Re: [RFC PATCH 0/5] perf: Intel uncore pmu counting support

["Yan, Zheng" <zheng.z.yan@intel.com>]

On 03/28/2012 04:49 PM, Peter Zijlstra wrote:
> On Wed, 2012-03-28 at 08:49 +0200, Ingo Molnar wrote:
>> * Yan, Zheng <zheng.z.yan@intel.com> wrote:
>>
>>> Hi, all
>>>
>>> Here is the RFC patches to add uncore counting support for Nehalem,
>>> Sandy Bridge and Sandy Bridge-EP, applied on top of current tip.
>>> The code is based on Lin Ming's old patches.
>>>
>>> You can use 'perf stat' to access to the uncore pmu. For example:
>>> perf stat -a -C 0 -e 'uncore_nhm/config=0xffff/' sleep 1
>>
>> My main complaint is that that's not user friendly *AT ALL*.
>>
>> You need to make this useful to mere mortals: go through the 
>> SDM, categorize interesting looking events, look at how it can 
>> be expressed via tooling, add a generic event where appropriate, 
>> provide examples, actually *USE* it to improve the kernel or an 
>> app and see the workflow as it happens and improve the tooling, 
>> etc.
> 
> Easiest way out here is add a /sys/bus/event_source/devices/*/events/
> directory which contains files who's name we can use as events and who's
> contents are of the form we would use given the format/ stuff.
> 
> Example, suppose a westmere,
> 
> $ cat /sys/bus/event_source/devices/cpu/events/frontend_stalled_cycles
> event=0x0e,umask=0x01,inv,cmask=1
> 
> I'll review the uncore patches later this week, but I suspect the whole
> cpu->node mapping stuff is still not done properly.
> 
> Also, quick question, did Intel fix the SNB uncore PMI?
No. furthermore there is completely no uncore PMI in Sandy Bridge-EP

Thanks.