On 19.03.2024 2:53 PM, Robin Murphy wrote: > On 2024-03-09 1:31 pm, Konrad Dybcio wrote: >> The smmu-v3 binding currently doesn't differentiate the SoCs it's >> implemented on. This is a poor design choice that may bite in the future, >> should any quirks surface. > > That doesn't seem entirely fair to say - the vast majority of bindings don't have separate compatibles for every known integration of the same implementation in different SoCs. And in this case we don't have per-implementation compatibles for quirks and errata because the implementation is architecturally discoverable from the SMMU_IIDR register. > > We have the whole mess for QCom SMMUv2 because the effective *implementation* is a mix of hardware and hypervisor, whose behaviour does seem to vary on almost a per-SoC basis. I'm not at all keen to start repeating that here without very good reason, and that of "documenting" a device which we typically expect to not even be accessible isn't really convincing me... >From my POV as an arch dts maintainer, this often ends up being the only way to retroactively add some conditional action into the code - the kernel is supposed to be backwards compatible with older device trees. And so far it's been almost by luck that all of the smmuv3 implementations have been a straight copy-and-paste of the reference design (or close enough), I don't believe this will be for much longer. Konrad
