bug#55444: elogind startup race between shepherd and dbus-daemon

["Ludovic Courtès" <ludo@gnu.org>]

Hi,

Liliana Marie Prikler <liliana.prikler@gmail.com> skribis:

> Am Montag, dem 16.05.2022 um 10:26 +0200 schrieb Ludovic Courtès:
>> [...]
>> So it would seem that the solution to this is to prevent dbus-daemon
>> from starting elogind.  We can do that by changing
>> org.freedesktop.login1.service so that it has “Exec=true” instead of
>> “Exec=elogind --daemon”.
>> 
>> “Exec=true” is a bit crude because it doesn’t guarantee that elogind
>> is really started; if that isn’t good enough, we could instead wait
>> for the PID file or something (as of Shepherd 0.9.0, invoking ‘herd
>> start elogind’ potentially leads shepherd to start a second instance
>> if the first one is still being started, so we can’t really do that).
> Why does shepherd race with itself here?  That sounds like a very evil
> bug.  Rather than waiting for a log file, I'd suggest writing an ad-hoc
> Guile script that communicates with shepherd and blocks until shepherd
> signals that elogind has been started, but this script too would have
> to work around shepherd racing against itself.

Right.  Currently services have two states: stopped, and started.
Fixing that needs non-trivial changes to how shepherd handles state.

We’ll have to do that (the way I see it, we’ll move state out of
<service> and have a fiber explicitly handle state, including
distinguishing between “stopped” and “starting”), but I think/hope we
can fix this bug without first addressing this issue.

>> Depending on what we end up with, we might also revisit whether
>> xorg-server needs to explicitly depend on elogind.
> At least in the case of GDM I think it does heavily depend on elogind.
> For the future, I think we also should take over dbus-daemon's
> autostart in the same way systemd already has.

Agreed, though that one is trickier: we’d need an implementation of the
D-Bus protocol.  There’s guile-ac-d-bus but it’s probably under-tested.

Thanks,
Ludo’.