[RFC PATCH 0/3]perf/core: extend perf_reg and perf_sample_regs_intr

[Michael Ellerman]

On Fri, 2015-11-06 at 10:24 +0100, Peter Zijlstra wrote:
> On Fri, Nov 06, 2015 at 12:57:17PM +0530, Madhavan Srinivasan wrote:
> > On Thursday 05 November 2015 06:37 PM, Peter Zijlstra wrote:
> > > On Thu, Nov 05, 2015 at 02:16:15AM +0530, Madhavan Srinivasan wrote:
> > > > Second patch updates struct arch_misc_reg for arch/powerpc with pmu registers
> > > > and adds offsetof macro for the same. It extends perf_reg_value()
> > > > to use reg idx to decide on struct to return value from.
> > > Why; what's in those regs?
> > 
> > Was out and did not have access to mail, so missed to respond in time.
> > 
> > In current implementation of patch 2, have added 
> > few pmu control/status and counter registers,
> > which give additional information about the PMU context
> > for the sample.
> 
> Yes, I saw that, you still haven't answered the question though. What is
> in those regs? Why is exposing that information like this the best
> option.

It's a perrenial request from our hardware PMU folks to be able to see the raw
values of the PMU registers.

I think partly it's so that they can verify that perf is doing what they want,
and some of it is that they're interested in some of the more obscure info that
isn't plumbed out through other perf interfaces.

We've used various internal hacks over the years to keep them happy. This is an
attempt to use a somewhat standard mechanism.

It would also be helpful for those of us working on the perf hardware backends,
to be able to verify that we're programming things correctly, without resorting
to debug printks etc.

Basically we want to sample regs at the time of the perf interrupt, so we
though PERF_SAMPLE_REGS_INTR made senes :)

But if you think this is the wrong mechanism within perf, then please let us
know.

I know perf's mission is to abstract as much of the arcane hardware details
into a generic interface and make PMUs actually useful for normal folks, and we
are committed to that, but it would also be useful to be able to get the raw
values for a different type of user.

Maddy's patch only exports PMC1-6 and MMCR0/1. I think we also need to export
some others, in particular MMCRA has a lot of stuff in it, half of which is not
even architected. So that would have to be exported as "POWER8_MMCRA". And then
there's the SIAR/SDAR/SIER which contain a bunch of info on sampled
instructions that is not currently plumbed out.

cheers