On Tue, Oct 19, 2021 at 03:02:26PM +0200, Christian König wrote:
> Am 13.10.21 um 16:23 schrieb Daniel Vetter:
> > On Tue, Oct 05, 2021 at 01:37:38PM +0200, Christian König wrote:
> > > Makes the handling a bit more complex, but avoids the use of
> > > dma_resv_get_excl_unlocked().
> > >
> > > v2: improve coding and documentation
> > >
> > > Signed-off-by: Christian König <christian.koenig@xxxxxxx>
> > > ---
> > > drivers/gpu/drm/drm_gem_atomic_helper.c | 13 +++++++++++--
> > > 1 file changed, 11 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/drm_gem_atomic_helper.c b/drivers/gpu/drm/drm_gem_atomic_helper.c
> > > index e570398abd78..8534f78d4d6d 100644
> > > --- a/drivers/gpu/drm/drm_gem_atomic_helper.c
> > > +++ b/drivers/gpu/drm/drm_gem_atomic_helper.c
> > > @@ -143,6 +143,7 @@
> > > */
> > > int drm_gem_plane_helper_prepare_fb(struct drm_plane *plane, struct drm_plane_state *state)
> > > {
> > > + struct dma_resv_iter cursor;
> > > struct drm_gem_object *obj;
> > > struct dma_fence *fence;
> > > @@ -150,9 +151,17 @@ int drm_gem_plane_helper_prepare_fb(struct drm_plane *plane, struct drm_plane_st
> > > return 0;
> > > obj = drm_gem_fb_get_obj(state->fb, 0);
> > > - fence = dma_resv_get_excl_unlocked(obj->resv);
> > > - drm_atomic_set_fence_for_plane(state, fence);
> > > + dma_resv_iter_begin(&cursor, obj->resv, false);
> > > + dma_resv_for_each_fence_unlocked(&cursor, fence) {
> > > + /* TODO: We only use the first write fence here and need to fix
> > > + * the drm_atomic_set_fence_for_plane() API to accept more than
> > > + * one. */
> > I'm confused, right now there is only one write fence. So no need to
> > iterate, and also no need to add a TODO. If/when we add more write fences
> > then I think this needs to be revisited, and ofc then we do need to update
> > the set_fence helpers to carry an entire array of fences.
>
> Well could be that I misunderstood you, but in your last explanation it
> sounded like the drm_atomic_set_fence_for_plane() function needs fixing
> anyway because a plane could have multiple BOs.
>
> So in my understanding what we need is a
> drm_atomic_add_dependency_for_plane() function which records that a certain
> fence needs to be signaled before a flip.
Yeah that's another issue, but in practice there's no libva which decodes
into planar yuv with different fences between the planes. So not a bug in
practice.
But this is entirely orthogonal to you picking up the wrong fence here if
there's not exclusive fence set:
- old code: Either pick the exclusive fence, or not fence if the exclusive
one is not set.
- new code: Pick the exclusive fence or the first shared fence
New behaviour is busted, because scanning out and reading from a buffer at
the same time (for the next frame, e.g. to copy over damaged areas or some
other tricks) is very much a supported thing. Atomic _only_ wants to look
at the exclusive fence slot, which mean "there is an implicitly synced
write to this buffers". Implicitly synced reads _must_ be ignored.
Now amdgpu doesn't have this distinction in its uapi, but many drivers do.
-Daniel
> Support for more than one write fence then comes totally naturally.
>
> Christian.
>
> > -Daniel
> >
> > > + dma_fence_get(fence);
> > > + break;
> > > + }
> > > + dma_resv_iter_end(&cursor);
> > > + drm_atomic_set_fence_for_plane(state, fence);
> > > return 0;
> > > }
> > > EXPORT_SYMBOL_GPL(drm_gem_plane_helper_prepare_fb);
> > > --
> > > 2.25.1
> > >
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
