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

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Thu, Mar 09, 2023 at 12:09:55PM +0100, Maarten Lankhorst wrote:
> 
> 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.

Every driver does that when you run X w/o a compositor. Assuming
there is an actual GPU in there.

> It's fine if DirtyFB hangs instead until the 
> job it waits on completes.

No one tried to make it just wait for the fence(s) w/o doing
a full blown atomic commit. It might work, but might also
still suck too much. I guess depends on how overloaded the GPU
is.

What we could do is do a frontbuffer invalidate on dirtyfb
invocation, and then once the fence(s) signal we do a frontbuffer
flush. That would most closely match the gem hook behaviour, except
the invalidate comes in a bit later. The alternative would be to
skip the invalidate, which should still guarantee correctness in
the end, just with possibly jankier interactivity.

-- 
Ville Syrjälä
Intel