[RFC v2 5/8] drm/fence: add in-fences support

[Ville Syrjälä]

On Thu, Apr 28, 2016 at 07:56:19PM +0300, Ville Syrjälä wrote:
> On Thu, Apr 28, 2016 at 11:36:44AM -0300, Gustavo Padovan wrote:
> > 2016-04-27 Daniel Stone <daniel at fooishbar.org>:
> > 
> > > Hi,
> > > 
> > > On 26 April 2016 at 21:48, Greg Hackmann <ghackmann at google.com> wrote:
> > > > On 04/26/2016 01:05 PM, Daniel Vetter wrote:
> > > >> On Tue, Apr 26, 2016 at 09:55:06PM +0300, Ville Syrjälä wrote:
> > > >>> What are they doing that can't stuff the fences into an array
> > > >>> instead of props?
> > > >>
> > > >> The hw composer interface is one in-fence per plane. That's really the
> > > >> major reason why the kernel interface is built to match. And I really
> > > >> don't think we should diverge just because we have a slight different
> > > >> color preference ;-)
> > > >
> > > > The relationship between layers and fences is only fuzzy and indirect
> > > > though.  The relationship is really between the buffer you're displaying on
> > > > that layer, and the fence representing the work done to render into that
> > > > buffer.  SurfaceFlinger just happens to bundle them together inside the same
> > > > struct hwc_layer_1 as an API convenience.
> > > 
> > > Right, and when using implicit fencing, this comes as a plane
> > > property, by virtue of plane -> fb -> buffer -> fence.
> > > 
> > > > Which is kind of splitting hairs as long as you have a 1-to-1 relationship
> > > > between layers and DRM planes.  But that's not always the case.
> > > 
> > > Can you please elaborate?
> > > 
> > > > A (per-CRTC?) array of fences would be more flexible.  And even in the cases
> > > > where you could make a 1-to-1 mapping between planes and fences, it's not
> > > > that much more work for userspace to assemble those fences into an array
> > > > anyway.
> > > 
> > > As Ville says, I don't want to go down the path of scheduling CRTC
> > > updates separately, because that breaks MST pretty badly. If you don't
> > > want your updates to display atomically, then don't schedule them
> > > atomically ... ? That's the only reason I can see for making fencing
> > > per-CRTC, rather than just a pile of unassociated fences appended to
> > > the request. Per-CRTC fences also forces userspace to merge fences
> > > before submission when using multiple planes per CRTC, which is pretty
> > > punitive.
> > > 
> > > I think having it semantically attached to the plane is a little bit
> > > nicer for tracing (why was this request delayed? -> a fence -> which
> > > buffer was that fence for?) at a glance. Also the 'pile of appended
> > > fences' model is a bit awkward for more generic userspace, which
> > > creates a libdrm request and builds it (add a plane, try it out, wind
> > > back) incrementally. Using properties makes that really easy, but
> > > without properties, we'd have to add separate codepaths - and thus
> > > separate ABI, which complicates distribution - to libdrm to account
> > > for a separate plane array which shares a cursor with the properties.
> > > So for that reason if none other, I'd really prefer not to go down
> > > that route.
> > 
> > I also agree to have it as FENCE_FD prop on the plane. Summarizing the
> > arguments on this thread, they are:
> 
> Your "summary" forgot to include any counter arguments.
> 
> > 
> >  - implicit fences also needs one fence per plane/fb, so it will be good to     
> >    match with that.                                                             
> 
> We would actually need a fence per object rather than per fb.

I guess you could overcome this by automagically creating a merged fence
for a multi-obj fb?

-- 
Ville Syrjälä
Intel OTC