init(5) File Formats Manual init(5)
init - Upstart init daemon job configuration
On startup, the Upstart init(8) daemon reads its job configuration from files in the /etc/init/ directory, and watches for future changes
to these files using inotify(7).
If D-Bus has been configured to allow non-privileged users to invoke all Upstart D-Bus methods, Upstart is also able to manage User Jobs.
See User Jobs for further details.
To be considered by Upstart, files in this directory must have a recognized suffix and may also be present in sub-directories. There are
two recognized suffixes:
o Files ending in .conf are called configuration files, or simply "conf files" for short. These are the primary vehicle for specifying a
o Files ending in .override are called override files. If an override file is present, the stanzas it contains take precedence over
those equivalently named stanzas in the corresponding configuration file contents for a particular job. The main use for override
files is to modify how a job will run without having to modify its configuration file directly. See the section Override File Handling
below for further details.
A job can thus be defined by either:
o A single configuration file.
o A single configuration file and a single override file.
Unless explicitly stated otherwise, any reference to a jobs configuration can refer both to a configuration file or an override file.
Each configuration file defines the template for a single service (long-running process or daemon) or task (short-lived process).
Note that a configuration file is not itself a job: it is a description of an environmenta job could be run in. A job is the runtime
embodiment of a configuration file.
The configuration file name as displayed by Upstart and associated tooling is taken from its relative path within the directory without the
extension. For example a configuration file /etc/init/rc-sysinit.conf is named rc-sysinit, while a configuration file
/etc/init/net/apache.conf is named net/apache. Since override files only modify the way a configuration file is interpreted, they are not
Configuration files are plain text and should not be executable.
A User Job is a job configuration file created by a non-privileged user in their $HOME/.init/ directory. Job configuration files in this
directory have the same syntax as system job configuration files. Files in this directory will be read and an inotify(7) watch created the
first time a user runs initctl(8).
Any user can create user jobs, but that user can control only jobs they create.
Users are able to manage their jobs using the standard initctl(8) facility.
Note that stanzas which manipulate resources limits may cause a job to fail to start should the value provided to such a stanza attempt to
exceed the maximum value the users privilege level allows.
Note that a user job configuration file cannot have the same name as a system job configuration file.
Upstart is able to manage jobs within a chroot(2). To control jobs within the chroot environment, use the standard initctl(8) facility.
Note that it is not necessary to install D-Bus within the chroot (in fact it is not recommended).
Note that User Jobs can be created within a chroot environment.
Configuration File Format
Each line begins with a configuration stanza and continues until either the end of the line or a line containing a closing stanza. Line
breaks within a stanza are permitted within single or double quotes, or if preceeded by a blackslash.
If a stanza is duplicated, the last occurence will be used. Unrecognized stanzas will generate parse errors, which will stop a job from
Stanzas and their arguments are delimited by whitespace, which consists of one or more space or tab characters which are otherwise ignored
unless placed within single or double quotes.
Comments begin with a `#' and continue until the end of the line. Blank lines and lines consisting only of whitespace or comments are
The primary use of jobs is to define services or tasks to be run by the init(8) daemon. Each job may have one or more different processes
run as part of its lifecycle, with the most common known as the main process.
The main process is defined using either the exec or script stanzas, only one of which is permitted. These specify the executable or shell
script that will be run when the job is considered to be running. Once this process terminates, the job stops.
All processes are run with the full job environment available as environment variables in their process.
exec COMMAND [ ARG ]...
This stanza defines the process to be run as the name of an executable on the filesystem, and zero or more arguments to be passed to
it. Any special characters, e.g. quotes or `$' specified will result in the entire command being passed to a shell for expansion.
exec /usr/sbin/acpid -c $EVENTSDIR -s $SOCKET
script ... end script
This stanza defines the process to be run as a shell script that will be executed using sh(1). The -e shell option is always used,
so any command that fails will terminate the script.
