On Tue, 2019-12-31 at 11:38 +0100, Krzysztof Kozlowski wrote:
> On Tue, 31 Dec 2019 at 11:23, Artur Świgoń <a.swigon@xxxxxxxxxxx> wrote:
> > >
> > > The order of patches should reflect first of all real dependency.
> > > Whether it compiles, works at all and does not break anything. Logical
> > > dependency of "when the feature will start working" is
> > > irrelevant to DTS because DTS goes in separate way and driver is
> > > independent of it.
> >
> > The order of patches does indeed reflect real dependency. I can also reorder
> > them (preserving the dependencies) so that DTS patches go first in the series
> > if this is the more preferred way.
>
> It looks wrong then. Driver should not depend on DTS. I cannot find
> the patch changing bindings (should be first in patchset) which could
> also point to this problem.
>
> It seems you added requirement for interconnect properties while it
> should be rather optional.
No, there is no requirement for interconnect properties (other than that it
simply does not make any sense to use the interconnect driver code and not the
DTS properties for it in the long run).
In case of the exynos-bus driver (code: patch 05, DTS: patch 04) if the DTS
properties ('exynos,interconnect-parent-node') are missing, the new code handles
it gracefully returning NULL from exynos_bus_icc_get_parent() (it is not an
error condition).
In case of the exynos-mixer driver (code: patch 07, DTS: patch 06) if the DTS
property ('interconnects') is missing, of_icc_get() returns NULL and the code does
not try to set any contraints for a NULL path. Same thing happens if
CONFIG_INTERCONNECT is 'n'.
The only case when something breaks is when you try to use the interconnect
consumer (implemented in patches 06 & 07) when there is no interconnect provider
(patches 04 & 05), in which case of_icc_get() returns an error (since it cannot
find a path). From what I understand, it probably makes sense to merge any
interconnect consumers one cycle later than the provider.
> > > > I still think the order of these patches is the most logical one for someone
> > > > reading this RFC as a whole.
> > >
> > > I am sorry but it brings only confusion. DTS is orthogonal of the
> > > driver code. You could even post the patchset without DTS (although then
> > > it would raise questions where is the user of it, but still, you
> > > could).
> > >
> > > Further, DTS describes also hardware so you could send certain DTS
> > > patches without driver implementation to describe the hardware.
> > >
> > > Driver code and DTS are kind of different worlds so mixing them up for
> > > logical review does not really make any sense.
> > >
> > > Not mentioning it is different than most of other patches on mailing
> > > lists.
> > >
> > > BTW, it is the same as bindings which should (almost) always go first as
> > > separate patches.
> >
> > Thanks for elaborating on this, I appreciate it.
> > Regarding your original concern, patches 04 & 06 are separate for several
> > reasons, one of which is that they are related to two different drivers
> > (exynos-bus vs. exynos-mixer).
>
> It's okay then (for them to be split).
>
> >
> > > >
> > > > > In certain cases dependency on DTS changes is ok:
> > > > > 1. Cleaning up deprecated properties,
> > > > > 2. Ignoring the backward compatibility for e.g. new platforms.
> > > > >
> > > > > None of these are applicable here.
> > > > >
> > > > > You need to rework it, put DTS changes at the end. This clearly shows
> > > > > that there is no wrong dependency.
> > > > >
> > > > > >
> > > > > > > Adjust the title to match the contents - you are not adding bindings but
> > > > > > > properties to bus nodes. Also the prefix is ARM: (look at recent
> > > > > > > commits).
> > > > > >
> > > > > > OK.
> > > > > >
> > > > > > > >
> > > > > > > > diff --git a/arch/arm/boot/dts/exynos4412-odroid-common.dtsi b/arch/arm/boot/dts/exynos4412-odroid-common.dtsi
> > > > > > > > index 4ce3d77a6704..d9d70eacfcaf 100644
> > > > > > > > --- a/arch/arm/boot/dts/exynos4412-odroid-common.dtsi
> > > > > > > > +++ b/arch/arm/boot/dts/exynos4412-odroid-common.dtsi
> > > > > > > > @@ -90,6 +90,7 @@
> > > > > > > > &bus_dmc {
> > > > > > > > exynos,ppmu-device = <&ppmu_dmc0_3>, <&ppmu_dmc1_3>;
> > > > > > > > vdd-supply = <&buck1_reg>;
> > > > > > > > + #interconnect-cells = <0>;
> > > > > > >
> > > > > > > This does not look like property of Odroid but Exynos4412 or Exynos4.
> > > > > >
> > > > > > Strangely enough, this file is where the 'exynos,parent-bus' (aka. 'devfreq')
> > > > > > properties are located (and everything in this RFC concerns devfreq).
> > > > >
> > > > > I cannot find exynos,parent-bus in exynos4412-odroid-common.dtsi. Can
> > > > > you elaborate?
> > > >
> > > > Currently a name change is being made: 'devfreq' -> 'exynos,parent-bus'
> > > > https://patchwork.kernel.org/patch/11304549/
> > > > (a dependency of this RFC; also available in devfreq-testing branch)
> > >
> > > I see. That property also does not look like board (Odroid) specific so
> > > it should be moved to Exynos4412 DTSI.
> >
> > Makes sense to me. Just from looking at the patch I referenced above, there is
> > a significant level of code duplication between
> > * arch/arm/boot/dts/exynos4412-itop-scp-core.dtsi
> > * arch/arm/boot/dts/exynos4412-midas.dtsi
> > * arch/arm/boot/dts/exynos4412-odroid-common.dtsi
> > with relation to the devfreq*/exynos,* properties.
>
> If you have in mind all the nodes with "status=okay", it's fine to
> duplicate them.
OK.
Regards,
--
Artur Świgoń
Samsung R&D Institute Poland
Samsung Electronics
