Re: [PATCH 3/9] perf/x86: Add is_hardware_event

[Mark Rutland <mark.rutland@arm.com>]

On Fri, Jul 17, 2015 at 04:03:36PM +0100, Liang, Kan wrote:
> > 
> > On Thu, Jul 16, 2015 at 09:33:45PM +0100, kan.liang@intel.com wrote:
> > > From: Kan Liang <kan.liang@intel.com>
> > >
> > > Using is_hardware_event to replace !is_software_event to indicate a
> > > hardware event.
> > 
> > Why...?
> 
> First, the comments of is_software_event is not correct. 
> 0 or !is_software_event is not for a hardware event.
> is_hardware_event is for a hardware event.

Circular logic is fantastic.

> Also, the following patch make mix core_misc event be part of hw/sw
> event, !is_software_event could be either hw event or core_misc event.

!is_software_event is also true for an uncore event currently, and the
code relies on this fact. Blindly replacing !is_software_event with
is_hardware_event changes the behaviour of the code for uncore events.

> > For an uncore event e, is_hardware_event(e) != !is_software_event(e),
> > so this will be a change of behaviour...
> 
> Uncore event cannot be part of hw/sw event group. So it doesn't change the behavior. 

My complaint had _nothing_ to do with groups. It had to do with the
accounting for throttling, where it _does_ change the behaviour.

However, now that you mention the group logic...

> > >  /*
> > > - * Return 1 for a software event, 0 for a hardware event
> > > + * Return 1 for a software event, 0 for other event
> > >   */
> > >  static inline int is_software_event(struct perf_event *event)  {
> > >  	return event->pmu->task_ctx_nr == perf_sw_context;  }
> > >
> > > +static inline int is_hardware_event(struct perf_event *event) {
> > > +	return event->pmu->task_ctx_nr == perf_hw_context; }
> > > +
> > >  extern struct static_key
> > perf_swevent_enabled[PERF_COUNT_SW_MAX];
> > >
> > >  extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64); diff
> > > --git a/kernel/events/core.c b/kernel/events/core.c index
> > > d3dae34..9077867 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -1347,7 +1347,7 @@ static void perf_group_attach(struct
> > perf_event *event)
> > >  	WARN_ON_ONCE(group_leader->ctx != event->ctx);
> > >
> > >  	if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
> > > -			!is_software_event(event))
> > > +			is_hardware_event(event))
> > >  		group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
> > >

...this changes the behaviour of attaching an uncore event to a software
group.

Before, we'd correctly clear the PERF_GROUP_SOFTWARE flag on the leader.
After this patch, we don't. That is a bug.

My original complaint was with the changes below.

> > >  	list_add_tail(&event->group_entry, &group_leader->sibling_list);
> > @@
> > > -1553,7 +1553,7 @@ event_sched_out(struct perf_event *event,
> > >  	event->pmu->del(event, 0);
> > >  	event->oncpu = -1;
> > >
> > > -	if (!is_software_event(event))
> > > +	if (is_hardware_event(event))
> > >  		cpuctx->active_oncpu--;
> > >  	if (!--ctx->nr_active)
> > >  		perf_event_ctx_deactivate(ctx);

Previously we'd call perf_event_ctx_deactivate() for an uncore PMU's
contexts, but now we never will.

> > > @@ -1881,7 +1881,7 @@ event_sched_in(struct perf_event *event,
> > >  		goto out;
> > >  	}
> > >
> > > -	if (!is_software_event(event))
> > > +	if (is_hardware_event(event))
> > >  		cpuctx->active_oncpu++;
> > >  	if (!ctx->nr_active++)
> > >  		perf_event_ctx_activate(ctx);

Similarly for perf_event_ctx_deactivate().

As I mention below, That means we will no longer perform throttling for
an uncore PMU's cpu context (see perf_event_task_tick()).

> > ... whereby we won't accuont uncore events as active, and thereforef will
> > never perform throttling.
> > 
> > That doesn't sound right.
> 
> I think active_oncpu should only impact if the group is exclusive.
> The changes will make pure perf_invalid_context event group never exclusive.
> If that's a problem, I will change this part back.

I'm not sure what you mean here -- I can't see what a group being
exclusive has to do with any of the points above.

What am I missing?

Thanks,
Mark.