systemd Developer Lennart Poettering lennart@poettering.net daemon 7 daemon Writing and Packaging System Daemons A daemon is a service process that runs in the background and supervises the system or provides functionality to other processes. Traditionally, daemons are implemented following a scheme originating in SysV Unix. Modern daemons should follow a simpler yet more powerful scheme (here called "new-style" daemons), as implemented by systemd1. When a traditional SysV daemon starts, it should execute the following steps as part of the initialization. Note that these steps are unnecessary for new-style daemons (see below), and should only be implemented if compatibility with SysV is essential. Close all open file descriptors except STDIN, STDOUT, STDERR (i.e. the first three file descriptors 0, 1, 2). This ensures that no accidentally passed file descriptor stays around in the daemon process. On Linux this is best implemented by iterating through /proc/self/fd, with a fallback of iterating from file descriptor 3 to the value returned by getrlimit() for RLIMIT_NOFILE. Reset all signal handlers to their default. This is best done by iterating through the available signals up to the limit of _NSIG and resetting them to SIG_DFL. Reset the signal mask using sigprocmask(). Sanitize the environment block, removing or resetting environment variables that might negatively impact daemon runtime. Call fork(), to create a background process. In the child, call setsid() to detach from any terminal and create an independent session. In the child, call fork() again, to ensure the daemon can never re-aquire a terminal again. Call exit() in the first child, so that only the second child (the actual daemon process) stays around. This ensures that the daemon process is reparented to init/PID 1, as all daemons should be. In the daemon process, connect /dev/null to STDIN, STDOUT, STDERR. In the daemon process, reset the umask to 0, so that the file modes passed to open(), mkdir() and suchlike directly control the access mode of the created files and directories. In the daemon process, change the current directory to the root directory (/), in order to avoid that the daemon involuntarily blocks mount points from being unmounted. In the daemon process, write the daemon PID (as returned by getpid()) to a PID file, for example /var/run/foobar.pid (for a hypothetical daemon "foobar"), to ensure that the daemon cannot be started more than once. This must be implemented in race-free fashion so that the PID file is only updated when at the same time it is verified that the PID previously stored in the PID file no longer exists or belongs to a foreign process. Commonly some kind of file locking is employed to implement this logic. In the daemon process, drop privileges, if possible and applicable. From the daemon process notify the original process started that initialization is complete. This can be implemented via an unnamed pipe or similar communication channel that is created before the first fork() and hence available in both the original and the daemon process. Call exit() in the original process. The process that invoked the daemon must be able to rely that this exit() happens after initialization is complete and all external communication channels established and accessible. The BSD daemon() function should not be used, as it implements only a subset of these steps. A daemon that needs to provide compatibility with SysV systems should implement the scheme pointed out above. However, it is recommended to make this behaviour optional and configurable via a command line argument, to ease debugging as well as to simplify integration into systems using systemd. Modern services for Linux should be implemented as new-style daemons. This makes it easier to supervise and control them at runtime and simplifies their implementation. For developing a new-style daemon none of the initialization steps recommended for SysV daemons need to be implemented. New-style init systems such as systemd make all of them redundant. Moreover, since some of these steps interfere with process monitoring, file descriptor passing and other functionality of the init system it is recommended not to execute them when run as new-style service. Note that new-style init systems guarantee execution of daemon processes in clean process contexts: it is guaranteed that the environment block is sanitized, that the signal handlers and mask is reset and that no left-over file descriptors are passed. Daemons will be executed in their own session, and STDIN/STDOUT/STDERR connected to /dev/null unless otherwise configured. The umask is reset. It is recommended for new-style daemons to implement the following: If SIGTERM is received, shut down the daemon and exit cleanly. If SIGHUP is received, reload the configuration files, if this applies. Provide a correct exit code from the main daemon process, as this is used by the init system to detect service errors and problems. It is recommended to follow the exit code scheme as defined in the LSB recommendations for SysV init scripts. As much as possible, rely on systemd's functionality to limit the access of the daemon to files, services and other resources. i.e. rely on systemd's resource limit control instead of implementing your own, rely on systemd's privilege dropping code instead of implementing it in the daemon, and similar. See systemd.exec5 for the available controls. If possible and applicable expose the daemon's control interface via the D-Bus IPC system and grab a bus name as last step of initialization. If D-Bus is used, make your daemon bus-activatable, via supplying a D-Bus service activation configuration file. This has multiple advantages: your daemon may be started lazily on-demand; it may be started in parallel to other daemons requiring it -- which maximizes parallelization and boot-up speed; your daemon can be restarted on failure, without losing any bus requests, as the bus queues requests for activatable services. See below for details. If your daemon provides services to other local processes or remote clients via a socket, it should be made socket-activatable following the scheme pointed out below. Like D-Bus activation this enables on-demand starting of services as well as it allows improved parallelization of service start-up. Also, for state-less protocols (such as syslog, DNS) a daemon implementing socket-based activation can be restarted