NAME

mailmunge-multiplexor - Process pool controller for mail filters.

SYNOPSIS

mailmunge-multiplexor [options]

DESCRIPTION

mailmunge-multiplexor manages a pool of Perl processes for scanning e-mail. It is designed to work in conjunction with mailmunge and your Perl filter program.

mailmunge-multiplexor opens a UNIX-domain socket and listens for requests for work from mailmunge. As requests come in, mailmunge-multiplexor creates Perl processes as needed to scan mail. The Perl processes are not killed when scanning is completed, but continue to run in a loop. Perl processes are re-used for subsequent e-mail messages. This eliminates the large overhead of starting a new Perl process for each incoming message.

To avoid memory leaks, the Perl processes are killed after they have handled some number of scans.

OPTIONS

-U user

Runs the multiplexor as user rather than root. This option is mandatory, and must match the -U option supplied to mailmunge.

-m minWorkers

The minimum number of Perl processes to keep running at all times. The default is zero.

-x maxWorkers

The maximum number of Perl processes to run simultaneously. If a request comes in and all processes are busy, a temporary failure is signalled to the SMTP peer. The default is 2.

-r maxRequests

The maximum number of requests a given process handles before it is killed and a replacement started. The default is 500.

-i idleTime

The idle time in seconds after which to kill of excess Perl processes. That is, if the process is idle for longer than this time, and there are more than minWorkers running, the process is killed. Note that this is implemented as a timer that ticks every idleTime seconds; therefore, processes may be idle for up to twice this time before they are killed. The default for idleTime is 300 seconds.

-V maxLifetime

The maximum lifetime in seconds of a worker before it is killed and a replacement started. The default is -1, which signifies no maximum lifetime. Note that the lifetime check is done only when a worker becomes idle after a request, or every time the idle-timeout check is made. On a very quiet system, workers may live for longer than maxLifetime by as much as idleTime. Note also that the lifetime is measured not from when the worker started running, but from when it was assigned its very first request. A completely-idle worker that has never processed any requests will not be terminated by the maxLifetime setting.

-b busyTime

The longest a Perl process is allowed to spend scanning an e-mail before it is declared hung up and killed. The default is 120 seconds.

-Z

This option specifies that the multiplexor should accept and process "status updates" from busy workers. Note that this consumes one extra file descriptor per worker, plus a small amount of CPU time per status update.

-c cmdTime

The timeout for communication between mailmunge-multiplexor and mailmunge, or between mailmunge-multiplexor and a Perl scanning process. The default is 10 seconds. This timeout should be kept quite short.

-w waitTime

When mailmunge-multiplexor starts the initial workers, or needs to bring the number of running workers up to the number defined by the -m option, it does not start all the workers at once, because this could overload your server. Instead, it starts one worker every waitTime seconds. The default value for waitTime is 1.

-W waitTime

If you use this option, mailmunge-multiplexor will never activate a worker until waitTime seconds have elapsed since the last worker activation. This could result in mail being tempfailed if worker activations do not keep pace with incoming mail. However, it may be preferable to tempfail mail rather than allow the load on your server to spike up too quickly. The default value for this option is 0, meaning that mailmunge-multiplexor will start workers as quickly as necessary to keep up with incoming mail. Except in very unusual situations, you should not need to use this option and we do not recommend that it be used.

-z spooldir

Set the spool directory to spooldir. If this option is omitted, the spool directory defaults to /var/spool/mailmunge.

-s pathName

The UNIX-domain socket on which mailmunge-multiplexor listens for requests. This should be specified as an absolute pathname. If this option is not supplied, it defaults to mailmunge-multiplexor.sock under the spool directory.

-a socket

A socket for listening for requests. This is similar to the -s socket, except that a restricted set of requests are processed. On this socket, the multiplexor will only process requests asking for status; it will not accept any commands to do scanning or that would consume a worker. See the "SOCKET SPECIFICATION" section for the format of socket.

-p fileName

Causes mailmunge-multiplexor to write its process-ID (after becoming a daemon) to the specified file. The file will be owned by root.

-o fileName

Causes mailmunge-multiplexor to use fileName as a lock file to avoid multiple instances from running. If you supply -p but not -o, then mailmunge-multiplexor constructs a lock file by appending ".lock" to the pid file. However, this is less secure than having a root-owned pid file in a root-owned directory and a lock file writable by the user named by the -U option. (The lock file must be writable by the -U user.)

-f filter_path

Normally, mailmunge-multiplexor executes a Perl filter script called /etc/mailmunge/mailmunge-filter.pl to scan the e-mail. However, you can have it execute any program you like by specifying the full path to the program with the -f option. This program must obey the protocol documented in mailmunge-protocol; see that manual page for details.

-l

Log certain events, including the output of the Perl workers' standard-error, using syslog. Normally, the multiplexor does not log much information.

-d

Write debugging information about event-handling code in /var/log/mailmunge-event-debug.log. This is only of use to people debugging mailmunge-multiplexor.