The script stanza appears on its own on a line, the script is everything up until the first end script stanza appearing on its own
on a line.
dd bs=1 if=/proc/kmsg of=$KMSGSINK
exec /sbin/klogd -P $KMSGSINK
There are an additional four processes that may be run as part of the job's lifecycle. These are specified as the process name, followed
by an exec or script stanza.
This process will be run after the job's starting(7) event has finished, but before the main process is run. It is typically used
to prepare the environment, such as making necessary directories.
This process will be run before the job's started(7) event is emitted, but after the main process has been spawned. It is typically
used to send necessary commands to the main process, or to delay the started(7) event until the main process is ready to receive
This process is run if the job is stopped by an event listed in its stop on stanza or by the stop(8) command. It will be run before
the job's stopping(7) event is emitted and before the main process is killed. It is typically used to send any necessary shutdown
commands to the main process, and it may also call the start(8) command without arguments to cancel the stop.
This process is run after the main process has been killed and before the job's stopped(7) event is emitted. It is typically used
to clean up the environment, such as removing temporary directories.
All of these processes, including the main process, are optional. Services without a main process will appear to be running until they are
stopped: this is commonly used to define states such as runlevels. It is permissable to have no main process, but to have pre-start and
post-stop processes for the state.
pre-start exec ifup -a
post-stop exec ifdown -a
Jobs can be manually started and stopped at any time by a system adminstrator using the start(8) and stop(8) tools, however it is far more
useful for jobs to be started and stopped automatically by the init(8) daemon when necessary.
This is done by specifying which events should cause your job to be started, and which cause your process to be stopped again.
The set of possible events is limitless, however there are a number of standard events defined by the init(8) daemon and telinit(8) tools
that you will want to use.
When first started, the init(8) daemon will emit the startup(7) event. This will activate jobs that implement System V compatibility and
the runlevel(7) event. As jobs are started and stopped, the init(8) daemon will emit the starting(7), started(7), stopping(7) and
stopped(7) events on their behalf.
start on EVENT [[KEY=]VALUE]... [and|or...]
The start on stanza defines the set of events that will cause the job to be automatically started. Each EVENT is given by its name.
Multiple events are permitted using the and & or operators, and complex expressions may be performed with parentheses (within which
line breaks are permitted).
You may also match on the environment variables contained within the event by specifying the KEY and expected VALUE. If you know
the order in which the variables are given to the event you may omit the KEY.
VALUE may contain wildcard matches and globs as permitted by fnmatch(3) and may expand the value of any variable defined with the
Negation is permitted by using != between the KEY and VALUE.
start on started gdm or started kdm
start on device-added SUBSYSTEM=tty DEVPATH=ttyS*
start on net-device-added INTERFACE!=lo
stop on EVENT [[KEY=]VALUE]... [and|or...]
The stop on stanza defines the set of events that will cause the job to be automatically stopped. It has the same syntax as start
VALUE may additionally expand the value of any variable that came from the job's start environment (either the event or the command
that started it).
stop on stopping gdm or stopping kdm
stop on device-removed DEVPATH=$DEVPATH
manual This stanza will disregard any previously seen start on definition. By adding this stanza on any line below the start on defini-
tion, it provides the ability to stop a job from being automatically started. When specified, the only way to start such a job is
via start (8).
Each job is run with the environment from the events or commands that started it. In addition, you may define defaults in the job which
may be overridden later and specify which environment variables are exported into the events generated for the job.
The special UPSTART_EVENTS environment variable contains the list of events that started the job, it will not be present if the job was
In addition, the pre-stop and post-stop scripts are run with the environment of the events or commands that stopped the job. The
UPSTART_STOP_EVENTS environment variable contains the list of events that stopped the job, it will not be present if the job was stopped
All jobs also contain the UPSTART_JOB and UPSTART_INSTANCE environment variables, containing the name of the job and instance. These are
mostly used by the initctl(8) utility to default to acting on the job the commands are called from.
