Re: [PATCH 5/6] perf_counter: add more context information

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

Peter Zijlstra wrote:
> On Mon, 2009-04-06 at 14:33 -0700, Corey Ashford wrote:
>> Peter Zijlstra wrote:
>>> On Mon, 2009-04-06 at 14:15 -0700, Corey Ashford wrote:
>>>> Peter Zijlstra wrote:
>>>>> On Mon, 2009-04-06 at 13:16 -0700, Corey Ashford wrote:
>>>>>
>>>>>>>> One downside of this approach is that you if you specify "no header" 
>>>>>>>> (currently not possible, but maybe later?), you will not be able to get 
>>>>>>>> the level bits.
>>>>>>> Would this be desirable? 
>>>>>> I think it would.  For one use case I'm working on right now, simple 
>>>>>> profiling, all I need are ip's.  If I could omit the header, that would 
>>>>>> reduce the frequency of sigio's by a factor of three, and make it faster 
>>>>>> to read up the ip's when the SIGIO's occur.
>>>>> Self-profiling?
>>>>>
>>>>> So you're interested in getting the smallest possible record size, that
>>>>> would still be 2 u64, right? Otherwise you don't get the IP context that
>>>>> started this.
>>>>>
>>>>>
>>>> Self-profiling mainly, yes.  PAPI specs an ability for remote monitoring 
>>>> of processes and threads, but I think it's only partially implemented.
>>>>
>>>> So when you are talking about IP context, you mean pid/tid?
>>> Ah, we called it level before, the hv/kernel/user thing. For remote
>>> profiling you'd want to have the mmap thing too.
>> Oh I see.  In PAPI, the user specifies the range(s) of addresses he's 
>> interested in profiling (any sampled IP's outside the requested ranges 
>> are discarded), and so as long as the kernel space IP's don't overlap 
>> with user space IP's, we should be fine.
> 
> Ah, while this would be true for most 'sane' architectures, Paul was
> right in pointing out that this is not true for all architectures -- and
> we should therefore not rely on address range alone.
> 
> You could of course use: hw_event.exclude_{hv,kernel} = 1 to ensure you
> only get userspace thingies I suppose (but then you have no way of
> telling how many you missed I guess).

That's a good point.  PAPI's profiling API doesn't have a way for the 
caller to distinguish which address spaces (user/kernel/hv) he wants to 
profile.  It does have a way to designate which levels to ignore, but if 
you enable them all, you cannot specify the profiling address ranges 
pertaining to each.  That may be something I could propose adding to 
PAPI.  I suspect it would be pretty rarely used, though.

- Corey