);

#ifndef DISABLE_EVENT
    deliver_host_address = addr->host_used->address;
    deliver_host_port =    addr->host_used->port;
    deliver_host =         addr->host_used->name;

    /* DNS lookup status */
    lookup_dnssec_authenticated = addr->host_used->dnssec==DS_YES ? US"yes"
                              : addr->host_used->dnssec==DS_NO ? US"no"
                              : NULL;
#endif
    }

#ifdef SUPPORT_TLS
  s = d_tlslog(s, &size, &ptr, addr);
#endif

  if (addr->authenticator)
    {
    s = string_append(s, &size, &ptr, 2, US" A=", addr->authenticator);
    if (addr->auth_id)
      {
      s = string_append(s, &size, &ptr, 2, US":", addr->auth_id);
      if (LOGGING(smtp_mailauth) && addr->auth_sndr)
        s = string_append(s, &size, &ptr, 2, US":", addr->auth_sndr);
      }
    }

#ifndef DISABLE_PRDR
  if (addr->flags & af_prdr_used)
    s = string_append(s, &size, &ptr, 1, US" PRDR");
#endif
  }

/* confirmation message (SMTP (host_used) and LMTP (driver_name)) */

if (  LOGGING(smtp_confirmation)
   && addr->message
   && (addr->host_used || Ustrcmp(addr->transport->driver_name, "lmtp") == 0)
   )
  {
  unsigned i;
  unsigned lim = big_buffer_size < 1024 ? big_buffer_size : 1024;
  uschar *p = big_buffer;
  uschar *ss = addr->message;
  *p++ = '\"';
  for (i = 0; i < lim && ss[i] != 0; i++)       /* limit logged amount */
    {
    if (ss[i] == '\"' || ss[i] == '\\') *p++ = '\\'; /* quote \ and " */
    *p++ = ss[i];
    }
  *p++ = '\"';
  *p = 0;
  s = string_append(s, &size, &ptr, 2, US" C=", big_buffer);
  }

/* Time on queue and actual time taken to deliver */

if (LOGGING(queue_time))
  s = string_append(s, &size, &ptr, 2, US" QT=",
    readconf_printtime( (int) ((long)time(NULL) - (long)received_time)) );

if (LOGGING(deliver_time))
  s = string_append(s, &size, &ptr, 2, US" DT=",
    readconf_printtime(addr->more_errno));

/* string_cat() always leaves room for the terminator. Release the
store we used to build the line after writing it. */

s[ptr] = 0;
log_write(0, flags, "%s", s);

#ifndef DISABLE_EVENT
if (!msg) msg_event_raise(US"msg:delivery", addr);
#endif

store_reset(reset_point);
return;
}



/*************************************************
*    Actions at the end of handling an address   *
*************************************************/

/* This is a function for processing a single address when all that can be done
with it has been done.

Arguments:
  addr         points to the address block
  result       the result of the delivery attempt
  logflags     flags for log_write() (LOG_MAIN and/or LOG_PANIC)
  driver_type  indicates which type of driver (transport, or router) was last
                 to process the address
  logchar      '=' or '-' for use when logging deliveries with => or ->

Returns:       nothing
*/

static void
post_process_one(address_item *addr, int result, int logflags, int driver_type,
  int logchar)
{
uschar *now = tod_stamp(tod_log);
uschar *driver_kind = NULL;
uschar *driver_name = NULL;
uschar *log_address;

int size = 256;         /* Used for a temporary, */
int ptr = 0;            /* expanding buffer, for */
uschar *s;              /* building log lines;   */
void *reset_point;      /* released afterwards.  */

DEBUG(D_deliver) debug_printf("post-process %s (%d)\n", addr->address, result);

/* Set up driver kind and name for logging. Disable logging if the router or
transport has disabled it. */

if (driver_type == DTYPE_TRANSPORT)
  {
  if (addr->transport)
    {
    driver_name = addr->transport->name;
    driver_kind = US" transport";
    disable_logging = addr->transport->disable_logging;
    }
  else driver_kind = US"transporting";
  }
else if (driver_type == DTYPE_ROUTER)
  {
  if (addr->router)
    {
    driver_name = addr->router->name;
    driver_kind = US" router";
    disable_logging = addr->router->disable_logging;
    }
  else driver_kind = US"routing";
  }

/* If there's an error message set, ensure that it contains only printing
characters - it should, but occasionally things slip in and this at least
stops the log format from getting wrecked. We also scan the message for an LDAP
expansion item that has a password setting, and flatten the password. This is a
fudge, but I don't know a cleaner way of doing this. (If the item is badly
malformed, it won't ever have gone near LDAP.) */

if (addr->message)
  {
  const uschar * s = string_printing(addr->message);

  /* deconst cast ok as string_printing known to have alloc'n'copied */
  addr->message = expand_hide_passwords(US s);
  }

/* If we used a transport that has one of the "return_output" options set, and
if it did in fact generate some output, then for return_output we treat the
message as failed if it was not already set that way, so that the output gets
returned to the sender, provided there is a sender to send it to. For
return_fail_output, do this only if the delivery failed. Otherwise we just
unlink the file, and remove the name so that if the delivery failed, we don't
try to send back an empty or unwanted file. The log_output options operate only
on a non-empty file.

In any case, we close the message file, because we cannot afford to leave a
file-descriptor for one address while processing (maybe very many) others. */

if (addr->return_file >= 0 && addr->return_filename)
  {
  BOOL return_output = FALSE;
  struct stat statbuf;
  (void)EXIMfsync(addr->return_file);

  /* If there is no output, do nothing. */

  if (fstat(addr->return_file, &statbuf) == 0 && statbuf.st_size > 0)
    {
    transport_instance *tb = addr->transport;

    /* Handle logging options */

    if (  tb->log_output
       || result == FAIL  && tb->log_fail_output
       || result == DEFER && tb->log_defer_output
       )
      {
      uschar *s;
      FILE *f = Ufopen(addr->return_filename, "rb");
      if (!f)
        log_write(0, LOG_MAIN|LOG_PANIC, "failed to open %s to log output "
          "from %s transport: %s", addr->return_filename, tb->name,
          strerror(errno));
      else
        if ((s = US Ufgets(big_buffer, big_buffer_size, f)))
          {
          uschar *p = big_buffer + Ustrlen(big_buffer);
          const uschar * sp;
          while (p > big_buffer && isspace(p[-1])) p--;
          *p = 0;
          sp = string_printing(big_buffer);
          log_write(0, LOG_MAIN, "<%s>: %s transport output: %s",
            addr->address, tb->name, sp);
          }
        (void)fclose(f);
      }

    /* Handle returning options, but only if there is an address to return
    the text to. */

    if (sender_address[0] != 0 || addr->prop.errors_address)
      if (tb->return_output)
        {
        addr->transport_return = result = FAIL;
        if (addr->basic_errno == 0 && !addr->message)
          addr->message = US"return message generated";
        return_output = TRUE;
        }
      else
        if (tb->return_fail_output && result == FAIL) return_output = TRUE;
    }

  /* Get rid of the file unless it might be returned, but close it in
  all cases. */

  if (!return_output)
    {
    Uunlink(addr->return_filename);
    addr->return_filename = NULL;
    addr->return_file = -1;
    }

  (void)close(addr->return_file);
  }

/* The success case happens only after delivery by a transport. */

if (result == OK)
  {
  addr->next = addr_succeed;
  addr_succeed = addr;

  /* Call address_done() to ensure that we don't deliver to this address again,
  and write appropriate things to the message log. If it is a child address, we
  call child_done() to scan the ancestors and mark them complete if this is the
  last child to complete. */

  address_done(addr, now);
  DEBUG(D_deliver) debug_printf("%s delivered\n", addr->address);

  if (!addr->parent)
    deliver_msglog("%s %s: %s%s succeeded\n", now, addr->address,
      driver_name, driver_kind);
  else
    {
    deliver_msglog("%s %s <%s>: %s%s succeeded\n", now, addr->address,
      addr->parent->address, driver_name, driver_kind);
    child_done(addr, now);
    }

  /* Certificates for logging (via events) */
#ifdef SUPPORT_TLS
  tls_out.ourcert = addr->ourcert;
  addr->ourcert = NULL;
  tls_out.peercert = addr->peercert;
  addr->peercert = NULL;

  tls_out.cipher = addr->cipher;
  tls_out.peerdn = addr->peerdn;
  tls_out.ocsp = addr->ocsp;
# ifdef EXPERIMENTAL_DANE
  tls_out.dane_verified = testflag(addr, af_dane_verified);
# endif
#endif

  delivery_log(LOG_MAIN, addr, logchar, NULL);

#ifdef SUPPORT_TLS
  tls_free_cert(&tls_out.ourcert);
  tls_free_cert(&tls_out.peercert);
  tls_out.cipher = NULL;
  tls_out.peerdn = NULL;
  tls_out.ocsp = OCSP_NOT_REQ;
# ifdef EXPERIMENTAL_DANE
  tls_out.dane_verified = FALSE;
# endif
#endif
  }


/* Soft failure, or local delivery process failed; freezing may be
requested. */

else if (result == DEFER || result == PANIC)
  {
  if (result == PANIC) logflags |= LOG_PANIC;

  /* This puts them on the chain in reverse order. Do not change this, because
  the code for handling retries assumes that the one with the retry
  information is last. */

  addr->next = addr_defer;
  addr_defer = addr;

  /* The only currently implemented special action is to freeze the
  message. Logging of this is done later, just before the -H file is
  updated. */

  if (addr->special_action == SPECIAL_FREEZE)
    {
    deliver_freeze = TRUE;
    deliver_frozen_at = time(NULL);
    update_spool = TRUE;
    }

  /* If doing a 2-stage queue run, we skip writing to either the message
  log or the main log for SMTP defers. */

  if (!queue_2stage || addr->basic_errno != 0)
    {
    uschar ss[32];

    /* For errors of the type "retry time not reached" (also remotes skipped
    on queue run), logging is controlled by L_retry_defer. Note that this kind
    of error number is negative, and all the retry ones are less than any
    others. */

    unsigned int use_log_selector = addr->basic_errno <= ERRNO_RETRY_BASE
      ? L_retry_defer : 0;

    /* Build up the line that is used for both the message log and the main
    log. */

    s = reset_point = store_get(size);

    /* Create the address string for logging. Must not do this earlier, because
    an OK result may be changed to FAIL when a pipe returns text. */

    log_address = string_log_address(addr, LOGGING(all_parents), result == OK);

    s = string_cat(s, &size, &ptr, log_address);

    if (*queue_name)
      s = string_append(s, &size, &ptr, 2, US" Q=", queue_name);

    /* Either driver_name contains something and driver_kind contains
    " router" or " transport" (note the leading space), or driver_name is
    a null string and driver_kind contains "routing" without the leading
    space, if all routing has been deferred. When a domain has been held,
    so nothing has been done at all, both variables contain null strings. */

    if (driver_name)
      {
      if (driver_kind[1] == 't' && addr->router)
        s = string_append(s, &size, &ptr, 2, US" R=", addr->router->name);
      Ustrcpy(ss, " ?=");
      ss[1] = toupper(driver_kind[1]);
      s = string_append(s, &size, &ptr, 2, ss, driver_name);
      }
    else if (driver_kind)
      s = string_append(s, &size, &ptr, 2, US" ", driver_kind);

    sprintf(CS ss, " defer (%d)", addr->basic_errno);
    s = string_cat(s, &size, &ptr, ss);

    if (addr->basic_errno > 0)
      s = string_append(s, &size, &ptr, 2, US": ",
        US strerror(addr->basic_errno));

    if (addr->host_used)
      {
      s = string_append(s, &size, &ptr, 5,
                        US" H=", addr->host_used->name,
                        US" [",  addr->host_used->address, US"]");
      if (LOGGING(outgoing_port))
        {
        int port = addr->host_used->port;
        s = string_append(s, &size, &ptr, 2,
              US":", port == PORT_NONE ? US"25" : string_sprintf("%d", port));
        }
      }

    if (addr->message)
      s = string_append(s, &size, &ptr, 2, US": ", addr->message);

    s[ptr] = 0;

    /* Log the deferment in the message log, but don't clutter it
    up with retry-time defers after the first delivery attempt. */

    if (deliver_firsttime || addr->basic_errno > ERRNO_RETRY_BASE)
      deliver_msglog("%s %s\n", now, s);

