Re: [PATCH -v3] perf, x86: try to handle unknown nmis with running perfctrs

[Don Zickus <dzickus@redhat.com>]

On Thu, Aug 26, 2010 at 12:24:58AM +0400, Cyrill Gorcunov wrote:
> On Wed, Aug 25, 2010 at 04:11:06PM -0400, Don Zickus wrote:
> ...
> > >  Uhhuh. NMI received for unknown reason 00 on CPU 15.
> > >  Do you have a strange power saving mode enabled?
> > >  Dazed and confused, but trying to continue
> > 
> > So I found a Nehalem box that can reliably reproduce Ingo's problem using
> > something as simple 'perf top'.  But like above, I am noticing the
> > samething, an extra NMI(PMI??) that comes out of nowhere.
> > 
> > Looking at the data above the delta between nmis is very small compared to
> > the other nmis.  It almost suggests that this is an extra PMI.
> > Considering there is already two cpu errata discussing extra PMIs under
> > certain configurations, I wouldn't be surprised if this was a third.
> > 
> > Cheers,
> > Don
> > 
> 
> Oh. I'm not sure if it would be a good idea at all but maybe we could
> use kind of Robert's idea about "pmu nmi relaxing time" ie some time
> slice in which we treat nmi's as being from pmu, but not arbitrary number
> but equal to the number of PMI turned off. Say we handle NMI and found
> that 4 events are overflowed, we clear them, arm timer and wait for
> 3 unknow nmis to happen, if they are not happening during some time
> period we clear this waitqueue, if they happen or partially happen
> - we destroy the timer. Ie almost the same as Robert's idea but
> without tsc? Just a thought.

The only problem is only one counter is overflowing in these cases, so we
would have to do it all the time, which may not be hard.  But I was
thinking of something similar.

For now, I am trying to force counter0 off, seeing that most of the perf
errata on nehalem have been on counter0.  Or maybe I can get 'perf top' to
use something other than counter0 by running 'perf record' first?

Cheers,
Don

> 
> 	-- Cyrill