[systemd-devel] unit been dead before been running

[Lennart Poettering]

On Mo, 18.02.19 16:46, aleivag (aleivag at gmail.com) wrote:

> 1.- Is this change in behaviour, intended, or a side effect of something
> else?

When a unit is first created it is in "dead" state. It only changes to
"active" state as soon as it actually starts running.

Note that PropertiesChanged is currently sent a bit too eagerly: when
some property changes we'll update you with the full data, not just
what changed. In an older verison we forgot to send out these change
messages in some conditions, which was fixed (iirc something about the
job assignment changing).

Note that a unit may stay in "dead" state for a while actually before
it moves to "active", since dependencies might need to be started first.

In general: if you want to properly watch a transient service starting
and then running and exiting, the way to do this is like this:

1. First watch for the JobRemoved event for the job you got back when
   you create the transient unit.

2. Now wait until the unit itself for PropertiesChanged, and wait
   until it transitions back do "dead" or "failed".

3. In some cases a unit might go directly to "dead" again (without
   going through "active"; for example when a condition is defined
   that results in the job to be shortcut). Because of this, after you
   saw the JobRemoved, and you subscribed to PropertieChanged you
   should explicitly query the ActiveStatus property to see where are
   at.

> 2.- was what i was doing the best way to decide if a transient unit has
> finished? i'm super temped to instead check when the
> `org.freedesktop.systemd1.Service.MainPID == 0` i think its the best idea,
> but maybe you guys have one that is better.

Not all processes have a main PID (i.e. Type=forking where no main
process can be identified, or those where RemaiNAfterExit= is set and
the service exited), hence checking that is not the best approach.

But yeah, the right approach is to wait until you saw JobRemoved for
the job you installed. And then wait until the unit is dead/failed
again, see above.

Lennart

--
Lennart Poettering, Red Hat