Re: RFC: Copying Device Tree File into reserved area of VMLINUX before deployment

From: Ulf Samuelsson
Date: Tue Nov 21 2017 - 15:03:18 EST




On 2017-11-21 18:09, Rob Herring wrote:
On Mon, Nov 20, 2017 at 3:39 PM, Frank Rowand <frowand.list@xxxxxxxxx> wrote:
Hi Ulf, Rob,

On 11/20/17 15:19, Ulf Samuelsson wrote:


On 2017-11-20 05:32, Frank Rowand wrote:
Hi Ulf,


On 11/19/17 23:23, Frank Rowand wrote:
adding devicetree list, devicetree maintainers

On 11/18/17 12:59, Ulf Samuelsson wrote:
I noticed when checking out the OpenWRT support for the board that they have a method to avoid having to pass the device tree address to the kernel, and can thus boot device tree based kernels with U-boots that
does not support device trees.

Is this something that would be considered useful for including in mainstream:

BACKGROUND:
Trying to load a yocto kernel into a MIPS target (MT7620A based),
and the U-Boot is more than stupid.
Does not support the "run" command as an example.
They modified the U-Boot MAGIC Word to complicate things.
The U-Boot is not configured to use device tree files.
The board runs a 2.6 kernel right now.

Several attempts by me a and others to rebuild U-Boot according to
the H/W vendors source code and build instructions results in a
bricked unit. Bricked units cannot be recovered.

Hopefully you have brought this to the attention of the vendor. U-Boot
is GPL v2 (or in some ways possibly GPL v2 or later), so if you can not
build U-Boot that is equivalent to the binary U-Boot they shipped, the
vendor may want to ensure that they are shipping the proper source and
build instructions.


I am not the one in contact with the H/W vendor.
The U-boot is pretty old, and from comments from those
in contact with them, the U-Boot knowledge at the H/W vendor
is minimal at best.
It might even be that they program an U-boot where the upgrade of the U-boot is broken...



Not my choice of H/W, so I cannot change it.


===================================================================
OPENWRT:
I noticed when checking out the OpenWRT support for the board that
they have a method to avoid having to pass the device tree address
to the kernel, and can thus boot device tree based kernels with
U-boots that does not support device trees.

What they do is to reserve 16 kB of kernel space, and tag it with
an ASCII string "OWRTDTB:". After the kernel and dtb is built, a
utility "patch-dtb" will update the vmlinux binary, copying in the
device tree file.

===================================================================
It would be useful to me, and I could of course patch the
mainstream kernel, but first I would like to check if this is of
interest for mainstream.

Not in this form. Hard coding a fixed size area in the boot image
to contain the FDT (aka DTB) is a non-starter.

OK, Is it the fixed size, which is a problem?

Yes, it is the fixed size which is a problem.

Is generally combining an image with a DTB into a single file also a non-starter?

Can you jump in here Rob? My understanding is that CONFIG_ARM_APPENDED_DTB,
which is the ARM based solution that Mark mentioned, was envisioned as a
temporary stop gap until boot loaders could add devicetree support. I don't
know if there is a desire to limit this approach or to remove it in the
future.

Yes, but I doubt we'll ever remove it. Most android devices use it (or
the multiple appended dtb stuff Qcom did) and there's out of tree
patches to do appended dtb on arm64. However, arm64 is a done a bit
differently in that the bootloader has to find the dtb rather than the
kernel (or really the decompressor). The purpose for arm64 is people
like the single kernel+dtb image, not legacy bootloader support as DT
support was there from day 1.

Another option is to do what's called an impedance matcher like this[1].

I'm not sure why this feature should not be permanently supported. I'm being
cautious, just in case I'm overlooking or missing an important issue, thus
asking for Rob's input. I do know that this feature does not advance the
desires of people who want a single kernel (single boot image?) that runs on
many different systems, instead of a boot image that is unique to each
target platform. But I don't see why that desire precludes also having
an option to have a target specific boot image.

MIPS already supports built-in dtbs which is an in-kernel way to get
the dtb rather than the decompressor. The addition in this case is
just putting in a dummy dtb and putting in the real dtb later. I'm not
sure what MIPS image header looks like, but seems like it should be
possible to have the "built-in dtb" point to an appended dtb and just
update sizes in the image file. Then you don't have the fixed size.

But now that I finished writing this, I see there is already appended
DTB support in MIPS.

I just managed to complete a build using yocto where I made the following changes.

I build the dtbs before linking the uImage.

I add some configuration variables, which define
* if the dtb should be included
* which dtb file should be included

uImage always gziped

Checking the resulting uImage, shows that the ".appended_dtb" section now contains the dtb file.

I have not loaded it into a board yet, but the default init seems to check if there is a U-boot header in the ".appended_dtb" section.

My bootloader has a custom MAGIC so loading still won't work.



Rob

[1] https://github.com/zonque/pxa-impedance-matcher


--
Best Regards
Ulf Samuelsson