Barebox OF-Overlay Handling

[Michael Tretter]

On Wed, 13 Jan 2021 15:07:48 +0100, Rouven Czerwinski wrote:
> On Wed, 2021-01-13 at 12:07 +0100, Marco Felsch wrote:
> > Hi there,
> > 
> > I have the following problem. My customer is using overlays for
> > external devices like: display, camera, etc. The current overlay support
> > is awesome and most of it works out of the box. There is only one
> > nitpick: If Barebox isn't build with CONFIG_OF_OVERLAY_LIVE enabled many
> > error's are printed during boot. Those error's are ignored but let the
> > user assume that something went really bad:
> > 
> > ERROR: of_resolver: __symbols__ missing from base devicetree
> > ERROR: of_overlay: fragment 30164b88: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 30164c84: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3012fb08: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3012fc34: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3012fd60: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3012fe8c: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3012ffb8: phandle 0xffffwfffffff not found
> > ERROR: of_overlay: fragment 301300e4: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 30130210: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 3013033c: phandle 0xffffffff not found
> > ERROR: of_resolver: __symbols__ missing from base devicetree
> > ERROR: of_overlay: fragment 30130e94: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 30130fc4: phandle 0xffffffff not found
> > ERROR: of_resolver: __symbols__ missing from base devicetree
> > ERROR: of_overlay: fragment 30131a14: phandle 0xffffffff not found
> > ERROR: of_overlay: fragment 301342d4: phandle 0xffffffff not found
> > 
> > The of_firmware_load_overlay() calls triggering those errors.
> > The error's by itself are correct and I don't wanna change them but the
> > context must be correct by context I mean:
> > 
> >  - barebox-dt on fpga-platform: An FPGA platform needs a barebox
> >    base devicetree with __symbols__ and those error are correct because
> >    the firmware manager needs to load the firmware.
> >  - barebox-dt on non fpga-platform: Those errors are not correct.
> >    We don't need __symbols__ for the barebox base devicetree.
> 
> I was wondering why you were getting this error even on a platform
> without FPGA, turns out of_firmware_load_overlay() is always called if
> a devicetree overlay is passed via the DT. This in itself isn't a
> problem, but IMO of_firmware_load_overlay() should check first whether
> the overlay contains the right compatible ("fpga-region") and property
> ("firmware-name") before doing of_resolve_phandles and
> of_process_overlay. This way we can skip a lot of unnecessary overlay
> handling if we don't end up loading a firmware.

At the moment, the check and firmware load path use the same code path,
because there is no difference except that firmware load requires a resolved
overlay and the check does not. This is a viable solution and there is even a
helper to iterate the fragments and call a check function for each fragment.
However the helper calls find_target, which is responsible for most of the
error messages, and therefore, does not completely resolve the issue.

(Indeed, the check if the fragment has the "fpga-region" compatible is
missing. That's a bug.)

I would rather change the error logging to only log errors if actually
something goes wrong. Because none of the aforementioned errors is actually
fatal from a system perspective, but only from a function perspective.
Furthermore, the messages are not helpful at all to actually find the problem
in case of a unresolved overlay.

>  
> >  - kernel-dt on any platform: Those error's are correct if
> >    the overlays are using phandles which should be the case most the
> >    time.
> 
> AFAICS this should also fix this case.
> 
> > My proposed solution would be a stub like this:
> > 
> > #ifdef CONFIG_OF_OVERLAY_LIVE
> > int of_firmware_load_overlay(struct device_node *overlay, const char *path);
> > #else
> > static inline int of_firmware_load_overlay(struct device_node *overlay, const char *path)
> > {
> >        return 0;
> > }
> > #endif /* CONFIG_OF_OVERLAY_LIVE */

Nack. of_firmware_load_overlay must fail if the overlay describes an
fpga-region. Otherwise the device tree that is passed to the kernel might have
an fpga-region with a firmware-name property and that per specification
requires that the fpga is programmed.

Michael

> 
> While correct in case all firmware overlays use phandles, in theory
> overlays could be purely path-based, with not dependency on the
> __symbols__ node. Not sure how common this is.

> 
> > The disadvantage of this solution is that it can happen to end in a
> > non-booting device for FPGA platforms using overlays to programm the
> > bit stream and forget to enable CONFIG_OF_OVERLAY_LIVE because the
> > kernel tries to access regions not existing.
> > 
> > I've discussed this with Michael in private but we didn't came to a
> > conclusion. Therefore I'm asking here any input would be helpfull :)
> 
> Regards,
> Rouven Czerwinski
> 
>