[PATCH v4 1/1] efi_loader: expose the device-tree file name

[Caleb Connolly]

Hi Heinrich,

On 3/28/25 15:18, Heinrich Schuchardt wrote:
> On 28.03.25 14:00, Caleb Connolly wrote:
>>
>>
>> On 3/28/25 13:01, Simon Glass wrote:
>>> Hi Caleb,
>>>
>>> On Sun, 23 Mar 2025 at 12:39, Caleb Connolly 
>>> <caleb.connolly at linaro.org> wrote:
>>>>
>>>> Hi all,
>>>>
>>>> Reviving this as it is still very much an issue, and especially 
>>>> relevant
>>>> for Qualcomm platforms.
>>>>
>>>> On 11/3/23 20:44, Simon Glass wrote:
>>>>> Hi Heinrich,
>>>>>
>>>>> On Wed, 25 Oct 2023 at 15:22, Heinrich Schuchardt
>>>>> <heinrich.schuchardt at canonical.com> wrote:
>>>>>>
>>>>>> On 10/25/23 23:13, Tom Rini wrote:
>>>>>>> On Wed, Oct 25, 2023 at 10:28:05PM +0200, Mark Kettenis wrote:
>>>>>>>>> Date: Wed, 25 Oct 2023 21:57:44 +0200
>>>>>>>>> From: Heinrich Schuchardt <heinrich.schuchardt at canonical.com>
>>>>>>>>>
>>>>>>>>> On 10/25/23 20:23, Simon Glass wrote:
>>>>>>>>>> Hi Heinrich,
>>>>>>>>>>
>>>>>>>>>> On Tue, 24 Oct 2023 at 18:02, Simon Glass <sjg at google.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>> Hi Heinrich,
>>>>>>>>>>>
>>>>>>>>>>> On Mon, 23 Oct 2023 at 23:20, Heinrich Schuchardt
>>>>>>>>>>> <heinrich.schuchardt at canonical.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>> Forward and backward compatibility of Linux kernel device- 
>>>>>>>>>>>> trees is
>>>>>>>>>>>> sometimes missing. One solution approach is to load a kernel 
>>>>>>>>>>>> specific
>>>>>>>>>>>> device-tree. This can either be done via a U-Boot scripts 
>>>>>>>>>>>> (like the one
>>>>>>>>>>>> generated by Debian package flash-kernel or by a boot loader 
>>>>>>>>>>>> like GRUB.
>>>>>>>>>>>> The boot loader approach currently requires to know the 
>>>>>>>>>>>> device-tree name
>>>>>>>>>>>> before first boot which makes it unusable for generic images.
>>>>>>>>>>>>
>>>>>>>>>>>> Expose the device-tree file name as EFI variable FdtFile.
>>>>>>>>>>>> This will allow bootloaders to load a kernel specific 
>>>>>>>>>>>> device- tree.
>>>>>>>>>>>
>>>>>>>>>>> kernel-specific
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> The variable will not be exposed on ACPI based systems or if 
>>>>>>>>>>>> the
>>>>>>>>>>>> environment variable fdtfile is not defined.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Heinrich Schuchardt 
>>>>>>>>>>>> <heinrich.schuchardt at canonical.com>
>>>>>>>>>>>> ---
>>>>>>>>>>>> v4:
>>>>>>>>>>>>             Generalize the description of the content of 
>>>>>>>>>>>> $fdtfile.
>>>>>>>>>>>> v3:
>>>>>>>>>>>>             Add documentation
>>>>>>>>>>>> v2:
>>>>>>>>>>>>             Use a unique GUID to enable future U-Boot 
>>>>>>>>>>>> independent
>>>>>>>>>>>>             standardization.
>>>>>>>>>>>>             Do not try to add the variable on ACPI based 
>>>>>>>>>>>> systems.
>>>>>>>>>>>> ---
>>>>>>>>>>>>      doc/develop/uefi/uefi.rst  | 14 ++++++++++++++
>>>>>>>>>>>>      include/efi_loader.h       |  5 +++++
>>>>>>>>>>>>      lib/efi_loader/efi_setup.c | 30 +++++++++++++++++++++++ 
>>>>>>>>>>>> + ++++++
>>>>>>>>>>>>      3 files changed, 49 insertions(+)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/doc/develop/uefi/uefi.rst b/doc/develop/uefi/ 
>>>>>>>>>>>> uefi.rst
>>>>>>>>>>>> index fb16ac743a..702c490831 100644
>>>>>>>>>>>> --- a/doc/develop/uefi/uefi.rst
>>>>>>>>>>>> +++ b/doc/develop/uefi/uefi.rst
>>>>>>>>>>>> @@ -916,6 +916,20 @@ So our final format of the 
