[RFC,1/3] drm: Introduce color fill properties for drm plane

Commit Message

Jessica Zhang Oct. 28, 2022, 10:59 p.m. UTC

Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
drm_plane. In addition, add support for setting and getting the values
of these properties.

COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
represents the format of the color fill. Userspace can set enable solid
fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
framebuffer to NULL.

Signed-off-by: Jessica Zhang <[e-mail]>
---
 drivers/gpu/drm/drm_atomic_uapi.c |  8 +++++++
 drivers/gpu/drm/drm_blend.c       | 38 +++++++++++++++++++++++++++++++
 include/drm/drm_blend.h           |  2 ++
 include/drm/drm_plane.h           | 28 +++++++++++++++++++++++
 4 files changed, 76 insertions(+)

Comments

Dmitry Baryshkov Oct. 29, 2022, 11:23 a.m. UTC | #1

On 29/10/2022 01:59, Jessica Zhang wrote:
> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> drm_plane. In addition, add support for setting and getting the values
> of these properties.
> 
> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> represents the format of the color fill. Userspace can set enable solid
> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> framebuffer to NULL.

I suppose that COLOR_FILL should override framebuffer rather than 
requiring that FB is set to NULL. In other words, if color_filL_format 
is non-zero, it would make sense to ignore the FB. Then one can use the 
color_fill_format property to quickly switch between filled plane and 
FB-backed one.

> 
> Signed-off-by: Jessica Zhang <[e-mail]>
> ---
>   drivers/gpu/drm/drm_atomic_uapi.c |  8 +++++++
>   drivers/gpu/drm/drm_blend.c       | 38 +++++++++++++++++++++++++++++++
>   include/drm/drm_blend.h           |  2 ++
>   include/drm/drm_plane.h           | 28 +++++++++++++++++++++++
>   4 files changed, 76 insertions(+)
> 
> diff --git a/drivers/gpu/drm/drm_atomic_uapi.c b/drivers/gpu/drm/drm_atomic_uapi.c
> index 79730fa1dd8e..e1664463fca4 100644
> --- a/drivers/gpu/drm/drm_atomic_uapi.c
> +++ b/drivers/gpu/drm/drm_atomic_uapi.c
> @@ -544,6 +544,10 @@ static int drm_atomic_plane_set_property(struct drm_plane *plane,
>   		state->src_w = val;
>   	} else if (property == config->prop_src_h) {
>   		state->src_h = val;
> +	} else if (property == plane->color_fill_format_property) {
> +		state->color_fill_format = val;
> +	} else if (property == plane->color_fill_property) {
> +		state->color_fill = val;
>   	} else if (property == plane->alpha_property) {
>   		state->alpha = val;
>   	} else if (property == plane->blend_mode_property) {
> @@ -616,6 +620,10 @@ drm_atomic_plane_get_property(struct drm_plane *plane,
>   		*val = state->src_w;
>   	} else if (property == config->prop_src_h) {
>   		*val = state->src_h;
> +	} else if (property == plane->color_fill_format_property) {
> +		*val = state->color_fill_format;
> +	} else if (property == plane->color_fill_property) {
> +		*val = state->color_fill;
>   	} else if (property == plane->alpha_property) {
>   		*val = state->alpha;
>   	} else if (property == plane->blend_mode_property) {
> diff --git a/drivers/gpu/drm/drm_blend.c b/drivers/gpu/drm/drm_blend.c
> index b4c8cab7158c..b8c2b263fa51 100644
> --- a/drivers/gpu/drm/drm_blend.c
> +++ b/drivers/gpu/drm/drm_blend.c
> @@ -616,3 +616,41 @@ int drm_plane_create_blend_mode_property(struct drm_plane *plane,
>   	return 0;
>   }
>   EXPORT_SYMBOL(drm_plane_create_blend_mode_property);
> +
> +int drm_plane_create_color_fill_property(struct drm_plane *plane)
> +{
> +	struct drm_property *prop;
> +
> +	prop = drm_property_create_range(plane->dev, 0, "color_fill",
> +					 0, 0xffffffff);
> +	if (!prop)
> +		return -ENOMEM;
> +
> +	drm_object_attach_property(&plane->base, prop, 0);
> +	plane->color_fill_property = prop;
> +
> +	if (plane->state)
> +		plane->state->color_fill = 0;
> +
> +	return 0;
> +}
> +EXPORT_SYMBOL(drm_plane_create_color_fill_property);
> +
> +int drm_plane_create_color_fill_format_property(struct drm_plane *plane)
> +{
> +	struct drm_property *prop;
> +
> +	prop = drm_property_create_range(plane->dev, 0, "color_fill_format",
> +					 0, 0xffffffff);
> +	if (!prop)
> +		return -ENOMEM;
> +
> +	drm_object_attach_property(&plane->base, prop, 0);
> +	plane->color_fill_format_property = prop;
> +
> +	if (plane->state)
> +		plane->state->color_fill_format = 0;

Don't you also need to reset these properties in 
__drm_atomic_helper_plane_state_reset() ?

> +
> +	return 0;
> +}
> +EXPORT_SYMBOL(drm_plane_create_color_fill_format_property);
> diff --git a/include/drm/drm_blend.h b/include/drm/drm_blend.h
> index 88bdfec3bd88..3e96f5e83cce 100644
> --- a/include/drm/drm_blend.h
> +++ b/include/drm/drm_blend.h
> @@ -58,4 +58,6 @@ int drm_atomic_normalize_zpos(struct drm_device *dev,
>   			      struct drm_atomic_state *state);
>   int drm_plane_create_blend_mode_property(struct drm_plane *plane,
>   					 unsigned int supported_modes);
> +int drm_plane_create_color_fill_property(struct drm_plane *plane);
> +int drm_plane_create_color_fill_format_property(struct drm_plane *plane);
>   #endif
> diff --git a/include/drm/drm_plane.h b/include/drm/drm_plane.h
> index 89ea54652e87..dcbfdb0e1f71 100644
> --- a/include/drm/drm_plane.h
> +++ b/include/drm/drm_plane.h
> @@ -116,6 +116,20 @@ struct drm_plane_state {
>   	/** @src_h: height of visible portion of plane (in 16.16) */
>   	uint32_t src_h, src_w;
>   
> +	/**
> +	 * @color_fill_format:
> +	 * Format of the color fill value.
> +	 */
> +	uint32_t color_fill_format;
> +
> +	/**
> +	 * @color_fill:
> +	 * Fill color of the plane with 0 as black and 0xffffffff as white.
> +	 * Can be set by user by setting the COLOR_FILL property. See
> +	 * drm_plane_create_color_fill_property() for more details.
> +	 */
> +	uint32_t color_fill;
> +
>   	/**
>   	 * @alpha:
>   	 * Opacity of the plane with 0 as completely transparent and 0xffff as
> @@ -699,6 +713,20 @@ struct drm_plane {
>   	 */
>   	struct drm_plane_state *state;
>   
> +	/*
> +	 * @color_fill_format_property:
> +	 * Optional color fill format property for this plane. See
> +	 * drm_plane_create_color_fill_format_property().
> +	 */
> +	struct drm_property *color_fill_format_property;
> +
> +	/*
> +	 * @color_fill_property:
> +	 * Optional color fill property for this plane. See
> +	 * drm_plane_create_color_fill_property().
> +	 */
> +	struct drm_property *color_fill_property;
> +
>   	/**
>   	 * @alpha_property:
>   	 * Optional alpha property for this plane. See

Dmitry Baryshkov Oct. 29, 2022, 12:08 p.m. UTC | #2

On 29/10/2022 01:59, Jessica Zhang wrote:
> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> drm_plane. In addition, add support for setting and getting the values
> of these properties.
> 
> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> represents the format of the color fill. Userspace can set enable solid
> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> framebuffer to NULL.
> 
> Signed-off-by: Jessica Zhang <[e-mail]>

Planes report supported formats using the drm_mode_getplane(). You'd 
also need to tell userspace, which formats are supported for color fill. 
I don't think one supports e.g. YV12.

A bit of generic comment for the discussion (this is an RFC anyway). 
Using color_fill/color_fill_format properties sounds simple, but this 
might be not generic enough. Limiting color_fill to 32 bits would 
prevent anybody from using floating point formats (e.g. 
DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with 
e.g. using 64-bit for the color_fill value, but then this doesn't sound 
extensible too much.

So, a question for other hardware maintainers. Do we have hardware that 
supports such 'color filled' planes? Do we want to support format 
modifiers for filling color/data? Because what I have in mind is closer 
to the blob structure, which can then be used for filling the plane:

struct color_fill_blob {
     u32 pixel_format;
     u64 modifiers4];
     u32 pixel_data_size; // fixme: is this necessary?
     u8 pixel_data[];
};

And then... This sounds a lot like a custom framebuffer.

So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL 
flag to the framebuffers, which would e.g. mean that the FB gets stamped 
all over the plane. This would also save us from changing if (!fb) 
checks all over the drm core.

Another approach might be using a format modifier instead of the FB flag.

What do you think?

Jessica Zhang Oct. 31, 2022, 9:58 p.m. UTC | #3

On 10/29/2022 4:23 AM, Dmitry Baryshkov wrote:
> On 29/10/2022 01:59, Jessica Zhang wrote:
>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>> drm_plane. In addition, add support for setting and getting the values
>> of these properties.
>>
>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>> represents the format of the color fill. Userspace can set enable solid
>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>> framebuffer to NULL.
> 
> I suppose that COLOR_FILL should override framebuffer rather than 
> requiring that FB is set to NULL. In other words, if color_filL_format 
> is non-zero, it would make sense to ignore the FB. Then one can use the 
> color_fill_format property to quickly switch between filled plane and 
> FB-backed one.

Hey Dmitry,

I think this is a good idea -- acked.

> 
>>
>> Signed-off-by: Jessica Zhang <[e-mail]>
>> ---
>>   drivers/gpu/drm/drm_atomic_uapi.c |  8 +++++++
>>   drivers/gpu/drm/drm_blend.c       | 38 +++++++++++++++++++++++++++++++
>>   include/drm/drm_blend.h           |  2 ++
>>   include/drm/drm_plane.h           | 28 +++++++++++++++++++++++
>>   4 files changed, 76 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/drm_atomic_uapi.c 
>> b/drivers/gpu/drm/drm_atomic_uapi.c
>> index 79730fa1dd8e..e1664463fca4 100644
>> --- a/drivers/gpu/drm/drm_atomic_uapi.c
>> +++ b/drivers/gpu/drm/drm_atomic_uapi.c
>> @@ -544,6 +544,10 @@ static int drm_atomic_plane_set_property(struct 
>> drm_plane *plane,
>>           state->src_w = val;
>>       } else if (property == config->prop_src_h) {
>>           state->src_h = val;
>> +    } else if (property == plane->color_fill_format_property) {
>> +        state->color_fill_format = val;
>> +    } else if (property == plane->color_fill_property) {
>> +        state->color_fill = val;
>>       } else if (property == plane->alpha_property) {
>>           state->alpha = val;
>>       } else if (property == plane->blend_mode_property) {
>> @@ -616,6 +620,10 @@ drm_atomic_plane_get_property(struct drm_plane 
>> *plane,
>>           *val = state->src_w;
>>       } else if (property == config->prop_src_h) {
>>           *val = state->src_h;
>> +    } else if (property == plane->color_fill_format_property) {
>> +        *val = state->color_fill_format;
>> +    } else if (property == plane->color_fill_property) {
>> +        *val = state->color_fill;
>>       } else if (property == plane->alpha_property) {
>>           *val = state->alpha;
>>       } else if (property == plane->blend_mode_property) {
>> diff --git a/drivers/gpu/drm/drm_blend.c b/drivers/gpu/drm/drm_blend.c
>> index b4c8cab7158c..b8c2b263fa51 100644
>> --- a/drivers/gpu/drm/drm_blend.c
>> +++ b/drivers/gpu/drm/drm_blend.c
>> @@ -616,3 +616,41 @@ int drm_plane_create_blend_mode_property(struct 
>> drm_plane *plane,
>>       return 0;
>>   }
>>   EXPORT_SYMBOL(drm_plane_create_blend_mode_property);
>> +
>> +int drm_plane_create_color_fill_property(struct drm_plane *plane)
>> +{
>> +    struct drm_property *prop;
>> +
>> +    prop = drm_property_create_range(plane->dev, 0, "color_fill",
>> +                     0, 0xffffffff);
>> +    if (!prop)
>> +        return -ENOMEM;
>> +
>> +    drm_object_attach_property(&plane->base, prop, 0);
>> +    plane->color_fill_property = prop;
>> +
>> +    if (plane->state)
>> +        plane->state->color_fill = 0;
>> +
>> +    return 0;
>> +}
>> +EXPORT_SYMBOL(drm_plane_create_color_fill_property);
>> +
>> +int drm_plane_create_color_fill_format_property(struct drm_plane *plane)
>> +{
>> +    struct drm_property *prop;
>> +
>> +    prop = drm_property_create_range(plane->dev, 0, "color_fill_format",
>> +                     0, 0xffffffff);
>> +    if (!prop)
>> +        return -ENOMEM;
>> +
>> +    drm_object_attach_property(&plane->base, prop, 0);
>> +    plane->color_fill_format_property = prop;
>> +
>> +    if (plane->state)
>> +        plane->state->color_fill_format = 0;
> 
> Don't you also need to reset these properties in 
> __drm_atomic_helper_plane_state_reset() ?

Ah, yes I believe so -- acked.

Thanks,

Jessica Zhang

> 
>> +
>> +    return 0;
>> +}
>> +EXPORT_SYMBOL(drm_plane_create_color_fill_format_property);
>> diff --git a/include/drm/drm_blend.h b/include/drm/drm_blend.h
>> index 88bdfec3bd88..3e96f5e83cce 100644
>> --- a/include/drm/drm_blend.h
>> +++ b/include/drm/drm_blend.h
>> @@ -58,4 +58,6 @@ int drm_atomic_normalize_zpos(struct drm_device *dev,
>>                     struct drm_atomic_state *state);
>>   int drm_plane_create_blend_mode_property(struct drm_plane *plane,
>>                        unsigned int supported_modes);
>> +int drm_plane_create_color_fill_property(struct drm_plane *plane);
>> +int drm_plane_create_color_fill_format_property(struct drm_plane 
>> *plane);
>>   #endif
>> diff --git a/include/drm/drm_plane.h b/include/drm/drm_plane.h
>> index 89ea54652e87..dcbfdb0e1f71 100644
>> --- a/include/drm/drm_plane.h
>> +++ b/include/drm/drm_plane.h
>> @@ -116,6 +116,20 @@ struct drm_plane_state {
>>       /** @src_h: height of visible portion of plane (in 16.16) */
>>       uint32_t src_h, src_w;
>> +    /**
>> +     * @color_fill_format:
>> +     * Format of the color fill value.
>> +     */
>> +    uint32_t color_fill_format;
>> +
>> +    /**
>> +     * @color_fill:
>> +     * Fill color of the plane with 0 as black and 0xffffffff as white.
>> +     * Can be set by user by setting the COLOR_FILL property. See
>> +     * drm_plane_create_color_fill_property() for more details.
>> +     */
>> +    uint32_t color_fill;
>> +
>>       /**
>>        * @alpha:
>>        * Opacity of the plane with 0 as completely transparent and 
>> 0xffff as
>> @@ -699,6 +713,20 @@ struct drm_plane {
>>        */
>>       struct drm_plane_state *state;
>> +    /*
>> +     * @color_fill_format_property:
>> +     * Optional color fill format property for this plane. See
>> +     * drm_plane_create_color_fill_format_property().
>> +     */
>> +    struct drm_property *color_fill_format_property;
>> +
>> +    /*
>> +     * @color_fill_property:
>> +     * Optional color fill property for this plane. See
>> +     * drm_plane_create_color_fill_property().
>> +     */
>> +    struct drm_property *color_fill_property;
>> +
>>       /**
>>        * @alpha_property:
>>        * Optional alpha property for this plane. See
> 
> -- 
> With best wishes
> Dmitry
>

Jessica Zhang Oct. 31, 2022, 10:24 p.m. UTC | #4

On 10/29/2022 5:08 AM, Dmitry Baryshkov wrote:
> On 29/10/2022 01:59, Jessica Zhang wrote:
>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>> drm_plane. In addition, add support for setting and getting the values
>> of these properties.
>>
>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>> represents the format of the color fill. Userspace can set enable solid
>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>> framebuffer to NULL.
>>
>> Signed-off-by: Jessica Zhang <[e-mail]> >
> Planes report supported formats using the drm_mode_getplane(). You'd 
> also need to tell userspace, which formats are supported for color fill. 
> I don't think one supports e.g. YV12.

Hey Dmitry,

Good point. Forgot to add this in the patch [3/3] commit message, but 
FWIW MSM DPU devices only support the RGB format variants [1].

[1] 
https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/msm/disp/dpu1/dpu_plane.c#L697

> 
> A bit of generic comment for the discussion (this is an RFC anyway). 
> Using color_fill/color_fill_format properties sounds simple, but this 
> might be not generic enough. Limiting color_fill to 32 bits would 
> prevent anybody from using floating point formats (e.g. 
> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with 
> e.g. using 64-bit for the color_fill value, but then this doesn't sound 
> extensible too much.

Hm... I can definitely change color_fill to use u64 instead. As for 
making it more extensible, do you have any suggestions?

> 
> So, a question for other hardware maintainers. Do we have hardware that 
> supports such 'color filled' planes? Do we want to support format 
> modifiers for filling color/data? Because what I have in mind is closer 
> to the blob structure, which can then be used for filling the plane:
> 
> struct color_fill_blob {
>      u32 pixel_format;
>      u64 modifiers4];
>      u32 pixel_data_size; // fixme: is this necessary?
>      u8 pixel_data[];
> };

Just a question about this blob struct -- what is the purpose of pixel_data?

At least for devices using the DPU driver, the only data needed to 
enable solid fill is color_fill and color_fill_format. I'd also be 
interested in knowing if there are other drivers support a similar 
feature and what is needed for them.

> 
> And then... This sounds a lot like a custom framebuffer.
> 
> So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL 
> flag to the framebuffers, which would e.g. mean that the FB gets stamped 
> all over the plane. This would also save us from changing if (!fb) 
> checks all over the drm core.

JFYI we did originally consider using a custom 1x1 FB to for color fill 
[1], but decided to go with a plane property instead. IIRC the 
conclusion was that having color fill as a plane property is a cleaner 
solution.

As for creating a new blob struct to hold all color fill info, I'm open 
to implementing that over having 2 separate properties.

[1] https://oftc.irclog.whitequark.org/dri-devel/2022-09-23#31409842

> 
> Another approach might be using a format modifier instead of the FB flag.
I like the idea of having a format modifier denoting if the driver 
supports color_fill for that specific format. This would allow userspace 
to know which formats are supported by solid fill planes.

Thanks,

Jessica Zhang

> 
> What do you think?
> 
> -- 
> With best wishes
> Dmitry
>

Dmitry Baryshkov Nov. 1, 2022, 12:11 a.m. UTC | #5

Hi,

On 01/11/2022 01:24, Jessica Zhang wrote:
> 
> 
> On 10/29/2022 5:08 AM, Dmitry Baryshkov wrote:
>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>> drm_plane. In addition, add support for setting and getting the values
>>> of these properties.
>>>
>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>> represents the format of the color fill. Userspace can set enable solid
>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>> framebuffer to NULL.
>>>
>>> Signed-off-by: Jessica Zhang <[e-mail]> >
>> Planes report supported formats using the drm_mode_getplane(). You'd 
>> also need to tell userspace, which formats are supported for color 
>> fill. I don't think one supports e.g. YV12.
> 
> Hey Dmitry,
> 
> Good point. Forgot to add this in the patch [3/3] commit message, but 
> FWIW MSM DPU devices only support the RGB format variants [1].
> 
> [1] 
> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/msm/disp/dpu1/dpu_plane.c#L697

Ack. So you'd need to tell this to userspace.

> 
>>
>> A bit of generic comment for the discussion (this is an RFC anyway). 
>> Using color_fill/color_fill_format properties sounds simple, but this 
>> might be not generic enough. Limiting color_fill to 32 bits would 
>> prevent anybody from using floating point formats (e.g. 
>> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with 
>> e.g. using 64-bit for the color_fill value, but then this doesn't 
>> sound extensible too much.
> 
> Hm... I can definitely change color_fill to use u64 instead. As for 
> making it more extensible, do you have any suggestions?

