[RFC PATCH 2/2] arm64: kernel: perf: add pmu CPU PM notifier

[Will Deacon]

On Fri, Mar 13, 2015 at 11:31:37PM +0000, Kevin Hilman wrote:
> Lorenzo Pieralisi <lorenzo.pieralisi at arm.com> writes:
> 
> > On Thu, Mar 12, 2015 at 01:18:54PM +0000, Ashwin Chaugule wrote:
> >> On 11 March 2015 at 12:02, Kevin Hilman <khilman at kernel.org> wrote:
> >> > Lorenzo Pieralisi <lorenzo.pieralisi at arm.com> writes:
> >> >
> >> >> When a CPU is being profiled through PMU events and it enters suspend
> >> >> or idle states, the PMU registers content can be lost, which means that
> >> >> counters that were relied upon on power down entry are reset on power
> >> >> up to values that are incosistent with the profile session.
> >> >>
> >> >> This patch adds a CPU PM notifier to arm64 perf code, that detects
> >> >> on entry if events are being monitored, and if so, it returns
> >> >> failure to the CPU PM notification chain, causing the suspend
> >> >> thread or the idle thread to abort power down, therefore preventing
> >> >> registers content loss.
> >> >>
> >> >> By triggering CPU PM notification failure this patch prevents
> >> >> suspending a system if the suspend thread is being profiled and
> >> >> it also prevents entering idle deep states on cores that have profile
> >> >> events in use, somehow limiting power management capabilities when
> >> >> there are active perf sessions.
> >> >
> >> > I guess that's one choice.  Couldn't you also stop the PMU and
> >> > save/restore it's context in the notifiers? so that you wouldn't affect
> >> > PM capabilities?
> >> >
> >> > That would imply that you lose the ability to profile after a certain
> >> > point in suspend/idle, but maybe that's a better trade off than having
> >> > profiling disable certain PM features?
> >> 
> >> I had something like that a few years ago on the Kraits and Scorpions [1].
> >
> > That's another option, but the point is understanding how we want to
> > tackle the issue, by preventing power down or by restoring the
> > PMU registers.
> 
> Personally, I think the save/restore approach is preferred.  IMO, it's
> more intuitive from the perspective of a user who doesn't understand all
> the mechanics and also actually allows you to profile most of the
> low-power paths and still actually hit the low power states.

I agree that save/restore is the nicest thing to do, but until we have a
generic description of the power-domains in device-tree I really don't want
PMU-specific hacks in the arch code to deal with this (since many platforms
do not lose PMU state over suspend).

What Lorenzo is proposing is a stop-gap to prevent perf silently losing its
state on platforms where the register contents is lost.

Will