without losing a single request. See below for details. If applicable a daemon should notify the init system about startup completion or status updates via the sd_notify3 interface. Instead of using the syslog() call to log directly to the system logger, a new-style daemon may choose to simply log to STDERR via fprintf(), which is then forwarded to syslog by the init system. If log priorities are necessary these can be encoded by prefixing individual log lines with strings like "<4>" (for log priority 4 "WARNING" in the syslog priority scheme), following a similar style as the Linux kernel's printk() priority system. In fact, using this style of logging also enables the init system to optionally direct all application logging to the kernel log buffer (kmsg), as accessible via dmesg1. This kind of logging may be enabled by setting StandardError=syslog in the service unit file. For details see sd-daemon7 and systemd.exec5. These recommendations are similar but not identical to the Apple MacOS X Daemon Requirements. When writing systemd unit files, it is recommended to consider the following suggestions: If possible do not use the Type=forking setting in service files. But if you do, make sure to set the PID file path using PIDFile=. See systemd.service5 for details. If your daemon registers a D-Bus name on the bus, make sure to use Type=dbus if possible. Make sure to set a good human-readable description string with Description=. Do not disable DefaultDependencies=, unless you really know what you do and your unit is involved in early boot or late system shutdown. Normally, little if any dependencies should need to be defined explicitly. However, if you do configure explicit dependencies, only refer to unit names listed on systemd.special7 or names introduced by your own package to keep the unit file operating system-independent. Make sure to include an [Install] section including installation information for the unit file. See systemd.unit5 for details. To activate your service on boot make sure to add a WantedBy=multi-user.target or WantedBy=graphical.target directive. At the build installation time (e.g. make install during package build) packages are recommended to install their systemd unit files in the directory returned by pkg-config systemd --variable=systemdsystemnunitdir (for system services), resp. pkg-config systemd --variable=systemdsessionunitdir (for session services). This will make the services available in the system on explicit request but not activate them automatically during boot. Optionally, during package installation (e.g. rpm -i by the administrator) symlinks should be created in the systemd configuration directories via the systemd-install1 tool, to activate them automatically on boot. Packages using autoconf1 are recommended to use a configure script excerpt like the following to determine the unit installation path during source configuration: PKG_PROG_PKG_CONFIG AC_ARG_WITH([systemdsystemunitdir], AS_HELP_STRING([--with-systemdsystemunitdir=DIR], [Directory for systemd service files]), [], [with_systemdsystemunitdir=$($PKG_CONFIG --variable=systemdsystemunitdir systemd)]) AC_SUBST([systemdsystemunitdir], [$with_systemdsystemunitdir]) AM_CONDITIONAL(HAVE_SYSTEMD, [test -n "$with_systemdsystemunitdir"]) This snippet allows automatic installation of the unit files on systemd machines, and optionally allows their installation even on machines lacking systemd. (Modification of this snippet for the session unit directory is left as excercise to the reader.) Additionally, to ensure that make distcheck continues to work, it is recommended to add the following to the top-level Makefile.am file in automake1-based projects: DISTCHECK_CONFIGURE_FLAGS = \ --with-systemdsystemunitdir=$$dc_install_base/$(systemdsystemunitdir) Finally, unit files should be installed in the system with an automake excerpt like the following: if HAVE_SYSTEMD systemdsystemunit_DATA = \ foobar.socket \ foobar.service endif In the rpm8 .spec file use a snippet like the following to enable/disable the service during installation/deinstallation. Consult the packaging guidelines of your distribution for details and the equivalent for other packaging managers: %post /usr/bin/systemd-install enable foobar.service foobar.socket >/dev/null 2>&1 || : %preun if [ "$1" -eq 0 ]; then /usr/bin/systemd-install disable foobar.service foobar.socket >/dev/null 2>&1 || : fi Since new-style init systems such as systemd are compatible with traditional SysV init systems it is not strictly necessary to port existing daemons to the new style. However doing this offers additional functionality to the daemons as well as it simplifies integration into new-style init systems. To port an existing SysV compatible daemon the following steps are recommended: If not already implemented, add an optional command line switch to the daemon to disable daemonization. This is useful not only for using the daemon in new-style init systems, but also to ease debugging. If the daemon offers interfaces to other software running on the local system via local AF_UNIX sockets, consider implementing socket-based activation (see above). Usually a minimal patch is sufficient to implement this: Extend the socket creation in the daemon code so that sd_listen_fds3 is checked for already passed sockets first. If sockets are passed (i.e. when sd_listen_fds() returns a positive value), skip the socket createn step and use the passed sockets. Secondly, ensure that the file-system socket nodes for local AF_UNIX sockets used in the socket-based activation are not removed when the daemon shuts down, if sockets have been passed. Third, if the daemon normally closes all remaining open file descriptors as part of its initialization, the sockets passed from the init system must be spared. Since new-style init systems guarantee that no left-over file descriptors are passed to executed processes, it might be a good choice to simply skip the closing of all remaining open file descriptors if file descriptors are passed. Write and install a systemd unit file for the service (and the sockets if socket-based activation is used, as well as a path unit file, if the daemon processes a spool directory), see above for details. If the daemon exposes interfaces via D-Bus, write and install a D-Bus activation file for the service, see above for details. systemd1, systemd-install1, sd-daemon7, sd_listen_fds3, sd_notify3, daemon3