Re: perf: h/w counters not counted and no error/info reported to user

[David Ahern <dsahern@gmail.com>]

On 05/30/2011 03:04 PM, Ashwin Chaugule wrote:
> Hi David,
> 
> On Fri, May 27, 2011 at 12:36 PM, David Ahern <dsahern@gmail.com> wrote:
>>
>> Ok, I got it now. On some occasions for simple programs -- like sleep 1
>> -- the counter reads show:
>>
>> cycles: 0 1066485 0
>>
>> While the counter value is non-0, the time enabled and running values
>> are 0 and those latter two are used by stat to say "not-counted".
> 
> I've seen time-enabled = 0 too, but I suspected that's because on my
> kernel the sched_clock implementation was missing. (defaults to a
> jiffies based timer)

I think in the case of 'perf stat -- sleep 1' it runs so quickly that
sometimes it does not cross timer ticks and the time_enabled comes out 0.

> I saw your patch for initializing the fd's to -1's, but I'm unable to
> see how that makes the value of time-enabled non zero ?

It doesn't. That patch addresses unsupported counters in the default
list (e.g., -ddd on a T9550 Core2Duo). The first open fails (cpu 0,
thread 0) and the entire FD array is not reset to -1 (from its
initialization to 0).

David

> 
> Cheers,
> Ashwin