>>>>>>>>>>>> FilePathList[] is::
>>>>>>>>>>>>
>>>>>>>>>>>>          Loaded image - end node (0xff) - VenMedia - 
>>>>>>>>>>>> initrd_1 - [end node (0x01) - initrd_n ...] - end node (0xff)
>>>>>>>>>>>>
>>>>>>>>>>>> +EFI variable FdtFile
>>>>>>>>>>>> +~~~~~~~~~~~~~~~~~~~~
>>>>>>>>>>>> +
>>>>>>>>>>>> +Ideally U-Boot would always expose a device-tree that can 
>>>>>>>>>>>> be used for booting
>>>>>>>>>>>> +any operating systems. Unfortunately operating systems like 
>>>>>>>>>>>> Linux sometimes
>>>>>>>>>>>> +break forward and backward compatibility. In this case 
>>>>>>>>>>>> there is a need to load
>>>>>>>>>>>> +an operating system version specific device-tree.
>>>>>>>>>>>
>>>>>>>>>>> This seems to be a strong statement. Given the effort that 
>>>>>>>>>>> goes into
>>>>>>>>>>> the DT, changes are supposed to be backwards-compatible. Is this
>>>>>>>>>>> generally true, or is it just that we want an up-to-date DT 
>>>>>>>>>>> for the
>>>>>>>>>>> kernel to enable new features?
>>>>>>>>>>
>>>>>>>>>> Did you see this comment?
>>>>>>>>>
>>>>>>>>> It would have been nice to put the person which made that 
>>>>>>>>> comment on copy.
>>>>>>>>>
>>>>>>>>> The truth lies in the world "supposed":
>>>>>>>>>
>>>>>>>>> The idea of a device-tree that never needs to change is quite 
>>>>>>>>> old and
>>>>>>>>> never became true on ARM devices.
>>>>>>>>>
>>>>>>>>> We all know Linux tends to break both forward and backward 
>>>>>>>>> compatibility
>>>>>>>>> of device-trees. Here is a nice example:
>>>>>>>>>
>>>>>>>>> d0c6707ca423 ("arm64: dts: allwinner: H5: NanoPi Neo Plus2: 
>>>>>>>>> phy- mode
>>>>>>>>> rgmii-id")
>>>>>>>>>
>>>>>>>>> Driver changes broke forward and backwards compatibility of a 
>>>>>>>>> lot of
>>>>>>>>> Allwinner boards.
>>>>>>>>
>>>>>>>> Well, that happened in 2020.  Things have gotten better over time.
>>>>
>>>> (kinda off-topic context on DT version compat)
>>>>
>>>>   From what I've seen, there is not yet very much infrastructure or
>>>> common practise in place in the kernel to handle parsing DTB in a
>>>> backwards compatible way. For example there has been efforts to 
>>>> simplify
>>>> the dwc3 devicetree for Qualcomm platforms with a series dating back
>>>> about as far as this U-Boot patch!
>>>>
>>>> https://lore.kernel.org/linux-arm-msm/20250318-dwc3-refactor- 
>>>> v5-1-90ea6e5b3ba4 at oss.qualcomm.com/
>>>>
>>>> Earlier versions attempted to convert the older DTS to the newer format
>>>> internally, but after much discussion it was decided that this wasn't
>>>> really feasible, instead the new approach is to duplicate the entire
>>>> dwc3 driver to maintain DT compatibility.
>>>>
>>>> It's obviously good to see compatibility taken seriously, since it 
>>>> seems
>>>> clear that if we ever want to treat DT as firmware, the kernel will 
>>>> need
>>>> to do this (and eventually vendors could ship laptops with DT out of 
>>>> the
>>>> box which works with upstream -- crazier things have happened).
>>>>
>>>> But I think in the mean time we still want to be able to drive distro
>>>> adoption, and the way I see it teaching GRUB and systemd-boot about a
>>>> new FdtFile EFI variable is going to make that way simpler...
> 
> There was a discussion in the context of EBBR if this is the right way 
> forward. And the general opinion was that the compatible string should 
> be used for matching the dtb files.

