[systemd-devel] [ANNOUNCE] systemd 198

[Lennart Poettering]

On Fri, 08.03.13 16:53, Oleksii Shevchuk (alxchk at gmail.com) wrote:

> 
> > In fact, the drop-in logic I added primarily to have a nice place
> > where "set-cgroup-attr" could be made persistent.
> 
> Maybe it's better to have generic interface for setting parameters,
> instead of this specific one?

The reason I made this explicitly specific to cgroups is that they are
more or less the only ones which really apply to a service as a
whole. ALl the other stuff, like nice levels, resource limits, env
blocks, and so on is (if it is settable at runtime at all) per-process,
and not per-service.

Now, for setting these settings we could certainly just go and apply
them to all processes of a service. But already that is not so nice,
since actually daemons might have worker daemons that run with different
settings than the main daemon, and we shouldn't blindly override
that. But it gets increasingly messy as soon as you think about reading
the current values from the processes, because you then have to pick one
process.

I mean, it's not inherently not doable, but it does get easily
non-obvious if we start to make all props dynamically changeable. And so
I wanted to side-step the whole issue, and just made the cgroup stuff
settable and nothing else.

Also for most cases the cgroup stuff is a ton more useful than the
per-process stuff. for examle the RLIMIT_xyz stuff is much less useful
thatn cgroup resource controllers. And the nice level and io prio and
suchlike is a lot less useful than cpu.shares and blk.weight and so
on... So I didn't really see the need to expose this more.

Lennart

-- 
Lennart Poettering - Red Hat, Inc.