Re: [PATCH 2/3] perf_events: fix validation of events using an extra reg (v4)

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2011-06-21 at 11:54 +0200, Stephane Eranian wrote:
> On Thu, Jun 9, 2011 at 10:48 PM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Thu, 2011-06-09 at 22:36 +0200, Stephane Eranian wrote:
> >> On Thu, Jun 9, 2011 at 8:40 PM, Peter Zijlstra <peterz@infradead.org> wrote:
> >> > On Mon, 2011-06-06 at 16:57 +0200, Stephane Eranian wrote:
> >> >> +static struct cpu_hw_events *allocate_fake_cpuc(void)
> >> >> +{
> >> >> +       struct cpu_hw_events *cpuc;
> >> >> +       int cpu = smp_processor_id();
> >> >
> >> > That's a boo-boo, clearly we are in a preemptible context here (see the
> >> > GFP_KERNEL allocation on the next line), so using smp_processor_id()
> >> > isn't valid.
> >> >
> >> Good point. I missed that.
> >
> > Yeah, I did too, Ingo found it during testing.
> >
> >> > Now since all that allocate_shared_regs() does with it is pick a NUMA
> >> > node, we should probably use raw_smp_processor_id() and leave it at
> >> > that, right?
> >> >
> Looked at that some more. It is more subtle than this.
> allocate_shared_regs() is used in two places:
>   - allocate_fake_cpuc()
>   - intel_pmu_cpu_prepare()
> 
> In the first case, it does not matter where the cpuc is allocated,
> it's not on any
> critical path. But for the other situation, it'd better be allocated
> on the cpu node.
> But I think that is what we get given it is called during the CPU
> hotplug prepare
> path, so it must be running on the CPU to prepare and thus a kzalloc() should
> allocate on the right node, right?

Yeah. Let me shove these patches mingo wards again.