Re: 2.6.28-rc9: oprofile regression

[Tim Blechmann]

On Wed, 2009-01-14 at 18:10 +0100, Thomas Gleixner wrote:
> On Fri, 2 Jan 2009, Tim Blechmann wrote:
> > On Fri, 2008-12-26 at 03:42 +0100, Andi Kleen wrote:
> > > Tim Blechmann wrote:
> > > >> i am experiencing an issue, similar to the one reported in
> > > >> http://lkml.org/lkml/2008/10/30/319.
> > > > 
> > > > bisecting showed, that commit b99170288421c79f0c2efa8b33e26e65f4bb7fb8
> > > > (oprofile: Implement Intel architectural perfmon support) caused the
> > > > problem. 
> > > > oddly, the newly introduced api is not used, since the model struct is
> > > > set during the ppro_init call ...
> > > 
> > > We're still investigating the problem. Thanks for the report.
> > 
> > btw, this issue still exists in tip/oprofile ... not sure, whether this
> > may be related, but i am running the machine in 64-bit mode ...
> 
> can you please apply the patch below and provide the output ?

[29030.863352] oprofile: using NMI interrupt.
[29051.826778] ppro counter_width: 40
[29051.826783] ppro counter_width: 40


> That's one of the subtle differences to the 2.6.27 code, where the
> counter width is fixed to 32bit, which is correct anyway as the
> counter MSRs can only write the lower 32bits and sign extend bit 31
> according to intel documentation.

this code (line 81/82), changes counter_width from 32 to 40.

if (counter_width < eax.split.bit_width)
        counter_width = eax.split.bit_width;

however when removing these lines, and thus keeping the value 32 for
counter_width, doesn't change the behavior, only one NMI per cpu.

best, tim

--
tim@xxxxxxxxxx
http://tim.klingt.org

It is better to make a piece of music than to perform one, better to
perform one than to listen to one, better to listen to one than to
misuse it as a means of distraction, entertainment, or acquisition of
'culture'.
  John Cage

Attachment: signature.asc
Description: This is a digitally signed message part