    /* Write the main log and reset the store */

    log_write(use_log_selector, logflags, "== %s", s);
    store_reset(reset_point);
    }
  }


/* Hard failure. If there is an address to which an error message can be sent,
put this address on the failed list. If not, put it on the deferred list and
freeze the mail message for human attention. The latter action can also be
explicitly requested by a router or transport. */

else
  {
  /* If this is a delivery error, or a message for which no replies are
  wanted, and the message's age is greater than ignore_bounce_errors_after,
  force the af_ignore_error flag. This will cause the address to be discarded
  later (with a log entry). */

  if (sender_address[0] == 0 && message_age >= ignore_bounce_errors_after)
    setflag(addr, af_ignore_error);

  /* Freeze the message if requested, or if this is a bounce message (or other
  message with null sender) and this address does not have its own errors
  address. However, don't freeze if errors are being ignored. The actual code
  to ignore occurs later, instead of sending a message. Logging of freezing
  occurs later, just before writing the -H file. */

  if (  !testflag(addr, af_ignore_error)
     && (  addr->special_action == SPECIAL_FREEZE
        || (sender_address[0] == 0 && !addr->prop.errors_address)
     )  )
    {
    frozen_info = addr->special_action == SPECIAL_FREEZE
      ? US""
      : sender_local && !local_error_message
      ? US" (message created with -f <>)"
      : US" (delivery error message)";
    deliver_freeze = TRUE;
    deliver_frozen_at = time(NULL);
    update_spool = TRUE;

    /* The address is put on the defer rather than the failed queue, because
    the message is being retained. */

    addr->next = addr_defer;
    addr_defer = addr;
    }

  /* Don't put the address on the nonrecipients tree yet; wait until an
  error message has been successfully sent. */

  else
    {
    addr->next = addr_failed;
    addr_failed = addr;
    }

  /* Build up the log line for the message and main logs */

  s = reset_point = store_get(size);

  /* Create the address string for logging. Must not do this earlier, because
  an OK result may be changed to FAIL when a pipe returns text. */

  log_address = string_log_address(addr, LOGGING(all_parents), result == OK);

  s = string_cat(s, &size, &ptr, log_address);

  if (LOGGING(sender_on_delivery))
    s = string_append(s, &size, &ptr, 3, US" F=<", sender_address, US">");

  if (*queue_name)
    s = string_append(s, &size, &ptr, 2, US" Q=", queue_name);

  /* Return path may not be set if no delivery actually happened */

  if (used_return_path && LOGGING(return_path_on_delivery))
    s = string_append(s, &size, &ptr, 3, US" P=<", used_return_path, US">");

  if (addr->router)
    s = string_append(s, &size, &ptr, 2, US" R=", addr->router->name);
  if (addr->transport)
    s = string_append(s, &size, &ptr, 2, US" T=", addr->transport->name);

  if (addr->host_used)
    s = d_hostlog(s, &size, &ptr, addr);

#ifdef SUPPORT_TLS
  s = d_tlslog(s, &size, &ptr, addr);
#endif

  if (addr->basic_errno > 0)
    s = string_append(s, &size, &ptr, 2, US": ",
      US strerror(addr->basic_errno));

  if (addr->message)
    s = string_append(s, &size, &ptr, 2, US": ", addr->message);

  s[ptr] = 0;

  /* Do the logging. For the message log, "routing failed" for those cases,
  just to make it clearer. */

  if (driver_name)
    deliver_msglog("%s %s\n", now, s);
  else
    deliver_msglog("%s %s failed for %s\n", now, driver_kind, s);

  log_write(0, LOG_MAIN, "** %s", s);

#ifndef DISABLE_EVENT
  msg_event_raise(US"msg:fail:delivery", addr);
#endif

  store_reset(reset_point);
  }

/* Ensure logging is turned on again in all cases */

disable_logging = FALSE;
}




/*************************************************
*            Address-independent error           *
*************************************************/

/* This function is called when there's an error that is not dependent on a
particular address, such as an expansion string failure. It puts the error into
all the addresses in a batch, logs the incident on the main and panic logs, and
clears the expansions. It is mostly called from local_deliver(), but can be
called for a remote delivery via findugid().

Arguments:
  logit        TRUE if (MAIN+PANIC) logging required
  addr         the first of the chain of addresses
  code         the error code
  format       format string for error message, or NULL if already set in addr
  ...          arguments for the format

Returns:       nothing
*/

static void
common_error(BOOL logit, address_item *addr, int code, uschar *format, ...)
{
address_item *addr2;
addr->basic_errno = code;

if (format)
  {
  va_list ap;
  uschar buffer[512];
  va_start(ap, format);
  if (!string_vformat(buffer, sizeof(buffer), CS format, ap))
    log_write(0, LOG_MAIN|LOG_PANIC_DIE,
      "common_error expansion was longer than " SIZE_T_FMT, sizeof(buffer));
  va_end(ap);
  addr->message = string_copy(buffer);
  }

for (addr2 = addr->next; addr2; addr2 = addr2->next)
  {
  addr2->basic_errno = code;
  addr2->message = addr->message;
  }

if (logit) log_write(0, LOG_MAIN|LOG_PANIC, "%s", addr->message);
deliver_set_expansions(NULL);
}




/*************************************************
*         Check a "never users" list             *
*************************************************/

/* This function is called to check whether a uid is on one of the two "never
users" lists.

Arguments:
  uid         the uid to be checked
  nusers      the list to be scanned; the first item in the list is the count

Returns:      TRUE if the uid is on the list
*/

static BOOL
check_never_users(uid_t uid, uid_t *nusers)
{
int i;
if (!nusers) return FALSE;
for (i = 1; i <= (int)(nusers[0]); i++) if (nusers[i] == uid) return TRUE;
return FALSE;
}



/*************************************************
*          Find uid and gid for a transport      *
*************************************************/

/* This function is called for both local and remote deliveries, to find the
uid/gid under which to run the delivery. The values are taken preferentially
from the transport (either explicit or deliver_as_creator), then from the
address (i.e. the router), and if nothing is set, the exim uid/gid are used. If
the resulting uid is on the "never_users" or the "fixed_never_users" list, a
panic error is logged, and the function fails (which normally leads to delivery
deferral).

Arguments:
  addr         the address (possibly a chain)
  tp           the transport
  uidp         pointer to uid field
  gidp         pointer to gid field
  igfp         pointer to the use_initgroups field

Returns:       FALSE if failed - error has been set in address(es)
*/

static BOOL
findugid(address_item *addr, transport_instance *tp, uid_t *uidp, gid_t *gidp,
  BOOL *igfp)
{
uschar *nuname;
BOOL gid_set = FALSE;

/* Default initgroups flag comes from the transport */

*igfp = tp->initgroups;

/* First see if there's a gid on the transport, either fixed or expandable.
The expanding function always logs failure itself. */

if (tp->gid_set)
  {
  *gidp = tp->gid;
  gid_set = TRUE;
  }
else if (tp->expand_gid)
  {
  if (!route_find_expanded_group(tp->expand_gid, tp->name, US"transport", gidp,
    &(addr->message)))
    {
    common_error(FALSE, addr, ERRNO_GIDFAIL, NULL);
    return FALSE;
    }
  gid_set = TRUE;
  }

/* If the transport did not set a group, see if the router did. */

if (!gid_set && testflag(addr, af_gid_set))
  {
  *gidp = addr->gid;
  gid_set = TRUE;
  }

/* Pick up a uid from the transport if one is set. */

if (tp->uid_set) *uidp = tp->uid;

/* Otherwise, try for an expandable uid field. If it ends up as a numeric id,
it does not provide a passwd value from which a gid can be taken. */

else if (tp->expand_uid)
  {
  struct passwd *pw;
  if (!route_find_expanded_user(tp->expand_uid, tp->name, US"transport", &pw,
       uidp, &(addr->message)))
    {
    common_error(FALSE, addr, ERRNO_UIDFAIL, NULL);
    return FALSE;
    }
  if (!gid_set && pw)
    {
    *gidp = pw->pw_gid;
    gid_set = TRUE;
    }
  }

/* If the transport doesn't set the uid, test the deliver_as_creator flag. */

else if (tp->deliver_as_creator)
  {
  *uidp = originator_uid;
  if (!gid_set)
    {
    *gidp = originator_gid;
    gid_set = TRUE;
    }
  }

/* Otherwise see if the address specifies the uid and if so, take it and its
initgroups flag. */

else if (testflag(addr, af_uid_set))
  {
  *uidp = addr->uid;
  *igfp = testflag(addr, af_initgroups);
  }

/* Nothing has specified the uid - default to the Exim user, and group if the
gid is not set. */

else
  {
  *uidp = exim_uid;
  if (!gid_set)
    {
    *gidp = exim_gid;
    gid_set = TRUE;
    }
  }

/* If no gid is set, it is a disaster. We default to the Exim gid only if
defaulting to the Exim uid. In other words, if the configuration has specified
a uid, it must also provide a gid. */

if (!gid_set)
  {
  common_error(TRUE, addr, ERRNO_GIDFAIL, US"User set without group for "
    "%s transport", tp->name);
  return FALSE;
  }

/* Check that the uid is not on the lists of banned uids that may not be used
for delivery processes. */

nuname = check_never_users(*uidp, never_users)
  ? US"never_users"
  : check_never_users(*uidp, fixed_never_users)
  ? US"fixed_never_users"
  : NULL;
if (nuname)
  {
  common_error(TRUE, addr, ERRNO_UIDFAIL, US"User %ld set for %s transport "
    "is on the %s list", (long int)(*uidp), tp->name, nuname);
  return FALSE;
  }

/* All is well */

return TRUE;
}




/*************************************************
*   Check the size of a message for a transport  *
*************************************************/

/* Checks that the message isn't too big for the selected transport.
This is called only when it is known that the limit is set.

Arguments:
  tp          the transport
  addr        the (first) address being delivered

Returns:      OK
              DEFER   expansion failed or did not yield an integer
              FAIL    message too big
*/

int
check_message_size(transport_instance *tp, address_item *addr)
{
int rc = OK;
int size_limit;

deliver_set_expansions(addr);
size_limit = expand_string_integer(tp->message_size_limit, TRUE);
deliver_set_expansions(NULL);

if (expand_string_message)
  {
  rc = DEFER;
  addr->message = size_limit == -1
    ? string_sprintf("failed to expand message_size_limit "
      "in %s transport: %s", tp->name, expand_string_message)
    : string_sprintf("invalid message_size_limit "
      "in %s transport: %s", tp->name, expand_string_message);
  }
else if (size_limit > 0 && message_size > size_limit)
  {
  rc = FAIL;
  addr->message =
    string_sprintf("message is too big (transport limit = %d)",
      size_limit);
  }

return rc;
}



/*************************************************
*  Transport-time check for a previous delivery  *
*************************************************/

/* Check that this base address hasn't previously been delivered to its routed
transport. If it has been delivered, mark it done. The check is necessary at
delivery time in order to handle homonymic addresses correctly in cases where
the pattern of redirection changes between delivery attempts (so the unique
fields change). Non-homonymic previous delivery is detected earlier, at routing
time (which saves unnecessary routing).

Arguments:
  addr      the address item
  testing   TRUE if testing wanted only, without side effects

Returns:    TRUE if previously delivered by the transport
*/

static BOOL
previously_transported(address_item *addr, BOOL testing)
{
(void)string_format(big_buffer, big_buffer_size, "%s/%s",
  addr->unique + (testflag(addr, af_homonym)? 3:0), addr->transport->name);

if (tree_search(tree_nonrecipients, big_buffer) != 0)
  {
  DEBUG(D_deliver|D_route|D_transport)
    debug_printf("%s was previously delivered (%s transport): discarded\n",
    addr->address, addr->transport->name);
  if (!testing) child_done(addr, tod_stamp(tod_log));
  return TRUE;
  }

return FALSE;
}



/******************************************************
*      Check for a given header in a header string    *
******************************************************/

/* This function is used when generating quota warnings. The configuration may
specify any header lines it likes in quota_warn_message. If certain of them are
missing, defaults are inserted, so we need to be able to test for the presence
of a given header.

