[RFC PATCH] drm: disable WC optimization for cache coherent devices on non-x86

[Michel Dänzer]

On 2019-01-24 12:45 p.m., Ard Biesheuvel wrote:
> On Thu, 24 Jan 2019 at 12:37, Koenig, Christian
> <Christian.Koenig at amd.com> wrote:
>> Am 24.01.19 um 12:26 schrieb Ard Biesheuvel:
>>> On Thu, 24 Jan 2019 at 12:23, Koenig, Christian
>>> <Christian.Koenig at amd.com> wrote:
>>>> Am 24.01.19 um 10:59 schrieb Ard Biesheuvel:
>>>>> [SNIP]
>>>>> This is *exactly* my point the whole time.
>>>>>
>>>>> The current code has
>>>>>
>>>>> static inline bool drm_arch_can_wc_memory(void)
>>>>> {
>>>>> #if defined(CONFIG_PPC) && !defined(CONFIG_NOT_COHERENT_CACHE)
>>>>>      return false;
>>>>>
>>>>> which means the optimization is disabled *unless the system is
>>>>> non-cache coherent*
>>>>>
>>>>> So if you have reports that the optimization works on some PowerPC, it
>>>>> must be non-cache coherent PowerPC, because that is the only place
>>>>> where it is enabled in the first place.
>>>>>
>>>>>> The only problematic here actually seems to be ARM, so you should
>>>>>> probably just add an "#ifdef .._ARM return false;".
>>>>>>
>>>>> ARM/arm64 does not have a Kconfig symbol like
>>>>> CONFIG_NOT_COHERENT_CACHE, so we can only disable it everywhere. If
>>>>> there are non-coherent ARM systems that are currently working in the
>>>>> same way as those non-coherent PowerPC systems, we will break them by
>>>>> doing this.
>>>> Summing the things I've read so far for ARM up I actually think it
>>>> depends on a runtime configuration and not on compile time one.
>>>>
>>>> So the whole idea of providing the device to the drm_*_can_wc_memory()
>>>> function isn't so far fetched.
>>>>
>>> Thank you.
>>>
>>>> But for now I do prefer working and slightly slower system over broken
>>>> one, so I think we should just disable this on ARM for now.
>>>>
>>> Again, this is not about non-cache coherent being slower without the
>>> optimization, it is about non-cache coherent likely not working *at
>>> all* unless the optimization is enabled.
>>
>> As Michel tried to explain this CAN'T happen. The optimization is a
>> purely optional request from userspace.
>>
> 
> Right.
> 
> So in that case, we can assume that the following test
> 
> static inline bool drm_arch_can_wc_memory(void)
> {
> #if defined(CONFIG_PPC) && !defined(CONFIG_NOT_COHERENT_CACHE)
>     return false;
> 
> is bogus, and it was just unnecessary caution on the part of the
> author to disregard non-cache coherent devices.

This is driver-independent core code, meaning "non-snooped PCIe
transfers don't work on cache coherent PPC". It doesn't imply anything
about whether or not amdgpu or any other driver works on
non-cache-coherent PPC in general.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer