DINIT-SERVICE(5) Dinit - service management system DINIT-SERVICE(5)

Dinit service description files

/etc/dinit.d/service-name, $HOME/.config/dinit.d/service-name

The service description files for Dinit each describe a service. The name of the file corresponds to the name of the service it describes.

Service description files specify the various attributes of a service. A service description file is named after the service it represents, and is a plain-text file with simple key-value format. The description files are located in a service description directory; by default, the system dinit process searches /etc/dinit.d, /usr/local/lib/dinit.d and /lib/dinit.d, while a user process searches $HOME/.config/dinit.d. See dinit(8) for more details of how and when service descriptions are loaded.

All services have a type and a set of dependencies. These are discussed in the following subsections. The type, dependencies, and other attributes are specified via property settings, the format of which are documented in the SERVICE PROPERTIES subsection, which also lists the available properties.

There are five basic types of service:

  • Process services. This kind of service runs as a single process; starting the service simply requires starting the process; stopping the service is accomplished by stopping the process (via sending it a signal).
  • Bgprocess services ("background process" services). This kind of service is similar to a regular process service, but the process daemonizes or otherwise forks from the original process which starts it, and the process ID is written to a file. Dinit can read the process ID from the file and, if it is running as the system init process, can supervise it.
  • Scripted services are services which are started and stopped by a command (which need not actually be a script, despite the name). They can not be supervised.
  • Internal services do not run as an external process at all. They can be started and stopped without any external action. They are useful for grouping other services (via service dependencies).
  • Triggered services are similar to internal processes, but an external trigger is required before they will start (i.e. their startup will pause until the trigger occurs). The dinitctl trigger command can be used to trigger such a service; see dinitctl(8).

Independent of their type, the state of services can be linked to other services via dependency relationships, which are discussed in the next section.

A service dependency relationship, broadly speaking, specifies that for one service to run, another must also be running. The first service is the dependent service and the latter is the dependency service (we will henceforth generally refer to the the dependency relationship as the relationship and use dependency to refer to the service). A dependency relationship is specified via the properties of the dependent. There are different relationship types, as follows:

  • A need (or "hard") relationship specifies that the dependent must wait for the dependency to be started before it starts, and that the dependency must remain started while the dependent is started. Starting the dependent will start the dependency, and stopping the dependency will stop the dependent. This type of relationship is specified using a depends-on property.
  • A milestone relationship specifies that the dependency must start successfully before the dependent starts. Starting the dependent will therefore start the dependency. Once started, the relationship is satisfied; if the dependency then stops, it has no effect on the dependent. However, if the dependency fails to start or has its startup cancelled, the dependent will not start (and will return to the stopped state). This type of relationship is specified using a depends-ms property.
  • A waits-for relationship specifies that the dependency must start successfully, or fail to start, before the dependent starts. Starting the dependent will attempt to first start the dependency, but failure will not prevent the dependent from starting. If the dependency starts, stopping it will have no effect on the dependent. This type of relationship is specified using a waits-for property.

Note that process-based services always wait for their dependency relationships to be satisfied (by the dependency starting, or failing to start in case of a waits-for relationship) before their process is launched. Conversely, a termination signal will not in general be sent to a service process until the service has no active dependents.

Since in general dependencies should remain started so long as their dependent does, an attachment forms between the two once both are started. This attachment is released when the dependent stops, and the dependency will then stop, unless it has other attachments or it has been explicitly started independently. Attachments between a dependent and dependency are re-created if a dependency starts (or starts again) while the dependent is still started.

This section described the various service properties that can be specified in a service description file. The properties specify the type of the service, dependencies of the service, and other service configuration.

Each line of the file can specify a single property value, expressed as `property-name = value'. Comments begin with a hash mark (#) and extend to the end of the line (they must be separated from setting values by at least one whitespace character). Values are interpreted literally, except that:

  • White space (comprised of spaces, tabs, etc) is collapsed to a single space, except leading or trailing white space around the property value, which is stripped.
  • For settings which specify a command with arguments, the value is interpreted as a series of tokens separated by white space, rather than a single string of characters.
  • Double quotes (") can be used around all or part of a property value, to prevent whitespace collapse and prevent interpretation of other special characters (such as "#") inside the quotes. The quote characters are not considered part of the property value. White space appearing inside quotes does not act as a delimiter for tokens.
  • A backslash (\) can be used to escape the next character, causing it to lose any special meaning and become part of the property value. A double backslash (\\) is collapsed to a single backslash within the parameter value. White space preceded by a backslash will not separate tokens.

