All generators are executed in parallel. That means all executables are started at the very same time and need to be able to cope with this parallelism.

Generators are run very early at boot and cannot rely on any external services. They may not talk to any other process. That includes simple things such as logging to syslog(3), or systemd itself (this means: no systemctl(1))! Non-essential file systems like /var/ and /home/ are mounted after generators have run. Generators can however rely on the most basic kernel functionality to be available, as well as mounted /sys/, /proc/, /dev/, /usr/ and /run/ file systems.

Units written by generators are removed when the configuration is reloaded. That means the lifetime of the generated units is closely bound to the reload cycles of systemd itself.

Generators should only be used to generate unit files, .d/*.conf drop-ins for them and symlinks to them, not any other kind of non-unit related configuration. Due to the lifecycle logic mentioned above, generators are not a good fit to generate dynamic configuration for other services. If you need to generate dynamic configuration for other services, do so in normal services you order before the service in question.

Note that using the StandardInputData=/StandardInputText= settings of service unit files (see systemd.exec(5)), it is possible to make arbitrary input data (including daemon-specific configuration) part of the unit definitions, which often might be sufficient to embed data or configuration for other programs into unit files in a native fashion.

Since syslog(3) is not available (see above), log messages have to be written to /dev/kmsg instead.

The generator should always include its own name in a comment at the top of the generated file, so that the user can easily figure out which component created or amended a particular unit.

The SourcePath= directive should be used in generated files to specify the source configuration file they are generated from. This makes things more easily understood by the user and also has the benefit that systemd can warn the user about configuration files that changed on disk but have not been read yet by systemd. The SourcePath= value does not have to be a file in a physical filesystem. For example, in the common case of the generator looking at the kernel command line, SourcePath=/proc/cmdline should be used.

Generators may write out dynamic unit files or just hook unit files into other units with the usual .wants/ or .requires/ symlinks. Often, it is nicer to simply instantiate a template unit file from /usr/ with a generator instead of writing out entirely dynamic unit files. Of course, this works only if a single parameter is to be used.

If you are careful, you can implement generators in shell scripts. We do recommend C code however, since generators are executed synchronously and hence delay the entire boot if they are slow.

Regarding overriding semantics: there are two rules we try to follow when thinking about the overriding semantics:

Of these two rules the first rule is probably the more important one and breaks the second one sometimes. Hence, when deciding whether to use argv[1], argv[2], or argv[3], your default choice should probably be argv[1].

Instead of heading off now and writing all kind of generators for legacy configuration file formats, please think twice! It is often a better idea to just deprecate old stuff instead of keeping it artificially alive.