Exim Internet Mailer

1. Concurrent delivery

If two messages arrive at almost the same time, and both are routed to a pipe delivery, the two pipe transports may be run concurrently. You must ensure that any pipe commands you set up are robust against this happening. If the commands write to a file, the exim_lock utility might be of use. Alternatively the max_parallel option could be used with a value of "1" to enforce serialization.

2. Returned status and data

If the command exits with a non-zero return code, the delivery is deemed to have failed, unless either the ignore_status option is set (in which case the return code is treated as zero), or the return code is one of those listed in the temp_errors option, which are interpreted as meaning “try again later”. In this case, delivery is deferred. Details of a permanent failure are logged, but are not included in the bounce message, which merely contains “local delivery failed”.

If the command exits on a signal and the freeze_signal option is set then the message will be frozen in the queue. If that option is not set, a bounce will be sent as normal.

If the return code is greater than 128 and the command being run is a shell script, it normally means that the script was terminated by a signal whose value is the return code minus 128. The freeze_signal option does not apply in this case.

If Exim is unable to run the command (that is, if execve() fails), the return code is set to 127. This is the value that a shell returns if it is asked to run a non-existent command. The wording for the log line suggests that a non-existent command may be the problem.

The return_output option can affect the result of a pipe delivery. If it is set and the command produces any output on its standard output or standard error streams, the command is considered to have failed, even if it gave a zero return code or if ignore_status is set. The output from the command is included as part of the bounce message. The return_fail_output option is similar, except that output is returned only when the command exits with a failure return code, that is, a value other than zero or a code that matches temp_errors.

3. How the command is run

The command line is (by default) broken down into a command name and arguments by the pipe transport itself. The allow_commands and restrict_to_path options can be used to restrict the commands that may be run.

Unquoted arguments are delimited by white space. If an argument appears in double quotes, backslash is interpreted as an escape character in the usual way. If an argument appears in single quotes, no escaping is done.

String expansion is applied to the command line except when it comes from a traditional .forward file (commands from a filter file are expanded). The expansion is applied to each argument in turn rather than to the whole line. For this reason, any string expansion item that contains white space must be quoted so as to be contained within a single argument. A setting such as

command = /some/path ${if eq{$local_part}{postmaster}{xx}{yy}}

will not work, because the expansion item gets split between several arguments. You have to write

command = /some/path "${if eq{$local_part}{postmaster}{xx}{yy}}"

to ensure that it is all in one argument. The expansion is done in this way, argument by argument, so that the number of arguments cannot be changed as a result of expansion, and quotes or backslashes in inserted variables do not interact with external quoting. However, this leads to problems if you want to generate multiple arguments (or the command name plus arguments) from a single expansion. In this situation, the simplest solution is to use a shell. For example:

command = /bin/sh -c ${lookup{$local_part}lsearch{/some/file}}

Special handling takes place when an argument consists of precisely the text $pipe_addresses (no quotes). This is not a general expansion variable; the only place this string is recognized is when it appears as an argument for a pipe or transport filter command. It causes each address that is being handled to be inserted in the argument list at that point as a separate argument. This avoids any problems with spaces or shell metacharacters, and is of use when a pipe transport is handling groups of addresses in a batch.

If force_command is enabled on the transport, special handling takes place for an argument that consists of precisely the text $address_pipe. It is handled similarly to $pipe_addresses above. It is expanded and each argument is inserted in the argument list at that point as a separate argument. The $address_pipe item does not need to be the only item in the argument; in fact, if it were then force_command should behave as a no-op. Rather, it should be used to adjust the command run while preserving the argument vector separation.

After splitting up into arguments and expansion, the resulting command is run in a subprocess directly from the transport, not under a shell. The message that is being delivered is supplied on the standard input, and the standard output and standard error are both connected to a single pipe that is read by Exim. The max_output option controls how much output the command may produce, and the return_output and return_fail_output options control what is done with it.

Not running the command under a shell (by default) lessens the security risks in cases when a command from a user’s filter file is built out of data that was taken from an incoming message. If a shell is required, it can of course be explicitly specified as the command to be run. However, there are circumstances where existing commands (for example, in .forward files) expect to be run under a shell and cannot easily be modified. To allow for these cases, there is an option called use_shell, which changes the way the pipe transport works. Instead of breaking up the command line as just described, it expands it as a single string and passes the result to /bin/sh. The restrict_to_path option and the $pipe_addresses facility cannot be used with use_shell, and the whole mechanism is inherently less secure.