Setting a property generally overrides any previous setting (from prior lines). However some properties are set additively; these include dependency relationships and options properties.

The following properties can be specified:

Specifies the service type; see the SERVICE TYPES section.
Specifies the command, including command-line arguments, for starting the process. Applies only to process, bgprocess and scripted services. The value is subject to variable substitution (see VARIABLE SUBSTITUTION).
Specifies the command to stop the service (optional). Applicable to process, bgprocess and scripted services. If specified for process or bgprocess services, the "stop command" will be executed in order to stop the service, instead of signalling the service process. The value is subject to variable substitution (see VARIABLE SUBSTITUTION).
Specifies the working directory for this service. For a scripted service, this affects both the start command and the stop command. The value is subject to variable substitution (see VARIABLE SUBSTITUTION).
Specifies which user to run the process(es) for this service as. Specify as a username or numeric ID. If specified by name, the group for the process will also be set to the primary group of the specified user, and supplementary groups will be initialised (unless support for them is disabled) according to the system's group database. If specified by number, the group for the process will remain the same as that of the running dinit process, and all supplementary groups will be dropped (unless support has been disabled).
Specifies a file containing value assignments for environment variables, in the same format recognised by the dinit command's --env-file option (see dinit(8)). The file is read when the service is loaded, therefore values from it can be used in variable substitutions (see VARIABLE SUBSTITUTION). Variable substitution is not performed on the env-file property value itself.
Indicates whether the service should automatically restart if it stops, including due to unexpected process termination or a dependency stopping. Note that if a service stops due to user request, automatic restart is inhibited. The default is to automatically restart.
Applies only to process and bgprocess services. When set true/yes, an automatic process restart can be performed without first stopping any dependent services. This setting is meaningless if the restart setting is set to false.
Specifies the minimum time (in seconds) between automatic restarts. Enforcing a sensible minimum prevents Dinit from consuming a large number of process cycles in case a process continuously fails immediately after it is started. The default is 0.2 (200 milliseconds).
Sets the interval (in seconds) over which restarts are limited. If a process automatically restarts more than a certain number of times (specified by the restart-limit-count setting) in this time interval, it will not be restarted again. The default value is 10 seconds.
Specifies the maximum number of times that a service can automatically restart over the interval specified by restart-limit-interval. Specify a value of 0 to disable the restart limit. The default value is 3.
Specifies the time in seconds allowed for the service to start. If the service takes longer than this, its process group is sent a SIGINT signal and enters the "stopping" state (this may be subject to a stop timeout, as specified via stop-timeout, after which the process group will be terminated via SIGKILL). The timeout period begins only when all dependencies have been satisfied. The default value is 60. A value of 0 allows unlimited start time.
Specifies the time in seconds allowed for the service to stop. If the service takes longer than this, its process group is sent a SIGKILL signal which should cause it to terminate immediately. The timeout period begins only when all dependent services have already stopped. The default value is 10. A value of 0 allows unlimited stop time.
For bgprocess type services only; specifies the path of the file where daemon will write its process ID before detaching. Dinit will read the contents of this file when starting the service, once the initial process exits, and will supervise the process with the discovered process ID. Dinit may also send signals to the process ID to stop the service; if dinit runs as a privileged user the path should therefore not be writable by unprivileged users.

The value is subject to variable substitution (see VARIABLE SUBSTITUTION).

This service depends on the named service. Starting this service will start the named service; the command to start this service will not be executed until the named service has started. If the named service is stopped then this service will also be stopped.
This service has a "milestone" dependency on the named service. Starting this service will start the named service; this service will not start until the named service has started, and will fail to start if the named service does not start. Once the named (dependent) service reaches the started state, however, the dependency may stop without affecting the dependent service.
When this service is started, wait for the named service to finish starting (or to fail starting) before commencing the start procedure for this service. Starting this service will automatically start the named service. If the named service fails to start, this service will start as usual (subject to other dependencies being met).
For each file name in directory-path which does not begin with a dot, add a waits-for dependency to the service with the same name. Note that contents of files in the specified directory are not significant; expected usage is to have symbolic links to the associated service description files, but this is not required. Failure to read the directory contents, or to find any of the services named within, is not considered fatal.