No. Not u64. It is a blob. Basically because when designing API you can 
not guarantee that all fill values would fit into u64. Also see below.

> 
>>
>> So, a question for other hardware maintainers. Do we have hardware 
>> that supports such 'color filled' planes? Do we want to support format 
>> modifiers for filling color/data? Because what I have in mind is 
>> closer to the blob structure, which can then be used for filling the 
>> plane:
>>
>> struct color_fill_blob {
>>      u32 pixel_format;
>>      u64 modifiers4];
>>      u32 pixel_data_size; // fixme: is this necessary?
>>      u8 pixel_data[];
>> };
> 
> Just a question about this blob struct -- what is the purpose of 
> pixel_data?
> 
> At least for devices using the DPU driver, the only data needed to 
> enable solid fill is color_fill and color_fill_format. I'd also be 
> interested in knowing if there are other drivers support a similar 
> feature and what is needed for them.

Yes. You are thinking from the DPU point of view. ARGB only. However as 
we are adding generic API, we should not limit ourselves to it. Other 
deivces might support other formats of fill data. For example using 
YUY2/UYVY for filling the plane. And such YUV data is not a colour 
anymore. It is a pixel data, just as I named it.

Another hardware might support some fill patterns. Or e.g. passing a 
compressed texels/macrotiles. So, pixel data. Note, I've added format 
modifiers. Maybe `u64 modifiers[4]` is an overkill, as we have just a 
single data plane. Maybe just `u64 modifier` would be enough.

> 
>>
>> And then... This sounds a lot like a custom framebuffer.
>>
>> So, maybe what should we do instead is to add new 
>> DRM_MODE_FB_COLOR_FILL flag to the framebuffers, which would e.g. mean 
>> that the FB gets stamped all over the plane. This would also save us 
>> from changing if (!fb) checks all over the drm core.
> 
> JFYI we did originally consider using a custom 1x1 FB to for color fill 
> [1], but decided to go with a plane property instead. IIRC the 
> conclusion was that having color fill as a plane property is a cleaner 
> solution.
> 
> As for creating a new blob struct to hold all color fill info, I'm open 
> to implementing that over having 2 separate properties.
> 
> [1] https://oftc.irclog.whitequark.org/dri-devel/2022-09-23#31409842

Let me cite the conclusion form the IRC chat: `22:20 <robclark> 
abhinav__: kinda.. the proposal was that userspace creates a blob 
property with the solid fill color, and then attaches the blob-prop id 
to the plane's FB_ID`.

It's not a pair of properties. It is a blob, because it is not that 
limited as the pair of range properties is.

I will reread the log later, but just my 2c. Attaching the blob property 
as the FB_ID might confuse userspace. I'd be slightly biased to being 
more conservative here. However as the final proposal was to attach the 
blob ID, let's do it this way.

> 
>>
>> Another approach might be using a format modifier instead of the FB flag.
> I like the idea of having a format modifier denoting if the driver 
> supports color_fill for that specific format. This would allow userspace 
> to know which formats are supported by solid fill planes.

Yes, exactly. It would come in a natural way.

[Rumbling: and then it's natural to have the fill data in FB. Dull mode 
off.]

Jessica Zhang Nov. 1, 2022, 5:35 p.m. UTC | #6

On 10/31/2022 5:11 PM, Dmitry Baryshkov wrote:
> Hi,
> 
> On 01/11/2022 01:24, Jessica Zhang wrote:
>>
>>
>> On 10/29/2022 5:08 AM, Dmitry Baryshkov wrote:
>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>> drm_plane. In addition, add support for setting and getting the values
>>>> of these properties.
>>>>
>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>> represents the format of the color fill. Userspace can set enable solid
>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>> framebuffer to NULL.
>>>>
>>>> Signed-off-by: Jessica Zhang <[e-mail]> >
>>> Planes report supported formats using the drm_mode_getplane(). You'd 
>>> also need to tell userspace, which formats are supported for color 
>>> fill. I don't think one supports e.g. YV12.
>>
>> Hey Dmitry,
>>
>> Good point. Forgot to add this in the patch [3/3] commit message, but 
>> FWIW MSM DPU devices only support the RGB format variants [1].
>>
>> [1] 
>> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/msm/disp/dpu1/dpu_plane.c#L697
> 
> Ack. So you'd need to tell this to userspace.
> 
>>
>>>
>>> A bit of generic comment for the discussion (this is an RFC anyway). 
>>> Using color_fill/color_fill_format properties sounds simple, but this 
>>> might be not generic enough. Limiting color_fill to 32 bits would 
>>> prevent anybody from using floating point formats (e.g. 
>>> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with 
>>> e.g. using 64-bit for the color_fill value, but then this doesn't 
>>> sound extensible too much.
>>
>> Hm... I can definitely change color_fill to use u64 instead. As for 
>> making it more extensible, do you have any suggestions?
> 
> No. Not u64. It is a blob. Basically because when designing API you can 
> not guarantee that all fill values would fit into u64. Also see below.
> 
>>
>>>
>>> So, a question for other hardware maintainers. Do we have hardware 
>>> that supports such 'color filled' planes? Do we want to support 
>>> format modifiers for filling color/data? Because what I have in mind 
>>> is closer to the blob structure, which can then be used for filling 
>>> the plane:
>>>
>>> struct color_fill_blob {
>>>      u32 pixel_format;
>>>      u64 modifiers4];
>>>      u32 pixel_data_size; // fixme: is this necessary?
>>>      u8 pixel_data[];
>>> };
>>
>> Just a question about this blob struct -- what is the purpose of 
>> pixel_data?
>>
>> At least for devices using the DPU driver, the only data needed to 
>> enable solid fill is color_fill and color_fill_format. I'd also be 
>> interested in knowing if there are other drivers support a similar 
>> feature and what is needed for them.
> 
> Yes. You are thinking from the DPU point of view. ARGB only. However as 
> we are adding generic API, we should not limit ourselves to it. Other 
> deivces might support other formats of fill data. For example using 
> YUY2/UYVY for filling the plane. And such YUV data is not a colour 
> anymore. It is a pixel data, just as I named it.
> 
> Another hardware might support some fill patterns. Or e.g. passing a 
> compressed texels/macrotiles. So, pixel data. Note, I've added format 
> modifiers. Maybe `u64 modifiers[4]` is an overkill, as we have just a 
> single data plane. Maybe just `u64 modifier` would be enough.

Got it, I think that's reasonable then.

> 
>>
>>>
>>> And then... This sounds a lot like a custom framebuffer.
>>>
>>> So, maybe what should we do instead is to add new 
>>> DRM_MODE_FB_COLOR_FILL flag to the framebuffers, which would e.g. 
>>> mean that the FB gets stamped all over the plane. This would also 
>>> save us from changing if (!fb) checks all over the drm core.
>>
>> JFYI we did originally consider using a custom 1x1 FB to for color 
>> fill [1], but decided to go with a plane property instead. IIRC the 
>> conclusion was that having color fill as a plane property is a cleaner 
>> solution.
>>
>> As for creating a new blob struct to hold all color fill info, I'm 
>> open to implementing that over having 2 separate properties.
>>
>> [1] https://oftc.irclog.whitequark.org/dri-devel/2022-09-23#31409842
> 
> Let me cite the conclusion form the IRC chat: `22:20 <robclark> 
> abhinav__: kinda.. the proposal was that userspace creates a blob 
> property with the solid fill color, and then attaches the blob-prop id 
> to the plane's FB_ID`.
> 
> It's not a pair of properties. It is a blob, because it is not that 
> limited as the pair of range properties is.
> 
> I will reread the log later, but just my 2c. Attaching the blob property 
> as the FB_ID might confuse userspace. I'd be slightly biased to being 
> more conservative here. However as the final proposal was to attach the 
> blob ID, let's do it this way.

Sounds good. Will spin up a v2 with the prop_blob implementation.

Thanks,

Jessica Zhang

> 
>>
>>>
>>> Another approach might be using a format modifier instead of the FB 
>>> flag.
>> I like the idea of having a format modifier denoting if the driver 
>> supports color_fill for that specific format. This would allow 
>> userspace to know which formats are supported by solid fill planes.
> 
> Yes, exactly. It would come in a natural way.
> 
> [Rumbling: and then it's natural to have the fill data in FB. Dull mode 
> off.]
> 
> -- 
> With best wishes
> Dmitry
>

Simon Ser Nov. 8, 2022, 6:25 p.m. UTC | #7

On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:

> On 29/10/2022 01:59, Jessica Zhang wrote:
> 
> > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > drm_plane. In addition, add support for setting and getting the values
> > of these properties.
> > 
> > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > represents the format of the color fill. Userspace can set enable solid
> > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > framebuffer to NULL.
> 
> I suppose that COLOR_FILL should override framebuffer rather than
> requiring that FB is set to NULL. In other words, if color_filL_format
> is non-zero, it would make sense to ignore the FB. Then one can use the
> color_fill_format property to quickly switch between filled plane and
> FB-backed one.

That would be inconsistent with the rest of the KMS uAPI. For instance,
the kernel will error out if CRTC has active=0 but a connector is still
linked to the CRTC. IOW, the current uAPI errors out if the KMS state
is inconsistent.

Simon Ser Nov. 8, 2022, 6:50 p.m. UTC | #8

cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.

On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:

> On 29/10/2022 01:59, Jessica Zhang wrote:
> > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > drm_plane. In addition, add support for setting and getting the values
> > of these properties.
> >
> > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > represents the format of the color fill. Userspace can set enable solid
> > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > framebuffer to NULL.
> >
> > Signed-off-by: Jessica Zhang <[e-mail]>
> 
> Planes report supported formats using the drm_mode_getplane(). You'd
> also need to tell userspace, which formats are supported for color fill.
> I don't think one supports e.g. YV12.
> 
> A bit of generic comment for the discussion (this is an RFC anyway).
> Using color_fill/color_fill_format properties sounds simple, but this
> might be not generic enough. Limiting color_fill to 32 bits would
> prevent anybody from using floating point formats (e.g.
> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
> e.g. using 64-bit for the color_fill value, but then this doesn't sound
> extensible too much.
> 
> So, a question for other hardware maintainers. Do we have hardware that
> supports such 'color filled' planes? Do we want to support format
> modifiers for filling color/data? Because what I have in mind is closer
> to the blob structure, which can then be used for filling the plane:
> 
> struct color_fill_blob {
>      u32 pixel_format;
>      u64 modifiers4];
>      u32 pixel_data_size; // fixme: is this necessary?
>      u8 pixel_data[];
> };
> 
> And then... This sounds a lot like a custom framebuffer.
> 
> So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
> flag to the framebuffers, which would e.g. mean that the FB gets stamped
> all over the plane. This would also save us from changing if (!fb)
> checks all over the drm core.
> 
> Another approach might be using a format modifier instead of the FB flag.
> 
> What do you think?

