DT overlay issues - dtc flags

[Frank Rowand]

On 04/10/17 09:16, Andreas Färber wrote:
> Hi,
> 
> I've tried to play around with Device Tree overlays (.dtbo files):
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/overlay-notes.txt
> 
> First of all, that document refers to a non-existing
> Documentation/devicetree/dt-object-internal.txt. Could someone please
> fix that one way or another?

I will.


> In my particular example I've tried to extend the &i2c1 and &gpio nodes
> of 4.11-rc5 arm64 broadcom/bcm2837-rpi-3-b.dts. The above documentation
> prominently claims that this can be done via target = <&foo> syntax, but
> U-Boot's fdt apply command fails for such a file. If instead I use the
> alternative target-path = "/soc/..." then it works just fine.

The "target = <&foo>" syntax is the internal dtb syntax that the kernel
expects.  The plan is to hide this internal format so it is not visble
at the source level, but that will require some more changes to the dtc
compiler.


> As mentioned in the very bottom of the documentation, resolution of
> phandle target references requires a __symbols__ node in the base .dtb.
> IIUC this is only generated when passing the -@ dtc command line flag.
> 
> At first I thought this were an issue with how we build the .dtb files
> in openSUSE [1], but by my reading of the kernel Makefiles not passing
> -@ in DTC_FLAGS or cmd_dtc, you should run into the exact same issue.

Yes, the kernel support for overlays is not complete.  This is one of
the issues that has not been addressed yet.

I don't have the rpi-3 documentation handy, so I'm going to take a wild
guess that the &i2c1 and &gpio node changes that you need are for
hardware on a hat.  If so, the plan for the kernel is to describe
the signals available via the hat connectors in the device tree and
then make those signals available through the connector node instead
of having the hat overlay directly reference the signal node.  The
connector abstraction has been discussed but is not implemented.


> I could think of a few ARMv7-M systems where such DT bloat might be
> undesired (small flash sector sizes), but then it would seem easier to
> suppress -@ where needed than to have a feature that by all practical
> means is half unusable by default.
> 
> U-Boot itself appears to face a similar issue in that its internal
> Device Trees are built without -@, and via Alex' distro boot extensions
> this internal DT is passed on via UEFI as fallback when no external .dtb
> file is found. So in the non-SPL case the DT should probably be built
> with -@, too.
> 
> Or am I misunderstanding something here?
> 
> Thanks,
> Andreas
> 
> [1]
> https://build.opensuse.org/package/view_file/Kernel:HEAD/dtb-aarch64/dtb-aarch64.spec?expand=1
>