Re: [Intel-gfx] [PATCH 2/2] drm/i915/dmabuf: drop the flush on discrete

[Matthew Auld <matthew.auld@intel.com>]

On 22/10/2021 10:26, Thomas Hellström wrote:
> Hi, Matt
> 
> On 10/21/21 14:53, Matthew Auld wrote:
>> We were overzealous here; even though discrete is non-LLC, it should
>> still be always coherent.
>>
>> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
>> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>> ---
>>   drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c 
>> b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c
>> index a45d0ec2c5b6..848e81368043 100644
>> --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c
>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c
>> @@ -251,7 +251,8 @@ static int i915_gem_object_get_pages_dmabuf(struct 
>> drm_i915_gem_object *obj)
>>           return PTR_ERR(pages);
>>       /* XXX: consider doing a vmap flush or something */
>> -    if (!HAS_LLC(i915) || i915_gem_object_can_bypass_llc(obj))
>> +    if ((!HAS_LLC(i915) && !IS_DGFX(i915)) ||
> 
> Q: I notice that DG1 at least has HAS_SNOOP. Would it be incorrect to 
> use that in this case?

AFAIK DG1 is special in that CACHE_NONE will still snoop transactions, 
which is not the case for other HAS_SNOOP platforms. AFAIK that is part 
of the reason why we also just force CACHE_LLC everywhere on DG1.

Could maybe do s/IS_DGFX/IS_DG1/ here? In case that changes on other 
discrete platforms. And then add a comment.

> 
> /Thomas
> 
> 
> 
>> +        i915_gem_object_can_bypass_llc(obj))
>>           wbinvd_on_all_cpus();
>>       sg_page_sizes = i915_sg_dma_sizes(pages->sgl);