The directory path, if not absolute, is relative to the directory containing the service description file.

When starting this service, if the named service is also starting, wait for the named service to finish starting before bringing this service up. This is similar to a waits-for dependency except no dependency relationship is implied; if the named service is not starting, starting this service will not cause it to start (nor wait for it in that case). It does not by itself cause the named service to be loaded (if loaded later, the "after" relationship will be enforced from that point).
When starting the named service, if this service is also starting, wait for this service to finish starting before bringing the named service up. This is largely equivalent to specifying an after relationship to this service from the named service. However, it does not by itself cause the named service to be loaded (if loaded later, the "before" relationship will be enforced from that point).
When this service terminates (i.e. starts successfully, and then stops of its own accord), the named service should be started. Note that the named service is not loaded until that time; naming an invalid service will not cause this service to fail to load.

This can be used for a service that supplies an interactive "recovery mode" for another service; once the user exits the recovery shell, the primary service (as named via this setting) will then start. It also supports multi-stage system startup where later service description files reside on a separate filesystem that is mounted during the first stage; such service descriptions will not be found at initial start, and so cannot be started directly, but can be chained via this directive.

The chain is not executed if the initial service was explicitly stopped, stopped due to a dependency stopping (for any reason), if it will restart (including due to a dependent restarting), or if its process terminates abnormally or with an exit status indicating an error. However, if the always-chain option is set the chain is started regardless of the reason and the status of this service termination.

Pre-open a socket for the service and pass it to the service using the systemd activation protocol. This by itself does not give so called "socket activation", but does allow any process trying to connect to the specified socket to do so immediately after the service is started (even before the service process is properly prepared to accept connections).

The path value is subject to variable substitution (see VARIABLE SUBSTITUTION).

