Re: [PATCH v2] perf/x86: fix SNB-EP CBOX and PCU uncore PMU filter management

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2012-08-21 at 13:59 +0200, Stephane Eranian wrote:
> The existing code had a bug whereby it would refuse to
> measure two events in a group for either CBO or PCU PMUs,
> if one of the events was using a filter. This was due to
> the fact that the kernel assumed all CBO and PCU events
> were using filters, and thus would detect false positive
> conflicts between attr->config1 values.
>     
> This patch fixes the problem by introducing the extra_reg
> event mapping tables for both CBO and PCU PMUs. Those
> tables associate an event code+umask with extra (filter)
> register. We used the same approach for the offcore_response
> core PMU event.
>     
> With this patch applied, it is possible to measure, for
> instance, CBO unc_c_tor_inserts:opcode:pcirdcur with
> unc_c_clockticks in the same group.
>     
> The filter for both CBO and PCU are more complex than what the
> current code support. Each filter is sub-divided into event
> specific filters. For now, we consider the filter as one single
> MSR value. We lose a bit of flexibility and force multiplexing
> when this is not really necessary in HW. We could later create
> the notion of virtual filters that would be aggregated at the
> last step (wrmsrl). But I think this is good enough for now.


Thanks, added Zheng Yan's ACK as well since its essentially the same
patch.