Re: [PATCH 3/4] perf: xgene: Add APM X-Gene SoC Performance Monitoring Unit driver

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

On Mon, Apr 04, 2016 at 04:42:11PM -0700, Tai Tri Nguyen wrote:
> On Fri, Apr 1, 2016 at 5:18 AM, Mark Rutland <mark.rutland@arm.com> wrote:
> >> +static int get_next_avail_cntr(struct xgene_pmu_dev *pmu_dev)
> >> +{
> >> +     int shift, cntr, retval;
> >> +     unsigned long flags;
> >> +
> >> +     raw_spin_lock_irqsave(&pmu_dev->lock, flags);
> >> +
> >> +     for (cntr = 0; cntr < pmu_dev->max_counters; cntr++) {
> >> +             shift = cntr;
> >> +             if (!(pmu_dev->cntr_assign_mask & (1ULL << shift))) {
> >> +                     pmu_dev->cntr_assign_mask |= (1ULL << shift);
> >> +                     retval = cntr;
> >> +                     goto out;
> >> +             }
> >> +     }
> >> +     retval = -ENOSPC;

> > Are the spinlocks necessary?
> >
> > I thought add and del couldn't be called in parallel for the same
> > context, and those are the only users of this mask.
> >
> 
> I'm trying to avoid the case where multiple events may claim the same
> available counter.
> There's a race condition here.

I don't think there is, so long as events are all associated with the same CPU,
and hence the same ctx.

As I mentioned, add and del are the only users of this mask. Both of those are
only called with ctx->lock held, so I couldn't see how these could race.

Were you considering events in different cpu contexts racing?

Is there something I've missed?

> >> +     hwc->config = config;
> >> +     if (config1)
> >> +             hwc->extra_reg.config = config1;
> >> +     else
> >> +             /* Enable all Agents */
> >> +             hwc->extra_reg.config = 0xFFFFFFFFFFFFFFFFULL;
> >
> > I'm not sure I follow what's going on here.
> >
> > It would be good to document precisely what this means.
> 
> These are X-Gene PMU specific for monitoring performance of a specific
> data path.
> X-Gene PMUs have 2 registers capable of masking the Agents from which
> the request come from. If the bit with the bit number corresponding to
> the AgentID
> is set, the event will be counted only if it is caused by a request
> from that agent.
> Each PMU has different set of Agents. By default, the event will be counted for
> all agent requests.
> 
> I'll have it commented better for next revision of the patch.

It might be worth having something under Documentation/ for this, similarly to
what we do for CCN in Documentation/arm/CCN.txt.

How is the user expected to determine agent IDs? Is there a listing somewhere?
Does this change between reivisions? This may be worth documenting.

Thanks,
Mark.