On Sat, Jan 07, 2023 at 04:54:57PM +0100, Andreas Kemnade wrote: > On Sat, 7 Jan 2023 16:07:35 +0100 > Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> wrote: > > > On 07/01/2023 16:01, Andreas Kemnade wrote: > > > On Sat, 7 Jan 2023 15:09:24 +0100 > > > Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> wrote: > > > > > >> On 07/01/2023 15:07, Andreas Kemnade wrote: > > >>> On Sat, 7 Jan 2023 15:00:56 +0100 > > >>> Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> wrote: > > >>> > > >>> [...] > > >>>>>> I asked to remove half-compatible. Not to enforce. > > >>>>>> > > >>> so you are saying that allowing > > >>> compatible = "A", "B" > > >>> is not ok, if B is not fully compatible. I agree with that > > >>> one. > > >> > > >> I did not say that. It's not related to this problem. > > >> > > > You said "I asked to remove half-compatible" that means to me > > > remove "B" if not fully compatible with A which sounds sane to me. > > > > > >> Again - you cannot have device which is and is not compatible with > > >> something else. It's not a Schroedinger's cat to be in two states, > > >> unless you explicitly document the cases (there are exception). If this > > >> is such exception, it requires it's own documentation. > > >> > > > so conclusion: > > > If having A and B half-compatible with A: > > > > > > compatible = "A" only: is allowed to specifiy it the binding (status quo), > > > but not allowed to make the actual dtsi match the binding documentation > > > https://lore.kernel.org/linux-devicetree/72e1194e10ccb4f87aed96265114f0963e805092.camel@xxxxxxxxxxxxxx/ > > > and > > > https://lore.kernel.org/linux-devicetree/20210924091439.2561931-5-andreas@xxxxxxxxxxxx/ > > > > > > compatible = "A", "B" in the binding definition: is not allowed ("I asked to remove > > > half-compatible" (= removing B)) > > > > No, half compatible is the A in such case. > > > I think that there is some misunderstanding in here. I try once again. > > Define compatible with "X" here: > To me it means: > > device fully works with flags defined in: > > static const struct esdhc_soc_data usdhc_X_data = { ... }; > > with usdhc_X_data referenced in > { .compatible = "X", .data = &usdhc_X_data, }, > > > So if there is only "A" matching with above definition of compatibility > compatible = "A" would sound sane to me. > > And scrutinizing the flags more and not just wanting to achieve error-free > dtbs_check, I think is this in most cases where there is only "A". > > If there is "A" and "B" which match that compatibility definition, you > say that only compatible = "A", "B" is allowed, but not compatible = "A". > In that case I would have no problem with that. > > But if there is only "A" but no "B" matching the above definition, I would expect > that only compatible = "A" is allowed but *not* compatible = "A", "B". A is either compatible with B or it isn't. You can look at that from the h/w perspective and client/OS perspective. From the h/w side, is the h/w interface the same or only has additions which can be ignored? On the client side, the question is whether a client that only understands B could use A's h/w without change. Looking at the match data is a good indicator of that for Linux. It's also possible the answer is different for different clients, but we only need 1 client that could benefit from compatibility. Rob