Arguments:
  hdr         the required header name
  hstring     the header string

Returns:      TRUE  the header is in the string
              FALSE the header is not in the string
*/

static BOOL
contains_header(uschar *hdr, uschar *hstring)
{
int len = Ustrlen(hdr);
uschar *p = hstring;
while (*p != 0)
  {
  if (strncmpic(p, hdr, len) == 0)
    {
    p += len;
    while (*p == ' ' || *p == '\t') p++;
    if (*p == ':') return TRUE;
    }
  while (*p != 0 && *p != '\n') p++;
  if (*p == '\n') p++;
  }
return FALSE;
}




/*************************************************
*           Perform a local delivery             *
*************************************************/

/* Each local delivery is performed in a separate process which sets its
uid and gid as specified. This is a safer way than simply changing and
restoring using seteuid(); there is a body of opinion that seteuid() cannot be
used safely. From release 4, Exim no longer makes any use of it. Besides, not
all systems have seteuid().

If the uid/gid are specified in the transport_instance, they are used; the
transport initialization must ensure that either both or neither are set.
Otherwise, the values associated with the address are used. If neither are set,
it is a configuration error.

The transport or the address may specify a home directory (transport over-
rides), and if they do, this is set as $home. If neither have set a working
directory, this value is used for that as well. Otherwise $home is left unset
and the cwd is set to "/" - a directory that should be accessible to all users.

Using a separate process makes it more complicated to get error information
back. We use a pipe to pass the return code and also an error code and error
text string back to the parent process.

Arguments:
  addr       points to an address block for this delivery; for "normal" local
             deliveries this is the only address to be delivered, but for
             pseudo-remote deliveries (e.g. by batch SMTP to a file or pipe)
             a number of addresses can be handled simultaneously, and in this
             case addr will point to a chain of addresses with the same
             characteristics.

  shadowing  TRUE if running a shadow transport; this causes output from pipes
             to be ignored.

Returns:     nothing
*/

static void
deliver_local(address_item *addr, BOOL shadowing)
{
BOOL use_initgroups;
uid_t uid;
gid_t gid;
int status, len, rc;
int pfd[2];
pid_t pid;
uschar *working_directory;
address_item *addr2;
transport_instance *tp = addr->transport;

/* Set up the return path from the errors or sender address. If the transport
has its own return path setting, expand it and replace the existing value. */

if(addr->prop.errors_address)
  return_path = addr->prop.errors_address;
#ifdef EXPERIMENTAL_SRS
else if (addr->prop.srs_sender)
  return_path = addr->prop.srs_sender;
#endif
else
  return_path = sender_address;

if (tp->return_path)
  {
  uschar *new_return_path = expand_string(tp->return_path);
  if (!new_return_path)
    {
    if (!expand_string_forcedfail)
      {
      common_error(TRUE, addr, ERRNO_EXPANDFAIL,
        US"Failed to expand return path \"%s\" in %s transport: %s",
        tp->return_path, tp->name, expand_string_message);
      return;
      }
    }
  else return_path = new_return_path;
  }

/* For local deliveries, one at a time, the value used for logging can just be
set directly, once and for all. */

used_return_path = return_path;

/* Sort out the uid, gid, and initgroups flag. If an error occurs, the message
gets put into the address(es), and the expansions are unset, so we can just
return. */

if (!findugid(addr, tp, &uid, &gid, &use_initgroups)) return;

/* See if either the transport or the address specifies a home directory. A
home directory set in the address may already be expanded; a flag is set to
indicate that. In other cases we must expand it. */

if (  (deliver_home = tp->home_dir)             /* Set in transport, or */
   || (  (deliver_home = addr->home_dir)        /* Set in address and */
      && !testflag(addr, af_home_expanded)      /*   not expanded */
   )  )
  {
  uschar *rawhome = deliver_home;
  deliver_home = NULL;                      /* in case it contains $home */
  if (!(deliver_home = expand_string(rawhome)))
    {
    common_error(TRUE, addr, ERRNO_EXPANDFAIL, US"home directory \"%s\" failed "
      "to expand for %s transport: %s", rawhome, tp->name,
      expand_string_message);
    return;
    }
  if (*deliver_home != '/')
    {
    common_error(TRUE, addr, ERRNO_NOTABSOLUTE, US"home directory path \"%s\" "
      "is not absolute for %s transport", deliver_home, tp->name);
    return;
    }
  }

/* See if either the transport or the address specifies a current directory,
and if so, expand it. If nothing is set, use the home directory, unless it is
also unset in which case use "/", which is assumed to be a directory to which
all users have access. It is necessary to be in a visible directory for some
operating systems when running pipes, as some commands (e.g. "rm" under Solaris
2.5) require this. */

working_directory = tp->current_dir ? tp->current_dir : addr->current_dir;
if (working_directory)
  {
  uschar *raw = working_directory;
  if (!(working_directory = expand_string(raw)))
    {
    common_error(TRUE, addr, ERRNO_EXPANDFAIL, US"current directory \"%s\" "
      "failed to expand for %s transport: %s", raw, tp->name,
      expand_string_message);
    return;
    }
  if (*working_directory != '/')
    {
    common_error(TRUE, addr, ERRNO_NOTABSOLUTE, US"current directory path "
      "\"%s\" is not absolute for %s transport", working_directory, tp->name);
    return;
    }
  }
else working_directory = deliver_home ? deliver_home : US"/";

/* If one of the return_output flags is set on the transport, create and open a
file in the message log directory for the transport to write its output onto.
This is mainly used by pipe transports. The file needs to be unique to the
address. This feature is not available for shadow transports. */

if (  !shadowing
   && (  tp->return_output || tp->return_fail_output
      || tp->log_output || tp->log_fail_output || tp->log_defer_output
   )  )
  {
  uschar * error;

  addr->return_filename =
    spool_fname(US"msglog", message_subdir, message_id,
      string_sprintf("-%d-%d", getpid(), return_count++));
  
  if ((addr->return_file = open_msglog_file(addr->return_filename, 0400, &error)) < 0)
    {
    common_error(TRUE, addr, errno, US"Unable to %s file for %s transport "
      "to return message: %s", error, tp->name, strerror(errno));
    return;
    }
  }

/* Create the pipe for inter-process communication. */

if (pipe(pfd) != 0)
  {
  common_error(TRUE, addr, ERRNO_PIPEFAIL, US"Creation of pipe failed: %s",
    strerror(errno));
  return;
  }

/* Now fork the process to do the real work in the subprocess, but first
ensure that all cached resources are freed so that the subprocess starts with
a clean slate and doesn't interfere with the parent process. */

search_tidyup();

if ((pid = fork()) == 0)
  {
  BOOL replicate = TRUE;

  /* Prevent core dumps, as we don't want them in users' home directories.
  HP-UX doesn't have RLIMIT_CORE; I don't know how to do this in that
  system. Some experimental/developing systems (e.g. GNU/Hurd) may define
  RLIMIT_CORE but not support it in setrlimit(). For such systems, do not
  complain if the error is "not supported".

  There are two scenarios where changing the max limit has an effect.  In one,
  the user is using a .forward and invoking a command of their choice via pipe;
  for these, we do need the max limit to be 0 unless the admin chooses to
  permit an increased limit.  In the other, the command is invoked directly by
  the transport and is under administrator control, thus being able to raise
  the limit aids in debugging.  So there's no general always-right answer.

  Thus we inhibit core-dumps completely but let individual transports, while
  still root, re-raise the limits back up to aid debugging.  We make the
  default be no core-dumps -- few enough people can use core dumps in
  diagnosis that it's reasonable to make them something that has to be explicitly requested.
  */

#ifdef RLIMIT_CORE
  struct rlimit rl;
  rl.rlim_cur = 0;
  rl.rlim_max = 0;
  if (setrlimit(RLIMIT_CORE, &rl) < 0)
    {
# ifdef SETRLIMIT_NOT_SUPPORTED
    if (errno != ENOSYS && errno != ENOTSUP)
# endif
      log_write(0, LOG_MAIN|LOG_PANIC, "setrlimit(RLIMIT_CORE) failed: %s",
        strerror(errno));
    }
#endif

  /* Reset the random number generator, so different processes don't all
  have the same sequence. */

  random_seed = 0;

  /* If the transport has a setup entry, call this first, while still
  privileged. (Appendfile uses this to expand quota, for example, while
  able to read private files.) */

  if (addr->transport->setup)
    switch((addr->transport->setup)(addr->transport, addr, NULL, uid, gid,
           &(addr->message)))
      {
      case DEFER:
        addr->transport_return = DEFER;
        goto PASS_BACK;

      case FAIL:
        addr->transport_return = PANIC;
        goto PASS_BACK;
      }

  /* Ignore SIGINT and SIGTERM during delivery. Also ignore SIGUSR1, as
  when the process becomes unprivileged, it won't be able to write to the
  process log. SIGHUP is ignored throughout exim, except when it is being
  run as a daemon. */

  signal(SIGINT, SIG_IGN);
  signal(SIGTERM, SIG_IGN);
  signal(SIGUSR1, SIG_IGN);

  /* Close the unwanted half of the pipe, and set close-on-exec for the other
  half - for transports that exec things (e.g. pipe). Then set the required
  gid/uid. */

  (void)close(pfd[pipe_read]);
  (void)fcntl(pfd[pipe_write], F_SETFD, fcntl(pfd[pipe_write], F_GETFD) |
    FD_CLOEXEC);
  exim_setugid(uid, gid, use_initgroups,
    string_sprintf("local delivery to %s <%s> transport=%s", addr->local_part,
      addr->address, addr->transport->name));

  DEBUG(D_deliver)
    {
    address_item *batched;
    debug_printf("  home=%s current=%s\n", deliver_home, working_directory);
    for (batched = addr->next; batched; batched = batched->next)
      debug_printf("additional batched address: %s\n", batched->address);
    }

  /* Set an appropriate working directory. */

  if (Uchdir(working_directory) < 0)
    {
    addr->transport_return = DEFER;
    addr->basic_errno = errno;
    addr->message = string_sprintf("failed to chdir to %s", working_directory);
    }

  /* If successful, call the transport */

  else
    {
    BOOL ok = TRUE;
    set_process_info("delivering %s to %s using %s", message_id,
     addr->local_part, addr->transport->name);

    /* Setting this global in the subprocess means we need never clear it */
    transport_name = addr->transport->name;

    /* If a transport filter has been specified, set up its argument list.
    Any errors will get put into the address, and FALSE yielded. */

    if (addr->transport->filter_command)
      {
      ok = transport_set_up_command(&transport_filter_argv,
        addr->transport->filter_command,
        TRUE, PANIC, addr, US"transport filter", NULL);
      transport_filter_timeout = addr->transport->filter_timeout;
      }
    else transport_filter_argv = NULL;

    if (ok)
      {
      debug_print_string(addr->transport->debug_string);
      replicate = !(addr->transport->info->code)(addr->transport, addr);
      }
    }

  /* Pass the results back down the pipe. If necessary, first replicate the
  status in the top address to the others in the batch. The label is the
  subject of a goto when a call to the transport's setup function fails. We
  pass the pointer to the transport back in case it got changed as a result of
  file_format in appendfile. */

  PASS_BACK:

  if (replicate) replicate_status(addr);
  for (addr2 = addr; addr2; addr2 = addr2->next)
    {
    int i;
    int local_part_length = Ustrlen(addr2->local_part);
    uschar *s;
    int ret;

    if(  (ret = write(pfd[pipe_write], &addr2->transport_return, sizeof(int))) != sizeof(int)
      || (ret = write(pfd[pipe_write], &transport_count, sizeof(transport_count))) != sizeof(transport_count)
      || (ret = write(pfd[pipe_write], &addr2->flags, sizeof(addr2->flags))) != sizeof(addr2->flags)
      || (ret = write(pfd[pipe_write], &addr2->basic_errno,    sizeof(int))) != sizeof(int)
      || (ret = write(pfd[pipe_write], &addr2->more_errno,     sizeof(int))) != sizeof(int)
      || (ret = write(pfd[pipe_write], &addr2->special_action, sizeof(int))) != sizeof(int)
      || (ret = write(pfd[pipe_write], &addr2->transport,
        sizeof(transport_instance *))) != sizeof(transport_instance *)

