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@xxxxxxxxxx> > >> > >>>>>> 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@xxxxxxxxxx/ > > 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
