Re: [PATCH] oprofile: Implement Intel architectural perfmon support

[Robert Richter <robert.richter@amd.com>]

On 25.09.08 17:44:39, Andi Kleen wrote:
> Robert Richter wrote:
>> On 20.08.08 18:40:31, Andi Kleen wrote:
>>> From: Andi Kleen <ak@linux.intel.com>
>>>
>>> Newer Intel CPUs (Core1+) have support for architectural
>>> events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
>>>
>>> The advantage of this is that it can be done without knowing about
>>> the specific CPU, because the CPU describes by itself what
>>> performance events are supported. This is only a fallback
>>> because only a limited set of 6 events are supported.
>>> This allows to do profiling on Nehalem and on Atom systems
>>> (later not tested)
>>>
>>> This patch implements support for that in oprofile's Intel
>>> Family 6 profiling module. It also has the advantage of supporting
>>> an arbitary number of events now as reported by the CPU.
>>> Also allow arbitary counter widths >32bit while we're at it.
>>>
>>> Requires a patched oprofile userland to support the new
>>> architecture.
>>>
>>> Signed-off-by: Andi Kleen <ak@linux.intel.com>
>>> ---
>>>  Documentation/kernel-parameters.txt |    5 ++
>>>  arch/x86/oprofile/nmi_int.c         |   32 +++++++++--
>>>  arch/x86/oprofile/op_model_ppro.c   |  104 
>>> +++++++++++++++++++++++++++-------
>>>  arch/x86/oprofile/op_x86_model.h    |    3 +
>>>  4 files changed, 116 insertions(+), 28 deletions(-)
>>>
>>> diff --git a/Documentation/kernel-parameters.txt 
>>> b/Documentation/kernel-parameters.txt
>>> index 056742c..10c8b1b 100644
>>> --- a/Documentation/kernel-parameters.txt
>>> +++ b/Documentation/kernel-parameters.txt
>>> @@ -1486,6 +1486,11 @@ and is between 256 and 4096 characters. It is 
>>> defined in the file
>>>  	oprofile.timer=	[HW]
>>>  			Use timer interrupt instead of performance counters
>>>  +	oprofile.force_arch_perfmon=1 [X86]
>>> +			Force use of architectural perfmon performance counters
>>> +			in oprofile on Intel CPUs.  The kernel selects the
>>> +			correct default on its own.
>>> +
>> Could you create a separate patch that introduces this new kernel
>> parameter? 
>
> The parameter only makes sense together with something which uses it.
> So an additional one liner patch ( + docs) would be a patch depending on
> the earlier arch perfmon patch. If you want that really I can do it, but
> frankly it doesn't make sense to me.
>
> It's only really a debugging feature, I can also just take it out
> if it's a problem.
>
>> This would make it easier to send all other changes
>> upstream. We already discussed the need of this parameter.
>
> I thought the result of the discussion was that it was not useful
> because there's no equivalent on arch perfmon on any other x86 CPUs?
> IBS is still not architectural, but family/model specific.
>
>> Maybe it
>> would fit better to have a more generalized paramater for this that
>> could be reused then by other archs/models as well. Something like
>> force_pmu_detection that could be used for all new CPUs (also other
>> models) that do not yet have a specific kernel implementation.
>
> You mean something like pmu=<oprofile arch string> to force
> use of that?

I think this would be the best solution, providing a parameter

 oprofile.force_pmu=<oprofile arch string>

This can easily be implemented and also reused by others. I would be
fine with this solution. No separate patch needed then.

>
>> Even better would a sysfs entry instead with that we can specify which
>> cpu type to use:
>
> module param is already in sysfs.
>
>>  echo "i386/arch_perfmon" > /sys/module/oprofile/parameters/cpu_type
>> That would allow us to switch the pmu at runtime and also from the
>> userland.
>
> Switching at runtime would be complicated changes I think

Right, this is overhead nobody will use.

-Robert

-- 
Advanced Micro Devices, Inc.
Operating System Research Center
email: robert.richter@amd.com