Gives the permissions for the socket specified using socket-listen. Normally this will be 600 (user access only), 660 (user and group access), or 666 (all users). The default is 666.
Specifies the user (name or numeric ID) that should own the activation socket. If socket-uid is specified as a name without also specifying socket-gid, then the socket group is the primary group of the specified user (as found in the system user database, normally /etc/passwd). If the socket-uid setting is not provided, the socket will be owned by the user id of the dinit process.
Specifies the group of the activation socket. See discussion of socket-uid.
Specifies the signal to send to the process when requesting it to terminate (applies to `process' and `bgprocess' services only). Signal names are specified as the POSIX signal name without the SIG- prefix. At least HUP, TERM, and KILL are supported (use dinitctl signal --list for the full list of supported signals). The default is TERM (the SIGTERM signal). See also the discussion of stop-timeout.
Specifies the mechanism, if any, by which a process service will notify that it is ready (successfully started). If not specified, a process service is considered started as soon as it has begun execution. The two options are:
  • pipefd:fd-number — the service will write a message to the specified file descriptor, which dinit sets up as the write end of a pipe before execution. This mechanism is compatible with the S6 supervision suite.
  • pipevar:env-var-name — the service will write a message to file descriptor identified using the contents of the specified environment variable, which will be set by dinit before execution to a file descriptor (chosen arbitrarily) attached to the write end of a pipe.
Specifies how the output of this service is logged. This setting is valid only for process-based services (including scripted services).
  • file: output will be written to a file; see the logfile setting.
  • buffer: output will be buffered in memory, up to a limit specified via the log-buffer-size setting. The buffer contents can be examined via the dinitctl(8) catlog subcommand.
  • pipe: output will be written to a pipe, and may be consumed by another service (see the consumer-of setting); note that, if output is not consumed promptly, the pipe buffer may become full which may cause the service process to stall.
  • none: output is discarded.
The default log type is none, unless the logfile setting is specified in which case the default log type is file. For pipe (and buffer, which uses a pipe internally) note that the pipe created may outlive the service process and be re-used if the service is stopped and restarted.
Specifies the log file for the service. Output from the service process (standard output and standard error streams) will be appended to this file, which will be created if it does not already exist. The file ownership and permissions are adjusted according to the logfile-uid, logfile-gid and logfile-permissions settings. This setting has no effect if the service is set to run on the console (via the runs-on-console, starts-on-console, or shares-console options). The value is subject to variable substitution (see VARIABLE SUBSTITUTION). Note that if the directory in which the logfile resides does not exist (or is not otherwise accessible to dinit) when the service is started, the service will not start successfully. If this settings is specified and log-type is not specified or is currently none, then the log type will be changed to file.
Gives the permissions for the log file specified using logfile. Normally this will be 600 (user access only), 640 (also readable by the group), or 644 (readable by all users). If the log file already exists when the service starts, its permissions will be changed in accordance with the value of this setting. The default is value 600 (accessible to only the owning user).
Specifies the user (name or numeric ID) that should own the log file. If logfile-uid is specified as a name without also specifying logfile-gid, then the log file group is the primary group of the specified user (as found in the system user database, normally /etc/passwd). If the log file already exists when the service starts, its ownership will be changed in accordance with the value of this setting. The default value is the user id of the dinit process.
Specifies the group of the log file. See discussion of logfile-uid.
If the log type (see log-type) is set to buffer, this setting controls the maximum size of the buffer used to store process output. If the buffer becomes full, further output from the service process will be discarded.
Specifies that this service consumes (as its standard input) the output of another service. For example, this allows this service to act as a logging agent for another service. The named service must be a process-based service with log-type set to pipe. This setting is only valid for process and bgprocess services.
Specifies various options for this service. See the OPTIONS section. This directive can be specified multiple times to set additional options.
Specifies options for interpreting other settings when loading this service description. Currently there are two available options. One is export-passwd-vars, which specifies that the environment variables `USER', `LOGNAME' (same as `USER'), `HOME', `SHELL', `UID', and `GID' should be exported into the service's load environment (that is, overriding any global environment including the global environment file, but being overridable by the service's environment file). The other is export-service-name, which will set the environment variable `DINIT_SERVICE' containing the name of the current service.
When this service is started, if this setting (or the inittab-line setting) has a specified value, an entry will be created in the system "utmp" database which tracks processes and logged-in users. Typically this database is used by the "who" command to list logged-in users. The entry will be cleared when the service terminates.

The inittab-id setting specifies the "inittab id" to be written in the entry for the process. The value is normally quite meaningless. However, it should be distinct (or unset) for separate processes. It is typically limited to a very short length.

The "utmp" database is mostly a historical artifact. Access to it on some systems is prone to denial-of-service by unprivileged users. It is therefore recommended that this setting not be used. However, "who" and similar utilities may not work correctly without this setting (or inittab-line) enabled appropriately.

This setting has no effect if Dinit was not built with support for writing to the "utmp" database. It applies only to process services.

This specifies the tty line that will be written to the "utmp" database when this service is started. Normally, for a terminal login service, it would match the terminal device name on which the login process runs, without the "/dev/" prefix.

See the description of the inittab-id setting for details.

Specifies the number of file descriptors that a process may have open simultaneously. See the RESOURCE LIMITS section.
Specifies the maximum size of the core dump file that will be generated for the process if it crashes (in a way that would result in a core dump). See the RESOURCE LIMITS section.
Specifies the maximum size of the data segment for the process, including statically allocated data and heap allocations. Precise meaning may vary between operating systems. See the RESOURCE LIMITS section.
Specifies the maximum size of the address space of the process. See the RESOURCE LIMITS section. Note that some operating systems (notably, OpenBSD) do not support this limit; the setting will be ignored on such systems.
Run the service process(es) in the specified cgroup (see cgroups(7)). The cgroup is specified as a path; if it has a leading slash, the remainder of the path is interpreted as relative to /sys/fs/cgroup, and otherwise the entire path is interpreted relative to the cgroup in which dinit is running (as determined at startup or specified by options). The latter can only be used if there is only a single cgroup hierarchy (either the cgroups v2 hierarchy with no cgroups v1 hierarchies, or a single cgroups v1 hierarchy).

