Re: [PATCH 1/2] dt-bindings: ufs: qcom: allow 'dma-coherent' property

[Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>]

On Tue, Dec 06, 2022 at 09:14:30AM +0100, Krzysztof Kozlowski wrote:
> On 05/12/2022 14:37, Manivannan Sadhasivam wrote:
> > On Mon, Dec 05, 2022 at 02:12:48PM +0100, Johan Hovold wrote:
> >> On Mon, Dec 05, 2022 at 06:30:48PM +0530, Manivannan Sadhasivam wrote:
> >>> On Mon, Dec 05, 2022 at 01:27:34PM +0100, Johan Hovold wrote:
> >>>> On Mon, Dec 05, 2022 at 05:50:18PM +0530, Manivannan Sadhasivam wrote:
> >>>>> On Mon, Dec 05, 2022 at 01:07:16PM +0100, Johan Hovold wrote:
> >>>>>> On Mon, Dec 05, 2022 at 05:29:06PM +0530, Manivannan Sadhasivam wrote:
> >>>>>>> On Mon, Dec 05, 2022 at 11:08:36AM +0100, Johan Hovold wrote:
> >>>>>>>> UFS controllers may be cache coherent and must be marked as such in the
> >>>>>>>> devicetree to avoid data corruption.
> >>>>>>>>
> >>>>>>>> This is specifically needed on recent Qualcomm platforms like SC8280XP.
> >>>>>>>>
> >>>>>>>> Signed-off-by: Johan Hovold <johan+linaro@kernel.org>
> >>
> >>>>>> Yes, it would be a valid, but it will only be added to the DTs of SoCs
> >>>>>> that actually require it. No need to re-encode the dtsi in the binding.
> >>>>>>
> >>>>>
> >>>>> But if you make a property valid in the binding then it implies that anyone
> >>>>> could add it to DTS which is wrong. You should make this property valid for
> >>>>> SoCs that actually support it.
> >>>>
> >>>> No, it's not wrong.
> >>>>
> >>>> Note that the binding only requires 'compatible' and 'regs', all other
> >>>> properties are optional, and you could, for example, add a
> >>>> 'reset' property to a node for a device which does not have a reset
> >>>> without the DT validation failing.
> 
> Optional properties are optional primarily looking at one variant. It
> means that on different boards with the same SoC, things can be routed a
> bit differently and some property can be skipped. E.g. sometimes
> regulators come from PMIC and sometimes are wired to some VBATT, so we
> do not have regulator in DTS for them. Or some interrupt/pin is not
> connected.
> 
> Now between variants of devices - different SoCs: I don't think that
> "optional" should be used in such context, except special cases or lack
> of knowledge about hardware. For given SoC/variant, the property is either:
> 1. valid and possible (can be required or optional),
> 2. not valid, not possible.
> And this we should express in constraints, if doable with reasonable
> complexity.
> 
> Therefore the question is: is dma-coherent not valid for other SoCs?
> 

Yes, it is not valid on older SoCs because they don't support I/O coherency.
So setting this property on those un-supported SoCs may lead to wierd behavior.
This was the concern I had for setting this property valid for all SoCs.

So far we only know that SC8280XP and newer SoCs support I/O coherency.

Thanks,
Mani

> If it is "not needed" for other SoCs, then I would leave it like this.
> Consider also what Rob said, that otherwise we would create DTS from the
> bindings.
> 
> Also, too many allOf:if:then: constraints in the bindings make them
> trickier to read.
> 
> >>>>
> >>>
> >>> Then what is the point of devicetree validation using bindings?
> >>
> >> You're still making sure that no properties are added that are not
> >> documented, number of clocks, names of clocks, etc.
> >>
> >>> There is also a comment from Krzysztof: https://lkml.org/lkml/2022/11/24/390
> >>
> >> Speaking of Krzysztof:
> >>
> >> 	https://lore.kernel.org/lkml/20221204094717.74016-5-krzysztof.kozlowski@linaro.org/
> 
> That's not the best example, because I just do not know where
> dma-coherent is applicable and where it is not, thus I added it as valid
> for all variants. Also, I think that all variants are capable of using
> IOMMU - it isn't restricted per variant. If they are capable of IOMMU,
> then dma-coherent is a possible choice.
> 
> 
> Best regards,
> Krzysztof
> 

-- 
மணிவண்ணன் சதாசிவம்