Re: [BUG] perf_events: PERF_FORMAT_GROUP not working correctly when monitoring another task

[Corey Ashford <cjashfor@linux.vnet.ibm.com>]

On 05/06/2010 01:22 PM, Peter Zijlstra wrote:
> On Thu, 2010-05-06 at 13:09 -0700, Corey Ashford wrote:
>> The patch works with the test case I provided, but for some reason it
>> breaks the normal operation of the libpfm4 "task" utility.  If I put
>> more than one event in a group, I get zero counts on all but the first
>> event.  That's even if I don't use the PERF_FORMAT_GROUP option.
>
> Bugger,.. weird though, I can't see how it would affect the state before
> the counters get detached from their context (exit() / close()).
>
>> So something appears to be messed up.  I will see if I can construct an
>> arch-independent test case which demonstrates the problem.
>
> That would be awesome, thanks!
>

Ok, it was pilot error.  We had made a few changes to the "task" program 
to expand its capabilities and introduced a couple of bugs at the same time.

Unfortunately, I have one remaining mystery.  The default behavior of 
task is to create groups, without the PERF_FORMAT_GROUP flag set, and 
then read up the counts and time enabled/running values for each counter 
in the group after the measurement has occurred.

Well, for some reason I haven't been able to figure out, all but first 
event in the group have their time running and time enable values set to 
zero, but their counts are non-zero.

If I replicate this setup with the show_fg_bug test case I posted, the 
time running and enabled values are correct for each counter.

The enabled and running flag bits are set for every event which is 
opened (in both show_fg_bug and task)... so I don't know what the 
difference would be.

If anything springs to mind there as to how I might be getting zero 
enabled/running values passed back from the kernel, I'd be very 
interested to know.

Otherwise, I will keep poking at it till I find the bug in task, or 
something I can point at in the kernel.

Thanks,

- Corey