[PATCH 2/4] drm/xe: store bind time pat index to xe_bo

[Matthew Auld]

On 23/01/2024 08:05, Ville Syrjälä wrote:
> On Fri, Jan 19, 2024 at 03:45:22PM +0000, Matthew Auld wrote:
>> On 18/01/2024 15:27, Juha-Pekka Heikkila wrote:
>>> Store pat index from xe_vma to xe_bo
>>>
>>> Signed-off-by: Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com>
>>> ---
>>>    drivers/gpu/drm/xe/xe_pt.c | 4 ++++
>>>    1 file changed, 4 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
>>> index de1030a47588..4b76db698878 100644
>>> --- a/drivers/gpu/drm/xe/xe_pt.c
>>> +++ b/drivers/gpu/drm/xe/xe_pt.c
>>> @@ -1252,6 +1252,10 @@ __xe_pt_bind_vma(struct xe_tile *tile, struct xe_vma *vma, struct xe_exec_queue
>>>    		return ERR_PTR(-ENOMEM);
>>>    	}
>>>    
>>> +	if (xe_vma_bo(vma)) {
>>> +		xe_vma_bo(vma)->pat_index = vma->pat_index;
>>
>> Multiple mappings will trash this I think. Is that OK for your usecase?
>> It can be useful to map the same resource as compressed and uncompressed
>> to facilitate in-place decompression/compression.
> 
> I thought the pat_index is set for the entire bo? The
> cache_level->pat_index stuff doesn't really work otherwise
> I don't think (assuming it works at all).

AFAIK it is mostly like that in i915 because it doesn't have a vm_bind 
interface. With Xe we have vm_bind. The pat_index is a property of the 
ppGTT binding and therefore vma. There seem to be legitimate reasons to 
map the same resource with different pat_index, like with 
compressed/uncompressed. See BSpec: 58797 "double map (alias) surfaces".

> 
> So dunno why this is doing anything using vmas. I think
> what we probably need is to check/set the bo pat_index
> at fb create time, and lock it into place (if there's
> some mechanism by which a random userspace client could
> change it after the fact, and thus screw up everything).

Maybe we can seal the pat_index on first bind or something if the BO 
underneath is marked with XE_BO_SCANOUT?

> 
>>
>> Also would be good to be clear about what happens if the KMD doesn't do
>> anything to prevent compression with non-tile4? Is it just a bit of
>> display corruption or something much worse that we need to prevent? Is
>> this just a best effort check to help userspace? Otherwise it is hard to
>> evaluate how solid we need to be here in our checking to prevent this
>> scenario. For example how is binding vs display races handled? What
>> happens if the bind appears after the display check?
>>
>>> +	}
>>> +
>>>    	fence = xe_migrate_update_pgtables(tile->migrate,
>>>    					   vm, xe_vma_bo(vma), q,
>>>    					   entries, num_entries,
>