4. Environment variables

The environment variables listed below are set up when the command is invoked. This list is a compromise for maximum compatibility with other MTAs. Note that the environment option can be used to add additional variables to this environment. The environment for the pipe transport is not subject to the add_environment and keep_environment main config options.

Note: Using enviroment variables loses track of tainted data. Writers of pipe transport commands should be wary of data supplied by potential attackers.

DOMAIN               the domain of the address
HOME                 the home directory, if set
HOST                 the host name when called from a router (see below)
LOCAL_PART           see below
LOCAL_PART_PREFIX    see below
LOCAL_PART_SUFFIX    see below
LOGNAME              see below
MESSAGE_ID           Exim’s local ID for the message
PATH                 as specified by the path option below
QUALIFY_DOMAIN       the sender qualification domain
RECIPIENT            the complete recipient address
SENDER               the sender of the message (empty if a bounce)
SHELL                /bin/sh
TZ                   the value of the timezone option, if set
USER                 see below

When a pipe transport is called directly from (for example) an accept router, LOCAL_PART is set to the local part of the address. When it is called as a result of a forward or alias expansion, LOCAL_PART is set to the local part of the address that was expanded. In both cases, any affixes are removed from the local part, and made available in LOCAL_PART_PREFIX and LOCAL_PART_SUFFIX, respectively. LOGNAME and USER are set to the same value as LOCAL_PART for compatibility with other MTAs.

HOST is set only when a pipe transport is called from a router that associates hosts with an address, typically when using pipe as a pseudo-remote transport. HOST is set to the first host name specified by the router.

If the transport’s generic home_directory option is set, its value is used for the HOME environment variable. Otherwise, a home directory may be set by the router’s transport_home_directory option, which defaults to the user’s home directory if check_local_user is set.

5. Private options for pipe

The string is expanded, and is then interpreted as a colon-separated list of permitted commands. If restrict_to_path is not set, the only commands permitted are those in the allow_commands list. They need not be absolute paths; the path option is still used for relative paths. If restrict_to_path is set with allow_commands, the command must either be in the allow_commands list, or a name without any slashes that is found on the path. In other words, if neither allow_commands nor restrict_to_path is set, there is no restriction on the command, but otherwise only commands that are permitted by one or the other are allowed. For example, if

allow_commands = /usr/bin/vacation

and restrict_to_path is not set, the only permitted command is /usr/bin/vacation. The allow_commands option may not be set if use_shell is set.

See the description of local delivery batching in chapter 25.

This limits the number of addresses that can be handled in a single delivery. See the description of local delivery batching in chapter 25.

As pipe writes the message, the start of each line is tested for matching check_string, and if it does, the initial matching characters are replaced by the contents of escape_string, provided both are set. The value of check_string is a literal string, not a regular expression, and the case of any letters it contains is significant. When use_bsmtp is set, the contents of check_string and escape_string are forced to values that implement the SMTP escaping protocol. Any settings made in the configuration file are ignored.

This option need not be set when pipe is being used to deliver to pipes obtained directly from address redirections. In other cases, the option must be set, to provide a command to be run. It need not yield an absolute path (see the path option below). The command is split up into separate arguments by Exim, and each argument is separately expanded, as described in section 29.3 above.

No part of the resulting command may be tainted.

This option is used to add additional variables to the environment in which the command runs (see section 29.4 for the default list). Its value is a string which is expanded, and then interpreted as a colon-separated list of environment settings of the form <name>=<value>.

See check_string above.

Failure to exec the command in a pipe transport is by default treated like any other failure while running the command. However, if freeze_exec_fail is set, failure to exec is treated specially, and causes the message to be frozen, whatever the setting of ignore_status.

Normally if the process run by a command in a pipe transport exits on a signal, a bounce message is sent. If freeze_signal is set, the message will be frozen in Exim’s queue instead.

