[RFC PATCH 5/7] kbuild: fit: Add opt-in FIT_FACTOR_DTBS support using fdtfactor

[Simon Glass]

Hi Doug,

Doug Anderson <dianders at chromium.org> wrote:
> I think that was in the the "after the cut" part, where I said:
>
> You can see the fdtfactor post at:
>  http://lore.kernel.org/r/20261015182221.2931742-1-dianders@chromium.org

OK, that covers where the tool comes from, and the question goes away
once fdtfactor is imported. Text below the cut doesn't reach the git
log, though, so a sentence or two in the commit message on why this is
wanted would still help.

> The goal was to make sure that we could fully reconstruct
> __local_fixups__ after applying the overlay to the base. v2 of the
> fdtfactor series handled this more properly, and I'm planning to see
> this through and land it unless there are major objections.
>
> By default here, I'm running fdtfactor with "-S" to strip the
> __local_fixups__ out with the assumption that today people don't
> really need them. However, someone could easily override that locally.
>
> I'm a little unclear how things would silently produce a bad overlay, though.

Sorry, I should have been clearer. This is about overlays which kbuild
applies at build time, not runtime ones like Beaglebone capes. An
entry such as:

    lemans-evk-camera-dtbs := lemans-evk.dtb lemans-evk-camera.dtbo

is merged by fdtoverlay into a plain lemans-evk-camera.dtb, which the
make_fit.py path then puts in the FIT like any other DTB. There are
several hundred of these on arm64.

My concern is the input to fdtfactor, not its output. In the in-tree
libfdt, fdt_overlay_apply() uses the .dtbo's __local_fixups__ and
__fixups__ to resolve its phandles, then overlay_merge() copies only
the __overlay__ contents into the base. Only __symbols__ is merged, so
the __local_fixups__ in lemans-evk-camera.dtb cover just the
lemans-evk.dtb part. Cells from the .dtso such as:

    avdd-supply = <&vreg_cam1_1p8>;
    remote-endpoint = <&imx577_ep1>;

end up as plain numbers, with nothing marking them as phandles. If
fdtfactor needs to renumber them, or puts those nodes in a per-board
overlay, it cannot tell they are phandles, so the overlay applies
cleanly but the links are wrong. Stripping the fixups from the output
with -S doesn't change that, since the information is already gone
from the input.

Does v2 make fdt_overlay_apply() keep the overlay's __local_fixups__?
If so, composite DTBs would be fine once that is synced into
scripts/dtc, and my concern goes away. If not, one option is for
FIT_FACTOR_DTBS to imply FIT_DECOMPOSE_DTBS, so that make_fit.py puts
the original .dtb and .dtbo files in the FIT separately and the .dtbo
keeps its own fixups. Leaving composite DTBs out would be a simpler
fallback. What do you think?

Regards,
Simon