On Wed, May 22, 2024 at 05:52:56PM +0000, Klymenko, Anatoliy wrote:
> On Wednesday, May 22, 2024 8:32 AM, Laurent Pinchart wrote:
> > On Mon, May 20, 2024 at 08:22:31PM -0700, Anatoliy Klymenko wrote:
> > > Unconditionally enable the DPSUB layer in the corresponding atomic plane
> > > update callback. Setting the new display mode may require disabling and
> > > re-enabling the CRTC. This effectively resets DPSUB to the default state
> > > with all layers disabled.
> >
> > Ah, I went through the code and I see that. Oops.
> >
> > > The original implementation of the plane atomic
> > > update enables the corresponding DPSUB layer only if the framebuffer
> > > format has changed. This would leave the layer disabled after switching to
> > > a different display mode with the same framebuffer format.
> >
> > Do we need a Fixes: tag or has this issue been there since the beginning
> > ?
>
> Yes, this was introduced in the initial driver commit.
>
> > > Signed-off-by: Anatoliy Klymenko <anatoliy.klymenko@xxxxxxx>
> > > ---
> > > drivers/gpu/drm/xlnx/zynqmp_kms.c | 5 ++---
> > > 1 file changed, 2 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/xlnx/zynqmp_kms.c b/drivers/gpu/drm/xlnx/zynqmp_kms.c
> > > index 43bf416b33d5..c4f038e34814 100644
> > > --- a/drivers/gpu/drm/xlnx/zynqmp_kms.c
> > > +++ b/drivers/gpu/drm/xlnx/zynqmp_kms.c
> > > @@ -120,9 +120,8 @@ static void
> > zynqmp_dpsub_plane_atomic_update(struct drm_plane *plane,
> > > zynqmp_disp_blend_set_global_alpha(dpsub->disp, true,
> > > plane->state->alpha >> 8);
> > >
> > > - /* Enable or re-enable the plane if the format has changed. */
> > > - if (format_changed)
> > > - zynqmp_disp_layer_enable(layer);
> > > + /* Enable or re-enable the plane. */
> > > + zynqmp_disp_layer_enable(layer);
> >
> > This should be safe for now, as the function will just write the plane
> > registers with identical values. The waste of CPU cycles may not be a
> > big issue, even if it would be best to avoid it.
>
> The CPU time wasted on doubling down layer enablement is neglectable
> compared to DP link training time.
Good point.
> > What bothers me more is that we may be working around a larger
> > problem.
> > Resetting the CRTC when disabling it affects the hardware state of the
> > whole device, and thus the state of all software DRM objects. The
> > hardware and software states lose sync. Maybe this patch is all that is
> > needed for now, but other similar issues could pop up in the future.
>
> I had similar thoughts about proper HW state tracking, but that would be
> rather large rework.
>
> > Would it be possible, at atomic check time, to detect the objects whose
> > hardware state will need to be synced, and marked that in their state ?
> > Then in zynqmp_dpsub_plane_atomic_update() you could re-enable the
> > layer
> > based on that. You may need to subclass the drm_plane_state if there's
> > no field in that structure that is suitable to store the information.
> > The format_changed local variable would move there.
>
> Thank you for the idea! I'll check it out.
If it ends up being overkill I think this patch is probably OK. A
comment to explain the reasoning in the code would be nice though.
> > > }
> > >
> > > static const struct drm_plane_helper_funcs
> > zynqmp_dpsub_plane_helper_funcs = {
> > >
> > > ---
> > > base-commit: 673087d8b023faf34b84e8faf63bbeea3da87bab
> > > change-id: 20240520-dp-layer-enable-7b561af29ca8
--
Regards,
Laurent Pinchart
