Message-ID: <abb77afe-c285-46ba-88ac-08574bd67712@kernel.org>
Date: Thu, 13 Nov 2025 18:43:46 +0100
From: Krzysztof Kozlowski <krzk@...nel.org>
To: Doug Anderson <dianders@...omium.org>
Cc: Rob Herring <robh@...nel.org>, Krzysztof Kozlowski <krzk+dt@...nel.org>,
 Conor Dooley <conor+dt@...nel.org>, Peter Griffin
 <peter.griffin@...aro.org>, André Draszik
 <andre.draszik@...aro.org>, Tudor Ambarus <tudor.ambarus@...aro.org>,
 linux-samsung-soc@...r.kernel.org, Roy Luo <royluo@...gle.com>,
 devicetree@...r.kernel.org, linux-arm-kernel@...ts.infradead.org,
 Chen-Yu Tsai <wenst@...omium.org>, Julius Werner <jwerner@...omium.org>,
 William McVicker <willmcvicker@...gle.com>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/4] dt-bindings: arm: google: Add bindings for
 frankel/blazer/mustang

On 13/11/2025 18:16, Doug Anderson wrote:
> Hi,
> 
> On Thu, Nov 13, 2025 at 8:34 AM Krzysztof Kozlowski <krzk@...nel.org> wrote:
>>
>> On 13/11/2025 17:23, Doug Anderson wrote:
>>> Hi,
>>>
>>> On Wed, Nov 12, 2025 at 11:23 PM Krzysztof Kozlowski <krzk@...nel.org> wrote:
>>>>
>>>>>>> +      # Google Tensor G5 AKA lga (laguna) SoC and boards
>>>>>>> +      - description: Tensor G5 SoC (laguna)
>>>>>>> +        items:
>>>>>>> +          - enum:
>>>>>>> +              - google,soc-id-0005-rev-00  # A0
>>>>>>> +              - google,soc-id-0005-rev-10  # B0
>>>>>>
>>>>>> SoCs cannot be final compatibles.
>>>>>
>>>>> Right. I talked about this at length "after the cut" in my patch. See
>>>>> above. I wish to relitigate this policy and wish to know more details
>>>>> about where it is documented, the reasons for decision, and where the
>>>>> boundary exactly lies between something that's allowed to be a final
>>>>> compatible and something that's not. I made several arguments above
>>>>> for why I think the SoC should be allowed as a final compatible, so it
>>>>
>>>> Because this represents a actual device users run. It is electronically,
>>>> physically impossible to run the SoC alone.
>>>
>>> I'm not convinced that this definition is as clear as you're making it
>>> out to be. It's physically impossible to run many "boards" alone.
>>>
>>> Want to boot up a Raspberry Pi? Provide it with power. Hook up a
>>> display to it. Hook up a keyboard to it. Plug in an Ethernet cable.
>>> Plug an SD card in it. Without those things it doesn't run.
>>
>> But I can plug them...
> 
> I can plug my lga SoC into my dev board too. My dev board literally

So this will be dev board... If you plug this into dev board, then I
claim that your compatible must be specific for this system thus
describe that product, which is SoC and devboard, not SoC alone.


> has a place for me to drop in the SoC. I unscrew the socket connector,
> carefully make sure that none of the balls of the SoC have dust
> particles on them (and we have instructions for cleaning the SoC),
> then drop the SoC into the socket (ideally using a vacuum pen tool). I
> then screw the top back together which uses compression to attach the
> balls on the SoC.
> 
> Yes, this is only true on dev boards and not phones, but we want to be
> able to support dev boards too and it would be silly to have a
> different split between DTB and overlays for dev boards and phones
> that are based on the same architecture
You still have dev board thus need dev board compatible.

