[PATCH 3/3] drm/sun4i: Fix layer zpos change/atomic modesetting

[Ondřej Jirman]

On Thu, Feb 22, 2024 at 09:02:53PM +0100, Jernej Škrabec wrote:
> Dne sreda, 21. februar 2024 ob 14:45:20 CET je Maxime Ripard napisal(a):
> > Hi,
> > 
> > On Fri, Feb 16, 2024 at 08:04:26PM +0100, Ondřej Jirman wrote:
> > > From: Ondrej Jirman <megi at xff.cz>
> > > 
> > > Identical configurations of planes can lead to different (and wrong)
> > > layer -> pipe routing at HW level, depending on the order of atomic
> > > plane changes.
> > > 
> > > For example:
> > > 
> > > - Layer 1 is configured to zpos 0 and thus uses pipe 0. No other layer
> > >   is enabled. This is a typical situation at boot.
> > > 
> > > - When a compositor takes over and layer 3 is enabled,
> > >   sun8i_ui_layer_enable() will get called with old_zpos=0 zpos=1, which
> > >   will lead to incorrect disabling of pipe 0 and enabling of pipe 1.
> > > 
> > > What happens is that sun8i_ui_layer_enable() function may disable
> > > blender pipes even if it is no longer assigned to its layer.
> > > 
> > > To correct this, move the routing setup out of individual plane's
> > > atomic_update into crtc's atomic_update, where it can be calculated
> > > and updated all at once.
> > > 
> > > Remove the atomic_disable callback because it is no longer needed.
> > > 
> > > Signed-off-by: Ondrej Jirman <megi at xff.cz>
> > 
> > I don't have enough knowledge about the mixers code to comment on your
> > patch, so I'll let Jernej review it. However, this feels to me like the
> > pipe assignment is typically the sort of things that should be dealt
> > with device-wide, and in atomic_check.
> 
> In DE2 and DE3.0, you cannot move planes between mixers (crtcs), because each
> one is hardwired to specific mixer. Movable planes are the feature of DE3.3
> and one of the pain points for upstreaming the code. Anyway, this commit only
> addresses current issue of enabling and disabling planes and handling zpos.
> 
> In atomic check you can only precalculate final register values, but I don't
> see any benefit doing that. I think that this code elegantly solves current
> issue of enabling or disabling wrong plane in certain situations, so:
> 
> Reviewed-by: Jernej Skrabec <jernej.skrabec at gmail.com>
> 
> Note, if there is new revision, please rewrite blender regmap_update_bits()
> to regmap_write(). Since there is HW issue with reads, I would like to
> get rid of regmap_update_bits() calls eventually.

I've looked into it and I can probably rewrite these quite readily:

+	regmap_update_bits(mixer->engine.regs,
+			   SUN8I_MIXER_BLEND_ROUTE(bld_base),
+			   SUN8I_MIXER_BLEND_ROUTE_PIPE_MSK(0) |
+			   SUN8I_MIXER_BLEND_ROUTE_PIPE_MSK(1) |
+			   SUN8I_MIXER_BLEND_ROUTE_PIPE_MSK(2) |
+			   SUN8I_MIXER_BLEND_ROUTE_PIPE_MSK(3),
+			   route);

The mask here covers all implemented bits in the register.

+	regmap_update_bits(mixer->engine.regs,
+			   SUN8I_MIXER_BLEND_PIPE_CTL(bld_base),
+			   SUN8I_MIXER_BLEND_PIPE_CTL_EN(0) |
+			   SUN8I_MIXER_BLEND_PIPE_CTL_EN(1) |
+			   SUN8I_MIXER_BLEND_PIPE_CTL_EN(2) |
+			   SUN8I_MIXER_BLEND_PIPE_CTL_EN(3),
+			   pipe_en);
+

The mask here doesn't cover BLD_FILL_COLOR_CTL.Px_FCEN bits that implement solid
color filling. But those can be 0 anyway except for pipe0 which is hardcoded by
the driver to 1, I think:

631         /*
632          * Set fill color of bottom plane to black. Generally not needed
633          * except when VI plane is at bottom (zpos = 0) and enabled.
634          */
635         regmap_write(mixer->engine.regs, SUN8I_MIXER_BLEND_PIPE_CTL(base),
636                      SUN8I_MIXER_BLEND_PIPE_CTL_FC_EN(0));

I will not be able to get rid of regmap_update_bits in sun8i_layer_enable
because that register there has other important things in it like framebuffer
pixel format, etc.

kind regards,
	o.

> Best regards,
> Jernej
> 
> > 
> > If I'm talking non-sense, it would be great to mention at least why that
> > can't be an option in the commit log.
> > 
> > Maxime
> > 
> 
> 
> 
>