Re: drm/amdgpu: replace readq/writeq with atomic64 operations

["Koenig, Christian" <Christian.Koenig@amd.com>]

Am 08.08.19 um 21:33 schrieb Alex Deucher:
> On Thu, Aug 8, 2019 at 3:31 PM Guenter Roeck <linux@roeck-us.net> wrote:
>> On Wed, Aug 07, 2019 at 10:56:40AM +0800, Tao Zhou wrote:
>>> readq/writeq are not supported on all architectures
>>>
>>> Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
>>> Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
>> Regarding the claim that this would work for 32-bit x86 builds:
> I wasn't talking about readq/writeq, I was talking about the atomic64
> interfaces.

On quite a bunch of architectures atomic64 operations are actually 
implemented with spinlocks or other architecture depended special handling.

So the approach of casting an iomem pointer to an atomic64 and then hope 
for the best is actually completely nonsense.

If the hardware really needs a single 64bit write for doorbells or 
registers, then we absolutely need to limit the driver to 64bit 
architectures.

If the hardware doesn't need 64bit writes we should actually always use 
two 32bit writes to not run into random and hard to debug problems 
because of this.

Christian.

>
> Alex
>
>> make ARCH=i386 allmodconfig
>> make ARCH=i386 drivers/gpu/drm/amd/amdgpu/amdgpu_device.o
>>
>> results in:
>>
>>    ...
>>    CC [M]  drivers/gpu/drm/amd/amdgpu/amdgpu_device.o
>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c: In function ‘amdgpu_mm_rreg64’:
>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:279:9: error: implicit declaration of function ‘readq’;
>>
>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c: In function ‘amdgpu_mm_wreg64’:
>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:298:3: error: implicit declaration of function ‘writeq’
>>
>> This is with next-20190808.
>>
>> Guenter
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx