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

Lorenzo Pieralisi lorenzo.pieralisi at arm.com
Tue Nov 24 09:53:43 PST 2015

On Mon, Nov 23, 2015 at 05:45:10PM +0000, Punit Agrawal wrote:
> Lorenzo Pieralisi <lorenzo.pieralisi at arm.com> writes:
> > On Mon, Oct 26, 2015 at 02:12:31PM +0000, Jon Medhurst (Tixy) wrote:
> >> On Thu, 2015-10-22 at 14:22 +0100, Punit Agrawal wrote:
> >> > "Jon Medhurst (Tixy)" <tixy at linaro.org> writes:
> >> > 
> >> > > From: Jon Medhurst <tixy at linaro.org>
> >> > >
> >> > > Signed-off-by: Jon Medhurst <tixy at linaro.org>
> >> > 
> >> > Apologies for resurrecting an old thread.
> >> > 
> >> > Following the discussion on this thread, even though certain concerns
> >> > were raised, there wasn't any objection to $SUBJECT being merged.
> >> > 
> >> > I don't see this patch in any tree; perhaps it's slipped through the
> >> > cracks.
> >> 
> >> It did slip through the cracks. Lorenzo's last comment was "I am fine
> >> with enabling the idle states, I need to review and test the idle states
> >> DT data in the patch first though." and I didn't chase things up.
> >> 
> >> The patch will need refreshing to add idle for Juno r1. Which will then
> >> probably resurrect the discussion about where the numbers come from for
> >> residency times, and are the same ones for r0 valid on r1 (and r2?).
> >> 
> >> In an effort to forestall that I would say: does anyone actually care if
> >> the values are optimal? Juno is a reference platform and powered off
> >> mains, so tuning for the optimum power consumption is pretty pointless.
> >> But because it _is_ used as a reference by people it should at least
> >> have these features enabled, to serve as an example, and for test
> >> coverage.
> >
> > I agree with you here, let me check the entry/exit latencies again
> > to make sure they are reasonably set-up, it is 4.5 material anyway.
> Hi Lorenzo,
> Have you had a chance to sanity check the values we've got here? I
> didn't want to miss this merge window if possible.

Yes and they need updating. This data is really really FW dependent,
that's the reason why I still think that these values should be provided
by FW and not shoved into the kernel. Anyway, here we go (I assume the
worst case and that's the only safe option I can see):

- the entry-latency should be set to 300us for _both_ idle states
- the exit-latency should be set to 1.2ms for _both_ idle states

The min-residency values looks reasonable, a tad optimistic (but again
that's workload and FW dependent so either we settle for the worst
possible case or we do not merge this at all).

With the changes above:

Acked-by: Lorenzo Pieralisi <lorenzo.pieralisi at arm.com>

More information about the linux-arm-kernel mailing list