On Mon Nov 4, 2024 at 4:37 PM CET, Rob Herring wrote: > On Thu, Oct 31, 2024 at 04:52:51PM +0100, Théo Lebrun wrote: > > Some compatibles expose a single clock. For those, we used to let them > > using `#clock-cells = <0>` (ie <&olb> reference rather than <&olb 0>). > > > > Switch away from that: enforce a cell for all compatibles. This is more > > straight forward, and avoids devicetree changes whenever a compatible > > goes from exposing a single clock to multiple ones. > > Your reasoning is flawed. Changing #clock-cells is an ABI break. So you > should only be changing this if it was just wrong. And if it's not wrong > in some cases, you shouldn't be changing those. The h/w either has 1 > clock or multiple and #clocks-cells should match. I see your reasoning, and I agree that changing #clock-cells is an ABI break. However, there are two things to take into account: - We do not (yet?) have an omniscient view of the hardware. We do not know what every single register in those memory regions do. Some clocks might be lurking in the shadows, especially as we don't support many HW capabilities yet. - The earlier the better. If we discover later down the road that, indeed, some more clocks were hiding, we'll have to do an ABI break. At that point, some people might actually be using the platform. Seeing what we currently have supported upstream versus the amount of HW blocks available in the SoC, I cannot imagine anyone using the platform with an upstream kernel. So the choice is: - potential ABI break in the future, once people use the platform, or, - guaranteed ABI break now, when no one is using it. I pick option two! Do you agree with the thought process? Thanks Rob, -- Théo Lebrun, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
