Re: [PATCH v4 08/11] drivers: perf: hisi: use poll method to avoid L3C counter overflow

From: Will Deacon
Date: Tue Feb 21 2017 - 07:09:42 EST

On Mon, Feb 20, 2017 at 11:09:43AM +0000, Mark Rutland wrote:
> On Sun, Feb 19, 2017 at 01:51:03PM -0500, Anurup M wrote:
> > The L3 cache PMU use N-N SPI interrupt which has no support
> > in kernel mainline.
> Could you elaborate on what you mean by this?
> I don't understand what is meant here. How exactly are the interrupts
> wired up in HW, and what exactly is not supported by Linux?
> > So use hrtimer to poll and update event
> > counter to avoid overflow condition for L3 cache PMU.
> > A interval of 10 seconds is used for the hrtimer.
> > The time interval can be configured in the sysfs.
> I'm not too keen on giving userspace the ability to control this, since
> it gives an awful lot of rope for userspace to tie around itself.

Agreed. I'd also go a step further and say that for PMUs with either
terminally broken interrupts (like this one) or just missing interrupts
(like the CPU PMU on raspberry pi iirc), then the perf core should take
care of an hrtimer in an attempt to generate samples often enough. We
already have PERF_PMU_CAP_NO_INTERRUPT, but it currently just disables
sampling events.

The fiddly part is knowing how to program the timer, and I think you'd
need the PMU driver to provide an upper-bound on events per nanosecond.
I'm pretty sure that would be highly unreliable (especially for shared
resources such as the L3), at which point, is it worth the hassle?