On Tue, Dec 17, 2024 at 8:01 PM Brad Griffis <bgriffis@xxxxxxxxxx> wrote: > > > From: Thierry Reding <treding@xxxxxxxxxx> > > > > The device tree specification (v0.4) suggests that #address-cells is > > mandatory for interrupt parent nodes. If this property is missing, Linux > > will default to the value of 0. > > Just to clarify, this relates to interrupt-map specifically. Yes. That is the only case that really cares and needs it. We used to warn on all interrupt-controller nodes if #address-cells was missing. But that turns out to be way too many cases to fix, so we dropped that part. > In that > scenario the device tree spec requires that both the child node and > parent node specify #address-cells and #interrupt-cells. It further > specifies if a unit address component is not needed then it must be > explicitly defined as zero. In other words, this does not seem to be > just a suggestion, but more of a firm requirement. unit-address being 0 and #address-cells being 0 are 2 different things. > > A number of device tree files rely on Linux' fallback and don't specify > > an explicit #address-cells as suggested by the specification. This can > > cause issues when these device trees are passed to software with a more > > pedantic interpretation of the DT spec. > > The device tree spec also says that in the case where #address-cells is > not specified that a value of 2 should be assumed. So in this context, I > find the kernel's current practice of assuming #address-cells = <0> also > violates the spec. Relying on defaults is not good practice and deprecated for longer than we've had DT support on Arm. The default for the kernel on PPC is 1 which doesn't agree with dtc nor OpenFirmware. The kernel walking up all parents to find #address-cells/#size-cells is not behavior defined in any spec either. 6.13 now warns if parent nodes or defaults are used, but not in the case of interrupt parsing. > > Add a warning when this case is detected so that device tree files can > > be fixed. > > I think a warning is reasonable, but perhaps we should consider making > it an outright error. Though given the number of impacted device trees, > perhaps that needs to be done in a couple of steps. I'm more inclined to change the spec if needed. We only need #address-cells if we have interrupt-map, so that's all we should check. If a given binding wants to be stricter, we can do that in its schema. Rob
