Re: [PATCH 3/4] perf, nmi: Move LVT un-masking into irq handlers

[Don Zickus <dzickus@redhat.com>]

On Mon, Apr 25, 2011 at 06:15:25PM +0400, Cyrill Gorcunov wrote:
> 
>   Don't get me wrong please but the whole picture of what is happening can be seen only when
> all caller sequence is taken into account and once (for some reason) the sequence
> get changed the "detailed" comment would simply mess the comment reader so I think
> the former comment is detailed enough and what is more important it's "general" enough
> so it doesn't depend on when code is called but points a reader on hw details and it's
> up to a reader to check "current" calling sequence because kernel code changes too
> damn fast ;)

I think Ingo is looking for is something like

/*
 * It has been observed that quirks in the P4 perf hw has forced the
 * following sequence of events to happen in the order below
 *
 * - clear the OVF bit (as it will continue to assert the NMI line)
 * - unmask the apic LVTPC bit to allow NMIs from the PMU again
 * - optionally re-enable the PMU to count events again
 *
 * Un-masking the apic prematurely (before clearing the OVF bit) has led
 * to the creation of a second NMI event (which led to the unknown NMI
 * warnings) due to the fact that the PMU will continue to generate an
 * interrupt until its OVF bit is cleared.
 */

Something that specifically documents what we saw, why we saw it and what
we are doing to avoid it.

Cheers,
Don