Note that due to the "no internal processes" rule in cgroups v2, a relative path must typically begin with ".." if cgroups v2 are used.

The named cgroup must already exist prior to the service starting; it will not be created by dinit.

This setting is only available if dinit was built with cgroups support.

These options are specified via the options parameter.

Specifies that this service uses the console; its input and output should be directed to the console (or precisely, to the device to which dinit's standard output stream is connected). A service running on the console prevents other services from running on the console (they will queue for the console).

Proper operation of this option (and related options) assumes that dinit is itself attached correctly to the console device (or a terminal, in which case that terminal will be used as the "console").

The interrupt key (normally control-C) may be active for process / scripted services that run on the console, depending on terminal configuration and operating-system specifics. The interrupt signal (SIGINT), however, is masked by default (but see unmask-intr).

Specifies that this service uses the console during service startup. This is identical to runs-on-console except that the console will be released (available for running other services) once the service has started. It is applicable only for bgprocess and scripted services.

As for the runs-on-console option, the interrupt key will be enabled while the service has the console.

Specifies that this service should be given access to the console (input and output will be connected to the console), but that it should not exclusively hold the console. A service given access to the console in this way will not delay the startup of services which require exclusive access to the console (see starts-on-console, runs-on-console) nor will it be itself delayed if such services are already running.

This is mutually exclusive with both starts-on-console and runs-on-console; setting this option unsets both those options, and setting either of those options unsets this option.

For services that run or start on the console, specifies that the terminal interrupt signal (SIGINT, normally invoked by control-C) should be unmasked. Handling of an interrupt is determined by the service process, but typically will cause it to terminate. This option may therefore be used to allow a service to be terminated by the user via a keypress combination. In combination with skippable, it may allow service startup to be skipped.

A service with this option will typically also have the start-interruptible option set.

Note that whether an interrupt can be generated, and the key combination required to do so, depends on the operating system's handling of the console device and, if it is a terminal, how the terminal is configured; see stty(1).

Note also that a process may choose to mask or unmask the interrupt signal of its own accord, once it has started. Shells, in particular, may unmask the signal; it might not be possible to reliably run a shell script on the console without allowing a user to interrupt it.

This service mounts the root filesystem read/write (or at least mounts the normal writable filesystems for the system). This prompts Dinit to attempt to create its control socket, if it has not already managed to do so, and similarly log boot time to the system wtmp(5) database (if supported) if not yet done. This option may be specified on multiple services, which may be useful if the wtmp database becomes writable at a different stage than the control socket location becomes writable, for example. If the control socket has already been created, this option currently causes Dinit to check that the socket "file" still exists and re-create it if not. It is not recommended to rely on this behaviour.
This service starts the system log daemon. Dinit will begin logging via the /dev/log socket.
Pass an open Dinit control socket to the process when launching it (the DINIT_CS_FD environment variable will be set to the file descriptor of the socket). This allows the service to issue commands to Dinit even if the regular control socket is not available yet.

Using this option has security implications! The service which receives the control socket must close it before launching any untrusted processes. You should not use this option unless the service is designed to receive a Dinit control socket.

Indicates that this service can have its startup interrupted (cancelled), by sending it the SIGINT signal. If service state changes such that this service will stop, but it is currently starting, and this option is set, then Dinit will attempt to interrupt it rather than waiting for its startup to complete. This is meaningful only for bgprocess and scripted services.
For scripted services, indicates that if the service startup process terminates via an interrupt signal (SIGINT), then the service should be considered started. Note that if the interrupt was issued by Dinit to cancel startup, the service will instead be considered stopped.

This can be combined with options such as starts-on-console to allow skipping certain non-essential services (such as filesystem checks) using the interrupt key (typically control-C).

Signal the service process only, rather than its entire process group, whenever sending it a signal for any reason.
Alters behaviour of the chain-to property, forcing the chained service to always start on termination of this service (instead of only when this service terminates with an exit status indicating success).
Before stopping this service, send a TERM signal and then (after a short pause) a KILL signal to all other processes in the system, forcibly terminating them. This option is intended to allow system shutdown scripts to run without any possible interference from "leftover" or orphaned processes (for example, unmounting file systems usually requires that the file systems are no longer in use).