    /* For a file delivery, pass back the local part, in case the original
    was only part of the final delivery path. This gives more complete
    logging. */

      || (testflag(addr2, af_file)
          && (  (ret = write(pfd[pipe_write], &local_part_length, sizeof(int))) != sizeof(int)
             || (ret = write(pfd[pipe_write], addr2->local_part, local_part_length)) != local_part_length
             )
         )
      )
      log_write(0, LOG_MAIN|LOG_PANIC, "Failed writing transport results to pipe: %s\n",
        ret == -1 ? strerror(errno) : "short write");

    /* Now any messages */

    for (i = 0, s = addr2->message; i < 2; i++, s = addr2->user_message)
      {
      int message_length = s ? Ustrlen(s) + 1 : 0;
      if(  (ret = write(pfd[pipe_write], &message_length, sizeof(int))) != sizeof(int)
        || message_length > 0  && (ret = write(pfd[pipe_write], s, message_length)) != message_length
        )
        log_write(0, LOG_MAIN|LOG_PANIC, "Failed writing transport results to pipe: %s\n",
          ret == -1 ? strerror(errno) : "short write");
      }
    }

  /* OK, this process is now done. Free any cached resources that it opened,
  and close the pipe we were writing down before exiting. */

  (void)close(pfd[pipe_write]);
  search_tidyup();
  exit(EXIT_SUCCESS);
  }

/* Back in the main process: panic if the fork did not succeed. This seems
better than returning an error - if forking is failing it is probably best
not to try other deliveries for this message. */

if (pid < 0)
  log_write(0, LOG_MAIN|LOG_PANIC_DIE, "Fork failed for local delivery to %s",
    addr->address);

/* Read the pipe to get the delivery status codes and error messages. Our copy
of the writing end must be closed first, as otherwise read() won't return zero
on an empty pipe. We check that a status exists for each address before
overwriting the address structure. If data is missing, the default DEFER status
will remain. Afterwards, close the reading end. */

(void)close(pfd[pipe_write]);

for (addr2 = addr; addr2; addr2 = addr2->next)
  {
  len = read(pfd[pipe_read], &status, sizeof(int));
  if (len > 0)
    {
    int i;
    uschar **sptr;

    addr2->transport_return = status;
    len = read(pfd[pipe_read], &transport_count,
      sizeof(transport_count));
    len = read(pfd[pipe_read], &addr2->flags, sizeof(addr2->flags));
    len = read(pfd[pipe_read], &addr2->basic_errno,    sizeof(int));
    len = read(pfd[pipe_read], &addr2->more_errno,     sizeof(int));
    len = read(pfd[pipe_read], &addr2->special_action, sizeof(int));
    len = read(pfd[pipe_read], &addr2->transport,
      sizeof(transport_instance *));

    if (testflag(addr2, af_file))
      {
      int local_part_length;
      len = read(pfd[pipe_read], &local_part_length, sizeof(int));
      len = read(pfd[pipe_read], big_buffer, local_part_length);
      big_buffer[local_part_length] = 0;
      addr2->local_part = string_copy(big_buffer);
      }

    for (i = 0, sptr = &addr2->message; i < 2; i++, sptr = &addr2->user_message)
      {
      int message_length;
      len = read(pfd[pipe_read], &message_length, sizeof(int));
      if (message_length > 0)
        {
        len = read(pfd[pipe_read], big_buffer, message_length);
        big_buffer[big_buffer_size-1] = '\0';           /* guard byte */
        if (len > 0) *sptr = string_copy(big_buffer);
        }
      }
    }

  else
    {
    log_write(0, LOG_MAIN|LOG_PANIC, "failed to read delivery status for %s "
      "from delivery subprocess", addr2->unique);
    break;
    }
  }

(void)close(pfd[pipe_read]);

/* Unless shadowing, write all successful addresses immediately to the journal
file, to ensure they are recorded asap. For homonymic addresses, use the base
address plus the transport name. Failure to write the journal is panic-worthy,
but don't stop, as it may prove possible subsequently to update the spool file
in order to record the delivery. */

if (!shadowing)
  {
  for (addr2 = addr; addr2; addr2 = addr2->next)
    if (addr2->transport_return == OK)
      {
      if (testflag(addr2, af_homonym))
        sprintf(CS big_buffer, "%.500s/%s\n", addr2->unique + 3, tp->name);
      else
        sprintf(CS big_buffer, "%.500s\n", addr2->unique);

      /* In the test harness, wait just a bit to let the subprocess finish off
      any debug output etc first. */

      if (running_in_test_harness) millisleep(300);

      DEBUG(D_deliver) debug_printf("journalling %s", big_buffer);
      len = Ustrlen(big_buffer);
      if (write(journal_fd, big_buffer, len) != len)
        log_write(0, LOG_MAIN|LOG_PANIC, "failed to update journal for %s: %s",
          big_buffer, strerror(errno));
      }

  /* Ensure the journal file is pushed out to disk. */

  if (EXIMfsync(journal_fd) < 0)
    log_write(0, LOG_MAIN|LOG_PANIC, "failed to fsync journal: %s",
      strerror(errno));
  }

/* Wait for the process to finish. If it terminates with a non-zero code,
freeze the message (except for SIGTERM, SIGKILL and SIGQUIT), but leave the
status values of all the addresses as they are. Take care to handle the case
when the subprocess doesn't seem to exist. This has been seen on one system
when Exim was called from an MUA that set SIGCHLD to SIG_IGN. When that
happens, wait() doesn't recognize the termination of child processes. Exim now
resets SIGCHLD to SIG_DFL, but this code should still be robust. */

while ((rc = wait(&status)) != pid)
  if (rc < 0 && errno == ECHILD)      /* Process has vanished */
    {
    log_write(0, LOG_MAIN, "%s transport process vanished unexpectedly",
      addr->transport->driver_name);
    status = 0;
    break;
    }

if ((status & 0xffff) != 0)
  {
  int msb = (status >> 8) & 255;
  int lsb = status & 255;
  int code = (msb == 0)? (lsb & 0x7f) : msb;
  if (msb != 0 || (code != SIGTERM && code != SIGKILL && code != SIGQUIT))
    addr->special_action = SPECIAL_FREEZE;
  log_write(0, LOG_MAIN|LOG_PANIC, "%s transport process returned non-zero "
    "status 0x%04x: %s %d",
    addr->transport->driver_name,
    status,
    msb == 0 ? "terminated by signal" : "exit code",
    code);
  }

/* If SPECIAL_WARN is set in the top address, send a warning message. */

if (addr->special_action == SPECIAL_WARN && addr->transport->warn_message)
  {
  int fd;
  uschar *warn_message;
  pid_t pid;

  DEBUG(D_deliver) debug_printf("Warning message requested by transport\n");

  if (!(warn_message = expand_string(addr->transport->warn_message)))
    log_write(0, LOG_MAIN|LOG_PANIC, "Failed to expand \"%s\" (warning "
      "message for %s transport): %s", addr->transport->warn_message,
      addr->transport->name, expand_string_message);

  else if ((pid = child_open_exim(&fd)) > 0)
    {
    FILE *f = fdopen(fd, "wb");
    if (errors_reply_to && !contains_header(US"Reply-To", warn_message))
      fprintf(f, "Reply-To: %s\n", errors_reply_to);
    fprintf(f, "Auto-Submitted: auto-replied\n");
    if (!contains_header(US"From", warn_message))
      moan_write_from(f);
    fprintf(f, "%s", CS warn_message);

    /* Close and wait for child process to complete, without a timeout. */

    (void)fclose(f);
    (void)child_close(pid, 0);
    }

  addr->special_action = SPECIAL_NONE;
  }
}




/* Check transport for the given concurrency limit.  Return TRUE if over
the limit (or an expansion failure), else FALSE and if there was a limit,
the key for the hints database used for the concurrency count. */

static BOOL
tpt_parallel_check(transport_instance * tp, address_item * addr, uschar ** key)
{
unsigned max_parallel;

if (!tp->max_parallel) return FALSE;

max_parallel = (unsigned) expand_string_integer(tp->max_parallel, TRUE);
if (expand_string_message)
  {
  log_write(0, LOG_MAIN|LOG_PANIC, "Failed to expand max_parallel option "
        "in %s transport (%s): %s", tp->name, addr->address,
        expand_string_message);
  return TRUE;
  }

if (max_parallel > 0)
  {
  uschar * serialize_key = string_sprintf("tpt-serialize-%s", tp->name);
  if (!enq_start(serialize_key, max_parallel))
    {
    address_item * next;
    DEBUG(D_transport)
      debug_printf("skipping tpt %s because concurrency limit %u reached\n",
                  tp->name, max_parallel);
    do
      {
      next = addr->next;
      addr->message = US"concurrency limit reached for transport";
      addr->basic_errno = ERRNO_TRETRY;
      post_process_one(addr, DEFER, LOG_MAIN, DTYPE_TRANSPORT, 0);
      } while ((addr = next));
    return TRUE;
    }
  *key = serialize_key;
  }
return FALSE;
}



/*************************************************
*              Do local deliveries               *
*************************************************/

/* This function processes the list of addresses in addr_local. True local
deliveries are always done one address at a time. However, local deliveries can
be batched up in some cases. Typically this is when writing batched SMTP output
files for use by some external transport mechanism, or when running local
deliveries over LMTP.

Arguments:   None
Returns:     Nothing
*/

static void
do_local_deliveries(void)
{
open_db dbblock;
open_db *dbm_file = NULL;
time_t now = time(NULL);

/* Loop until we have exhausted the supply of local deliveries */

while (addr_local)
  {
  time_t delivery_start;
  int deliver_time;
  address_item *addr2, *addr3, *nextaddr;
  int logflags = LOG_MAIN;
  int logchar = dont_deliver? '*' : '=';
  transport_instance *tp;
  uschar * serialize_key = NULL;

  /* Pick the first undelivered address off the chain */

  address_item *addr = addr_local;
  addr_local = addr->next;
  addr->next = NULL;

  DEBUG(D_deliver|D_transport)
    debug_printf("--------> %s <--------\n", addr->address);

  /* An internal disaster if there is no transport. Should not occur! */

  if (!(tp = addr->transport))
    {
    logflags |= LOG_PANIC;
    disable_logging = FALSE;  /* Jic */
    addr->message = addr->router
      ? string_sprintf("No transport set by %s router", addr->router->name)
      : string_sprintf("No transport set by system filter");
    post_process_one(addr, DEFER, logflags, DTYPE_TRANSPORT, 0);
    continue;
    }

  /* Check that this base address hasn't previously been delivered to this
  transport. The check is necessary at this point to handle homonymic addresses
  correctly in cases where the pattern of redirection changes between delivery
  attempts. Non-homonymic previous delivery is detected earlier, at routing
  time. */

  if (previously_transported(addr, FALSE)) continue;

  /* There are weird cases where logging is disabled */

  disable_logging = tp->disable_logging;

  /* Check for batched addresses and possible amalgamation. Skip all the work
  if either batch_max <= 1 or there aren't any other addresses for local
  delivery. */

  if (tp->batch_max > 1 && addr_local)
    {
    int batch_count = 1;
    BOOL uses_dom = readconf_depends((driver_instance *)tp, US"domain");
    BOOL uses_lp = (  testflag(addr, af_pfr)
                   && (testflag(addr, af_file) || addr->local_part[0] == '|')
                   )
                || readconf_depends((driver_instance *)tp, US"local_part");
    uschar *batch_id = NULL;
    address_item **anchor = &addr_local;
    address_item *last = addr;
    address_item *next;

    /* Expand the batch_id string for comparison with other addresses.
    Expansion failure suppresses batching. */

    if (tp->batch_id)
      {
      deliver_set_expansions(addr);
      batch_id = expand_string(tp->batch_id);
      deliver_set_expansions(NULL);
      if (!batch_id)
        {
        log_write(0, LOG_MAIN|LOG_PANIC, "Failed to expand batch_id option "
          "in %s transport (%s): %s", tp->name, addr->address,
          expand_string_message);
        batch_count = tp->batch_max;
        }
      }