First off, we only need to represent the value of a single pixel here. So I'm
not quite following why we need format modifiers. Format modifiers describe how
pixels are laid out in memory. Since there's a single pixel described, this
is non-sensical to me, the format modifier is always LINEAR.

Then, I can understand why putting the pixel_format in there is tempting to
guarantee future extensibility, but it also adds complexity. For instance, how
does user-space figure out which formats can be used for COLOR_FILL? Can
user-space use any format supported by the plane? What does it mean for
multi-planar formats? Do we really want the kernel to have conversion logic for
all existing formats? Do we need to also add a new read-only blob prop to
indicate supported COLOR_FILL formats?

We've recently-ish standardized a new Wayland protocol [1] which has the same
purpose as this new kernel uAPI. The conclusion there was that using 32-bit
values for each channel (R, G, B, A) would be enough for almost all use-cases.
The driver can convert these high-precision values to what the hardware expects.
The only concern was about sending values outside of the [0.0, 1.0] range,
which may have HDR use-cases.

So, there are multiple ways to go about this. I can think of:

- Put "RGBA32" in the name of the prop, and if we ever need a different
  color format, pick a different name.
- Define a struct with an enum of possible fill kinds:
  #define FILL_COLOR_RGBA32 1
  #define FILL_COLOR_F32 2
  struct color_fill_blob { u32 kind; u8 data[]; };
- Define a struct with a version and RGBA values:
  struct color_fill_blob { u32 version; u32 rgba[4]; };
  If we need to add more formats later, or new metadata:
  struct color_fill_blob2 { u32 version; /* new fields */ };
  where version must be set to 2.
- Define a struct with a "pixel_format" prop, but force user-space to use a
  fixed format for now. Later, if we need another format, add a new prop to
  advertise supported formats.
- More complicated solutions, e.g. advertise the list of supported formats from
  the start.

[1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104

Sebastian Wick Nov. 8, 2022, 10:01 p.m. UTC | #9

On Tue, Nov 8, 2022 at 7:51 PM Simon Ser <[e-mail]> wrote:
>
> cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.
>
> On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:
>
> > On 29/10/2022 01:59, Jessica Zhang wrote:
> > > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > > drm_plane. In addition, add support for setting and getting the values
> > > of these properties.
> > >
> > > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > > represents the format of the color fill. Userspace can set enable solid
> > > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > > framebuffer to NULL.
> > >
> > > Signed-off-by: Jessica Zhang <[e-mail]>
> >
> > Planes report supported formats using the drm_mode_getplane(). You'd
> > also need to tell userspace, which formats are supported for color fill.
> > I don't think one supports e.g. YV12.
> >
> > A bit of generic comment for the discussion (this is an RFC anyway).
> > Using color_fill/color_fill_format properties sounds simple, but this
> > might be not generic enough. Limiting color_fill to 32 bits would
> > prevent anybody from using floating point formats (e.g.
> > DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
> > e.g. using 64-bit for the color_fill value, but then this doesn't sound
> > extensible too much.
> >
> > So, a question for other hardware maintainers. Do we have hardware that
> > supports such 'color filled' planes? Do we want to support format
> > modifiers for filling color/data? Because what I have in mind is closer
> > to the blob structure, which can then be used for filling the plane:
> >
> > struct color_fill_blob {
> >      u32 pixel_format;
> >      u64 modifiers4];
> >      u32 pixel_data_size; // fixme: is this necessary?
> >      u8 pixel_data[];
> > };
> >
> > And then... This sounds a lot like a custom framebuffer.
> >
> > So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
> > flag to the framebuffers, which would e.g. mean that the FB gets stamped
> > all over the plane. This would also save us from changing if (!fb)
> > checks all over the drm core.
> >
> > Another approach might be using a format modifier instead of the FB flag.
> >
> > What do you think?
>
> First off, we only need to represent the value of a single pixel here. So I'm
> not quite following why we need format modifiers. Format modifiers describe how
> pixels are laid out in memory. Since there's a single pixel described, this
> is non-sensical to me, the format modifier is always LINEAR.
>
> Then, I can understand why putting the pixel_format in there is tempting to
> guarantee future extensibility, but it also adds complexity. For instance, how
> does user-space figure out which formats can be used for COLOR_FILL? Can
> user-space use any format supported by the plane? What does it mean for
> multi-planar formats? Do we really want the kernel to have conversion logic for
> all existing formats? Do we need to also add a new read-only blob prop to
> indicate supported COLOR_FILL formats?
>
> We've recently-ish standardized a new Wayland protocol [1] which has the same
> purpose as this new kernel uAPI. The conclusion there was that using 32-bit
> values for each channel (R, G, B, A) would be enough for almost all use-cases.
> The driver can convert these high-precision values to what the hardware expects.
> The only concern was about sending values outside of the [0.0, 1.0] range,
> which may have HDR use-cases.
>
> So, there are multiple ways to go about this. I can think of:
>
> - Put "RGBA32" in the name of the prop, and if we ever need a different
>   color format, pick a different name.
> - Define a struct with an enum of possible fill kinds:
>   #define FILL_COLOR_RGBA32 1
>   #define FILL_COLOR_F32 2
>   struct color_fill_blob { u32 kind; u8 data[]; };
> - Define a struct with a version and RGBA values:
>   struct color_fill_blob { u32 version; u32 rgba[4]; };
>   If we need to add more formats later, or new metadata:
>   struct color_fill_blob2 { u32 version; /* new fields */ };
>   where version must be set to 2.
> - Define a struct with a "pixel_format" prop, but force user-space to use a
>   fixed format for now. Later, if we need another format, add a new prop to
>   advertise supported formats.
> - More complicated solutions, e.g. advertise the list of supported formats from
>   the start.
>
> [1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104
>

Agreeing with most of what you said here. However, what's the idea
behind a format anyway? The 4 values provided here are fed directly
into the color pipeline which seems to define the color channels it's
working on as RGBA (or doesn't define anything at all). The only
reason I can think of is that hardware might support only ingesting
values either in a format with high bit depth color channels and no
alpha or a format with low bit depth color but with alpha, so choosing
between the formats provides a real trade-off. Is that actually
something hardware might be restricted to or do they all just support
ingesting the color data with enough precision on every channel?

Pekka Paalanen Nov. 9, 2022, 9:18 a.m. UTC | #10

On Tue, 8 Nov 2022 23:01:47 +0100
Sebastian Wick <[e-mail]> wrote:

> On Tue, Nov 8, 2022 at 7:51 PM Simon Ser <[e-mail]> wrote:
> >
> > cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.

Hi all,

thanks! Comments below.

> >
> > On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:
> >  
> > > On 29/10/2022 01:59, Jessica Zhang wrote:  
> > > > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > > > drm_plane. In addition, add support for setting and getting the values
> > > > of these properties.
> > > >
> > > > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > > > represents the format of the color fill. Userspace can set enable solid
> > > > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > > > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > > > framebuffer to NULL.
> > > >
> > > > Signed-off-by: Jessica Zhang <[e-mail]>  
> > >
> > > Planes report supported formats using the drm_mode_getplane(). You'd
> > > also need to tell userspace, which formats are supported for color fill.
> > > I don't think one supports e.g. YV12.
> > >
> > > A bit of generic comment for the discussion (this is an RFC anyway).
> > > Using color_fill/color_fill_format properties sounds simple, but this
> > > might be not generic enough. Limiting color_fill to 32 bits would
> > > prevent anybody from using floating point formats (e.g.
> > > DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
> > > e.g. using 64-bit for the color_fill value, but then this doesn't sound
> > > extensible too much.
> > >
> > > So, a question for other hardware maintainers. Do we have hardware that
> > > supports such 'color filled' planes? Do we want to support format
> > > modifiers for filling color/data? Because what I have in mind is closer
> > > to the blob structure, which can then be used for filling the plane:
> > >
> > > struct color_fill_blob {
> > >      u32 pixel_format;
> > >      u64 modifiers4];
> > >      u32 pixel_data_size; // fixme: is this necessary?
> > >      u8 pixel_data[];
> > > };
> > >
> > > And then... This sounds a lot like a custom framebuffer.
> > >
> > > So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
> > > flag to the framebuffers, which would e.g. mean that the FB gets stamped
> > > all over the plane. This would also save us from changing if (!fb)
> > > checks all over the drm core.
> > >
> > > Another approach might be using a format modifier instead of the FB flag.
> > >
> > > What do you think?  
> >
> > First off, we only need to represent the value of a single pixel here. So I'm
> > not quite following why we need format modifiers. Format modifiers describe how
> > pixels are laid out in memory. Since there's a single pixel described, this
> > is non-sensical to me, the format modifier is always LINEAR.

Agreed.

> >
> > Then, I can understand why putting the pixel_format in there is tempting to
> > guarantee future extensibility, but it also adds complexity. For instance, how
> > does user-space figure out which formats can be used for COLOR_FILL? Can
> > user-space use any format supported by the plane? What does it mean for
> > multi-planar formats? Do we really want the kernel to have conversion logic for
> > all existing formats? Do we need to also add a new read-only blob prop to
> > indicate supported COLOR_FILL formats?

Right. This does not seem to require pixel formats at all.

The point of pixel formats is to be able to feed large amounts of data
as-is into hardware and avoid the CPU ever touching it. You do that
with DRM FBs pointing to suitably allocated hardware buffers. But here
we have exactly one pixel, which I imagine will always be read by the
CPU so the driver will convert it into a hardware-specific format and
program it; probably the driver will not create an internal DRM FB for
it.