Defines a default environment variable, the value of which may be overriden by the event or command that starts the job. If
'KEY=VALUE' is specified, the variable KEY is given the value VALUE. If only 'KEY' is given, then the value is taken from the
init(8) daemon's own environment.
Exports the value of an environment variable into the starting(7), started(7), stopping(7) and stopped(7) events for this job and to
all resultant events (not just those relating to the current job).
Services, tasks and respawning
Jobs are services by default. This means that the act of starting the job is considered to be finished when the job is running, and that
even exiting with a zero exit status means the service will be respawned.
task This stanza may be used to specify that the job is a task instead. This means that the act of starting the job is not considered to
be finished until the job itself has been run and stopped again, but that exiting with a zero exit status means the task has com-
pleted successfully and will not be respawned.
The start(8) command, and any starting(7) or stopping(7) events will block only until a service is running or until a task has finished.
A service or task with this stanza will be automatically started if it should stop abnormally. All reasons for a service stopping,
except the stop(8) command itself, are considered abnormal. Tasks may exit with a zero exit status to prevent being respawned.
respawn limit COUNT INTERVAL
Respawning is subject to a limit, if the job is respawned more than COUNT times in INTERVAL seconds, it will be considered to be
having deeper problems and will be stopped. Default COUNT is 10. Default INTERVAL is 5 seconds.
This only applies to automatic respawns and not the restart(8) command.
normal exit STATUS|SIGNAL...
Additional exit statuses or even signals may be added, if the job process terminates with any of these it will not be considered to
have failed and will not be respawned.
normal exit 0 1 TERM HUP
By default, only one instance of any job is permitted to exist at one time. Attempting to start a job when it's already starting or run-
ning results in an error. Note that a job is considered to be running if its pre-start process is running.
Multiple instances may be permitted by defining the names of those instances. If an instance with the same name is not already starting or
running, a new instance will be started instead of returning an error.
This stanza defines the names of instances, on its own its not particularly useful since it would just define the name of the single
permitted instance, however NAME expands any variable defined in the job's environment.
These will often be variables that you need to pass to the process anyway, so are an excellent way to limit the instances.
exec /sbin/httpd -c $CONFFILE
exec /sbin/getty -8 38300 $TTY
These jobs appear in the initctl(8) output with the instance name in parentheses, and have the INSTANCE environment variable set in
Upstart provides several stanzas useful for documentation and external tools.
This stanza may contain a description of the job.
description "This does neat stuff"
This stanza may contain the author of the job, often used as a contact for bug reports.
author "Scott James Remnant <email@example.com>"
This stanza may contain version information about the job, such as revision control or package version number. It is not used or
interpreted by init(8) in any way.
All processes on the system are free to emit their own events by using the initctl(8) tool, or by communicating directly with the
This stanza allows a job to document in its job configuration what events it emits itself, and may be useful for graphing possible
The initctl(8) check-config command attempts to use this stanza to resolve events.
EVENT can be either a literal string or a string including shell wildcard meta-characters (asterisk ('*'), question mark ('?'), and
square brackets ('[' and ']')). Meta-characters are useful to allow initctl(8) check-config to resolve a class of events, such as
those emitted by upstart-udev-bridge(8).
Many common adjustments to the process environment, such as resource limits, may be configured directly in the job rather than having to
handle them yourself.
By default the standard input, output and error file descriptors of jobs are connected to /dev/null
If this stanza is specified, they are connected to /dev/console instead.
console owner is special, it not only connects the job to the system console but sets the job to be the owner of the system console,
which means it will receive certain signals from the kernel when special key combinations such as Control-C are pressed.
A common configuration is to set the file mode creation mask for the process. UMASK should be an octal value for the mask, see
umask(2) for more details.
Another common configuration is to adjust the process's nice value, see nice(1) for more details.
oom score ADJUSTMENT|never
Normally the OOM killer regards all processes equally, this stanza advises the kernel to treat this job differently.
