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

[Cyrill Gorcunov <gorcunov@gmail.com>]

On Tue, Aug 24, 2010 at 04:27:18PM -0400, Don Zickus wrote:
> On Tue, Aug 24, 2010 at 11:52:40PM +0400, Cyrill Gorcunov wrote:
> > > I use 2.6.34 atm. Letme try 2.6.36 (which might require some time to recompile).
> > > 
> > > 	-- Cyrill
> > 
> > Don, for me it fails with somehow unrelated page handling fault - in
> > reiserfs_evict_inode O_o Fails in __get_cpu_var I suspect might means
> > a problem either in per-cpu allocator itself or we screw pointer somehow.
> > Weird.
> 
> I just found out (with the help of the crash utility and Dave A.) that
> Robert's percpu struct nmi clashes with the exception entry point .nmi.
> I only see this problem in 2.6.36, so I am not sure what changed with
> regards to compiler flags to confuse variables with text segments.
> 

yeah, I suspect name clashes here but then I did grep over per-cpu variables
in whole kernel and didn't find match so I thought the assumption was wrong,
but eventually it happens to be true but via other way :) good to know, thanks!

> But renaming the percpu struct nmi to nmidon fixed the problem for me (I
> am open to other suggestions :-) ).

nmi_don_zickus ;) well, I think nmi_pmu or something like that
might be a bit modest ;)

> 
> Regarding your reiserfs, what was the variable's name?
> 

It seems to be different, it's pity that I had only 80x25 vga
mode and was unable to snap the whole log. But actually I didn't
even check precisely all .config options I had set since I was
more interested in early stage where per-cpu access should already
happen rather then real init'ed environmen. But I think I'll be
moving completely to .36 this week so we will see how it goes.

> Cheers,
> Don
> 
	-- Cyrill