Oh, I also see that my claws-mail filters are in chaos and that there's additional email to this mail-thread. On Thu, 21 May 2020 16:14:34 +0100 Stefan Hajnoczi <stefanha@redhat.com> wrote:

On Wed, May 20, 2020 at 07:23:32PM +0200, Paolo Bonzini wrote:

...

On 20/05/20 19:13, Stefan Hajnoczi wrote:

...

Hi, libvirt refuses to set KVM_HINTS_DEDICATED when the CPU model is not host-passthrough.

Is there a reason for this limitation?

My understanding is that KVM_HINTS_DEDICATED means the vCPU is pinned to a host CPU that is not shared with other tasks. Any KVM vCPU should be able to support this feature, regardless of whether host-passthrough is used or not.

Perhaps it's because KVM_HINTS_DEDICATED enables mwait passthrough, and the details of mwait vary per CPU model?

KVM_HINTS_DEDICATED (now known as KVM_HINTS_REALTIME) is also used by cpuidle-haltpoll. In that case we don't care about mwait.

CPU passthrough can prevent live migration, so I think we should relax this check, if possible.

I agree, the check for <cpu mode='host-passthrough'/> should be relaxed.

Wim: Can you let us know what the original intention of checking for cpu-passthrough was?

From our, my proposed, point of view it was. But that seems silly indeed.

Rgds, - Wim10H.

Thanks, Stefan

On Mon, Jun 29, 2020 at 01:58:13PM +0200, Wim ten Have wrote:

Hi Stefan, sorry for my very late response as I was temporarily out and unfortunately missed this email thread.

For your question (remark), yes, you're right. This feature should not apply to specific <cpu mode='host-passthrough'... /> scenario. Something else to look into is that the feature can only work if the serving layer underneath by capabilities does support the request.

Short, yes, this needs correction.

Thank you! I have filed a bugzilla so that this issue is tracked: https://bugzilla.redhat.com/show_bug.cgi?id=1857671 Maybe someone will have time to work on it. Stefan