[Intel-gfx] [Intel-xe] [PATCH] drm/xe/display: Do not use i915 frontbuffer tracking implementation

Maarten Lankhorst maarten.lankhorst at linux.intel.com
Thu Mar 9 11:09:55 UTC 2023


On 2023-03-09 12:04, Hogander, Jouni wrote:
> On Mon, 2023-03-06 at 22:58 +0200, Ville Syrjälä wrote:
>> On Mon, Mar 06, 2023 at 09:23:50PM +0100, Maarten Lankhorst wrote:
>>> Hey,
>>>
>>> On 2023-03-06 16:23, Souza, Jose wrote:
>>>> On Mon, 2023-03-06 at 15:16 +0100, Maarten Lankhorst wrote:
>>>>> As a fallback if we decide not to merge the frontbuffer
>>>>> tracking, allow
>>>>> i915 to keep its own implementation, and do the right thing in
>>>>> Xe.
>>>>>
>>>>> The frontbuffer tracking for Xe is still done per-fb, while
>>>>> i915 can
>>>>> keep doing the weird intel_frontbuffer + i915_active thing
>>>>> without
>>>>> blocking Xe.
>>>> Please also disable PSR and FBC with this or at least add a way
>>>> for users to disable those features.
>>>> Without frontbuffer tracker those two features will break in some
>>>> cases.
>>> FBC and PSR work completely as expected. I don't remove frontbuffer
>>> tracking; I only remove the GEM parts.
>>>
>>> Explicit invalidation using pageflip or CPU rendering + DirtyFB
>>> continue
>>> to work, as I validated on my laptop with FBC.
>> Neither of which are relevant to the removal of the gem hooks.
>>
>> Like I already said ~10 times in the last meeting, we need a proper
>> testcase. Here's a rough idea what it should do:
>>
>> prepare a batch with
>> 1. spinner
>> 2. something that clobbers the fb
>>
>> Then
>> 1. grab reference crc
>> 2. execbuffer
>> 3. dirtyfb
>> 4. wait long enough for fbc to recompress
>> 5. terminate spinner
>> 6. gem_sync
>> 7. grab crc and compare with reference
>>
>> No idea what the current status of PSR+CRC is, so not sure
>> whether we can actually test PSR or not.
>>
> CRC calculation doesn't work with PSR currently. PSR is disabled if CRC
> capture is requested.
>
> Are we supposed to support frontbuffer rendering using GPU?

No other driver does that. It's fine if DirtyFB hangs instead until the 
job it waits on completes.

~Maarten




More information about the dri-devel mailing list