Re: [PATCH RFC 1/4] arm64: kernel: implement DT based idle states infrastructure

[Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>]

On Mon, Mar 24, 2014 at 05:46:47PM +0000, Rob Herring wrote:
> On Thu, Mar 20, 2014 at 1:19 PM, Sebastian Capella
> <sebastian.capella@linaro.org> wrote:
> > On 19 March 2014 10:23, Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> wrote:
> >> Rob is asking a property describing hw, and honestly unless we define
> >> a power consumption property value (defining what it means will be rather
> >> complicated though) I'd rather remove index altogether and use min_residency
> >> as a comparison value instead.
> >
> > I like having a separate ranking for power to sort the states.  To me
> > it seems that min residency could be ambiguous for this.  We could
> > have states with inverted min_residencies: where a lower consumption
> > state has a lower target residency because the cost is recovered more
> > quickly.  Such a state might have higher latency for exit, so even
> > though the lower min_residency/lower power state is clearly
> > preferable, selection of the lower power state may be prevented by
> > kernel latency constraints.
> 
> Perhaps you can add the min residency and exit latency times.

I do not think we should mix exit latency with power consumption
ranking, they are used and define different things.

The problem is that min_residency is extremely hard to define (see
discussion below).

http://www.spinics.net/lists/arm-kernel/msg316147.html

What should min_residency represent ? As Anti correctly pointed out, it should
be compared to a shallower idle state, but to define "shallower" we need
a ranking, do not we ?

If we say "min_residency is the time required for the processor to stay
in the idle state in order to maximize power savings" (and I tried to do that),
that's wrong by definition, because there might be a deeper idle state saving
more power when entered for a time within the next event.

state 1 - min residency 10 us
state 2 - min residency 20 us

If we sleep for 1 second, state 2 must be entered, not 1 and the problem
is always the same, state 1 10us compares to state 0 (break even time
against state 0).

So, we are back to square one. How should min_residency be defined ?
If it is the break even time against a shallower idle state, we have
to know what's the shallower idle state, so we need a power ranking for
idle states.

> Another piece to consider is the affinity level of the state should
> probably be the basis for 1st order sorting.

I would like to see the idle states as a flat list, when you say
affinity you mean a cpumask ? Affinity level in the PSCI power state
parameter ? In the MPIDR ?

Thanks,
Lorenzo