Re: [PATCH v1 2/2] drivers/gpu/drm: don't select DMA_CMA or CMA from aspeed or etnaviv

[David Hildenbrand <david@redhat.com>]

On 08.04.21 22:29, Arnd Bergmann wrote:
> On Thu, Apr 8, 2021 at 6:45 PM David Hildenbrand <david@redhat.com> wrote:
>> On 08.04.21 14:49, Linus Walleij wrote:
>>> On Thu, Apr 8, 2021 at 2:01 PM David Hildenbrand <david@redhat.com> wrote:
>>>
>>>>> This is something you could do using a hidden helper symbol like
>>>>>
>>>>> config DRMA_ASPEED_GFX
>>>>>           bool "Aspeed display driver"
>>>>>           select DRM_WANT_CMA
>>>>>
>>>>> config DRM_WANT_CMA
>>>>>           bool
>>>>>           help
>>>>>              Select this from any driver that benefits from CMA being enabled
>>>>>
>>>>> config DMA_CMA
>>>>>           bool "Use CMA helpers for DRM"
>>>>>           default DRM_WANT_CMA
>>>>>
>>>>>             Arnd
>>>>>
>>>>
>>>> That's precisely what I had first, with an additional "WANT_CMA" --  but
>>>> looking at the number of such existing options (I was able to spot 1 !)
>>>
>>> If you do this it probably makes sense to fix a few other drivers
>>> Kconfig in the process. It's not just a problem with your driver.
>>> "my" drivers:
>>>
>>
>> :) I actually wanted to convert them to "depends on DMA_CMA" but ran
>> into recursive dependencies ...
>>
>>> drivers/gpu/drm/mcde/Kconfig
>>> drivers/gpu/drm/pl111/Kconfig
>>> drivers/gpu/drm/tve200/Kconfig
> 
> Right, this is the main problem caused by using 'select' to
> force-enable symbols that other drivers depend on.
> 
> Usually, the answer is to be consistent about the use of 'select'
> and 'depends on', using the former only to enable symbols that
> are hidden, while using 'depends on' for anything that is an
> actual build time dependency.
> 
>> I was assuming these are "real" dependencies. Will it also work without
>> DMA_CMA?
> 
> I think in this case, it is fairly likely to work without DMA_CMA when the
> probe function gets called during a fresh boot, but fairly likely to fail if
> it gets called after the system has run for long enough to fragment the
> free memory.
> 
> The point of DMA_CMA is to make it work reliably.

Right, and even at runtime there is no guarantee that DMA_CMA will do 
anything -- especially if we don't reserve a CMA region (e.g., "cma=X").

So this really sounds like a

"desires DMA_CMA"

and achieving that via an additional symbol or via "default y if ..." 
for DMA_CMA sounds reasonable.

Thanks!

-- 
Thanks,

David / dhildenb