Re: Perf events warning..

[David Ahern <dsahern@gmail.com>]

On 5/15/12 4:49 AM, Peter Zijlstra wrote:
> On Mon, 2012-05-14 at 15:25 -0700, Linus Torvalds wrote:
>> On Mon, May 14, 2012 at 3:20 PM, Peter Zijlstra<a.p.zijlstra@chello.nl>  wrote:
>>>
>>> So far I can't make any of the things I came up stick. You ran something
>>> simple like: 'perf record -e cycles:pp -F 20000 make test' ? Or did you
>>> do something more interesting?
>>
>> It was not much more complex than that.
>>
>> It did use "make -j64 test" to make the load a *bit* more interesting
>> (and go noticeably faster), but other than that you got it.
>
> OK, that limits the scope of crazy scenarios I have to consider, still
> no immediate clue though..
>
> I think I've found a possible race, but I can't make it work with that
> workload. I've also let your workload run for 2+ hours in trying to
> reproduce, but no luck, it must be a very narrow window indeed.
>
> I'll keep prodding at it..

Perhaps it is specific to processor generation? Yesterday I noted that 
perf-stat -g trips a WARNING only on Nehalem. Westmere works fine - 
perf-stat -g generates output and no warning is triggered. Arnaldo is 
using Sandy Bridge - though it's not clear if his success (3.4.0-rc3 on 
server named sandy) or fail (3.4.0-rc4-uprobes on felicio) was on a SNB.

David