> 
> 
>>> Want to boot up a lga-B0 SoC? Hook up power to the power pins. Connect
>>> a MIPI panel to the MIPI pins. Connect a UFS chip to the UFS pins.
>>> Without those things it doesn't run.
>>
>> These I cannot plug, it's impossible for me.
>>
>> My clumsy fingers are too big for these pins.
> 
> As per above, sockets do exist. They are pluggable. I have confidence
> that even with clumsy fingers you could drop the SoC into the slot and
> screw the connector down.
> 
> I'm still not totally convinced that it should require the existence
> of a socket to justify this but, yes, they do exist.
> 
> 
>> And following your logic, we should have the compatible for the
>> transistors, because that's basically what SoC is made of.
> 
> My logic (stated later in my email) is that the minimum requirement
> for a compatible should be something with a CPU able to execute
> instructions. A single transistor can't do that. If you want to
> combine a bunch of transistors together to make a CPU then absolutely
> you should be able to have a DTB representing this CPU.

Linux cannot execute instructions on this SoC alone. You need dev board.

My clumsy fingers were just example, but you simply cannot run any
software on this SoC without something else.

Virtual models and virtual platforms could be run without that
"something else" and they are allowed to have just the compatible. See
all the FVPs and QEMU platforms.


 >
>>> Yes, the complexity of just "hooking up" the components on an SoC is
>>> an order of magnitude harder than a Raspberry Pi, but it's still just
>>> hooking it up to external components. In both cases, we are modeling
>>> the core "brains" (the part that contains the processor) as the DTB
>>> and everything else just "hooks up" to interfaces.
>>
>> You mix the topics, so I don't follow. I speak here about bindings - you
>> cannot have the that compatible alone, because it is incomplete, just
>> like compatible for "transistor" is not correct in that context. You
>> speak what could or could be DTB, different topic.
> 
> A "SoC" is "complete". It has a processor that can run instructions.

Then show me executing any piece of software, which is the consumer of
the bindings, and runs on the SoC without the rest of the hardware system.

> 
> 
>>> If I had to make a definition for what the base DTB should be it
>>> should be the component with the boot CPU. _Why_ is that the wrong
>>> definition?
>>>
>>>
>>>> There are few - one or two - exceptions for the SoMs, but never for SoC.
>>>
>>> OK, but the big question: _WHY???_
>>>
>>> Where does it say that a DTB has to be something that can run "alone"
>>
>> We don't discuss DTB here, but the top-level compatibles.
>>
>> Why? Because DT spec says so.
>>
>> "Specifies a list of platform architectures with which this platform is
>> compatible. "
>>
>> And when you combine it with the standard definition of the
>> "compatible", it is not *a* "platform architecture" but *list* of
>> platform architectures describing this device as a whole.
> 
> I still don't understand why a SoC doesn't qualify for your
> definition. Even if it did, there is _no benefit_ from excluding a SoC
> from this definition. I'm trying to figure out what the benefit is for
> holding to this stance.

We write bindings for users. There is no user of SoC and such compatible
alone. There is no such real use case.

Your use case is broken bootloader which pretends there is no board, but
there is a board, someone just did not want to admit it.

We don't develop for such use cases.

> 
> In any case, maybe we can approach this a different way that I alluded
> to in one of my other posts. Can we just call the SoC thing something
> different and make everyone happy?
> 
> 1. Rename the SoC file to lga-b0.dtf (device tree fragment) and
> _REMOVE_ the top-level compatible. Problem solved--we're not adding a
> top-level compatible.
> 
> 2. Add a special node at the top level of the "dtf" file describing it
> (so someone could figure it's useful for). Like:
> 
> fragment-info {
>   compatible = "google,soc-id";
>   google,product-id = <0x5>;
>   google,major-rev = <0x1>;
>   google,minor-rev = <0x0>;
>   google,package-mode = <0x0>;
> };
> 
> 3. We can compile the "dtf" file using existing tools into a "dtfb".
> This looks just like a "dtb" but has no top-level compatible but
> instead has "fragment-info".
> 
> Now we're not violating any spec because we're not adding any
> top-level compatible.

Didn't you just describe an overlay or DTSI file?

Best regards,
Krzysztof

Powered by blists - more mailing lists