    /* Until we reach the batch_max limit, pick off addresses which have the
    same characteristics. These are:

      same transport
      not previously delivered (see comment about 50 lines above)
      same local part if the transport's configuration contains $local_part
        or if this is a file or pipe delivery from a redirection
      same domain if the transport's configuration contains $domain
      same errors address
      same additional headers
      same headers to be removed
      same uid/gid for running the transport
      same first host if a host list is set
    */

    while ((next = *anchor) && batch_count < tp->batch_max)
      {
      BOOL ok =
           tp == next->transport
        && !previously_transported(next, TRUE)
        && (addr->flags & (af_pfr|af_file)) == (next->flags & (af_pfr|af_file))
        && (!uses_lp  || Ustrcmp(next->local_part, addr->local_part) == 0)
        && (!uses_dom || Ustrcmp(next->domain, addr->domain) == 0)
        && same_strings(next->prop.errors_address, addr->prop.errors_address)
        && same_headers(next->prop.extra_headers, addr->prop.extra_headers)
        && same_strings(next->prop.remove_headers, addr->prop.remove_headers)
        && same_ugid(tp, addr, next)
        && (  !addr->host_list && !next->host_list
           ||    addr->host_list
              && next->host_list
              && Ustrcmp(addr->host_list->name, next->host_list->name) == 0
           );

      /* If the transport has a batch_id setting, batch_id will be non-NULL
      from the expansion outside the loop. Expand for this address and compare.
      Expansion failure makes this address ineligible for batching. */

      if (ok && batch_id)
        {
        uschar *bid;
        address_item *save_nextnext = next->next;
        next->next = NULL;            /* Expansion for a single address */
        deliver_set_expansions(next);
        next->next = save_nextnext;
        bid = expand_string(tp->batch_id);
        deliver_set_expansions(NULL);
        if (!bid)
          {
          log_write(0, LOG_MAIN|LOG_PANIC, "Failed to expand batch_id option "
            "in %s transport (%s): %s", tp->name, next->address,
            expand_string_message);
          ok = FALSE;
          }
        else ok = (Ustrcmp(batch_id, bid) == 0);
        }

      /* Take address into batch if OK. */

      if (ok)
        {
        *anchor = next->next;           /* Include the address */
        next->next = NULL;
        last->next = next;
        last = next;
        batch_count++;
        }
      else anchor = &next->next;        /* Skip the address */
      }
    }

  /* We now have one or more addresses that can be delivered in a batch. Check
  whether the transport is prepared to accept a message of this size. If not,
  fail them all forthwith. If the expansion fails, or does not yield an
  integer, defer delivery. */

  if (tp->message_size_limit)
    {
    int rc = check_message_size(tp, addr);
    if (rc != OK)
      {
      replicate_status(addr);
      while (addr)
        {
        addr2 = addr->next;
        post_process_one(addr, rc, logflags, DTYPE_TRANSPORT, 0);
        addr = addr2;
        }
      continue;    /* With next batch of addresses */
      }
    }

  /* If we are not running the queue, or if forcing, all deliveries will be
  attempted. Otherwise, we must respect the retry times for each address. Even
  when not doing this, we need to set up the retry key string, and determine
  whether a retry record exists, because after a successful delivery, a delete
  retry item must be set up. Keep the retry database open only for the duration
  of these checks, rather than for all local deliveries, because some local
  deliveries (e.g. to pipes) can take a substantial time. */

  if (!(dbm_file = dbfn_open(US"retry", O_RDONLY, &dbblock, FALSE)))
    {
    DEBUG(D_deliver|D_retry|D_hints_lookup)
      debug_printf("no retry data available\n");
    }

  addr2 = addr;
  addr3 = NULL;
  while (addr2)
    {
    BOOL ok = TRUE;   /* to deliver this address */
    uschar *retry_key;

    /* Set up the retry key to include the domain or not, and change its
    leading character from "R" to "T". Must make a copy before doing this,
    because the old key may be pointed to from a "delete" retry item after
    a routing delay. */

    retry_key = string_copy(
      tp->retry_use_local_part ? addr2->address_retry_key :
        addr2->domain_retry_key);
    *retry_key = 'T';

    /* Inspect the retry data. If there is no hints file, delivery happens. */

    if (dbm_file)
      {
      dbdata_retry *retry_record = dbfn_read(dbm_file, retry_key);

      /* If there is no retry record, delivery happens. If there is,
      remember it exists so it can be deleted after a successful delivery. */

      if (retry_record)
        {
        setflag(addr2, af_lt_retry_exists);

        /* A retry record exists for this address. If queue running and not
        forcing, inspect its contents. If the record is too old, or if its
        retry time has come, or if it has passed its cutoff time, delivery
        will go ahead. */

        DEBUG(D_retry)
          {
          debug_printf("retry record exists: age=%s ",
            readconf_printtime(now - retry_record->time_stamp));
          debug_printf("(max %s)\n", readconf_printtime(retry_data_expire));
          debug_printf("  time to retry = %s expired = %d\n",
            readconf_printtime(retry_record->next_try - now),
            retry_record->expired);
          }

        if (queue_running && !deliver_force)
          {
          ok = (now - retry_record->time_stamp > retry_data_expire)
            || (now >= retry_record->next_try)
            || retry_record->expired;

          /* If we haven't reached the retry time, there is one more check
          to do, which is for the ultimate address timeout. */

          if (!ok)
            ok = retry_ultimate_address_timeout(retry_key, addr2->domain,
                retry_record, now);
          }
        }
      else DEBUG(D_retry) debug_printf("no retry record exists\n");
      }

    /* This address is to be delivered. Leave it on the chain. */

    if (ok)
      {
      addr3 = addr2;
      addr2 = addr2->next;
      }

    /* This address is to be deferred. Take it out of the chain, and
    post-process it as complete. Must take it out of the chain first,
    because post processing puts it on another chain. */

    else
      {
      address_item *this = addr2;
      this->message = US"Retry time not yet reached";
      this->basic_errno = ERRNO_LRETRY;
      addr2 = addr3 ? (addr3->next = addr2->next)
                    : (addr = addr2->next);
      post_process_one(this, DEFER, logflags, DTYPE_TRANSPORT, 0);
      }
    }

  if (dbm_file) dbfn_close(dbm_file);

  /* If there are no addresses left on the chain, they all deferred. Loop
  for the next set of addresses. */

  if (!addr) continue;

  /* If the transport is limited for parallellism, enforce that here.
  We use a hints DB entry, incremented here and decremented after
  the transport (and any shadow transport) completes. */

  if (tpt_parallel_check(tp, addr, &serialize_key))
    {
    if (expand_string_message)
      {
      logflags |= LOG_PANIC;
      do
        {
        addr = addr->next;
        post_process_one(addr, DEFER, logflags, DTYPE_TRANSPORT, 0);
        } while ((addr = addr2));
      }
    continue;                   /* Loop for the next set of addresses. */
    }


  /* So, finally, we do have some addresses that can be passed to the
  transport. Before doing so, set up variables that are relevant to a
  single delivery. */

  deliver_set_expansions(addr);
  delivery_start = time(NULL);
  deliver_local(addr, FALSE);
  deliver_time = (int)(time(NULL) - delivery_start);

  /* If a shadow transport (which must perforce be another local transport), is
  defined, and its condition is met, we must pass the message to the shadow
  too, but only those addresses that succeeded. We do this by making a new
  chain of addresses - also to keep the original chain uncontaminated. We must
  use a chain rather than doing it one by one, because the shadow transport may
  batch.

  NOTE: if the condition fails because of a lookup defer, there is nothing we
  can do! */

  if (  tp->shadow
     && (  !tp->shadow_condition
        || expand_check_condition(tp->shadow_condition, tp->name, US"transport")
     )  )
    {
    transport_instance *stp;
    address_item *shadow_addr = NULL;
    address_item **last = &shadow_addr;

    for (stp = transports; stp; stp = stp->next)
      if (Ustrcmp(stp->name, tp->shadow) == 0) break;

    if (!stp)
      log_write(0, LOG_MAIN|LOG_PANIC, "shadow transport \"%s\" not found ",
        tp->shadow);

    /* Pick off the addresses that have succeeded, and make clones. Put into
    the shadow_message field a pointer to the shadow_message field of the real
    address. */

    else for (addr2 = addr; addr2; addr2 = addr2->next)
      if (addr2->transport_return == OK)
        {
        addr3 = store_get(sizeof(address_item));
        *addr3 = *addr2;
        addr3->next = NULL;
        addr3->shadow_message = (uschar *) &(addr2->shadow_message);
        addr3->transport = stp;
        addr3->transport_return = DEFER;
        addr3->return_filename = NULL;
        addr3->return_file = -1;
        *last = addr3;
        last = &(addr3->next);
        }

    /* If we found any addresses to shadow, run the delivery, and stick any
    message back into the shadow_message field in the original. */

    if (shadow_addr)
      {
      int save_count = transport_count;

      DEBUG(D_deliver|D_transport)
        debug_printf(">>>>>>>>>>>>>>>> Shadow delivery >>>>>>>>>>>>>>>>\n");
      deliver_local(shadow_addr, TRUE);

      for(; shadow_addr; shadow_addr = shadow_addr->next)
        {
        int sresult = shadow_addr->transport_return;
        *(uschar **)shadow_addr->shadow_message =
          sresult == OK
          ? string_sprintf(" ST=%s", stp->name)
          : string_sprintf(" ST=%s (%s%s%s)", stp->name,
              shadow_addr->basic_errno <= 0
              ? US""
              : US strerror(shadow_addr->basic_errno),
              shadow_addr->basic_errno <= 0 || !shadow_addr->message
              ? US""
              : US": ",
              shadow_addr->message
              ? shadow_addr->message
              : shadow_addr->basic_errno <= 0
              ? US"unknown error"
              : US"");

        DEBUG(D_deliver|D_transport)
          debug_printf("%s shadow transport returned %s for %s\n",
            stp->name,
            sresult == OK ?    "OK" :
            sresult == DEFER ? "DEFER" :
            sresult == FAIL ?  "FAIL" :
            sresult == PANIC ? "PANIC" : "?",
            shadow_addr->address);
        }

      DEBUG(D_deliver|D_transport)
        debug_printf(">>>>>>>>>>>>>>>> End shadow delivery >>>>>>>>>>>>>>>>\n");

      transport_count = save_count;   /* Restore original transport count */
      }
    }

  /* Cancel the expansions that were set up for the delivery. */

  deliver_set_expansions(NULL);

  /* If the transport was parallelism-limited, decrement the hints DB record. */

  if (serialize_key) enq_end(serialize_key);

  /* Now we can process the results of the real transport. We must take each
  address off the chain first, because post_process_one() puts it on another
  chain. */

  for (addr2 = addr; addr2; addr2 = nextaddr)
    {
    int result = addr2->transport_return;
    nextaddr = addr2->next;

    DEBUG(D_deliver|D_transport)
      debug_printf("%s transport returned %s for %s\n",
        tp->name,
        result == OK ?    "OK" :
        result == DEFER ? "DEFER" :
        result == FAIL ?  "FAIL" :
        result == PANIC ? "PANIC" : "?",
        addr2->address);

    /* If there is a retry_record, or if delivery is deferred, build a retry
    item for setting a new retry time or deleting the old retry record from
    the database. These items are handled all together after all addresses
    have been handled (so the database is open just for a short time for
    updating). */

    if (result == DEFER || testflag(addr2, af_lt_retry_exists))
      {
      int flags = result == DEFER ? 0 : rf_delete;
      uschar *retry_key = string_copy(tp->retry_use_local_part
        ? addr2->address_retry_key : addr2->domain_retry_key);
      *retry_key = 'T';
      retry_add_item(addr2, retry_key, flags);
      }

    /* Done with this address */

    if (result == OK) addr2->more_errno = deliver_time;
    post_process_one(addr2, result, logflags, DTYPE_TRANSPORT, logchar);

    /* If a pipe delivery generated text to be sent back, the result may be
    changed to FAIL, and we must copy this for subsequent addresses in the
    batch. */

    if (addr2->transport_return != result)
      {
      for (addr3 = nextaddr; addr3; addr3 = addr3->next)
        {
        addr3->transport_return = addr2->transport_return;
        addr3->basic_errno = addr2->basic_errno;
        addr3->message = addr2->message;
        }
      result = addr2->transport_return;
      }

    /* Whether or not the result was changed to FAIL, we need to copy the
    return_file value from the first address into all the addresses of the
    batch, so they are all listed in the error message. */

    addr2->return_file = addr->return_file;

    /* Change log character for recording successful deliveries. */

    if (result == OK) logchar = '-';
    }
  }        /* Loop back for next batch of addresses */
}




/*************************************************
*           Sort remote deliveries               *
*************************************************/

/* This function is called if remote_sort_domains is set. It arranges that the
chain of addresses for remote deliveries is ordered according to the strings
specified. Try to make this shuffling reasonably efficient by handling
sequences of addresses rather than just single ones.

Arguments:  None
Returns:    Nothing
*/

static void
sort_remote_deliveries(void)
{
int sep = 0;
address_item **aptr = &addr_remote;
const uschar *listptr = remote_sort_domains;
uschar *pattern;
uschar patbuf[256];

while (  *aptr
      && (pattern = string_nextinlist(&listptr, &sep, patbuf, sizeof(patbuf)))
      )
  {
  address_item *moved = NULL;
  address_item **bptr = &moved;

  while (*aptr)
    {
    address_item **next;
    deliver_domain = (*aptr)->domain;   /* set $domain */
    if (match_isinlist(deliver_domain, (const uschar **)&pattern, UCHAR_MAX+1,
          &domainlist_anchor, NULL, MCL_DOMAIN, TRUE, NULL) == OK)
      {
      aptr = &(*aptr)->next;
      continue;
      }

    next = &(*aptr)->next;
    while (  *next
          && (deliver_domain = (*next)->domain,  /* Set $domain */
            match_isinlist(deliver_domain, (const uschar **)&pattern, UCHAR_MAX+1,
              &domainlist_anchor, NULL, MCL_DOMAIN, TRUE, NULL)) != OK
          )
      next = &(*next)->next;

    /* If the batch of non-matchers is at the end, add on any that were
    extracted further up the chain, and end this iteration. Otherwise,
    extract them from the chain and hang on the moved chain. */

    if (!*next)
      {
      *next = moved;
      break;
      }

    *bptr = *aptr;
    *aptr = *next;
    *next = NULL;
    bptr = next;
    aptr = &(*aptr)->next;
    }

  /* If the loop ended because the final address matched, *aptr will
  be NULL. Add on to the end any extracted non-matching addresses. If
  *aptr is not NULL, the loop ended via "break" when *next is null, that
  is, there was a string of non-matching addresses at the end. In this
  case the extracted addresses have already been added on the end. */

  if (!*aptr) *aptr = moved;
  }

DEBUG(D_deliver)
  {
  address_item *addr;
  debug_printf("remote addresses after sorting:\n");
  for (addr = addr_remote; addr; addr = addr->next)
    debug_printf("  %s\n", addr->address);
  }
}



/*************************************************
*  Read from pipe for remote delivery subprocess *
*************************************************/

/* This function is called when the subprocess is complete, but can also be
called before it is complete, in order to empty a pipe that is full (to prevent
deadlock). It must therefore keep track of its progress in the parlist data
block.

We read the pipe to get the delivery status codes and a possible error message
for each address, optionally preceded by unusability data for the hosts and
also by optional retry data.

Read in large chunks into the big buffer and then scan through, interpreting
the data therein. In most cases, only a single read will be necessary. No
individual item will ever be anywhere near 2500 bytes in length, so by ensuring
that we read the next chunk when there is less than 2500 bytes left in the
non-final chunk, we can assume each item is complete in the buffer before
handling it. Each item is written using a single write(), which is atomic for
small items (less than PIPE_BUF, which seems to be at least 512 in any Unix and
often bigger) so even if we are reading while the subprocess is still going, we
should never have only a partial item in the buffer.

Argument:
  poffset     the offset of the parlist item
  eop         TRUE if the process has completed

Returns:      TRUE if the terminating 'Z' item has been read,
              or there has been a disaster (i.e. no more data needed);
              FALSE otherwise
*/

static BOOL
par_read_pipe(int poffset, BOOL eop)
{
host_item *h;
pardata *p = parlist + poffset;
address_item *addrlist = p->addrlist;
address_item *addr = p->addr;
pid_t pid = p->pid;
int fd = p->fd;
uschar *endptr = big_buffer;
uschar *ptr = endptr;
uschar *msg = p->msg;
BOOL done = p->done;
BOOL unfinished = TRUE;
/* minimum size to read is header size including id, subid and length */
int required = PIPE_HEADER_SIZE;

/* Loop through all items, reading from the pipe when necessary. The pipe
is set up to be non-blocking, but there are two different Unix mechanisms in
use. Exim uses O_NONBLOCK if it is defined. This returns 0 for end of file,
and EAGAIN for no more data. If O_NONBLOCK is not defined, Exim uses O_NDELAY,
which returns 0 for both end of file and no more data. We distinguish the
two cases by taking 0 as end of file only when we know the process has
completed.

Each separate item is written to the pipe in a single write(), and as they are
all short items, the writes will all be atomic and we should never find
ourselves in the position of having read an incomplete item. "Short" in this
case can mean up to about 1K in the case when there is a long error message
associated with an address. */

DEBUG(D_deliver) debug_printf("reading pipe for subprocess %d (%s)\n",
  (int)p->pid, eop? "ended" : "not ended");

while (!done)
  {
  retry_item *r, **rp;
  int remaining = endptr - ptr;
  uschar header[PIPE_HEADER_SIZE + 1];
  uschar id, subid;
  uschar *endc;

  /* Read (first time) or top up the chars in the buffer if necessary.
  There will be only one read if we get all the available data (i.e. don't
  fill the buffer completely). */

  if (remaining < required && unfinished)
    {
    int len;
    int available = big_buffer_size - remaining;

    if (remaining > 0) memmove(big_buffer, ptr, remaining);

    ptr = big_buffer;
    endptr = big_buffer + remaining;
    len = read(fd, endptr, available);

    DEBUG(D_deliver) debug_printf("read() yielded %d\n", len);

    /* If the result is EAGAIN and the process is not complete, just
    stop reading any more and process what we have already. */

    if (len < 0)
      {
      if (!eop && errno == EAGAIN) len = 0; else
        {
        msg = string_sprintf("failed to read pipe from transport process "
          "%d for transport %s: %s", pid, addr->transport->driver_name,
          strerror(errno));
        break;
        }
      }

    /* If the length is zero (eof or no-more-data), just process what we
    already have. Note that if the process is still running and we have
    read all the data in the pipe (but less that "available") then we
    won't read any more, as "unfinished" will get set FALSE. */

    endptr += len;
    remaining += len;
    unfinished = len == available;
    }

  /* If we are at the end of the available data, exit the loop. */
  if (ptr >= endptr) break;

  /* copy and read header */
  memcpy(header, ptr, PIPE_HEADER_SIZE);
  header[PIPE_HEADER_SIZE] = '\0';
  id = header[0];
  subid = header[1];
  required = Ustrtol(header + 2, &endc, 10) + PIPE_HEADER_SIZE;     /* header + data */
  if (*endc)
    {
    msg = string_sprintf("failed to read pipe from transport process "
      "%d for transport %s: error reading size from header", pid, addr->transport->driver_name);
    done = TRUE;
    break;
    }

  DEBUG(D_deliver)
    debug_printf("header read  id:%c,subid:%c,size:%s,required:%d,remaining:%d,unfinished:%d\n",
                    id, subid, header+2, required, remaining, unfinished);

  /* is there room for the dataset we want to read ? */
  if (required > big_buffer_size - PIPE_HEADER_SIZE)
    {
    msg = string_sprintf("failed to read pipe from transport process "
      "%d for transport %s: big_buffer too small! required size=%d buffer size=%d", pid, addr->transport->driver_name,
      required, big_buffer_size - PIPE_HEADER_SIZE);
    done = TRUE;
    break;
    }

  /* we wrote all datasets with atomic write() calls
     remaining < required only happens if big_buffer was too small
     to get all available data from pipe. unfinished has to be true
     as well. */
  if (remaining < required)
    {
    if (unfinished)
      continue;
    msg = string_sprintf("failed to read pipe from transport process "
      "%d for transport %s: required size=%d > remaining size=%d and unfinished=false",
      pid, addr->transport->driver_name, required, remaining);
    done = TRUE;
    break;
    }

  /* step behind the header */
  ptr += PIPE_HEADER_SIZE;

  /* Handle each possible type of item, assuming the complete item is
  available in store. */

  switch (id)
    {
    /* Host items exist only if any hosts were marked unusable. Match
    up by checking the IP address. */

    case 'H':
    for (h = addrlist->host_list; h; h = h->next)
      {
      if (!h->address || Ustrcmp(h->address, ptr+2) != 0) continue;
      h->status = ptr[0];
      h->why = ptr[1];
      }
    ptr += 2;
    while (*ptr++);
    break;

    /* Retry items are sent in a preceding R item for each address. This is
    kept separate to keep each message short enough to guarantee it won't
    be split in the pipe. Hopefully, in the majority of cases, there won't in
    fact be any retry items at all.

    The complete set of retry items might include an item to delete a
    routing retry if there was a previous routing delay. However, routing
    retries are also used when a remote transport identifies an address error.
    In that case, there may also be an "add" item for the same key. Arrange
    that a "delete" item is dropped in favour of an "add" item. */

    case 'R':
    if (!addr) goto ADDR_MISMATCH;

    DEBUG(D_deliver|D_retry)
      debug_printf("reading retry information for %s from subprocess\n",
        ptr+1);

    /* Cut out any "delete" items on the list. */

    for (rp = &(addr->retries); (r = *rp); rp = &r->next)
      if (Ustrcmp(r->key, ptr+1) == 0)           /* Found item with same key */
        {
        if ((r->flags & rf_delete) == 0) break;  /* It was not "delete" */
        *rp = r->next;                           /* Excise a delete item */
        DEBUG(D_deliver|D_retry)
          debug_printf("  existing delete item dropped\n");
        }

    /* We want to add a delete item only if there is no non-delete item;
    however we still have to step ptr through the data. */

    if (!r || (*ptr & rf_delete) == 0)
      {
      r = store_get(sizeof(retry_item));
      r->next = addr->retries;
      addr->retries = r;
      r->flags = *ptr++;
      r->key = string_copy(ptr);
      while (*ptr++);
      memcpy(&(r->basic_errno), ptr, sizeof(r->basic_errno));
      ptr += sizeof(r->basic_errno);
      memcpy(&(r->more_errno), ptr, sizeof(r->more_errno));
      ptr += sizeof(r->more_errno);
      r->message = (*ptr)? string_copy(ptr) : NULL;
      DEBUG(D_deliver|D_retry)
        debug_printf("  added %s item\n",
          ((r->flags & rf_delete) == 0)? "retry" : "delete");
      }

    else
      {
      DEBUG(D_deliver|D_retry)
        debug_printf("  delete item not added: non-delete item exists\n");
      ptr++;
      while(*ptr++);
      ptr += sizeof(r->basic_errno) + sizeof(r->more_errno);
      }

    while(*ptr++);
    break;

    /* Put the amount of data written into the parlist block */

    case 'S':
    memcpy(&(p->transport_count), ptr, sizeof(transport_count));
    ptr += sizeof(transport_count);
    break;

    /* Address items are in the order of items on the address chain. We
    remember the current address value in case this function is called
    several times to empty the pipe in stages. Information about delivery
    over TLS is sent in a preceding X item for each address. We don't put
    it in with the other info, in order to keep each message short enough to
    guarantee it won't be split in the pipe. */

#ifdef SUPPORT_TLS
    case 'X':
    if (!addr) goto ADDR_MISMATCH;          /* Below, in 'A' handler */
    switch (subid)
      {
      case '1':
      addr->cipher = NULL;
      addr->peerdn = NULL;

      if (*ptr)
        addr->cipher = string_copy(ptr);
      while (*ptr++);
      if (*ptr)
        addr->peerdn = string_copy(ptr);
      break;

      case '2':
      if (*ptr)
        (void) tls_import_cert(ptr, &addr->peercert);
      else
        addr->peercert = NULL;
      break;

      case '3':
      if (*ptr)
        (void) tls_import_cert(ptr, &addr->ourcert);
      else
        addr->ourcert = NULL;
      break;