The above might also be a reason to not model this as a special-case
DRM FB in UAPI. Or, at least you need a whole new ioctl to create such
DRM FB to avoid the need to allocate e.g. a dumb buffer or a
GPU-specific buffer.

What one does need is what Sebastian brought up: does it support alpha
or not?

Userspace would also be interested in the supported precision of the
values, but the hardware pixel component order is irrelevant because the
driver will always convert the one pixel with CPU anyway.

YUV vs. RGB is a another question. The KMS color pipeline is defined in
terms of RGBA as far as I know, and alpha-blending YUV values makes no
sense. So will there ever be any need to set an YUV fill? I have a hard
time imagining it.

If you do set an YUV fill, the KMS color pipeline most likely needs to
convert it to RGBA for blending, and then you need the plane properties
COLOR_ENCODING and COLOR_RANGE.

But why bother when userspace can convert that one pixel to RGBA itself
anyway?

> > We've recently-ish standardized a new Wayland protocol [1] which has the same
> > purpose as this new kernel uAPI. The conclusion there was that using 32-bit
> > values for each channel (R, G, B, A) would be enough for almost all use-cases.
> > The driver can convert these high-precision values to what the hardware expects.
> > The only concern was about sending values outside of the [0.0, 1.0] range,
> > which may have HDR use-cases.

This is what I would suggest, yes. This representation has enough
precision to be future-proof, and the driver will be converting the
value anyway.

The question about values outside of the unit range is a good one, too.
With Wayland, we can simply add another request to set a value in
floating-point if that turns up necessary.

Whether that will ever be necessary is connected to how the DRM KMS
abstract color pipeline is modelled, and that you must define from the
beginning:

If DRM KMS gets color processing related plane properties like CTM,
GAMMA or DEGAMMA (they already exist for CRTC, and these have been
proposed for planes quite some time ago), does the fill color go
through all these operations, or will the fill color skip all these
operations and go straight to plane blending?

Whether values outside of the unit range will ever be needed depends
also on the userspace design. Userspace could choose the value range
freely if the KMS color pipeline elements happen to support that range.
However things like LUTs are naturally limited to unit range input, so
using values outside of the unit range might be difficult if not
impossible. Therefore I'm not concerned about this question much.

> > So, there are multiple ways to go about this. I can think of:
> >
> > - Put "RGBA32" in the name of the prop, and if we ever need a different
> >   color format, pick a different name.

I could see problems with that if the same device supports more than
one kind. How to turn off all but one color fill property?

What if userspace understands an old color fill property but not the
new one, and the new one happens to be set for some reason? Userspace
will attempt to use the old property without setting the new property
to nil, and fail.

> > - Define a struct with an enum of possible fill kinds:
> >   #define FILL_COLOR_RGBA32 1
> >   #define FILL_COLOR_F32 2
> >   struct color_fill_blob { u32 kind; u8 data[]; };

This could work.

Btw. maybe call it RGBA32323232 to make it follow the drm_fourcc naming
convention. Some userspace libraries already use RGBA32 to mean 32 bit
per pixel instead of per channel.

How will userspace know what kinds are supported?

> > - Define a struct with a version and RGBA values:
> >   struct color_fill_blob { u32 version; u32 rgba[4]; };
> >   If we need to add more formats later, or new metadata:
> >   struct color_fill_blob2 { u32 version; /* new fields */ };
> >   where version must be set to 2.

This could work.

> > - Define a struct with a "pixel_format" prop, but force user-space to use a
> >   fixed format for now. Later, if we need another format, add a new prop to
> >   advertise supported formats.
> > - More complicated solutions, e.g. advertise the list of supported formats from
> >   the start.

Feels more complicated than necessary.

Anyway, the idea of creating a blob and then setting that into some KMS
plane property sounds a very good mechanism.

> >
> > [1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104
> >  
> 
> Agreeing with most of what you said here. However, what's the idea
> behind a format anyway? The 4 values provided here are fed directly
> into the color pipeline which seems to define the color channels it's
> working on as RGBA (or doesn't define anything at all). The only
> reason I can think of is that hardware might support only ingesting
> values either in a format with high bit depth color channels and no
> alpha or a format with low bit depth color but with alpha, so choosing
> between the formats provides a real trade-off. Is that actually
> something hardware might be restricted to or do they all just support
> ingesting the color data with enough precision on every channel?

Right.


Thanks,
pq

Daniel Vetter Nov. 9, 2022, 1:52 p.m. UTC | #11

On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
> On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
> 
> > On 29/10/2022 01:59, Jessica Zhang wrote:
> > 
> > > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > > drm_plane. In addition, add support for setting and getting the values
> > > of these properties.
> > > 
> > > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > > represents the format of the color fill. Userspace can set enable solid
> > > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > > framebuffer to NULL.
> > 
> > I suppose that COLOR_FILL should override framebuffer rather than
> > requiring that FB is set to NULL. In other words, if color_filL_format
> > is non-zero, it would make sense to ignore the FB. Then one can use the
> > color_fill_format property to quickly switch between filled plane and
> > FB-backed one.
> 
> That would be inconsistent with the rest of the KMS uAPI. For instance,
> the kernel will error out if CRTC has active=0 but a connector is still
> linked to the CRTC. IOW, the current uAPI errors out if the KMS state
> is inconsistent.

So if the use-case here really is to solid-fill a plane (and not just
provide a background color for the crtc overall), then I guess we could
also extend addfb to make that happen. We've talked in the past about
propertery-fying framebuffer objects, and that would sort out this uapi
wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.

But if the use-cases are all background color then just doing the crtc
background color would be tons simpler (and likely also easier to support
for more hardware).
-Daniel

Dmitry Baryshkov Nov. 9, 2022, 1:53 p.m. UTC | #12

On 09/11/2022 16:52, Daniel Vetter wrote:
> On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
>> On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
>>
>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>
>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>> drm_plane. In addition, add support for setting and getting the values
>>>> of these properties.
>>>>
>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>> represents the format of the color fill. Userspace can set enable solid
>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>> framebuffer to NULL.
>>>
>>> I suppose that COLOR_FILL should override framebuffer rather than
>>> requiring that FB is set to NULL. In other words, if color_filL_format
>>> is non-zero, it would make sense to ignore the FB. Then one can use the
>>> color_fill_format property to quickly switch between filled plane and
>>> FB-backed one.
>>
>> That would be inconsistent with the rest of the KMS uAPI. For instance,
>> the kernel will error out if CRTC has active=0 but a connector is still
>> linked to the CRTC. IOW, the current uAPI errors out if the KMS state
>> is inconsistent.
> 
> So if the use-case here really is to solid-fill a plane (and not just
> provide a background color for the crtc overall), then I guess we could
> also extend addfb to make that happen. We've talked in the past about
> propertery-fying framebuffer objects, and that would sort out this uapi
> wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.
> 
> But if the use-cases are all background color then just doing the crtc
> background color would be tons simpler (and likely also easier to support
> for more hardware).

No. The hardware supports multiple color-filled planes, which do not 
have to cover the whole CRTC.

Daniel Vetter Nov. 9, 2022, 1:59 p.m. UTC | #13

On Wed, Nov 09, 2022 at 04:53:45PM +0300, Dmitry Baryshkov wrote:
> On 09/11/2022 16:52, Daniel Vetter wrote:
> > On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
> > > On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
> > > 
> > > > On 29/10/2022 01:59, Jessica Zhang wrote:
> > > > 
> > > > > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > > > > drm_plane. In addition, add support for setting and getting the values
> > > > > of these properties.
> > > > > 
> > > > > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > > > > represents the format of the color fill. Userspace can set enable solid
> > > > > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > > > > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > > > > framebuffer to NULL.
> > > > 
> > > > I suppose that COLOR_FILL should override framebuffer rather than
> > > > requiring that FB is set to NULL. In other words, if color_filL_format
> > > > is non-zero, it would make sense to ignore the FB. Then one can use the
> > > > color_fill_format property to quickly switch between filled plane and
> > > > FB-backed one.
> > > 
> > > That would be inconsistent with the rest of the KMS uAPI. For instance,
> > > the kernel will error out if CRTC has active=0 but a connector is still
> > > linked to the CRTC. IOW, the current uAPI errors out if the KMS state
> > > is inconsistent.
> > 
> > So if the use-case here really is to solid-fill a plane (and not just
> > provide a background color for the crtc overall), then I guess we could
> > also extend addfb to make that happen. We've talked in the past about
> > propertery-fying framebuffer objects, and that would sort out this uapi
> > wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.
> > 
> > But if the use-cases are all background color then just doing the crtc
> > background color would be tons simpler (and likely also easier to support
> > for more hardware).
> 
> No. The hardware supports multiple color-filled planes, which do not have to
> cover the whole CRTC.

The use case here means the userspace use-case. What the hw can do on any
given chip kinda doesnt matter, which is why I'm asking. KMD uapi is not
meant to reflect 100% exactly what a specific chip can do, but instead:
- provide features userspace actually needs. If you want per-plane fill,
  you need userspace that makes use of per-plane fill, and if all you have
  is crtc background, then that's it.
- we should create uapi with an eye towards what's actually possible on a
  reasonable set of drivers and hw. Sometimes that means a slightly more
  restricted set so that it's possible to implement in more places,
  especially if that restricted feature set still gets the job done for
  userspace.

