[PATCH RFC 01/10] dt-bindings: gpu: Add PowerVR Series5 SGX GPUs

Wed Dec 6 21:43:27 UTC 2023

> Am 06.12.2023 um 17:02 schrieb Conor Dooley <conor at kernel.org>:
> 
> On Tue, Dec 05, 2023 at 07:04:05PM +0100, H. Nikolaus Schaller wrote:
>>> Am 05.12.2023 um 18:33 schrieb Andrew Davis <afd at ti.com>:
>>> 
>>> On 12/5/23 2:17 AM, H. Nikolaus Schaller wrote:
>>>>> +          - enum:
>>>>> +              - ti,omap3430-gpu # Rev 121
>>>>> +              - ti,omap3630-gpu # Rev 125
>>>> Is the "Rev 121" and "Rev 125" a property of the SoC integration (clock/reset/power
>>>> hookup etc.) or of the integrated SGX core?
>>> 
>>> The Rev is a property of the SGX core, not the SoC integration.
>> 
>> Then, it should belong there and not be a comment of the ti,omap*-gpu record.
>> In this way it does not seem to be a proper hardware description.
>> 
>> BTW: there are examples where the revision is part of the compatible string, even
>> if the (Linux) driver makes no use of it:
>> 
>> drivers/net/ethernet/xilinx/xilinx_emaclite.c
> 
> AFAICT these Xilinx devices that put the revisions in the compatible are
> a different case - they're "soft" IP intended for use in the fabric of
> an FPGA, and assigning a device specific compatible there does not make
> sense.
> 
> In this case it appears that the revision is completely known once you
> see "ti,omap3630-gpu", so encoding the extra "121" into the compatible
> string is not required.

Well, I would not put my hand in the fire for this assumption.

And I am a friend of explicitly stating what is instead ot encoding indirectly.

> 
>> 
>>> But it seems that
>>> compatible string is being used to define both (as we see being debated in the other
>>> thread on this series).
>>> 
>>>> In my understanding the Revs are different variants of the SGX core (errata
>>>> fixes, instruction set, pipeline size etc.). And therefore the current driver code
>>>> has to be configured by some macros to handle such cases.
>>>> So the Rev should IMHO be part of the next line:
>>>>> +          - const: img,powervr-sgx530
>>>> +          - enum:
>>>> +              - img,powervr-sgx530-121
>>>> +              - img,powervr-sgx530-125
>>>> We have a similar definition in the openpvrsgx code.
>>>> Example: compatible = "ti,omap3-sgx530-121", "img,sgx530-121", "img,sgx530";
>>>> (I don't mind about the powervr- prefix).
>>>> This would allow a generic and universal sgx driver (loaded through just matching
>>>> "img,sgx530") to handle the errata and revision specifics at runtime based on the
>>>> compatible entry ("img,sgx530-121") and know about SoC integration ("ti,omap3-sgx530-121").
> 
> The "raw" sgx530 compatible does not seem helpful if the sgx530-121 or
> sgx530-125 compatibles are also required to be present for the driver to
> actually function.

Indeed. This seems to be redundant (but may need some pattern processing).

> The revision specific compatibles I would not object
> to, but everything in here has different revisions anyway - does the
> same revision actually appear in multiple devices? If it doesn't then I
> don't see any value in the suffixed compatibles either.

Well, we don't know.

So far only a subset of SoC with the SGX GPU core variants has been considered
(mainly because lack of user space code and device owners).

Maybe someone with insider knowledge can give a hint if the SGX version numbers
were assigned uniquely for each SoC integration project.

> 
>>>> And user-space can be made to load the right firmware variant based on "img,sgx530-121"
>>>> I don't know if there is some register which allows to discover the revision long
>>>> before the SGX subsystem is initialized and the firmware is up and running.
>>>> What I know is that it is possible to read out the revision after starting the firmware
>>>> but it may just echo the version number of the firmware binary provided from user-space.
>>> 
>>> We should be able to read out the revision (register EUR_CR_CORE_REVISION), the problem is
>>> today the driver is built for a given revision at compile time.
>> 
>> Yes, that is something we had planned to get rid of for a long time by using different compatible
>> strings and some variant specific struct like many others drivers are doing it.
>> But it was a to big task so nobody did start with it.
>> 
>>> That is a software issue,
>>> not something that we need to encode in DT. While the core ID (SGX5xx) can be also detected
>>> (EUR_CR_CORE_ID), the location of that register changes, and so it does need encoded in
>>> DT compatible.
>> 
>> Ok, I didn't know about such registers as there is not much public information available.
>> Fair enough, there are some error reports about in different forums.
>> 
>> On the other hand we then must read out this register in more or less early initialization
>> stages. Even if we know this information to be static and it could be as simple as a list
>> of compatible strings in the driver.
>> 
>>> The string "ti,omap3430-gpu" tells us the revision if we cannot detect it (as in the current
>>> driver), and the SoC integration is generic anyway (just a reg and interrupt).
>> 
>> It of course tells, but may need a translation table that needs to be maintained in a
>> different format. Basically the same what the comments show in a non-machine readable
>> format.
>> 
>> I just wonder why the specific version can or should not become simply part of the DTS
>> and needs this indirection.
>> 
>> Basically it is a matter of openness for future (driver) development and why it needs
>> careful decisions.
>> 
>> So in other words: I would prefer to see the comments about versions encoded in the device
>> tree binary to make it machine readable.
> 
> It's already machine readable if it is invariant on an SoC given the
> patch had SoC-specific compatibles.

But needs a translation table to get to the revision number.

I have not yet brought into discussion that there are different firmwares for sgx530-121,
sgx530-125, sgx544-116 etc. And user-space code may also depend on to be able to chose
the right one if multiple firmware packages are installed. Currently this is not the case
but would be a major benfit for OS packages.

To automate this we need a mechanism to scan the device tree for a compatible string that
tells which firmware variant to load.

But why force this to depend on the SoC compatible if it only depends indirectly?

By the way, there is a tested and working driver using the scheme with the sub-versions:

https://github.com/openpvrsgx-devgroup/linux_openpvrsgx/blob/11cc7876ba39b6172d19ee0bf0a872c1d3d745e1/drivers/gpu/drm/pvrsgx/pvr-drv.c#L306

On the other hand As far as I see this will can of course be adapted to the newly
proposed scheme.

But it still seems a bit twisted to me. Maybe because we for example define LCD panel
compatibles not by the compatible of the device they are installed in.

BR,
Nikolaus