To follow up on my post, it seems that the Vserver scheduler has no effect
on the KVM processes. You can have a very intensive task in a KVM (QEMU
running in a Vserver) and the token counters in /proc/virtual/<xid>/sched
do no go up at all.
I've read something about the CPU time accounting not working for KVM, but
I think it's been fixed in the latest kernels.
I did my tests on 2.6.27.5-vs2.3.0.35.9 (I used the patch in Experimental
for 2.6.27.4, which applies to 2.6.27.5 cleanly as well).
I think it'd be neat to be able to control KVM guests with the VServer
token-bucket scheduler.
Grisha
On Thu, 6 Nov 2008, Gregory (Grisha) Trubetskoy wrote:
>
> I've been able to start the KVM QEMU in a VServer guest without any problems,
> and with prettu much default configuration, i.e. no special [b]caps were
> needed (haven't quite figured out the networking part though).
>
> All I needed was a /dev/kvm in a guest (and the kvm kernel modules loaded, as
> well as a VT capable CPU, of course).
>
> I wonder if anyone else here has played with this? What are security
> implications of a guest having a /dev/kvm? Can a guest with access to
> /dev/kvm wreak havoc on other KVM's (possibly running in other guests or in
> the host)?
>
> Grisha
>
>