-R kbytes

Limits the resident-set size of the worker filter processes to kbytes kilobytes. This limit is not supported on all operating systems; it is known to work on Linux.

-M kbytes

Limits the total memory space of worker filter processes to kbytes kilobytes. This limit is supported on all operating systems that support the setrlimit system call. This should include most modern UNIX systems.

We recommend that you monitor your worker filter processes and get a feel for how much memory they use. You should then limit the memory to two or three times the worst-case that you have observed. This can help mitigate denial-of-service attacks that use complicated MIME messages to force your Perl filter to consume lots of memory.

-h

Print usage information and exit.

-t filename

Log statistical information to filename. See the section "STATISTICS" for more information.

-T

Log statistical information using syslog. You may use -t and -T together, in which case statistical information is logged in a file and using syslog.

-u

Flush the statistics file after every write. Normally, mailmunge-multiplexor does not flush the file; this is the best choice for minimizing disk I/O on a busy mail server. However, if you wish to watch statistics entries in real-time, you should enable flushing.

-D

Do not fork into the background and become a daemon. Instead, stay in the foreground. Useful mainly for debugging or if you have a supervisory process such as systemd managing mailmunge-multiplexor.

-q queue_size

Normally, if all workers are busy and mailmunge-multiplexor receives another request, it fails it with the error "No free workers." However, if you use the -q option, then up to queue_size requests will be queued. As soon as a worker becomes free, the queued requests will be handed off in FIFO order. If the queue is full and another request comes in, then the request is failed with "No free workers".

-Q queue_timeout

Queued requests should not stay on the queue indefinitely. If a queued request cannot be processed within queue_timeout (default 30) seconds of being placed on the queue, it is failed with a "Queued request timed out" message. See the section "QUEUEING REQUESTS" for more discussion.

-O sock

Listen on a notification socket for connections from listeners. mailmunge-multiplexor can inform external programs of state changes by sending messages over a notification socket. The external programs connect to this socket and then listen for notifications. See the section "SOCKET SPECIFICATION" for the format of sock.

See the mailmunge-notify man page for details of the notification protocol.

-N map_sock

Listen on a map socket for Sendmail SOCKETMAP connections. As of Sendmail 8.13, you can define a Sendmail map type that talks to a daemon over a socket. mailmunge-multiplexor implements that protocol; see "filter_map" in Mailmunge::Filter for details.

See the section "SOCKET SPECIFICATION" for the format of map_sock.

-I backlog

When mailmunge-multiplexor creates a listening socket, it calculates the "backlog" argument to listen based on the maximum number of workers. However, you can explicitly set this backlog with the -I option. Setting the backlog to a high value (around 30-50) may help on a very busy server. If you see mail log messages saying "MXCommand: socket: Connection refused" during busy periods, then that's an indication you need a higher listen backlog.

-L interval

Log the worker status every interval seconds. This logs a line using syslog; the line looks like this:

Worker status: Stopped=s Idle=i Busy=b Killed=k Queued=q Msgs=m Activations=a

Here, "Stopped" is the number of non-running workers, "Idle" is the number of idle workers, "Busy" is the number of busy workers, "Killed" is the number of killed workers yet to be reaped, "Queued" is the number of queued requests, "Msgs" is the total number of messages processed since the multiplexor began running, and "Activations" is the number of times a Perl process has been started since the multiplexor began running.

If you supply an interval of 0 (which is the default), no periodic status logging is performed. If you supply an interval of less than 5 seconds, it is silently reset to 5 seconds.

-S facility

Specifies the syslog facility for log messages. The default is mail. See openlog for a list of valid facilities. You can use either the short name ("mail") or long name ("LOG_MAIL") for the facility name.

-E

Specifies that the multiplexor should create an embedded Perl interpreter. This can improve performance dramatically. But see the section "EMBEDDING PERL" for more information.

-X n

Specifies that the multiplexor should initiate a "tick" request every n seconds. This causes your filter's tick method (if defined) to be called. Note that you have no control over which worker executes tick. If all workers are busy when a tick occurs, that tick request is skipped and a warning message is logged.

-P n

Specifies that the multiplexor should run n tick requests in parallel. Each tick is run as often as specified with the -X argument. (If you omit the -P option, then the multiplexor behaves as if -P 1 had been specified.)

If you run parallel ticks, each tick is assigned an integer identifying its "type". The type ranges from 0 to n-1. While there may be as many as n tick requests running at a time, only one tick of each type will be active at any time.

-Y label

Sets the tag used in the multiplexor's syslog messages to label instead of mailmunge-multiplexor.

-G

Normally, mailmunge-multiplexor uses a umask of 027 when creating listening sockets. If you would like the sockets to be readable and writeable by the group as well as the owner, supply the -G option. This causes the umask to be 007 whenever UNIX-domain sockets are created.

-y n

Limits the maximum number of concurrent recipok checks to n on a per-domain basis. The value of n can range from 0 (in which case no limit is applied) to maxWorkers, where maxWorkers is the argument to the -x option. If n is outside that range, it is ignored (and no limit is applied.)

