[Intel-gfx] [PATCH] RFC drm/i915: Expose a PMU interface for perf queries

[Chris Wilson]

On Wed, Aug 21, 2013 at 03:14:22PM +0200, Daniel Vetter wrote:
> On Tue, Aug 20, 2013 at 10:17 PM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > The first goal is to be able to measure GPU (and invidual ring) busyness
> > without having to poll registers from userspace. (Which not only incurs
> > holding the forcewake lock indefinitely, perturbing the system, but also
> > runs the risk of hanging the machine.) As an alternative we can use the
> > perf event counter interface to sample the ring registers periodically
> > and send those results to userspace.
> >
> > To be able to do so, we need to export the two symbols from
> > kernel/events/core.c to register and unregister a PMU device.
> >
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> Make sense imo to move the perf gathering into the kernel, a few
> random (and rather uninformed) points I've thought of:
> - What about the instdone busy bits intel_gpu_top currently presents?

I'm contemplating adding 64 u64 (and the rest!), and converting intel_gpu_top
as my next use case.

> - Shouldn't we have one polling hrtimer that checks all enabled pmu event bits?

I had already done so.

> - Plans for integrating the various status/debug bits in the display
> block (like fetches, fbc compression stuff, ...)? I know that this
> will tickle the all-encompassing paranoia about releasing performance
> counter details, but we can still hope. Iirc there's also some very
> minimal perf counter stuff on the GT arbiter.

I'd been adding things as I use them. If we have other counters, yes
they will be useful and I strongly advise people working on
power/performance sensitive areas start thinking about how we can
monitor them in real time..
 
> Overall I think this is great and should fill the gap until we're
> allowed to release some more of the integrated global perf counters
> (and even then those should neatly fit in ...).
> 
> For testing I think just running your overlay tool with all possible
> events should be good enough (in some headless mode ofc) - functional
> testing of perf counter is imo too much fuss.

My aim is to have complete integration with perf as well as having a
pretty tool for my own use. (It has already shown me more bugs than I
thought existed...)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre