Re: [PATCH v2] KVM: x86: avoid redundant REQ_EVENT

From: Roman Kagan
Date: Thu Dec 15 2016 - 09:57:35 EST

On Thu, Dec 15, 2016 at 03:32:45PM +0100, Paolo Bonzini wrote:
> On 15/12/2016 15:30, Radim KrÄmÃÅ wrote:
> >
> > One useless round of KVM_REQ_EVENT is not going change nested
> > performance by much and it is not the only thing we could improve wrt.
> > TPR ... I would just leave it for now and take care of it when we
> > * don't to update PPR at all with APICv -- it is already correct
> > * drop the KVM_REQ_EVENT with flex priority, because lower TPR cannot
> > unmask an interrupt
> I agree. I still don't like the patch very much, because I feel like an
> explicit state machine ("can KVM_REQ_EVENT do anything?") would be more
> maintainable.

We all seem to share that feeling towards this patch :) That's the
reason why it was baking here internally for a long time: Denis
discovered this scenario over a month ago while analyzing the
performance regressions in KVM against our proprietary hypervisor, but
pinning down a palatable and safe fix turned out to be a challenge.

I think we did our best to stay safe; I agree that it ended up no so
beautiful indeed.