This option must be used with care since the signal broadcast does not discriminate and potentially kills other services (or their shutdown scripts); a strict dependency ordering is suggested, i.e. every other service should either be a (possibly transitive) dependency or dependent of the service with this option set.

This option can be used for scripted and internal services only.

There are several settings for specifying process resource limits: rlimit-nofile, rlimit-core, rlimit-data and rlimit-addrspace. See the descriptions of each above. These settings place a limit on resource usage directly by the process. Note that resource limits are inherited by subprocesses, but that usage of a resource and subprocess are counted separately (in other words, a process can effectively bypass its resource limits by spawning a subprocess and allocating further resources within it).

Resources have both a hard and soft limit. The soft limit is the effective limit, but note that a process can raise its soft limit up to the hard limit for any given resource. Therefore the soft limit acts more as a sanity-check; a process can exceed the soft limit only by deliberately raising it first.

Resource limits are specified in the following format:

soft-limit:hard-limit

Either the soft limit or the hard limit can be omitted (in which case it will be unchanged). A limit can be specified as a dash, `-', in which case the limit will be removed. If only one value is specified with no colon separator, it affects both the soft and hard limit.

Some service properties specify a path to a file or directory, or a command line. For these properties, the specified value may contain one or more environment variable names, each preceded by a single `$' character, as in `$NAME'. In each case the value of the named environment variable will be substituted. The name must begin with a non-punctuation, non-space, non-digit character, and ends before the first control character, space, or punctuation character other than `_'. To avoid substitution, a single `$' can be escaped with a second, as in `$$'.

Variable substitution also supports a limited subset of shell syntax. You can use curly braces to enclose the variable, as in `${NAME}'. Limited parameter expansion is also supported, specifically the forms `${NAME:-word}' (substitute `word' if variable is unset or empty), `${NAME-word}' (substitute `word' if variable is unset), `${NAME:+word}' (substitute `word' if variable is set and non-empty), and `${NAME+word}' (substitute `word' if variable is set). Unlike in shell expansion, the substituted word does not itself undergo expansion and cannot contain closing brace characters or whitespace, even if quoted.

Note command-line variable substitution occurs after splitting the line into separate arguments and so a single environment variable cannot be used to add multiple arguments to a command line. If a designated variable is not defined, it is replaced with an empty (zero-length) string, possibly producing a zero-length argument.

Variable substitution occurs when the service is loaded. Therefore, it is typically not useful for dynamically changing service parameters (including command line) based on a variable that is inserted into dinit's environment once it is running (for example via dinitctl setenv).

The effective environment for variable substitution in setting values matches the environment supplied to the process for a service when it is launched. The priority of environment variables, from highest to lowest, for both is:

  • variables from the service env-file
  • variables set by the export-passwd-vars and export-service-name load options
  • the process environment of dinit (which is established on launch by the process environment of the parent, amended by loading the environment file (if any) as specified in dinit(8), and further amended via dinitctl setenv commands or equivalent).

Note that since variable substitution is performed on service load, the values seen by a service process may differ from those used for substitution, if they have been changed in the meantime.

Here is an example service description for the mysql database server. It has a dependency on the rcboot service (not shown) which is expected to have set up the system to a level suitable for basic operation.

# mysqld service
type = process
command = /usr/bin/mysqld --user=mysql
logfile = /var/log/mysqld.log
smooth-recovery = true
restart = false
depends-on = rcboot # Basic system services must be ready



Here is an examples for a filesystem check "service",
    run by a script (/etc/dinit.d/rootfscheck.sh). The script may need to
    reboot the system, but the control socket may not have been created, so it
    uses the pass-cs-fd option to allow the reboot command to
    issue control commands to Dinit. It runs on the console, so that output is
    visible and the process can be interrupted using control-C, in which case
    the check is skipped but dependent services continue to start.






# rootfscheck service
type = scripted
command = /etc/dinit.d/rootfscheck.sh
restart = false
options = starts-on-console pass-cs-fd
options = start-interruptible skippable
depends-on = early-filesystems  # /proc and /dev
depends-on = device-node-daemon





More examples are provided with the Dinit distribution.








Dinit, and this manual, were written by Davin McCall.







  

    October 2022
    Dinit 0.17.0pre