RE: [PATCH v1 8/8] perf/x86: Add event owner check when PEBS output to Intel PT

[Kang, Luwei]

> > For PEBS output to Intel PT, a Intel PT event should be the group
> > leader of an PEBS counter event in host. For Intel PT virtualization
> > enabling in KVM guest, the PT facilities will be passthrough to guest
> > and do not allocate PT event from host perf event framework. This is
> > different with PMU virtualization.
> >
> > Intel new hardware feature that can make PEBS enabled in KVM guest by
> > output PEBS records to Intel PT buffer. KVM need to allocate a event
> > counter for this PEBS event without Intel PT event leader.
> >
> > This patch add event owner check for PEBS output to PT event that only
> > non-kernel event need group leader(PT).
> >
> > Signed-off-by: Luwei Kang <luwei.kang@xxxxxxxxx>
> > ---
> >  arch/x86/events/core.c     | 3 ++-
> >  include/linux/perf_event.h | 1 +
> >  kernel/events/core.c       | 2 +-
> >  3 files changed, 4 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c index
> > 7b21455..214041a 100644
> > --- a/arch/x86/events/core.c
> > +++ b/arch/x86/events/core.c
> > @@ -1014,7 +1014,8 @@ static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader,
> >  		 * away, the group was broken down and this singleton event
> >  		 * can't schedule any more.
> >  		 */
> > -		if (is_pebs_pt(leader) && !leader->aux_event)
> > +		if (is_pebs_pt(leader) && !leader->aux_event &&
> > +					!is_kernel_event(leader))
> 
> indent fail, but also, I'm not sure I buy this.
> 
> Surely pt-on-kvm has a perf event to claim PT for the vCPU context?

Hi Peter,
    PT on KVM will not allocate perf events from host (this is different from performance counter). The guest PT MSRs value will be load to hardware directly before VM-entry.
    A PT event is needed by PEBS event as the event group leader in native. In virtualization, we can allocate a counter for PEBS but can't assign a PT event as the leader of this PEBS event.

> 
> Even if not, this is not strictly correct. Not even now is KVM the sole user of perf_event_create_kernel_counter(), so saying any
> kernel event is excempt from this scheduling constraint is jsut wrong.

KVM is not the sole user of perf_event_create_kernel_counter() but I think KVM is the only user that needs group leader events. I found the kernel event check (63b6da39bb) is from you and hope can get some suggestions on this.

Thanks,
Luwei Kang

> 
> >  			return -EINVAL;
> >
> >  		/*