Normally when a router redirects an address directly to a pipe command the command option on the transport is ignored. If force_command is set, the command option will used. This is especially useful for forcing a wrapper or additional argument to be added to the command. For example:

command = /usr/bin/remote_exec myhost -- $address_pipe
force_command

Note that $address_pipe is handled specially in command when force_command is set, expanding out to the original argument vector as separate items, similarly to a Unix shell "$@" construct.

If this option is true, the status returned by the subprocess that is set up to run the command is ignored, and Exim behaves as if zero had been returned. Otherwise, a non-zero status or termination by signal causes an error return from the transport unless the status value is one of those listed in temp_errors; these cause the delivery to be deferred and tried again later.

Note: This option does not apply to timeouts, which do not return a status. See the timeout_defer option for how timeouts are handled.

If this option is set, and the status returned by the command is one of the codes listed in temp_errors (that is, delivery was deferred), and any output was produced on stdout or stderr, the first line of it is written to the main log.

If this option is set, and the command returns any output on stdout or stderr, and also ends with a return code that is neither zero nor one of the return codes listed in temp_errors (that is, the delivery failed), the first line of output is written to the main log. This option and log_output are mutually exclusive. Only one of them may be set.

If this option is set and the command returns any output on stdout or stderr, the first line of output is written to the main log, whatever the return code. This option and log_fail_output are mutually exclusive. Only one of them may be set.

This specifies the maximum amount of output that the command may produce on its standard output and standard error file combined. If the limit is exceeded, the process running the command is killed. This is intended as a safety measure to catch runaway processes. The limit is applied independently of the settings of the options that control what is done with such output (for example, return_output). Because of buffering effects, the amount of output may exceed the limit by a small amount before Exim notices.

The string specified here is expanded and output at the start of every message. The default is unset if use_bsmtp is set. Otherwise it is

message_prefix = \
  From ${if def:return_path{$return_path}{MAILER-DAEMON}}\
  ${tod_bsdinbox}\n

This is required by the commonly used /usr/bin/vacation program. However, it must not be present if delivery is to the Cyrus IMAP server, or to the tmail local delivery agent. The prefix can be suppressed by setting

message_prefix =

Note: If you set use_crlf true, you must change any occurrences of \n to \r\n in message_prefix.

The string specified here is expanded and output at the end of every message. The default is unset if use_bsmtp is set. Otherwise it is a single newline. The suffix can be suppressed by setting

message_suffix =

Note: If you set use_crlf true, you must change any occurrences of \n to \r\n in message_suffix.

This option is expanded and specifies the string that is set up in the PATH environment variable of the subprocess. If the command option does not yield an absolute path name, the command is sought in the PATH directories, in the usual way. Warning: This does not apply to a command specified as a transport filter.

Normally Exim inhibits core-dumps during delivery. If you have a need to get a core-dump of a pipe command, enable this command. This enables core-dumps during delivery and affects both the Exim binary and the pipe command run. It is recommended that this option remain off unless and until you have a need for it and that this only be enabled when needed, as the risk of excessive resource consumption can be quite high. Note also that Exim is typically installed as a setuid binary and most operating systems will inhibit coredumps of these by default, so further OS-specific action may be required.

If the generic user option is not set and this option is true, the delivery process is run under the uid that was in force when Exim was originally called to accept the message. If the group id is not otherwise set (via the generic group option), the gid that was in force when Exim was originally called to accept the message is used.

When this option is set, any command name not listed in allow_commands must contain no slashes. The command is searched for only in the directories listed in the path option. This option is intended for use in the case when a pipe command has been generated from a user’s .forward file. This is usually handled by a pipe transport called address_pipe.

If this option is true, and the command produced any output and ended with a return code other than zero or one of the codes listed in temp_errors (that is, the delivery failed), the output is returned in the bounce message. However, if the message has a null sender (that is, it is itself a bounce message), output from the command is discarded. This option and return_output are mutually exclusive. Only one of them may be set.

If this option is true, and the command produced any output, the delivery is deemed to have failed whatever the return code from the command, and the output is returned in the bounce message. Otherwise, the output is just discarded. However, if the message has a null sender (that is, it is a bounce message), output from the command is always discarded, whatever the setting of this option. This option and return_fail_output are mutually exclusive. Only one of them may be set.