ADJUSTMENT may be an integer value from -999 (very unlikely to be killed by the OOM killer) up to 1000 (very likely to be killed by
the OOM killer). It may also be the special value never to have the job ignored by the OOM killer entirely.
Runs the job's processes in a chroot(8) environment underneath DIR
Note that DIR must have all the necessary system libraries for the process to be run, often including /bin/sh
Runs the job's processes with a working directory of DIR instead of the root of the filesystem.
limit LIMIT SOFT|unlimited HARD|unlimited
Sets initial system resource limits for the job's processes. LIMIT may be one of core, cpu, data, fsize, memlock, msgqueue, nice,
nofile, nproc, rss, rtprio, sigpending or stack.
Limits are specified as both a SOFT value and a HARD value, both of which are integers. The special value unlimited may be speci-
fied for either.
Override File Handling
Override files allow a jobs environment to be changed without modifying the jobs configuration file. Rules governing override files:
o If a job is embodied with only a configuration file, the contents of this file define the job.
o If an override files exists where there is no existing cofiguration file, the override file is ignored.
o If both a configuration file and an override file exist for a job and both files are syntactically correct:
o stanzas in the override file will take precedence over stanzas present in the corresponding configuration file.
o stanzas in the override file which are not present in the corresponding configuration file will be honoured when the job runs.
o If both a configuration file and an override file exist for a job and subsequently the override file is deleted, the configuration file
is automatically reloaded with the effect that any changes introduced by the override file are undone and the configuration file alone
now defines the job.
o If both a configuration file and an override file exist for a job and subsequently the configuration file is deleted, a new instance of
the job can no longer be started (since without a corresponding configuration file an override file is ignored).
o If both a configuration file and an override file exist for a job and any of the contents of the override file are invalid, the override
file is ignored and only the contents of the configuration file are considered.
kill signal SIGNAL
Specifies the stopping signal, SIGTERM by default, a job's main process will receive when stopping the running job.
kill signal INT
kill timeout INTERVAL
Specifies the interval between sending the job's main process the "stopping" (see above) and SIGKILL signals when stopping the run-
ning job. Default is 5 seconds.
Specifies that the job's main process will raise the SIGSTOP signal to indicate that it is ready. init(8) will wait for this signal
before running the job's post-start script, or considering the job to be running.
init(8) will send the process the SIGCONT signal to allow it to continue.
Specifies that the job's main process is a daemon, and will fork twice after being run. init(8) will follow this daemonisation, and
will wait for this to occur before running the job's post-start script or considering the job to be running.
Without this stanza init(8) is unable to supervise daemon processes and will believe them to have stopped as soon as they daemonise
Specifies that the job's main process will fork once after being run. init(8) will follow this fork, and will wait for this to
occur before running the job's post-start script or considering the job to be running.
Without this stanza init(8) is unable to supervise forking processes and will believe them to have stopped as soon as they fork on
The use of symbolic links in job configuration file directories is not supported since it can lead to unpredicable behaviour resulting from
broken or inaccessible links (such as would be caused by a link crossing a filesystem boundary to a filesystem that has not yet been
The and and or operators allowed with start on and stop on do not work intuitively: operands to the right of either operator are only eval-
uated once and state information is then discarded. This can lead to jobs with complex start on or stop on conditions not behaving as
expected when restarted. For example, if a job encodes the following condition:
start on A and (B or C)
When 'A' and 'B' become true, the condition is satisfied so the job will be run. However, if the job ends and subsequently 'A' and 'C'
become true, the job will not be re-run even though the condtion is satisfied. Avoid using complex conditions with jobs which need to be
Manual page written by Scott James Remnant <firstname.lastname@example.org> and James Hunt <email@example.com>.
Report bugs at <https://launchpad.net/upstart/+bugs>
Copyright (C) 2009-2011 Canonical Ltd.
This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICU-
init(8) initctl(8) sh(1) upstart-events(7)
Upstart 2011-05-12 init(5)