[PATCH v3 0/5] Reset driver for IMX8MQ VPU hardware block

[Benjamin Gaignard]

Le 03/03/2021 à 17:25, Philipp Zabel a écrit :
> On Wed, 2021-03-03 at 16:20 +0100, Benjamin Gaignard wrote:
>> Le 03/03/2021 à 15:17, Philipp Zabel a écrit :
>>> Hi Benjamin,
>>>
>>> On Mon, 2021-03-01 at 16:17 +0100, Benjamin Gaignard wrote:
>>>> The two VPUs inside IMX8MQ share the same control block which can be see
>>>> as a reset hardware block.
>>> This isn't a reset controller though. The control block also contains
>>> clock gates of some sort and a filter register for the featureset fuses.
>>> Those shouldn't be manipulated via the reset API.
>> They are all part of the control block and of the reset process for this
>> hardware that why I put them here. I guess it is border line :-)
> I'm pushing back to keep the reset control framework focused on
> controlling reset lines. Every side effect (such as the asymmetric clock
> ungating) in a random driver makes it harder to reason about behaviour
> at the API level, and to review patches for hardware I am not familiar
> with.
>
>>>> In order to be able to add the second VPU (for HECV decoding) it will be
>>>> more handy if the both VPU drivers instance don't have to share the
>>>> control block registers. This lead to implement it as an independ reset
>>>> driver and to change the VPU driver to use it.
>>> Why not switch to a syscon regmap for the control block? That should
>>> also allow to keep backwards compatibility with the old binding with
>>> minimal effort.
>> I will give a try in this direction.
> Thank you.
>
>>>> Please note that this series break the compatibility between the DTB and
>>>> kernel. This break is limited to IMX8MQ SoC and is done when the driver
>>>> is still in staging directory.
>>> I know in this case we are pretty sure there are no users of this
>>> binding except for a staging driver, but it would still be nice to keep
>>> support for the deprecated binding, to avoid the requirement of updating
>>> kernel and DT in lock-step.
>> If I want to use a syscon (or a reset) the driver must not ioremap the "ctrl"
>> registers. It means that "ctrl" has to be removed from the driver requested
>> reg-names (imx8mq_reg_names[]). Doing that break the kernel/DT compatibility.
>> Somehow syscon and "ctrl" are exclusive.
> The way the driver is set up currently, yes. You could add a bit of
> platform specific probe code, though, that would set up the regmap
> either by calling
> 	syscon_regmap_lookup_by_phandle();
> for the new binding, or, if the phandle is not available, fall back to
> 	platform_get_resource_byname(..., "ctrl");
> 	devm_ioremap_resource();
> 	devm_regmap_init_mmio();
> for the old binding.
> The actual codec .reset and variant .runtime_resume ops could be
> identical then.

I made it works with syscon and your proposal.
The next version of the patches will be without reset and won't break
DT compatibility.

Thanks for your help,
Benjamin

>
> regards
> Philipp
>