Re: [Patch v3 0/2] cgroup: KVM: New Encryption IDs cgroup controller

[David Rientjes]

On Thu, 10 Dec 2020, Christian Borntraeger wrote:

> > * However, the boilerplate to usefulness ratio doesn't look too good and I
> >   wonder whether what we should do is adding a generic "misc" controller
> >   which can host this sort of static hierarchical counting. I'll think more
> >   on it.
> 
> We first dicussed to have
> encryption_ids.stat
> encryption_ids.max
> encryption_ids.current
> 
> and we added the sev in later, so that we can also have tdx, seid, sgx or whatever.
> Maybe also 2 or more things at the same time.
> 
> Right now this code has
> 
> encryption_ids.sev.stat
> encryption_ids.sev.max
> encryption_ids.sev.current
> 
> And it would be trivial to extend it to have
> encryption_ids.seid.stat
> encryption_ids.seid.max
> encryption_ids.seid.current
> on s390 instead (for our secure guests).
> 
> So in the end this is almost already a misc controller, the only thing that we
> need to change is the capability to also define things other than encryption.*.*
> And of course we would need to avoid adding lots of random garbage to such a thing.
> 
> But if you feel ok with the burden to keep things kind of organized a misc
> controller would certainly work for the encryption ID usecase as well. 
> So I would be fine with the thing as is or a misc controlĺer.
> 

Yeah, I think generalization of this would come in the form of either (1) 
the dumping ground of an actual "misc" controller, that you elude to, or 
(2) a kernel abstraction so you can spin up your own generic controller 
that has the {current, max, stat} support.  In the case of the latter, 
encryption IDs becomes a user of that abstraction.

Concern with a single misc controller would be that any subsystem that 
wants to use it has to exactly fit this support: current, max, stat, 
nothing more.  The moment a controller needs some additional support, and 
its controller is already implemented in previous kernel versionv as a 
part of "misc," we face a problem.

On the other hand, a kernel abstraction that provides just the basic 
{current, max, stat} support might be interesting if it can be extended by 
the subsystem instance using it.