Re: [PATCH v1 01/11] perf/x86/core: Support KVM to assign a dedicated counter for guest PEBS

[Peter Zijlstra]

On Fri, Mar 06, 2020 at 09:42:47AM -0500, Liang, Kan wrote:
> 
> 
> On 3/6/2020 8:53 AM, Peter Zijlstra wrote:
> > On Fri, Mar 06, 2020 at 01:56:55AM +0800, Luwei Kang wrote:
> > > From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > > 
> > > The PEBS event created by host needs to be assigned specific counters
> > > requested by the guest, which means the guest and host counter indexes
> > > have to be the same or fail to create. This is needed because PEBS leaks
> > > counter indexes into the guest. Otherwise, the guest driver will be
> > > confused by the counter indexes in the status field of the PEBS record.
> > > 
> > > A guest_dedicated_idx field is added to indicate the counter index
> > > specifically requested by KVM. The dedicated event constraints would
> > > constrain the counter in the host to the same numbered counter in guest.
> > > 
> > > A intel_ctrl_guest_dedicated_mask field is added to indicate the enabled
> > > counters for guest PEBS events. The IA32_PEBS_ENABLE MSR will be switched
> > > during the VMX transitions if intel_ctrl_guest_owned is set.
> > > 
> > 
> > > +	/* the guest specified counter index of KVM owned event, e.g PEBS */
> > > +	int				guest_dedicated_idx;
> > 
> > We've always objected to guest 'owned' counters, they destroy scheduling
> > freedom. Why are you expecting that to be any different this time?
> > 
> 
> The new proposal tries to 'own' a counter by setting the event constraint.
> It doesn't stop other events using the counter.
> If there is high priority event which requires the same counter, scheduler
> can still reject the request from KVM.
> I don't think it destroys the scheduling freedom this time.

Suppose your KVM thing claims counter 0/2 (ICL/SKL) for some random PEBS
event, and then the host wants to use PREC_DIST.. Then one of them will
be screwed for no reason what so ever.

How is that not destroying scheduling freedom? Any other situation we'd
have moved the !PREC_DIST PEBS event to another counter.