Re: [PATCH] external references for device tree overlays

On 06/08/17 00:08, Stefani Seibold wrote:

Hi Frank,


On Wed, 2017-06-07 at 20:07 -0700, Frank Rowand wrote:

quoted

Adding Jon.

Adding back the original distribution.

Hi Stefani,

On 06/06/17 23:03, Stefani Seibold wrote:

quoted

Am Dienstag, den 06.06.2017, 17:46 -0700 schrieb Frank Rowand:

quoted

On 06/06/17 12:22, Stefani Seibold wrote:

quoted

Hi Frank,

On 06.06.2017, 00:20 -0700 Frank Rowand wrote::

quoted

On 06/05/17 05:59, Stefani Seibold wrote:

quoted

From: Stefani Seibold <redacted>

This patch enables external references for symbols which
are
not
exported by the current device tree. For example

// RASPI example (only for testing)
/dts-v1/;
/plugin/;

/ {
    compatible = "brcm,bcm2835", "brcm,bcm2708",
"brcm,bcm2709";

    fragment@0 {
        target-path = "/soc/i2s@7e203000";
        __overlay__ {
            #address-cells = <0x00000001>;
            #size-cells = <0x00000001>;
            test = "test";
            timer = <&timer>;
        };
    };

    __external_symbols__ {
        timer = "/soc/timer@7e003000";
    };
};

My hope is that the dtc compiler will stop supporting 
specification of the __symbols__ node in dts source, and only
generate it automatically in the dtb. That change to dtc
would
not allow any node name specified in a dts to begin with an
underscore.  Thus node __external_symbols__ would not be 
allowed.

The name is not so important to me, only the solution.

quoted

quoted

In case of the RASPI device tree this could be simple fixed
by modifing the device tree source, but when the device
tree
is provided by a closed source BIOS this kind of missing
symbol could not be fixed.

Is there a real example of this issue, or is this a
theoretical concern? If this is a real example, we should be
discouraging such behavior.

Yes, I have a BIOS on some ARM64 servers which provides broken 
device tree. It also lacks some devices in this tree which
needs 
references to other devices which lacks a phandle.

I want to make sure I understand the use case, or cases.

My interpretation of the patch email is that there is an ARM64 server
with a device tree that is missing a single symbol in the __symbols__
node.  I could also read that email as saying that there is no
__symbols__ node in the device tree.  Which is the actual case?

quoted

quoted

Jon Masters is pushing a message that if the firmware on your
arm64 server is broken, then insist that the vendor fix it.

Well with this kind of arguments you can also sit and wait until a 
vendor decided to provide a kernel driver for an device. As we
know 
this would in 99% percent of the cases never happened.

So that is not a solution if the vendor does not do it. Fu....
closed source is as it is. So i need a solution and i think other
developer to.

quoted

I think he was talking about ACPI, but the same message should
also
apply to device tree.

Jon, are you also interested in broken device trees?

quoted

I talk about a device tree, not ACPI!

quoted

If you are having trouble getting your vendor to fix it, ask Jon
if he is willing to help apply pressure.

My vendor will not fix it. Believe me.

If you are not going to try to get your vendor to fix it (or have
already tried and failed), can you tell us the name and model of the
computer so that other people can be aware of the broken device tree?

Due an NDA i am not allowed to give you this kind of information.

quoted

And can you describe the specific way the device tree is broken?

In this case the device tree is broken due a missing device node which
must refer to the network interface(s) and interrupt controller.

Since there is no symbols for the network interfaces and there are also
missing a phandle, this information must be generated on the fly.

And that is what my patch do: provide an path to the node (symbol) and
generate the phandle on the fly.

quoted

quoted

And again: We need a solution when a closed source BIOS or
bootloader is not fixed by the vendor, which is also a case for
devices like single board computer, settop boxes, tv, smartphones
and
tablets.

The patch does not fix the general case of a broken device tree.  It
only potentially addresses a missing symbol from the base device tree
that is needed for an overlay.  The general case is an entirely
different discussion.

IMHO i think this is the last puzzle piece for an general solution. All
other things are there: Overlays and Symbols.

With this patch i can overwrite and modify any part of a device tree.

But here is another solution: decompile the broken device tree

quoted

and apply the fix to the decompiled tree.  This can be a bit
challenging since the original labels will be missing in the
decompiled tree (though you will be able to get them from the
__overlays__ node if that exists).

You are wrong. This will not work as a general solution due some
important information are provided by the BIOS, like the MAC addresses
and the memory layout and the NUMA nodes and the IDs for RoCE and so
one. 

In this case i need to decompile a lot of machines and patch them. And
every time the memory is changed i must do this again. Thats is not
feasible solution and very error prone.

I am guessing that every one of you machines uses the same device tree
and that the bootloader modifies the per machine information, such
as MAC addresses.  Why can't you just replace the bad device tree
with a good device tree?

quoted

Lack of me providing another solution is not a reason to apply a
patch that I disagree with.

The patch works exact the same way as the current __symbols__ node. It
is very tiny and easy to review and has absolut no side effects.

If you have no technical objections you should apply it until you have
a better solution.

My first reply provided a technical reason.  A leading underscore in a
node name is not valid device tree syntax.  It is currently allowed to
handle hand coded overlays, but patches have been submitted to the
device tree compiler to remove this temporary hack.  Hand coded
overlay nodes are overly complex and more likely to contain errors
than an auto-generated node.  If hand coded overlay nodes were
allowed then additional error checking would need to be added to
the kernel.

Assuming the compiler patches are accepted (which I believe is the
correct direction) the proposed special node name of __external_symbols__
will result in a compile error.

-Frank

- Stefani

--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html