Re: [PATCH 5/5] perf, x86: Add support for AMD family 15h core counters

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

On 02.02.11 17:44:22, Stephane Eranian wrote:
> On Wed, Feb 2, 2011 at 6:29 PM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Wed, 2011-02-02 at 18:24 +0100, Robert Richter wrote:
> >> On 02.02.11 12:03:18, Peter Zijlstra wrote:
> >> > Why this and not something like x86_pmu.perfctr + (index << 1)?
> >> > You could even use alternatives.
> >>
> >> I was thinking about this. The main reason is the implementation of
> >> northbridge counters, the range is in MSRC001_02[47:40]. This would
> >> add more complexity then. Using a table would be something like
> >>
> >> unsigned int eventsel_f15h[] = {
> >>       MSR_F15H_PERF_CTL,
> >>       MSR_F15H_PERF_CTL + 2,
> >>       MSR_F15H_PERF_CTL + 4,
> >>       MSR_F15H_PERF_CTL + 6,
> >>       MSR_F15H_PERF_CTL + 8,
> >>       MSR_F15H_PERF_CTL + 10,
> >>       MSR_F15H_NB_PERF_CTL,
> >>       MSR_F15H_NB_PERF_CTL + 2,
> >>       MSR_F15H_NB_PERF_CTL + 6,
> >>       MSR_F15H_NB_PERF_CTL + 8,
> >> };
> >>
> >> We don't need to change the address generation for this. Otherwise we
> >> need to introduce more logic for the calculation.
> >>
> >> Also, were could be potential easier implementations for fixed
> >> counters, BTS, P4, IBS, etc. But didn't look that close at it.
> >>
> >> (Btw, I am not yet sure if NB counters shouldn't better start at index
> >> 16 or so to reserve space for perf counter expansion.)
> >
> > Now that the NB PMU is completely separate from the core PMU, wouldn't
> > it make more sense to implement that as a separate entity just like the
> > intel uncore bits?
> 
> I agree on this.

Peter,

ok, nb events may be implemented independent from core events in a
separate struct pmu.

I still would prefer a lookup table for counter addresses. Adding a
shift parameter to struct x86_pmu to do a

      addr = base + (index << shift)

seems to me a quite special solution that may not be reused in other
implementations, while a lookup table is more generic. I also don't
see a performance or memory impact there.

Anyway, a shift parameter would work too. What do you think?

-Robert

-- 
Advanced Micro Devices, Inc.
Operating System Research Center