[OE-core] [yocto] Move device tree generation from include file to bbclass

[Bach, Pascal]

Hi Bruce, Hi Nikolay

> >>>>
> >>>
> >>> Adding oe-core, since that's the right place to have a discussion
> >>> like this.

Thanks I'm never sure where to ask what :)
> >>>
> >>>> As ARM now also moved to device tree it look like in future we
> will
> >>>> have more kernels that are using device tree then ones that are
> >>>> not.
> >>>
> >>>
> >>> True, but it has been like this for quite some time now :)
> >>>
> >>>> As far as I understand currently the generation of device trees is
> >>>> controlled via KERNEL_DEVICETREE and is handled in via an include
> >>>> file recipes-kernel/linux/linux-dtb.inc.
> >>>>
> >>>> I was thinking about moving this include into a class so it
> becomes
> >>>> easier to use. Before I dive into implementing something I would
> >>>> like some feedback from the community.
> >>>
> >>>
> >>> The big trick with changing anything like this is compatibility
> with
> >>> existing recipes. Whatever we do, existing recipes and layers
> >>> shouldn't be broken .. or if they are broken, there should be a
> >>> compelling technical reason to do so.
> >>>
> >>>>
> >>>> I have the following variant in mind.
> >>>>
> >>>> Add the device tree generation to the current kernel.bbclass (or
> >>>> let kernel.bblcass inherit from a kernel-dtb.bbclass).
> >>>> This way all kernels would automatically be DT enabled. The class
> >>>> would check if KERNEL_DEVICETREE is set and generate device trees
> >>>> based on this information. For boards that don't have
> >>>> KERNEL_DEVICETREE set the class would do nothing and the behavior
> >>>> is like before. The advantage I see with this approach is that the
> >>>> only thing a user needs to do is to set KERNEL_DEVICETREE in the
> >>>> board and make sure the device trees are available in the kernel
> >>>> they like to build.
> >>>
> >>>
> >>> That's pretty much the experience that most users have now, since
> >>> there's nearly always a kernel recipe created, that recipe includes
> >>> linux-dtb.inc, and sets KERNEL_DEVICETREE.
> >>
> >>
> >> As far as I understood, Pascal's idea is to remove the need for user
> >> recipes to include linux-dtb.inc, and provide this functionality via
> >> inheritance.
> >
> > That is obvious. My questions are around "why". There's no big
> > technical advantage, and if you remove that existing file, you break
> > existing recipes. Which means you need to leave a stub in place.
> >
> > So without a technical advantage, it's churn for the sake of churn.
> 
> Well, removing redundancy and simplifying users' recipes could be
> considered an advantage. Also, as the contents of linux-dtb.inc are
> going to be moved to bbclass, the file can be left empty, later
> maintainers remove the extra line from all users' recipes in following
> commits. I don't see breaking as an option.

I completely agree that it is not worth to breaking existing recipes because of that.
But I think the option with an empty linux-dtb.inc is acceptable.

> >>> Everything else happens to build and package the device tree.
> >>>
> >>> Was there something specifically that was causing issues with the
> >>> current way of building them ?

There was no specific issue except that it feels like an unnecessary includes.
And it seems a bit odd to me that most of the work of building the kernel is done in bbclasses, 
while just the dtb handling is done with an include. But of course it still is more of a cosmetic 
change than a real technical necessity.

Regards
Pascal