Cheers, Daniel

Jessica Zhang Nov. 10, 2022, 1:44 a.m. UTC | #14

On 11/9/2022 5:59 AM, Daniel Vetter wrote:
> On Wed, Nov 09, 2022 at 04:53:45PM +0300, Dmitry Baryshkov wrote:
>> On 09/11/2022 16:52, Daniel Vetter wrote:
>>> On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
>>>> On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
>>>>
>>>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>>>
>>>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>>>> drm_plane. In addition, add support for setting and getting the values
>>>>>> of these properties.
>>>>>>
>>>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>>>> represents the format of the color fill. Userspace can set enable solid
>>>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>>>> framebuffer to NULL.
>>>>>
>>>>> I suppose that COLOR_FILL should override framebuffer rather than
>>>>> requiring that FB is set to NULL. In other words, if color_filL_format
>>>>> is non-zero, it would make sense to ignore the FB. Then one can use the
>>>>> color_fill_format property to quickly switch between filled plane and
>>>>> FB-backed one.
>>>>
>>>> That would be inconsistent with the rest of the KMS uAPI. For instance,
>>>> the kernel will error out if CRTC has active=0 but a connector is still
>>>> linked to the CRTC. IOW, the current uAPI errors out if the KMS state
>>>> is inconsistent.
>>>
>>> So if the use-case here really is to solid-fill a plane (and not just
>>> provide a background color for the crtc overall), then I guess we could
>>> also extend addfb to make that happen. We've talked in the past about
>>> propertery-fying framebuffer objects, and that would sort out this uapi
>>> wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.
>>>
>>> But if the use-cases are all background color then just doing the crtc
>>> background color would be tons simpler (and likely also easier to support
>>> for more hardware).
>>
>> No. The hardware supports multiple color-filled planes, which do not have to
>> cover the whole CRTC.
> 
> The use case here means the userspace use-case. What the hw can do on any
> given chip kinda doesnt matter, which is why I'm asking. KMD uapi is not
> meant to reflect 100% exactly what a specific chip can do, but instead:
> - provide features userspace actually needs. If you want per-plane fill,
>    you need userspace that makes use of per-plane fill, and if all you have
>    is crtc background, then that's it.

Hey Daniel,

The userspace use case we're trying to support is the Android HWC 
SOLID_FILL hint here [1], which is specifying per-plane fill.

Thanks,

Jessica Zhang

[1] 
https://android.googlesource.com/platform/hardware/interfaces/+/refs/heads/master/graphics/composer/aidl/android/hardware/graphics/composer3/Composition.aidl#52

> - we should create uapi with an eye towards what's actually possible on a
>    reasonable set of drivers and hw. Sometimes that means a slightly more
>    restricted set so that it's possible to implement in more places,
>    especially if that restricted feature set still gets the job done for
>    userspace.
> 
> Cheers, Daniel
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch

Daniel Vetter Nov. 11, 2022, 9:45 a.m. UTC | #15

On Wed, Nov 09, 2022 at 05:44:37PM -0800, Jessica Zhang wrote:
> 
> 
> On 11/9/2022 5:59 AM, Daniel Vetter wrote:
> > On Wed, Nov 09, 2022 at 04:53:45PM +0300, Dmitry Baryshkov wrote:
> > > On 09/11/2022 16:52, Daniel Vetter wrote:
> > > > On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
> > > > > On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
> > > > > 
> > > > > > On 29/10/2022 01:59, Jessica Zhang wrote:
> > > > > > 
> > > > > > > Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> > > > > > > drm_plane. In addition, add support for setting and getting the values
> > > > > > > of these properties.
> > > > > > > 
> > > > > > > COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> > > > > > > represents the format of the color fill. Userspace can set enable solid
> > > > > > > fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> > > > > > > the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> > > > > > > framebuffer to NULL.
> > > > > > 
> > > > > > I suppose that COLOR_FILL should override framebuffer rather than
> > > > > > requiring that FB is set to NULL. In other words, if color_filL_format
> > > > > > is non-zero, it would make sense to ignore the FB. Then one can use the
> > > > > > color_fill_format property to quickly switch between filled plane and
> > > > > > FB-backed one.
> > > > > 
> > > > > That would be inconsistent with the rest of the KMS uAPI. For instance,
> > > > > the kernel will error out if CRTC has active=0 but a connector is still
> > > > > linked to the CRTC. IOW, the current uAPI errors out if the KMS state
> > > > > is inconsistent.
> > > > 
> > > > So if the use-case here really is to solid-fill a plane (and not just
> > > > provide a background color for the crtc overall), then I guess we could
> > > > also extend addfb to make that happen. We've talked in the past about
> > > > propertery-fying framebuffer objects, and that would sort out this uapi
> > > > wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.
> > > > 
> > > > But if the use-cases are all background color then just doing the crtc
> > > > background color would be tons simpler (and likely also easier to support
> > > > for more hardware).
> > > 
> > > No. The hardware supports multiple color-filled planes, which do not have to
> > > cover the whole CRTC.
> > 
> > The use case here means the userspace use-case. What the hw can do on any
> > given chip kinda doesnt matter, which is why I'm asking. KMD uapi is not
> > meant to reflect 100% exactly what a specific chip can do, but instead:
> > - provide features userspace actually needs. If you want per-plane fill,
> >    you need userspace that makes use of per-plane fill, and if all you have
> >    is crtc background, then that's it.
> 
> Hey Daniel,
> 
> The userspace use case we're trying to support is the Android HWC SOLID_FILL
> hint here [1], which is specifying per-plane fill.

Does surfaceflinger actually use this for more than background fills? Yes
I'm annoying, but if we can simplify the kernel driver implementation
burden by asking compositors to do the math and simplify things, then I
think we should.

We also need an open source implementation for this that works and is
tested end-to-end. There's the drm_hwc project, but last time I've checked
there's really not much happpening there unfortunately.
-Daniel

> 
> Thanks,
> 
> Jessica Zhang
> 
> [1] https://android.googlesource.com/platform/hardware/interfaces/+/refs/heads/master/graphics/composer/aidl/android/hardware/graphics/composer3/Composition.aidl#52
> 
> > - we should create uapi with an eye towards what's actually possible on a
> >    reasonable set of drivers and hw. Sometimes that means a slightly more
> >    restricted set so that it's possible to implement in more places,
> >    especially if that restricted feature set still gets the job done for
> >    userspace.
> > 
> > Cheers, Daniel
> > -- 
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > http://blog.ffwll.ch

Jessica Zhang Nov. 18, 2022, 7:15 p.m. UTC | #16

On 11/11/2022 1:45 AM, Daniel Vetter wrote:
> On Wed, Nov 09, 2022 at 05:44:37PM -0800, Jessica Zhang wrote:
>>
>>
>> On 11/9/2022 5:59 AM, Daniel Vetter wrote:
>>> On Wed, Nov 09, 2022 at 04:53:45PM +0300, Dmitry Baryshkov wrote:
>>>> On 09/11/2022 16:52, Daniel Vetter wrote:
>>>>> On Tue, Nov 08, 2022 at 06:25:29PM +0000, Simon Ser wrote:
>>>>>> On Saturday, October 29th, 2022 at 13:23, Dmitry Baryshkov <[e-mail]> wrote:
>>>>>>
>>>>>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>>>>>
>>>>>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>>>>>> drm_plane. In addition, add support for setting and getting the values
>>>>>>>> of these properties.
>>>>>>>>
>>>>>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>>>>>> represents the format of the color fill. Userspace can set enable solid
>>>>>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>>>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>>>>>> framebuffer to NULL.
>>>>>>>
>>>>>>> I suppose that COLOR_FILL should override framebuffer rather than
>>>>>>> requiring that FB is set to NULL. In other words, if color_filL_format
>>>>>>> is non-zero, it would make sense to ignore the FB. Then one can use the
>>>>>>> color_fill_format property to quickly switch between filled plane and
>>>>>>> FB-backed one.
>>>>>>
>>>>>> That would be inconsistent with the rest of the KMS uAPI. For instance,
>>>>>> the kernel will error out if CRTC has active=0 but a connector is still
>>>>>> linked to the CRTC. IOW, the current uAPI errors out if the KMS state
>>>>>> is inconsistent.
>>>>>
>>>>> So if the use-case here really is to solid-fill a plane (and not just
>>>>> provide a background color for the crtc overall), then I guess we could
>>>>> also extend addfb to make that happen. We've talked in the past about
>>>>> propertery-fying framebuffer objects, and that would sort out this uapi
>>>>> wart. And I agree the color fill vs PLANE_ID issue is a bit ugly at least.
>>>>>
>>>>> But if the use-cases are all background color then just doing the crtc
>>>>> background color would be tons simpler (and likely also easier to support
>>>>> for more hardware).
>>>>
>>>> No. The hardware supports multiple color-filled planes, which do not have to
>>>> cover the whole CRTC.
>>>
>>> The use case here means the userspace use-case. What the hw can do on any
>>> given chip kinda doesnt matter, which is why I'm asking. KMD uapi is not
>>> meant to reflect 100% exactly what a specific chip can do, but instead:
>>> - provide features userspace actually needs. If you want per-plane fill,
>>>     you need userspace that makes use of per-plane fill, and if all you have
>>>     is crtc background, then that's it.
>>
>> Hey Daniel,
>>
>> The userspace use case we're trying to support is the Android HWC SOLID_FILL
>> hint here [1], which is specifying per-plane fill.
> 
> Does surfaceflinger actually use this for more than background fills? Yes
> I'm annoying, but if we can simplify the kernel driver implementation
> burden by asking compositors to do the math and simplify things, then I
> think we should.

AFAIK surfaceflinger allows apps to use this for cases beyond just 
background fill -- an app, for example, can pass the hint for a plane 
that only partially covers a screen and the driver would be expected to 
fill just that ROI.

> 
> We also need an open source implementation for this that works and is
> tested end-to-end. There's the drm_hwc project, but last time I've checked
> there's really not much happpening there unfortunately.

FWIW, Simon mentioned in a separate reply that Wayland supports a 1x1 FB 
support protocol [1] for a similar purpose as this RFC series. I can 
also create an IGT test meanwhile showing an example of E2E usage.

Thanks,

Jessica

[1] 
https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104

> -Daniel
> 
>>
>> Thanks,
>>
>> Jessica Zhang
>>
>> [1] https://android.googlesource.com/platform/hardware/interfaces/+/refs/heads/master/graphics/composer/aidl/android/hardware/graphics/composer3/Composition.aidl#52
>>
>>> - we should create uapi with an eye towards what's actually possible on a
>>>     reasonable set of drivers and hw. Sometimes that means a slightly more
>>>     restricted set so that it's possible to implement in more places,
>>>     especially if that restricted feature set still gets the job done for
>>>     userspace.
>>>
>>> Cheers, Daniel
>>> -- 
>>> Daniel Vetter
>>> Software Engineer, Intel Corporation
>>> http://blog.ffwll.ch
> 
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch

Jessica Zhang Nov. 23, 2022, 11:27 p.m. UTC | #17

On 11/9/2022 1:18 AM, Pekka Paalanen wrote:
> On Tue, 8 Nov 2022 23:01:47 +0100
> Sebastian Wick <[e-mail]> wrote:
> 
>> On Tue, Nov 8, 2022 at 7:51 PM Simon Ser <[e-mail]> wrote:
>>>
>>> cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.
> 
> Hi all,
> 
> thanks! Comments below.