Ideally, the file name should contain the same information as the 
compatible string, or more. Currently in upstream there are devices with 
different display variants that have identical compatible strings but 
different dtb files.

> 
>>>
>>> GRUB is unlikely even to handle devicetree properly as it does not
>>> implement FIT, nor the best-match algorithm in FIT. So everything is a
>>> workaround.
> 
> For the Qualcomm laptop case the current patch would not solve anything 
> as the selection of the device-tree needs to be based on SMBIOS value.

I bought this idea up with the laptops team and the upstream Qualcomm 
maintainer, everyone agreed that this approach would be a much more 
scalable intermediate solution. Since the HWID matching stuff will just 
not scale or make it into distros easily (e.g. generic installer won't 
work on the laptops currently because of this).

If my understanding of EFI is correct, it would be possible to set this 
variable on the laptops via EFI shell (or even some custom EFI app for 
simplicity), and there is a non-zero chance that we could get some 
vendors to support setting this variable directly in their EFI firmware 
setup.

> 
> Systemd-boot allows to add a rule-table to the UKI header to select 
> device-trees from a UKI image.
> 
> That is a smarter solution than trying to hard-code logic in U-Boot to 
> set fdtfile and an FdtFile EFI variable.

Right, we try to guess fdtfile today in mach-snapdragon and that sucks. 
I brainstormed this a bit with Ilias and what we came up with is to have 
U-Boot embed the fdtfile value directly into the /chosen node of every 
DTB we build with OF_UPSTREAM. Since the paths are identical to the kernel.

This way you can have a single U-Boot binary (as we support for 
Qualcomm) boot on different boards and whichever DTB you give to U-Boot 
(say qcom/sdm845-db845c.dtb for example) will be the value that FdtFile 
gets. Then sd-boot/GRUB will look for a DTB under the same name relative 
to the devicetree-dir for the kernel version you're booting.

Yes long term all this is far from the ideal solution, but it's a lot 
better than what we have today and much less hacky than every other 
proposal I've seen particularly for the laptops.

Most importantly, it doesn't require distro changes, we already know for 
sure that distros are not going to adopt FIT. Arguing on the merits is 
pointless.

