Linux distributions with systemd support a “socket activation” feature that allows systemd to listen on the socket from early in the boot process and start the service when the first client connects.
Supporting this in PHP-FPM streamlines systems with many pools (so they can start on-demand), for administrators that prefer to have a PHP-FPM pool listen on a privileged port or path without having to start it initially as root, and for administrators wanting to resolve a dependency between the web server accepting requests and PHP-FPM's socket being able to queue or service them.
launchd and legacy internet superservers support socket activation in similar ways.
Socket activation creates no overhead once the daemon has started. systemd does not proxy any traffic; it just hands over the file descriptor. Once the daemon is running and using the systemd-provided socket(s), there's no distinction in daemon operation until shutdown, where it skips closing the socket(s).
Socket activation allows replacing the entire binary (as with a PHP-FPM security update) without interrupting listening on the socket. PHP-FPM supports reloading to a limited degree, but it doesn't seem possible to fully replace the executable. The suggested configuration below also allows restarting and reloading individual PHP-FPM pools without any effect on others.
There is work underway to have socket activation on the base system spawn or start full containers (like LXC) on-demand. Since PHP-FPM will, itself, be in the container, something else needs to listen on its behalf.
Finally, it's a platform consistency issue. As more services move to socket activation in Fedora, Arch, Suse, and Red Hat Enterprise Linux (and its derivatives), socket units will become a sort of “common currency” for configuring which services listen where.
When systemd starts a socket-activated service, it passes the bound, listening socket in as a file descriptor. PHP-FPM is already accustomed to such behavior using the FPM_SOCKETS environmental variable. It's actually possible to “socket activate” PHP-FPM once by setting it cleverly (so that it looks for the file descriptor(s) created by systemd). Unfortunately, PHP-FPM's shutdown behavior leaves the systemd socket in a state that's unusable for subsequent activations.
It's possible to provide reliable socket activation support by having a loop before or after the “import inherited sockets” one in fpm_sockets_init_main(). The loop can either inspect the environmental variables systemd sends in directly or call the reference implementations defined in systemd/sd-daemon.h. Jerry and I have a patch working with the latter approach, including the necessary linking against systemd's library. (Linking is also desirable to ease adding support later for sd_notify(), which can inform systemd of PHP-FPM's current lifecycle status, including startup, reloading, and shutdown.)
Adding the support also requires marking the systemd-imported sockets as persistent past shutdown so PHP-FPM does not perform various cleanup actions on them that break subsequent activations of the service.
If systemd socket activation support isn't enabled at build time, this feature will have no effect. Even if this feature is enabled for a build, it will have no effect on non-users of socket activation.
The recommended use for PHP-FPM with socket activation is a 1:1:1 relationship between the listening socket unit in systemd, the service unit in systemd, and the PHP-FPM pool. This approach allows each pool to spawn on-demand. It also has nice effects like allowing cgroups resource limitations (like CPU shares, memory, block I/O, and network I/O) and security isolations (like PrivateTmp and syscall restrictions) to be set per-pool in the corresponding systemd service unit. Logs will also aggregate per-pool in the systemd journal for the corresponding service.
Here's an example configuration. It's still possible to have PHP-FPM drop its own permissions or have it perform its own daemonization fork, but it's unnecessary. It's also possible to directly start the PHP-FPM pool without waiting for the first incoming connection; systemd will still start the necessary socket and pass it in.
[Socket] ListenStream=/var/run/my-php-fpm-pool.socket [Install] WantedBy=sockets.target
[Service] User=my-php-fpm-pool Group=my-php-fpm-pool ExecStart=/usr/sbin/php-fpm --fpm-config=/etc/php-fpm.d/my-php-fpm-pool.conf KillMode=process
It's possible the final implementation may move to “listen = fd:3” or similar syntax to directly request use of an inherited socket rather than requiring a perfect socket configuration match-up between the systemd socket configuration and the PHP-FPM pool configuration. This would make PHP's inheritance more consistent with my nginx proposal.
[global] pid = /var/run/my-php-fpm-pool.pid ; Not really used by anything with daemonize = no. error_log = syslog ; Will aggregate to the service's systemd journal. daemonize = no ; systemd handles the forking. [www] listen = /var/run/my-php-fpm-pool.socket ; Must match systemd socket unit. user = my-php-fpm-pool ; Ignored but required. group = my-php-fpm-pool ; Ignored but required. pm = static pm.max_children = 10 slowlog = syslog
systemctl enable my-php-fpm-pool.socket # Enables the socket at system startup. systemctl start my-php-fpm-pool.socket # Starts the socket listening.
The ondemand process manager still keeps considerable memory allocated, and PHP-FPM currently has some idle CPU load when not processing requests. It's <1% of a core per service, but it adds up when you manage 500+ pools, each as a service for security/resource isolation, on a box.
The ondemand process manager doesn't solve the dependency issue mentioned earlier (a web server requiring PHP-FPM to be ready) or allow privileges to be dropped before PHP-FPM gets invoked at all. The latter is useful for platform providers that let users configure PHP-FPM for their individual use cases but want to provide assigned “listening” sockets.
Upstart seems to have basic socket activation support, and integrating PHP-FPM with it would be a great follow-on project. All socket activation basically works the same way, in the sense of a file descriptor getting handed into the daemon. This RFC would pave the way for integration into additional superserver and init daemons.
In order for pools to share an opcode cache they must be forked from the same parent process. There are ways to work around that, but it requires some fancy fd passing footwork in APC and I guess in many instances you don't actually want to share across pools anyway. (Abbreviated from Rasmus on PHP internals)