Thanks for the feedback!

> 
>>>
>>> On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:
>>>   
>>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>>> drm_plane. In addition, add support for setting and getting the values
>>>>> of these properties.
>>>>>
>>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>>> represents the format of the color fill. Userspace can set enable solid
>>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>>> framebuffer to NULL.
>>>>>
>>>>> Signed-off-by: Jessica Zhang <[e-mail]>
>>>>
>>>> Planes report supported formats using the drm_mode_getplane(). You'd
>>>> also need to tell userspace, which formats are supported for color fill.
>>>> I don't think one supports e.g. YV12.
>>>>
>>>> A bit of generic comment for the discussion (this is an RFC anyway).
>>>> Using color_fill/color_fill_format properties sounds simple, but this
>>>> might be not generic enough. Limiting color_fill to 32 bits would
>>>> prevent anybody from using floating point formats (e.g.
>>>> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
>>>> e.g. using 64-bit for the color_fill value, but then this doesn't sound
>>>> extensible too much.
>>>>
>>>> So, a question for other hardware maintainers. Do we have hardware that
>>>> supports such 'color filled' planes? Do we want to support format
>>>> modifiers for filling color/data? Because what I have in mind is closer
>>>> to the blob structure, which can then be used for filling the plane:
>>>>
>>>> struct color_fill_blob {
>>>>       u32 pixel_format;
>>>>       u64 modifiers4];
>>>>       u32 pixel_data_size; // fixme: is this necessary?
>>>>       u8 pixel_data[];
>>>> };
>>>>
>>>> And then... This sounds a lot like a custom framebuffer.
>>>>
>>>> So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
>>>> flag to the framebuffers, which would e.g. mean that the FB gets stamped
>>>> all over the plane. This would also save us from changing if (!fb)
>>>> checks all over the drm core.
>>>>
>>>> Another approach might be using a format modifier instead of the FB flag.
>>>>
>>>> What do you think?
>>>
>>> First off, we only need to represent the value of a single pixel here. So I'm
>>> not quite following why we need format modifiers. Format modifiers describe how
>>> pixels are laid out in memory. Since there's a single pixel described, this
>>> is non-sensical to me, the format modifier is always LINEAR.
> 
> Agreed.
> 
>>>
>>> Then, I can understand why putting the pixel_format in there is tempting to
>>> guarantee future extensibility, but it also adds complexity. For instance, how
>>> does user-space figure out which formats can be used for COLOR_FILL? Can
>>> user-space use any format supported by the plane? What does it mean for
>>> multi-planar formats? Do we really want the kernel to have conversion logic for
>>> all existing formats? Do we need to also add a new read-only blob prop to
>>> indicate supported COLOR_FILL formats?

FWIW the formats supported by solid_fill wouldn't necessarily be all the 
formats supported by the plane (ex. for msm/dpu, solid_fill only 
supports all RGB color variants, though planes can normally support YUV 
formats too).

That being said, I'm ok with having the solid_fill take in only 
RGBA32323232 format based on the comments below.

> 
> Right. This does not seem to require pixel formats at all.
> 
> The point of pixel formats is to be able to feed large amounts of data
> as-is into hardware and avoid the CPU ever touching it. You do that
> with DRM FBs pointing to suitably allocated hardware buffers. But here
> we have exactly one pixel, which I imagine will always be read by the
> CPU so the driver will convert it into a hardware-specific format and
> program it; probably the driver will not create an internal DRM FB for
> it. >
> The above might also be a reason to not model this as a special-case
> DRM FB in UAPI. Or, at least you need a whole new ioctl to create such
> DRM FB to avoid the need to allocate e.g. a dumb buffer or a
> GPU-specific buffer. >
> What one does need is what Sebastian brought up: does it support alpha
> or not?
Hmm, the drm_plane struct already supports an alpha property so it seems 
a bit redundant to also have a separate alpha value in the solid fill color.

That being said, we could have it so that setting the alpha for the 
solid_fill property will also change the value of the plane's alpha 
property too.

> 
> Userspace would also be interested in the supported precision of the
> values, but the hardware pixel component order is irrelevant because the
> driver will always convert the one pixel with CPU anyway.
> 
> YUV vs. RGB is a another question. The KMS color pipeline is defined in
> terms of RGBA as far as I know, and alpha-blending YUV values makes no
> sense. So will there ever be any need to set an YUV fill? I have a hard
> time imagining it.
> 
> If you do set an YUV fill, the KMS color pipeline most likely needs to
> convert it to RGBA for blending, and then you need the plane properties
> COLOR_ENCODING and COLOR_RANGE.
> 
> But why bother when userspace can convert that one pixel to RGBA itself
> anyway?

Noted, I think this is reasonable.

> 
>>> We've recently-ish standardized a new Wayland protocol [1] which has the same
>>> purpose as this new kernel uAPI. The conclusion there was that using 32-bit
>>> values for each channel (R, G, B, A) would be enough for almost all use-cases.
>>> The driver can convert these high-precision values to what the hardware expects.
>>> The only concern was about sending values outside of the [0.0, 1.0] range,
>>> which may have HDR use-cases.
> 
> This is what I would suggest, yes. This representation has enough
> precision to be future-proof, and the driver will be converting the
> value anyway.
> 
> The question about values outside of the unit range is a good one, too.
> With Wayland, we can simply add another request to set a value in
> floating-point if that turns up necessary.
> 
> Whether that will ever be necessary is connected to how the DRM KMS
> abstract color pipeline is modelled, and that you must define from the
> beginning:
> 
> If DRM KMS gets color processing related plane properties like CTM,
> GAMMA or DEGAMMA (they already exist for CRTC, and these have been
> proposed for planes quite some time ago), does the fill color go
> through all these operations, or will the fill color skip all these
> operations and go straight to plane blending?

The fill color would still go through color processing operations, 
though FWIW the MSM driver doesn't support GAMMA/DEGAMMA.

> 
> Whether values outside of the unit range will ever be needed depends
> also on the userspace design. Userspace could choose the value range
> freely if the KMS color pipeline elements happen to support that range.
> However things like LUTs are naturally limited to unit range input, so
> using values outside of the unit range might be difficult if not
> impossible. Therefore I'm not concerned about this question much.
> 
>>> So, there are multiple ways to go about this. I can think of:
>>>
>>> - Put "RGBA32" in the name of the prop, and if we ever need a different
>>>    color format, pick a different name.
> 
> I could see problems with that if the same device supports more than
> one kind. How to turn off all but one color fill property?
> 
> What if userspace understands an old color fill property but not the
> new one, and the new one happens to be set for some reason? Userspace
> will attempt to use the old property without setting the new property
> to nil, and fail.
> 
>>> - Define a struct with an enum of possible fill kinds:
>>>    #define FILL_COLOR_RGBA32 1
>>>    #define FILL_COLOR_F32 2
>>>    struct color_fill_blob { u32 kind; u8 data[]; };
> 
> This could work.
> 
> Btw. maybe call it RGBA32323232 to make it follow the drm_fourcc naming
> convention. Some userspace libraries already use RGBA32 to mean 32 bit
> per pixel instead of per channel.
> 
> How will userspace know what kinds are supported?
> 
>>> - Define a struct with a version and RGBA values:
>>>    struct color_fill_blob { u32 version; u32 rgba[4]; };
>>>    If we need to add more formats later, or new metadata:
>>>    struct color_fill_blob2 { u32 version; /* new fields */ };
>>>    where version must be set to 2.
> 
> This could work.

Leaning towards this option.

Thanks,

Jessica Zhang

> 
>>> - Define a struct with a "pixel_format" prop, but force user-space to use a
>>>    fixed format for now. Later, if we need another format, add a new prop to
>>>    advertise supported formats.
>>> - More complicated solutions, e.g. advertise the list of supported formats from
>>>    the start.
> 
> Feels more complicated than necessary.
> 
> Anyway, the idea of creating a blob and then setting that into some KMS
> plane property sounds a very good mechanism.
> 
>>>
>>> [1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104
>>>   
>>
>> Agreeing with most of what you said here. However, what's the idea
>> behind a format anyway? The 4 values provided here are fed directly
>> into the color pipeline which seems to define the color channels it's
>> working on as RGBA (or doesn't define anything at all). The only
>> reason I can think of is that hardware might support only ingesting
>> values either in a format with high bit depth color channels and no
>> alpha or a format with low bit depth color but with alpha, so choosing
>> between the formats provides a real trade-off. Is that actually
>> something hardware might be restricted to or do they all just support
>> ingesting the color data with enough precision on every channel? >
> Right.
> 
> 
> Thanks,
> pq

Pekka Paalanen Nov. 24, 2022, 8:50 a.m. UTC | #18

On Wed, 23 Nov 2022 15:27:04 -0800
Jessica Zhang <[e-mail]> wrote:

