Geert Uytterhoeven wrote on Mon, Apr 19, 2021 at 11:03:24AM +0200: > > This is going to need quite some more work to be acceptable, in my > > opinion, but I think it should be possible. > > In general, this is very hard to do, IMHO. Some drivers may be used on > multiple platforms, some of them registering an SoC device, some of > them not registering an SoC device. So there is no way to know the > difference between "SoC device not registered, intentionally", and > "SoC device not yet registered".
Hm, good point, I was probably a bit too optimistic if there are devices which don't register any soc yet have drivers which want one; I don't see how to make the difference indeed... And that does mean we can't just defer all the time. > soc_device_match() should only be used as a last resort, to identify > systems that cannot be identified otherwise. Typically this is used for > quirks, which should only be enabled on a very specific subset of > systems. IMHO such systems should make sure soc_device_match() > is available early, by registering their SoC device early. I definitely agree there, my suggestion to defer was only because I know of no other way to influence the ordering of drivers loading reliably and gave up on soc being init'd early. In this particular case the problem is that since 7d981405d0fd ("soc: imx8m: change to use platform driver") the soc probe tries to use the nvmem driver for ocotp fuses for imx8m devices, which isn't ready yet. So soc loading gets pushed back to the end of the list because it gets defered and other drivers relying on soc_device_match get confused because they wrongly think a device doesn't match a quirk when it actually does. If there is a way to ensure the nvmem driver gets loaded before the soc, that would also solve the problem nicely, and avoid the need to mess with all the ~50 drivers which use it. Is there a way to control in what order drivers get loaded? Something in the dtb perhaps? Thanks, -- Dominique