Kind regards,

> 
>>
>> FIT has nothing to do with DT loading
>>>
>>>>>>>
>>>>>>> Well, yes and no.  Given the brief summary here, I bet this was just
>>>>>>> like when phy-mode and am335x platforms had DT compatibility 
>>>>>>> broken and
>>>>>>> the answer was that it was OK because the DT was incorrectly 
>>>>>>> describing
>>>>>>> hardware. So this is the reminder that there are cases of 
>>>>>>> breaking DT
>>>>>>> compatibility that are allowed. Even if the DT has been out (and 
>>>>>>> wrong)
>>>>>>> for several years.
>>>>>>>
>>>>>>> That's not the main point of this thread and I don't want to derail
>>>>>>> things further along this point, I just want to note that the 
>>>>>>> details
>>>>>>> here reminded me of when things are allowed to be incompatible with
>>>>>>> previous trees.
>>>>>>>
>>>>>>
>>>>>
>>>>> Unfortunately I was off the list for a week or so, but I see this one
>>>>> so will reply here.
>>>>>
>>>>>> You are conflating things:
>>>>>>
>>>>>> The EFI variable is a hint to the GRUB OS to find the correct
>>>>>> device-tree file without scanning hundreds of device-tree. You can 
>>>>>> not
>>>>>> build a compatibility check for it into U-Boot.
>>>>>
>>>>> Actually we can...and I think that is what we should do.
>>>>
>>>> No, how and where an EFI bootloader app chooses to load the FDT from is
>>>> entirely implementation dependent. It may be extremely costly to 
>>>> compare
>>>> compatible strings for every single dtb, and whether or not to do so is
>>>> the decision of the distro anyways, all we can do in U-Boot is try to
>>>> promote best practise (which clearly isn't compatible prop comparison
>>>> for hundreds of DTBs...)
>>>
>>> U-Boot promotes best practice, which is to check the compatible
>>> properties of hundreds of nodes in the FIT to find the right one.
>>
>> Distros don't ship FIT images (which would be absolutely huge by the 
>> way, 75M for all DTBs today), they ship DTBs in a directory structure 
>> following the kernel layout.
> 
> It might make sense to support compressed FIT images.
> 
> Whether all device-trees are stored as files in /lib/firmware/<kernel- 
> version>/device-tree/ or are part of the kernel image (FIT or UKI) does 
> not change the space requirement on disk.
> 
> Best regards
> 
> Heinrich
> 
>>
>> If the kernel started outputting FIT images with all DTBs it would 
>> have to be installed in addition to the directory layout, and only 
>> provide a considerably less efficient interface to pick a DTB.
>>
>> We want Fedora images that you download today to be able to boot and 
>> select an appropriate DTB. FIT is just not a good fit (haha) for this 
>> problem.
>>
>> A fit mapping compatible strings to filenames would be an improvement 
>> but that's so much additional complexity to come up with a value that 
>> U- Boot already has...
>>
>>>>
>>>> Providing FdtFile as an EFI variable will make it possible for grub and
>>>> systemd-boot to support the same kind of "fdtdir" property that 
>>>> extlinux
>>>> does, this would simply enable versioned DTB loading for distros and
>>>> make booting wayyyyy easier.
>>>
>>> It is perpetuating an incorrect approach, though. It won't lead to
>>> happiness. Perhaps systemd-boot should support FIT?
>>
>> What do i even say to this. You haven't explained what's incorrect
>>>
>>> As to versions, the compatible string needs to handle that. We cannot
>>> have a situation where we ship two different devices trees that have
>>> the same compatible strings but different contents as there is no
>>> (spec-correct) way to distinguish them.
>>>
>>>>
>>>> With regards to the design, I think a good follow-up patch would set a
>>>> default value for $fdtfile (and consequently FdtFile) from the
>>>> DEVICE_TREE build variable as a constant (maybe only if OF_UPSTREAM is
>>>> enabled).
>>>
>>> Please no. We should not continue down this broken path.
>>
>> ???
>>>
>>>>
>>>> This would be correct in almost all cases (and wayyy better than the
>>>> hacky fdtfile generating code we have in mach-snapdragon right now).
>>>
>>> Yes, but you really should figure out how to remove that code. It is
>>> present on other boards too. One of the things not on my list (but I
>>> wish it were) is to clean up how RISC-V does devicetree selection
>>> within U-Boot.
>>
>> That is why i revived this patch...
>>>
>>>>>
>>>>> For distro-boot, do you mean the scripts? We are trying to deprecate
>>>>> those. Of course we have brought in all the same work-arounds, etc.,
>>>>> but with bootstd we can start to clean things up, I hope.
>>>>>
>>>>> So let's think about how we can have U-Boot choose the right DT to 
>>>>> boot with.
>>>>>
>>>>> Regards,
>>>>> Simon
>>>>
>>>> Heinrich, since it's been a while, if you aren't super interested in
>>>> re-sending this patch I'd be happy to address the wording feedback and
>>>> re-spin it, along with a patch setting the default $fdtfile (else I can
>>>> send that as a follow-up).
>>>
>>> The best thing to do here is to use FIT. You can put all the FDTs in a
>>> FIT and leave the kernel out, if necessary.
>>
>> You cannot do this, distros will not do this for thousands of FDTs
>>>
>>> Another idea, which I may have suggested before (can't remember) would
>>> be to add an EFI service which U-Boot implements, which returns the
>>> correct FDT to use. U-Boot can then scan the files one by one to
>>> figure it out, although hopefully that would prompt someone to create
>>> a FIT (or a text file?) which has this information in it.
>>>
>>> But fdtfile is just not the right approach.
>>
>> You've pushed for FIT but never explained why fdtfile is a bad solution.
>>>
>>>>
>>>> I'll also open an issue for getting support for this added to 
>>>> systemd-boot.
>>>
>>> FIT?
>>>
>>> Regards,
>>> Simon
>>
> 

-- 
Caleb (they/them)