This option contains either a colon-separated list of numbers, or a single asterisk. If ignore_status is false and return_output is not set, and the command exits with a non-zero return code, the failure is treated as temporary and the delivery is deferred if the return code matches one of the numbers, or if the setting is a single asterisk. Otherwise, non-zero return codes are treated as permanent errors. The default setting contains the codes defined by EX_TEMPFAIL and EX_CANTCREAT in sysexits.h. If Exim is compiled on a system that does not define these macros, it assumes values of 75 and 73, respectively.

If the command fails to complete within this time, it is killed. This normally causes the delivery to fail (but see timeout_defer). A zero time interval specifies no timeout. In order to ensure that any subprocesses created by the command are also killed, Exim makes the initial process a process group leader, and kills the whole process group on a timeout. However, this can be defeated if one of the processes starts a new process group.

A timeout in a pipe transport, either in the command that the transport runs, or in a transport filter that is associated with it, is by default treated as a hard error, and the delivery fails. However, if timeout_defer is set true, both kinds of timeout become temporary errors, causing the delivery to be deferred.

This specifies the umask setting for the subprocess that runs the command.

If this option is set true, the pipe transport writes messages in “batch SMTP” format, with the envelope sender and recipient(s) included as SMTP commands. If you want to include a leading HELO command with such messages, you can do so by setting the message_prefix option. See section 49.10 for details of batch SMTP.

This option is available only when Exim is running on FreeBSD, NetBSD, or BSD/OS. If it is set true, the setclassresources() function is used to set resource limits when a pipe transport is run to perform a delivery. The limits for the uid under which the pipe is to run are obtained from the login class database.

This option causes lines to be terminated with the two-character CRLF sequence (carriage return, linefeed) instead of just a linefeed character. In the case of batched SMTP, the byte sequence written to the pipe is then an exact image of what would be sent down a real SMTP connection.

The contents of the message_prefix and message_suffix options are written verbatim, so must contain their own carriage return characters if these are needed. When use_bsmtp is not set, the default values for both message_prefix and message_suffix end with a single linefeed, so their values must be changed to end with \r\n if use_crlf is set.

If this option is set, it causes the command to be passed to /bin/sh instead of being run directly from the transport, as described in section 29.3. This is less secure, but is needed in some situations where the command is expected to be run under a shell and cannot easily be modified. The allow_commands and restrict_to_path options, and the $pipe_addresses facility are incompatible with use_shell. The command is expanded as a single string, and handed to /bin/sh as data for its -c option.

6. Using an external local delivery agent

The pipe transport can be used to pass all messages that require local delivery to a separate local delivery agent such as procmail. When doing this, care must be taken to ensure that the pipe is run under an appropriate uid and gid. In some configurations one wants this to be a uid that is trusted by the delivery agent to supply the correct sender of the message. It may be necessary to recompile or reconfigure the delivery agent so that it trusts an appropriate user. The following is an example transport and router configuration for procmail:

# transport
procmail_pipe:
  driver = pipe
  command = /usr/local/bin/procmail -d $local_part_data
  return_path_add
  delivery_date_add
  envelope_to_add
  check_string = "From "
  escape_string = ">From "
  umask = 077
  user = $local_part_data
  group = mail

# router
procmail:
  driver = accept
  check_local_user
  transport = procmail_pipe

In this example, the pipe is run as the local user, but with the group set to mail. An alternative is to run the pipe as a specific user such as mail or exim, but in this case you must arrange for procmail to trust that user to supply a correct sender address. If you do not specify either a group or a user option, the pipe command is run as the local user. The home directory is the user’s home directory by default.

Note: The command that the pipe transport runs does not begin with

IFS=" "

as shown in some procmail documentation, because Exim does not by default use a shell to run pipe commands.

The next example shows a transport and a router for a system where local deliveries are handled by the Cyrus IMAP server.

# transport
local_delivery_cyrus:
  driver = pipe
  command = /usr/cyrus/bin/deliver \
            -- $local_part_data
  user = cyrus
  group = mail
  return_output
  log_output
  message_prefix =
  message_suffix =

# router
local_user_cyrus:
  driver = accept
  check_local_user
  transport = local_delivery_cyrus

Note the unsetting of message_prefix and message_suffix, and the use of return_output to cause any text written by Cyrus to be returned to the sender.