[PATCH v2] ARM: perf: save/restore pmu registers in pm notifier

[will.deacon@arm.com (Will Deacon)]

On Tue, Apr 15, 2014 at 01:46:08PM +0100, Neil Zhang wrote:
> > On Tue, Apr 15, 2014 at 01:37:17PM +0100, Neil Zhang wrote:
> > > > On Mon, Apr 14, 2014 at 02:42:22AM +0100, Neil Zhang wrote:
> > > > > From: Sudeep KarkadaNagesha <sudeep.karkadanagesha@arm.com>
> > > > >
> > > > > This adds core support for saving and restoring CPU PMU registers
> > > > > for suspend/resume support i.e. deeper C-states in cpuidle terms.
> > > > > This patch adds support only to ARMv7 PMU registers save/restore.
> > > > > It needs to be extended to xscale and ARMv6 if needed.
> > > > >
> > > > > [Neil] We found that DS-5 not work on our CA7 based SoCs.
> > > > > After debuging, found PMU registers were lost because of core power
> > down.
> > > > > Then i found Sudeep had a patch to fix it about two years ago but
> > > > > not in the mainline, just port it.
> > > >
> > > > What I don't like about this patch is that we're introducing
> > > > significant overhead for SoCs that don't require save/restore of the
> > > > PMU state. I'd much rather see core power down disabled whilst the
> > > > PMU is in use but, if that's not possible, then I think we need to:
> > > >
> > > >  (1) Make this conditional for cores that really need it
> > > >
> > > >  (2) Only save/restore if the PMU is in use (even better, just save/restore
> > > >      the live registers, but that's probably not worth the effort
> > > >      initially).
> > > >
> > >
> > > The patch has check the ARMV7_PMNC_E bit when save / restore, so
> > > suppose only the core's that use PMU will do the save / restore work.
> > 
> > Seems pretty fragile to base our save/restore decision on the value of one of
> > the registers, though. What happens if the control register is zeroed by the
> > core power down?
> > 
> It will check the saved control value when restore, so is should be OK
> while control register is zeroed. 

Ah yes, I mixed up and save and restore functions. It's still horrible that
we have to read the control register unconditionally during the save though
-- it might be nicer if we simply register/unregister the notifier during
perf runs (in the same way that we request/free the PMU IRQs).

Will