# ifndef DISABLE_OCSP
      case '4':
      addr->ocsp = OCSP_NOT_REQ;
      if (*ptr)
        addr->ocsp = *ptr - '0';
      break;
# endif
      }
    while (*ptr++);
    break;
#endif  /*SUPPORT_TLS*/

    case 'C':   /* client authenticator information */
    switch (subid)
      {
      case '1':
        addr->authenticator = (*ptr)? string_copy(ptr) : NULL;
        break;
      case '2':
        addr->auth_id = (*ptr)? string_copy(ptr) : NULL;
        break;
      case '3':
        addr->auth_sndr = (*ptr)? string_copy(ptr) : NULL;
        break;
      }
    while (*ptr++);
    break;

#ifndef DISABLE_PRDR
    case 'P':
    addr->flags |= af_prdr_used;
    break;
#endif

    case 'D':
    if (!addr) goto ADDR_MISMATCH;
    memcpy(&(addr->dsn_aware), ptr, sizeof(addr->dsn_aware));
    ptr += sizeof(addr->dsn_aware);
    DEBUG(D_deliver) debug_printf("DSN read: addr->dsn_aware = %d\n", addr->dsn_aware);
    break;

    case 'A':
    if (!addr)
      {
      ADDR_MISMATCH:
      msg = string_sprintf("address count mismatch for data read from pipe "
        "for transport process %d for transport %s", pid,
          addrlist->transport->driver_name);
      done = TRUE;
      break;
      }

    switch (subid)
      {
#ifdef SUPPORT_SOCKS
      case '2': /* proxy information; must arrive before A0 and applies to that addr XXX oops*/
        proxy_session = TRUE;   /*XXX shouod this be cleared somewhere? */
        if (*ptr == 0)
          ptr++;
        else
          {
          proxy_local_address = string_copy(ptr);
          while(*ptr++);
          memcpy(&proxy_local_port, ptr, sizeof(proxy_local_port));
          ptr += sizeof(proxy_local_port);
          }
        break;
#endif

#ifdef EXPERIMENTAL_DSN_INFO
      case '1': /* must arrive before A0, and applies to that addr */
                /* Two strings: smtp_greeting and helo_response */
        addr->smtp_greeting = string_copy(ptr);
        while(*ptr++);
        addr->helo_response = string_copy(ptr);
        while(*ptr++);
        break;
#endif

      case '0':
        addr->transport_return = *ptr++;
        addr->special_action = *ptr++;
        memcpy(&(addr->basic_errno), ptr, sizeof(addr->basic_errno));
        ptr += sizeof(addr->basic_errno);
        memcpy(&(addr->more_errno), ptr, sizeof(addr->more_errno));
        ptr += sizeof(addr->more_errno);
        memcpy(&(addr->flags), ptr, sizeof(addr->flags));
        ptr += sizeof(addr->flags);
        addr->message = (*ptr)? string_copy(ptr) : NULL;
        while(*ptr++);
        addr->user_message = (*ptr)? string_copy(ptr) : NULL;
        while(*ptr++);

        /* Always two strings for host information, followed by the port number and DNSSEC mark */

        if (*ptr != 0)
          {
          h = store_get(sizeof(host_item));
          h->name = string_copy(ptr);
          while (*ptr++);
          h->address = string_copy(ptr);
          while(*ptr++);
          memcpy(&(h->port), ptr, sizeof(h->port));
          ptr += sizeof(h->port);
          h->dnssec = *ptr == '2' ? DS_YES
                    : *ptr == '1' ? DS_NO
                    : DS_UNK;
          ptr++;
          addr->host_used = h;
          }
        else ptr++;

        /* Finished with this address */

        addr = addr->next;
        break;
      }
    break;

    /* Local interface address/port */
    case 'I':
    if (*ptr) sending_ip_address = string_copy(ptr);
    while (*ptr++) ;
    if (*ptr) sending_port = atoi(CS ptr);
    while (*ptr++) ;
    break;

    /* Z marks the logical end of the data. It is followed by '0' if
    continue_transport was NULL at the end of transporting, otherwise '1'.
    We need to know when it becomes NULL during a delivery down a passed SMTP
    channel so that we don't try to pass anything more down it. Of course, for
    most normal messages it will remain NULL all the time. */

    case 'Z':
    if (*ptr == '0')
      {
      continue_transport = NULL;
      continue_hostname = NULL;
      }
    done = TRUE;
    DEBUG(D_deliver) debug_printf("Z0%c item read\n", *ptr);
    break;

    /* Anything else is a disaster. */

    default:
    msg = string_sprintf("malformed data (%d) read from pipe for transport "
      "process %d for transport %s", ptr[-1], pid,
        addr->transport->driver_name);
    done = TRUE;
    break;
    }
  }

/* The done flag is inspected externally, to determine whether or not to
call the function again when the process finishes. */

p->done = done;

/* If the process hadn't finished, and we haven't seen the end of the data
or suffered a disaster, update the rest of the state, and return FALSE to
indicate "not finished". */

if (!eop && !done)
  {
  p->addr = addr;
  p->msg = msg;
  return FALSE;
  }

/* Close our end of the pipe, to prevent deadlock if the far end is still
pushing stuff into it. */

(void)close(fd);
p->fd = -1;

/* If we have finished without error, but haven't had data for every address,
something is wrong. */

if (!msg && addr)
  msg = string_sprintf("insufficient address data read from pipe "
    "for transport process %d for transport %s", pid,
      addr->transport->driver_name);

/* If an error message is set, something has gone wrong in getting back
the delivery data. Put the message into each address and freeze it. */

if (msg)
  for (addr = addrlist; addr; addr = addr->next)
    {
    addr->transport_return = DEFER;
    addr->special_action = SPECIAL_FREEZE;
    addr->message = msg;
    }

/* Return TRUE to indicate we have got all we need from this process, even
if it hasn't actually finished yet. */

return TRUE;
}



/*************************************************
*   Post-process a set of remote addresses       *
*************************************************/

/* Do what has to be done immediately after a remote delivery for each set of
addresses, then re-write the spool if necessary. Note that post_process_one
puts the address on an appropriate queue; hence we must fish off the next
one first. This function is also called if there is a problem with setting
up a subprocess to do a remote delivery in parallel. In this case, the final
argument contains a message, and the action must be forced to DEFER.

Argument:
   addr      pointer to chain of address items
   logflags  flags for logging
   msg       NULL for normal cases; -> error message for unexpected problems
   fallback  TRUE if processing fallback hosts

Returns:     nothing
*/

static void
remote_post_process(address_item *addr, int logflags, uschar *msg,
  BOOL fallback)
{
host_item *h;

/* If any host addresses were found to be unusable, add them to the unusable
tree so that subsequent deliveries don't try them. */

for (h = addr->host_list; h; h = h->next)
  if (h->address)
    if (h->status >= hstatus_unusable) tree_add_unusable(h);

/* Now handle each address on the chain. The transport has placed '=' or '-'
into the special_action field for each successful delivery. */

while (addr)
  {
  address_item *next = addr->next;

  /* If msg == NULL (normal processing) and the result is DEFER and we are
  processing the main hosts and there are fallback hosts available, put the
  address on the list for fallback delivery. */

  if (  addr->transport_return == DEFER
     && addr->fallback_hosts
     && !fallback
     && !msg
     )
    {
    addr->host_list = addr->fallback_hosts;
    addr->next = addr_fallback;
    addr_fallback = addr;
    DEBUG(D_deliver) debug_printf("%s queued for fallback host(s)\n", addr->address);
    }

  /* If msg is set (=> unexpected problem), set it in the address before
  doing the ordinary post processing. */

  else
    {
    if (msg)
      {
      addr->message = msg;
      addr->transport_return = DEFER;
      }
    (void)post_process_one(addr, addr->transport_return, logflags,
      DTYPE_TRANSPORT, addr->special_action);
    }

  /* Next address */

  addr = next;
  }

/* If we have just delivered down a passed SMTP channel, and that was
the last address, the channel will have been closed down. Now that
we have logged that delivery, set continue_sequence to 1 so that
any subsequent deliveries don't get "*" incorrectly logged. */

if (!continue_transport) continue_sequence = 1;
}



/*************************************************
*     Wait for one remote delivery subprocess    *
*************************************************/

/* This function is called while doing remote deliveries when either the
maximum number of processes exist and we need one to complete so that another
can be created, or when waiting for the last ones to complete. It must wait for
the completion of one subprocess, empty the control block slot, and return a
pointer to the address chain.

Arguments:    none
Returns:      pointer to the chain of addresses handled by the process;
              NULL if no subprocess found - this is an unexpected error
*/

static address_item *
par_wait(void)
{
int poffset, status;
address_item *addr, *addrlist;
pid_t pid;

set_process_info("delivering %s: waiting for a remote delivery subprocess "
  "to finish", message_id);

/* Loop until either a subprocess completes, or there are no subprocesses in
existence - in which case give an error return. We cannot proceed just by
waiting for a completion, because a subprocess may have filled up its pipe, and
be waiting for it to be emptied. Therefore, if no processes have finished, we
wait for one of the pipes to acquire some data by calling select(), with a
timeout just in case.

The simple approach is just to iterate after reading data from a ready pipe.
This leads to non-ideal behaviour when the subprocess has written its final Z
item, closed the pipe, and is in the process of exiting (the common case). A
call to waitpid() yields nothing completed, but select() shows the pipe ready -
reading it yields EOF, so you end up with busy-waiting until the subprocess has
actually finished.

To avoid this, if all the data that is needed has been read from a subprocess
after select(), an explicit wait() for it is done. We know that all it is doing
is writing to the pipe and then exiting, so the wait should not be long.

The non-blocking waitpid() is to some extent just insurance; if we could
reliably detect end-of-file on the pipe, we could always know when to do a
blocking wait() for a completed process. However, because some systems use
NDELAY, which doesn't distinguish between EOF and pipe empty, it is easier to
use code that functions without the need to recognize EOF.

There's a double loop here just in case we end up with a process that is not in
the list of remote delivery processes. Something has obviously gone wrong if
this is the case. (For example, a process that is incorrectly left over from
routing or local deliveries might be found.) The damage can be minimized by
looping back and looking for another process. If there aren't any, the error
return will happen. */

for (;;)   /* Normally we do not repeat this loop */
  {
  while ((pid = waitpid(-1, &status, WNOHANG)) <= 0)
    {
    struct timeval tv;
    fd_set select_pipes;
    int maxpipe, readycount;

    /* A return value of -1 can mean several things. If errno != ECHILD, it
    either means invalid options (which we discount), or that this process was
    interrupted by a signal. Just loop to try the waitpid() again.

    If errno == ECHILD, waitpid() is telling us that there are no subprocesses
    in existence. This should never happen, and is an unexpected error.
    However, there is a nasty complication when running under Linux. If "strace
    -f" is being used under Linux to trace this process and its children,
    subprocesses are "stolen" from their parents and become the children of the
    tracing process. A general wait such as the one we've just obeyed returns
    as if there are no children while subprocesses are running. Once a
    subprocess completes, it is restored to the parent, and waitpid(-1) finds
    it. Thanks to Joachim Wieland for finding all this out and suggesting a
    palliative.

    This does not happen using "truss" on Solaris, nor (I think) with other
    tracing facilities on other OS. It seems to be specific to Linux.

    What we do to get round this is to use kill() to see if any of our
    subprocesses are still in existence. If kill() gives an OK return, we know
    it must be for one of our processes - it can't be for a re-use of the pid,
    because if our process had finished, waitpid() would have found it. If any
    of our subprocesses are in existence, we proceed to use select() as if
    waitpid() had returned zero. I think this is safe. */

    if (pid < 0)
      {
      if (errno != ECHILD) continue;   /* Repeats the waitpid() */

      DEBUG(D_deliver)
        debug_printf("waitpid() returned -1/ECHILD: checking explicitly "
          "for process existence\n");

      for (poffset = 0; poffset < remote_max_parallel; poffset++)
        {
        if ((pid = parlist[poffset].pid) != 0 && kill(pid, 0) == 0)
          {
          DEBUG(D_deliver) debug_printf("process %d still exists: assume "
            "stolen by strace\n", (int)pid);
          break;   /* With poffset set */
          }
        }

      if (poffset >= remote_max_parallel)
        {
        DEBUG(D_deliver) debug_printf("*** no delivery children found\n");
        return NULL;   /* This is the error return */
        }
      }

    /* A pid value greater than 0 breaks the "while" loop. A negative value has
    been handled above. A return value of zero means that there is at least one
    subprocess, but there are no completed subprocesses. See if any pipes are
    ready with any data for reading. */

    DEBUG(D_deliver) debug_printf("selecting on subprocess pipes\n");

    maxpipe = 0;
    FD_ZERO(&select_pipes);
    for (poffset = 0; poffset < remote_max_parallel; poffset++)
      {
      if (parlist[poffset].pid != 0)
        {
        int fd = parlist[poffset].fd;
        FD_SET(fd, &select_pipes);
        if (fd > maxpipe) maxpipe = fd;
        }
      }

    /* Stick in a 60-second timeout, just in case. */

    tv.tv_sec = 60;
    tv.tv_usec = 0;

    readycount = select(maxpipe + 1, (SELECT_ARG2_TYPE *)&select_pipes,
         NULL, NULL, &tv);

    /* Scan through the pipes and read any that are ready; use the count
    returned by select() to stop when there are no more. Select() can return
    with no processes (e.g. if interrupted). This shouldn't matter.

    If par_read_pipe() returns TRUE, it means that either the terminating Z was
    read, or there was a disaster. In either case, we are finished with this
    process. Do an explicit wait() for the process and break the main loop if
    it succeeds.

    It turns out that we have to deal with the case of an interrupted system
    call, which can happen on some operating systems if the signal handling is
    set up to do that by default. */

    for (poffset = 0;
         readycount > 0 && poffset < remote_max_parallel;
         poffset++)
      {
      if (  (pid = parlist[poffset].pid) != 0
         && FD_ISSET(parlist[poffset].fd, &select_pipes)
         )
        {
        readycount--;
        if (par_read_pipe(poffset, FALSE))    /* Finished with this pipe */
          {
          for (;;)                            /* Loop for signals */
            {
            pid_t endedpid = waitpid(pid, &status, 0);
            if (endedpid == pid) goto PROCESS_DONE;
            if (endedpid != (pid_t)(-1) || errno != EINTR)
              log_write(0, LOG_MAIN|LOG_PANIC_DIE, "Unexpected error return "
                "%d (errno = %d) from waitpid() for process %d",
                (int)endedpid, errno, (int)pid);
            }
          }
        }
      }

    /* Now go back and look for a completed subprocess again. */
    }

  /* A completed process was detected by the non-blocking waitpid(). Find the
  data block that corresponds to this subprocess. */

  for (poffset = 0; poffset < remote_max_parallel; poffset++)
    if (pid == parlist[poffset].pid) break;

  /* Found the data block; this is a known remote delivery process. We don't
  need to repeat the outer loop. This should be what normally happens. */

  if (poffset < remote_max_parallel) break;

  /* This situation is an error, but it's probably better to carry on looking
  for another process than to give up (as we used to do). */

  log_write(0, LOG_MAIN|LOG_PANIC, "Process %d finished: not found in remote "
    "transport process list", pid);
  }  /* End of the "for" loop */

/* Come here when all the data was completely read after a select(), and
the process in pid has been wait()ed for. */

PROCESS_DONE:

DEBUG(D_deliver)
  {
  if (status == 0)
    debug_printf("remote delivery process %d ended\n", (int)pid);
  else
    debug_printf("remote delivery process %d ended: status=%04x\n", (int)pid,
      status);
  }

set_process_info("delivering %s", message_id);

/* Get the chain of processed addresses */

addrlist = parlist[poffset].addrlist;

/* If the process did not finish cleanly, record an error and freeze (except
for SIGTERM, SIGKILL and SIGQUIT), and also ensure the journal is not removed,
in case the delivery did actually happen. */

if ((status & 0xffff) != 0)
  {
  uschar *msg;
  int msb = (status >> 8) & 255;
  int lsb = status & 255;
  int code = (msb == 0)? (lsb & 0x7f) : msb;

  msg = string_sprintf("%s transport process returned non-zero status 0x%04x: "
    "%s %d",
    addrlist->transport->driver_name,
    status,
    (msb == 0)? "terminated by signal" : "exit code",
    code);

  if (msb != 0 || (code != SIGTERM && code != SIGKILL && code != SIGQUIT))
    addrlist->special_action = SPECIAL_FREEZE;

  for (addr = addrlist; addr; addr = addr->next)
    {
    addr->transport_return = DEFER;
    addr->message = msg;
    }

  remove_journal = FALSE;
  }

/* Else complete reading the pipe to get the result of the delivery, if all
the data has not yet been obtained. */

else if (!parlist[poffset].done) (void)par_read_pipe(poffset, TRUE);

/* Put the data count and return path into globals, mark the data slot unused,
decrement the count of subprocesses, and return the address chain. */

transport_count = parlist[poffset].transport_count;
used_return_path = parlist[poffset].return_path;
parlist[poffset].pid = 0;
parcount--;
return addrlist;
}



/*************************************************
*      Wait for subprocesses and post-process    *
*************************************************/

/* This function waits for subprocesses until the number that are still running
is below a given threshold. For each complete subprocess, the addresses are
post-processed. If we can't find a running process, there is some shambles.
Better not bomb out, as that might lead to multiple copies of the message. Just
log and proceed as if all done.

Arguments:
  max         maximum number of subprocesses to leave running
  fallback    TRUE if processing fallback hosts

Returns:      nothing
*/

static void
par_reduce(int max, BOOL fallback)
{
while (parcount > max)
  {
  address_item *doneaddr = par_wait();
  if (!doneaddr)
    {
    log_write(0, LOG_MAIN|LOG_PANIC,
      "remote delivery process count got out of step");
    parcount = 0;
    }
  else
    {
    transport_instance * tp = doneaddr->transport;
    if (tp->max_parallel)
      enq_end(string_sprintf("tpt-serialize-%s", tp->name));

    remote_post_process(doneaddr, LOG_MAIN, NULL, fallback);
    }
  }
}




static void
rmt_dlv_checked_write(int fd, char id, char subid, void * buf, int size)
{
uschar  writebuffer[PIPE_HEADER_SIZE + BIG_BUFFER_SIZE];
int header_length;
int ret;

/* we assume that size can't get larger then BIG_BUFFER_SIZE which currently is set to 16k */
/* complain to log if someone tries with buffer sizes we can't handle*/

if (size > 99999)
  {
  log_write(0, LOG_MAIN|LOG_PANIC_DIE,
    "Failed writing transport result to pipe: can't handle buffers > 99999 bytes. truncating!\n");
  size = 99999;
  }

/* to keep the write() atomic we build header in writebuffer and copy buf behind */
/* two write() calls would increase the complexity of reading from pipe */

/* convert size to human readable string prepended by id and subid */
header_length = snprintf(CS writebuffer, PIPE_HEADER_SIZE+1, "%c%c%05d", id, subid, size);
if (header_length != PIPE_HEADER_SIZE)
  {
  log_write(0, LOG_MAIN|LOG_PANIC_DIE, "header snprintf failed\n");
  writebuffer[0] = '\0';
  }

DEBUG(D_deliver) debug_printf("header write id:%c,subid:%c,size:%d,final:%s\n",
                                 id, subid, size, writebuffer);

if (buf && size > 0)
  memcpy(writebuffer + PIPE_HEADER_SIZE, buf, size);

size += PIPE_HEADER_SIZE;
if ((ret = write(fd, writebuffer, size)) != size)
  log_write(0, LOG_MAIN|LOG_PANIC_DIE, "Failed writing transport result to pipe: %s\n",
    ret == -1 ? strerror(errno) : "short write");
}

/*************************************************
*           Do remote deliveries                 *
*************************************************/

/* This function is called to process the addresses in addr_remote. We must
pick off the queue all addresses that have the same transport, remote
destination, and errors address, and hand them to the transport in one go,
subject to some configured limitations. If this is a run to continue delivering
to an existing delivery channel, skip all but those addresses that can go to
that channel. The skipped addresses just get deferred.

If mua_wrapper is set, all addresses must be able to be sent in a single
transaction. If not, this function yields FALSE.

In Exim 4, remote deliveries are always done in separate processes, even
if remote_max_parallel = 1 or if there's only one delivery to do. The reason
is so that the base process can retain privilege. This makes the
implementation of fallback transports feasible (though not initially done.)

We create up to the configured number of subprocesses, each of which passes
back the delivery state via a pipe. (However, when sending down an existing
connection, remote_max_parallel is forced to 1.)

Arguments:
  fallback  TRUE if processing fallback hosts

Returns:    TRUE normally
            FALSE if mua_wrapper is set and the addresses cannot all be sent
              in one transaction
*/

static BOOL
do_remote_deliveries(BOOL fallback)
{
int parmax;
int delivery_count;
int poffset;

parcount = 0;    /* Number of executing subprocesses */

/* When sending down an existing channel, only do one delivery at a time.
We use a local variable (parmax) to hold the maximum number of processes;
this gets reduced from remote_max_parallel if we can't create enough pipes. */

if (continue_transport) remote_max_parallel = 1;
parmax = remote_max_parallel;

/* If the data for keeping a list of processes hasn't yet been
set up, do so. */

if (!parlist)
  {
  parlist = store_get(remote_max_parallel * sizeof(pardata));
  for (poffset = 0; poffset < remote_max_parallel; poffset++)
    parlist[poffset].pid = 0;
  }

/* Now loop for each remote delivery */

for (delivery_count = 0; addr_remote; delivery_count++)
  {
  pid_t pid;
  uid_t uid;
  gid_t gid;
  int pfd[2];
  int address_count = 1;
  int address_count_max;
  BOOL multi_domain;
  BOOL use_initgroups;
  BOOL pipe_done = FALSE;
  transport_instance *tp;
  address_item **anchor = &addr_remote;
  address_item *addr = addr_remote;
  address_item *last = addr;
  address_item *next;
  uschar * panicmsg;
  uschar * serialize_key = NULL;

  /* Pull the first address right off the list. */

  addr_remote = addr->next;
  addr->next = NULL;

  DEBUG(D_deliver|D_transport)
    debug_printf("--------> %s <--------\n", addr->address);

  /* If no transport has been set, there has been a big screw-up somewhere. */

  if (!(tp = addr->transport))
    {
    disable_logging = FALSE;  /* Jic */
    panicmsg = US"No transport set by router";
    goto panic_continue;
    }

  /* Check that this base address hasn't previously been delivered to this
  transport. The check is necessary at this point to handle homonymic addresses
  correctly in cases where the pattern of redirection changes between delivery
  attempts. Non-homonymic previous delivery is detected earlier, at routing
  time. */

  if (previously_transported(addr, FALSE)) continue;

  /* Force failure if the message is too big. */

  if (tp->message_size_limit)
    {
    int rc = check_message_size(tp, addr);
    if (rc != OK)
      {
      addr->transport_return = rc;
      remote_post_process(addr, LOG_MAIN, NULL, fallback);
      continue;
      }
    }

  /* Get the flag which specifies whether the transport can handle different
  domains that nevertheless resolve to the same set of hosts. If it needs
  expanding, get variables set: $address_data, $domain_data, $localpart_data,
  $host, $host_address, $host_port. */
  if (tp->expand_multi_domain)
    deliver_set_expansions(addr);

  if (exp_bool(addr, US"transport", tp->name, D_transport,
                US"multi_domain", tp->multi_domain, tp->expand_multi_domain,
                &multi_domain) != OK)
    {
    deliver_set_expansions(NULL);
    panicmsg = addr->message;
    goto panic_continue;
    }

  /* Get the maximum it can handle in one envelope, with zero meaning
  unlimited, which is forced for the MUA wrapper case. */

  address_count_max = tp->max_addresses;
  if (address_count_max == 0 || mua_wrapper) address_count_max = 999999;


  /************************************************************************/
  /*****    This is slightly experimental code, but should be safe.   *****/

  /* The address_count_max value is the maximum number of addresses that the
  transport can send in one envelope. However, the transport must be capable of
  dealing with any number of addresses. If the number it gets exceeds its
  envelope limitation, it must send multiple copies of the message. This can be
  done over a single connection for SMTP, so uses less resources than making
  multiple connections. On the other hand, if remote_max_parallel is greater
  than one, it is perhaps a good idea to use parallel processing to move the
  message faster, even if that results in multiple simultaneous connections to
  the same host.