[PATCH] arm64: dts: Add idle-states for Juno

[Lorenzo Pieralisi]

On Fri, May 01, 2015 at 02:30:13PM +0100, Catalin Marinas wrote:
> On Fri, May 01, 2015 at 11:12:11AM +0100, Lorenzo Pieralisi wrote:
> > On Fri, May 01, 2015 at 10:02:02AM +0100, Catalin Marinas wrote:
> > > On Thu, Apr 30, 2015 at 06:17:01PM +0100, Lorenzo Pieralisi wrote:
> > > > On Thu, Apr 30, 2015 at 05:40:35PM +0100, Jon Medhurst (Tixy) wrote:
> > > > > On Thu, 2015-04-30 at 17:00 +0100, Sudeep Holla wrote:
> > > > > > On 30/04/15 14:57, Jon Medhurst (Tixy) wrote:
> > > > > > > From: Jon Medhurst <tixy at linaro.org>
> > > > > > >
> > > > > > > Signed-off-by: Jon Medhurst <tixy at linaro.org>
> > > > > > > ---
> > > > > > >
> > > > > > > These have been kicking around out of tree for ages, any reason they
> > > > > > > shouldn't be in mainline?
> > > > > > 
> > > > > > One possible reason could be that these values are not tuned(e.g.
> > > > > > latency values, can they be same for both clusters ?)
> > > > > 
> > > > > I thought that both clusters being the same was questionable.
> > > > > 
> > > > > >  Though these
> > > > > > reasons are not blocking and this patch will not cause any
> > > > > > functionality break even if is merged as is.
> > > > > 
> > > > > My main purpose with trying to get this merged is so that people using
> > > > > Juno for general testing and validation will actually have cpuidle
> > > > > running and so potentially find more bugs.
> > > > 
> > > > I am reluctant to enable idle states in the default Juno dts, they
> > > > will affect latencies and performance tests significantly.
> > > 
> > > OTOH, I guess they will improve the power benchmarks. IMO, we should
> > > place in the DT whatever the hardware and firmware supports. It's up to
> > > those doing benchmarks to disable CPU suspend.
> > 
> > I am ok with DT defining whatever HW and FW support, the question is
> > whether we want CPUidle (and CPUfreq) enabled by default in the
> > defconfig then.
> 
> It's functionality we support, so I think it should be enabled in
> defconfig. This config is meant as a point of reference for what's
> supported by the arm64 kernel and not tuned for some specific
> benchmarks.

Point raised and taken, it makes sense so I am happy with this.

> > This will certainly trigger mainline regressions from a latency/performance
> > standpoint (true, some tests disable idle states by default ie cyclic
> > test), unless we disable CPUidle via command line parameter or we
> > force people carrying out tests to disable idle states through sysfs knobs.
> 
> It may affect performance for specific benchmarks but what if others
> test the power consumption?

See above.

> BTW, cannot the cpuidle governor be made smart enough (unless it is
> already) to figure out when the next timer is going to happen so that
> the CPU doesn't go in a deep sleep state with a high latency?

The menu governor already detects if/when core can sleep for long
enough to break even in terms of energy and to avoid breaking latency
constraints (which are tunable, with lax default values though).

My concern is the time required to exit the idle states (and the
cache refills required on power-up), we can certainly tune the DT residency
parameters (and user space can prevent idle by forcing latency constraints,
that's what eg cyclic test does by default), and that's what I am going
to do. And then there are IRQs that the governor can't predict, if a cpu
is powered off it will take a latency hit, that can only be prevented
by disabling the idle states.

Anyway, with all of the above in mind, I am fine with enabling the idle
states, I need to review and test the idle states DT data in the patch
first though.

Lorenzo