The recipok command ultimately invokes the filter_recipient function in your filter. If you are doing recipient verification against servers that may be slow or unreliable, you can use the -y option to limit the number of concurrent recipient verifications per domain. That way, if one domain's server becomes very slow, it won't consume all available workers for recipient verification. Instead, its RCPT commands will be tempfailed and there will be workers available to handle RCPT commands for other domains.

SOCKET SPECIFICATION

The -a, -N and -O options take a socket as an argument. This socket can be specified as:

/path/to/socket: A UNIX-domain socket
inet:portnum: A TCP socket bound to port portnum, but that accepts connections only from the IPv4 loopback address (127.0.0.1).
inet_any:portnum: A TCP socket bound to port portnum that will accept connections from any address. Use inet_any with caution!
inet6:portnum: A TCP socket bound to port portnum listening on the IPv6 loopback address.
inet6_any:portnum: A TCP socket bound to port portnum listening on the IPv6 wildcard address.

QUEUEING REQUESTS

Normally, if all workers are busy, any additional requests are failed immediately. However, the -q and -Q options allow you to queue requests for a short amount of time. This facility is intended to gracefully handle a temporary overload; most of the time, your queue should be empty.

Because mailmunge checks the number of free workers when a connection is opened and fails the connection if there are no free workers, the intent of the queue is to allow SMTP transactions that are already underway to continue if there is a slight overload. Any new connections will be failed if all workers are busy, but existing connections are allowed to continue. Queuing requests may improve throughput on extremely busy servers.

Note that if you supply the -q option to mailmunge, then even new connections are allowed to queue. This may improve throughput by keeping the worker utilization higher.

The -R option to mailmunge can be used to reserve a specified number of workers for connections from the loopback address. Using the -R option has the side-effect of permitting new connections from the loopback address to queue.

EMBEDDING PERL

Normally, when mailmunge-multiplexor activates a worker, it forks and execs your Perl filter program. However, if the multiplexor was compiled with embedded Perl support, and you supply the -E command-line option, the multiplexor works like this:

It creates an embedded Perl interpreter, and sources your Perl filter program with a special command-line argument telling it to read the filter, but not to enter the main loop.
Each time a worker is activated, the multiplexor calls fork() and calls the Perl function _mailmunge_do_main_loop. This invokes filter_initialize and then runs your filter's main loop.

On some very old platforms (for example, Red Hat Linux 7.3 with Perl 5.6.1), it is not safe to destroy and recreate a Perl interpreter without causing a memory leak. On those platforms, if you attempt to reread the filter file (by sending the multiplexor a HUP signal or reread command), the filter will not be re-read, and a message will be logged to syslog. On those platforms, you must kill and restart mailmunge-multiplexor if you change the filter file.

On most platforms, however, a filter reread is accomplished by destroying and re-creating the embedded interpreter, re-sourcing your Perl filter and killing workers as soon as they are idle.

STATISTICS

With the -t option, mailmunge-multiplexor logs certain events to a file. This file can be post-processed to gather statistics about the multiplexor. You can use it to tune the number of workers you run, adjust timeouts, and so on.

Each line of the file looks like this:

YYYY/MM/DD:HH:MM:SS timestamp event key=val key=val...

Here, YYYY/MM/DD:HH:MM:SS is the local time of day. Timestamp is the number of seconds since January 1, 1970. Event is the name of an event. The valid events are:

StartWorker: A worker process has been started.
KillWorker: A worker process has been killed.
ReapWorker: A dead worker process has been reaped. It is possible to have a ReapWorker event without a previous KillWorker event if the worker process terminated abnormally.
StartFilter: A worker process has begun filtering an e-mail message.
EndFilter: A worker process has finished filtering an e-mail message.

The possible keys in the key=value pairs are:

worker=n: The worker involved in the event. Every worker is identified by a small integer.
nworkers=n: The total number of running workers immediately after the event happened.
nbusy=n: The number of busy workers (workers that are processing an e-mail message) immediately after the event happened.
reason="string": The reason for a StartWorker or KillWorker event. (Present only for these events.)
numRequests=n: The number of e-mails processed by the worker. Present only for an EndFilter event.

If you send the mailmunge-multiplexor process a SIGHUP signal (kill -1 pid), it closes and reopens the statistics file. This is useful during log file rotation.

If you send the mailmunge-multiplexor process a SIGINT signal (kill -INT pid), it terminates all active-but-idle workers. Also, any active-and-busy workers terminate as soon as they finish filtering the current message. This is useful to force a reread of the filter rules file without stopping and restarting Sendmail.

If you send the mailmunge-multiplexor process a SIGTERM signal (kill pid), it terminates all workers and exits immediately.

AUTHOR

mailmunge-multiplexor was written by Dianne Skoll <dianne@skollsoft.com>.

LICENSE

This code is licensed under the terms of the GNU General Public License, version 2.