> On 11/9/2022 1:18 AM, Pekka Paalanen wrote:
> > On Tue, 8 Nov 2022 23:01:47 +0100
> > Sebastian Wick <[e-mail]> wrote:
> >   
> >> On Tue, Nov 8, 2022 at 7:51 PM Simon Ser <[e-mail]> wrote:  
> >>>
> >>> cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.  
> > 
> > Hi all,
> > 
> > thanks! Comments below.  
> 
> Thanks for the feedback!
> 
> >   
> >>>
> >>> On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:
> >>>     
> >>>> On 29/10/2022 01:59, Jessica Zhang wrote:  
> >>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
> >>>>> drm_plane. In addition, add support for setting and getting the values
> >>>>> of these properties.
> >>>>>
> >>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
> >>>>> represents the format of the color fill. Userspace can set enable solid
> >>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
> >>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
> >>>>> framebuffer to NULL.
> >>>>>
> >>>>> Signed-off-by: Jessica Zhang <[e-mail]>  
> >>>>
> >>>> Planes report supported formats using the drm_mode_getplane(). You'd
> >>>> also need to tell userspace, which formats are supported for color fill.
> >>>> I don't think one supports e.g. YV12.
> >>>>
> >>>> A bit of generic comment for the discussion (this is an RFC anyway).
> >>>> Using color_fill/color_fill_format properties sounds simple, but this
> >>>> might be not generic enough. Limiting color_fill to 32 bits would
> >>>> prevent anybody from using floating point formats (e.g.
> >>>> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
> >>>> e.g. using 64-bit for the color_fill value, but then this doesn't sound
> >>>> extensible too much.
> >>>>
> >>>> So, a question for other hardware maintainers. Do we have hardware that
> >>>> supports such 'color filled' planes? Do we want to support format
> >>>> modifiers for filling color/data? Because what I have in mind is closer
> >>>> to the blob structure, which can then be used for filling the plane:
> >>>>
> >>>> struct color_fill_blob {
> >>>>       u32 pixel_format;
> >>>>       u64 modifiers4];
> >>>>       u32 pixel_data_size; // fixme: is this necessary?
> >>>>       u8 pixel_data[];
> >>>> };
> >>>>
> >>>> And then... This sounds a lot like a custom framebuffer.
> >>>>
> >>>> So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
> >>>> flag to the framebuffers, which would e.g. mean that the FB gets stamped
> >>>> all over the plane. This would also save us from changing if (!fb)
> >>>> checks all over the drm core.
> >>>>
> >>>> Another approach might be using a format modifier instead of the FB flag.
> >>>>
> >>>> What do you think?  
> >>>
> >>> First off, we only need to represent the value of a single pixel here. So I'm
> >>> not quite following why we need format modifiers. Format modifiers describe how
> >>> pixels are laid out in memory. Since there's a single pixel described, this
> >>> is non-sensical to me, the format modifier is always LINEAR.  
> > 
> > Agreed.
> >   
> >>>
> >>> Then, I can understand why putting the pixel_format in there is tempting to
> >>> guarantee future extensibility, but it also adds complexity. For instance, how
> >>> does user-space figure out which formats can be used for COLOR_FILL? Can
> >>> user-space use any format supported by the plane? What does it mean for
> >>> multi-planar formats? Do we really want the kernel to have conversion logic for
> >>> all existing formats? Do we need to also add a new read-only blob prop to
> >>> indicate supported COLOR_FILL formats?  
> 
> FWIW the formats supported by solid_fill wouldn't necessarily be all the 
> formats supported by the plane (ex. for msm/dpu, solid_fill only 
> supports all RGB color variants, though planes can normally support YUV 
> formats too).
> 
> That being said, I'm ok with having the solid_fill take in only 
> RGBA32323232 format based on the comments below.
> 
> > 
> > Right. This does not seem to require pixel formats at all.
> > 
> > The point of pixel formats is to be able to feed large amounts of data
> > as-is into hardware and avoid the CPU ever touching it. You do that
> > with DRM FBs pointing to suitably allocated hardware buffers. But here
> > we have exactly one pixel, which I imagine will always be read by the
> > CPU so the driver will convert it into a hardware-specific format and
> > program it; probably the driver will not create an internal DRM FB for
> > it. >
> > The above might also be a reason to not model this as a special-case
> > DRM FB in UAPI. Or, at least you need a whole new ioctl to create such
> > DRM FB to avoid the need to allocate e.g. a dumb buffer or a
> > GPU-specific buffer. >
> > What one does need is what Sebastian brought up: does it support alpha
> > or not?  
> Hmm, the drm_plane struct already supports an alpha property so it seems 
> a bit redundant to also have a separate alpha value in the solid fill color.

Hi Jessica,

that's a good point! - Assuming that if hardware supports fill with
alpha, it supports plane-alpha with real FBs as well.

> That being said, we could have it so that setting the alpha for the 
> solid_fill property will also change the value of the plane's alpha 
> property too.

No! Definitely not. That would be confusing.

One must not have properties that change the value of other
non-immutable properties. It would become a real mess to handle in
userspace and for backward compatibility. Just like the kernel must not
spontaneously change the value of a non-immutable property. (Some
mistakes exist already, and I think they cause userspace to need
exceptional code for them.)


> > 
> > Userspace would also be interested in the supported precision of the
> > values, but the hardware pixel component order is irrelevant because the
> > driver will always convert the one pixel with CPU anyway.
> > 
> > YUV vs. RGB is a another question. The KMS color pipeline is defined in
> > terms of RGBA as far as I know, and alpha-blending YUV values makes no
> > sense. So will there ever be any need to set an YUV fill? I have a hard
> > time imagining it.
> > 
> > If you do set an YUV fill, the KMS color pipeline most likely needs to
> > convert it to RGBA for blending, and then you need the plane properties
> > COLOR_ENCODING and COLOR_RANGE.
> > 
> > But why bother when userspace can convert that one pixel to RGBA itself
> > anyway?  
> 
> Noted, I think this is reasonable.
> 
> >   
> >>> We've recently-ish standardized a new Wayland protocol [1] which has the same
> >>> purpose as this new kernel uAPI. The conclusion there was that using 32-bit
> >>> values for each channel (R, G, B, A) would be enough for almost all use-cases.
> >>> The driver can convert these high-precision values to what the hardware expects.
> >>> The only concern was about sending values outside of the [0.0, 1.0] range,
> >>> which may have HDR use-cases.  
> > 
> > This is what I would suggest, yes. This representation has enough
> > precision to be future-proof, and the driver will be converting the
> > value anyway.
> > 
> > The question about values outside of the unit range is a good one, too.
> > With Wayland, we can simply add another request to set a value in
> > floating-point if that turns up necessary.
> > 
> > Whether that will ever be necessary is connected to how the DRM KMS
> > abstract color pipeline is modelled, and that you must define from the
> > beginning:
> > 
> > If DRM KMS gets color processing related plane properties like CTM,
> > GAMMA or DEGAMMA (they already exist for CRTC, and these have been
> > proposed for planes quite some time ago), does the fill color go
> > through all these operations, or will the fill color skip all these
> > operations and go straight to plane blending?  
> 
> The fill color would still go through color processing operations, 
> though FWIW the MSM driver doesn't support GAMMA/DEGAMMA.

That's ok. The important bit is to define what must happen *if* such
plane properties are exposed by a driver. If they are not exposed, no
problem.

Btw. I could easily expect disagreement between different hardware
here, so I think this part will need many eyes to review.

If hardware is hard-wired to feed the fill color straight to blending,
then if fill color UAPI is defined to go through per-plane color
processing, the driver needs to apply that color processing on the CPU
before programming the hardware.

If hardware allows processing the fill color through per-plane color
processing, but fill color UAPI is defined to feed straight to blending,
then the driver can simply ignore the per-plane color properties and
program pass-through to the hardware.

For userspace, I think the choice makes little difference. Userspace
can compensate for the choice the same way a driver would, except
userspace can perhaps use more precise calculation methods. OTOH, if
fill color is intended to match the color on a real FB on another
plane, not going through the exact same computations might cause error.
Whether that error is significant depends on the use case and is
impossible to say here and now.

The important bit is to make that choice and document it.

...

> >>> - Define a struct with a version and RGBA values:
> >>>    struct color_fill_blob { u32 version; u32 rgba[4]; };
> >>>    If we need to add more formats later, or new metadata:
> >>>    struct color_fill_blob2 { u32 version; /* new fields */ };
> >>>    where version must be set to 2.  
> > 
> > This could work.  
> 
> Leaning towards this option.

Yes, it seems the best to me too. Just rgb[3] rather than rgba[4],
given the discussion about the plane alpha property. Or even u32 r;
u32 g; u32 b; to avoid having to think about the index in code.


Thanks,
pq

> 
> Thanks,
> 
> Jessica Zhang
> 
> >   
> >>> - Define a struct with a "pixel_format" prop, but force user-space to use a
> >>>    fixed format for now. Later, if we need another format, add a new prop to
> >>>    advertise supported formats.
> >>> - More complicated solutions, e.g. advertise the list of supported formats from
> >>>    the start.  
> > 
> > Feels more complicated than necessary.
> > 
> > Anyway, the idea of creating a blob and then setting that into some KMS
> > plane property sounds a very good mechanism.
> >   
> >>>
> >>> [1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104
> >>>     
> >>
> >> Agreeing with most of what you said here. However, what's the idea
> >> behind a format anyway? The 4 values provided here are fed directly
> >> into the color pipeline which seems to define the color channels it's
> >> working on as RGBA (or doesn't define anything at all). The only
> >> reason I can think of is that hardware might support only ingesting
> >> values either in a format with high bit depth color channels and no
> >> alpha or a format with low bit depth color but with alpha, so choosing
> >> between the formats provides a real trade-off. Is that actually
> >> something hardware might be restricted to or do they all just support
> >> ingesting the color data with enough precision on every channel? >  
> > Right.
> > 
> > 
> > Thanks,
> > pq

Jessica Zhang Nov. 29, 2022, 6:53 p.m. UTC | #19

On 11/24/2022 12:50 AM, Pekka Paalanen wrote:
> On Wed, 23 Nov 2022 15:27:04 -0800
> Jessica Zhang <[e-mail]> wrote:
> 
>> On 11/9/2022 1:18 AM, Pekka Paalanen wrote:
>>> On Tue, 8 Nov 2022 23:01:47 +0100
>>> Sebastian Wick <[e-mail]> wrote:
>>>    
>>>> On Tue, Nov 8, 2022 at 7:51 PM Simon Ser <[e-mail]> wrote:
>>>>>
>>>>> cc'ing Pekka and wayland-devel for userspace devs feedback on the new uAPI.
>>>
>>> Hi all,
>>>
>>> thanks! Comments below.
>>
>> Thanks for the feedback!
>>
>>>    
>>>>>
>>>>> On Saturday, October 29th, 2022 at 14:08, Dmitry Baryshkov <[e-mail]> wrote:
>>>>>      
>>>>>> On 29/10/2022 01:59, Jessica Zhang wrote:
>>>>>>> Add support for COLOR_FILL and COLOR_FILL_FORMAT properties for
>>>>>>> drm_plane. In addition, add support for setting and getting the values
>>>>>>> of these properties.
>>>>>>>
>>>>>>> COLOR_FILL represents the color fill of a plane while COLOR_FILL_FORMAT
>>>>>>> represents the format of the color fill. Userspace can set enable solid
>>>>>>> fill on a plane by assigning COLOR_FILL to a uint64_t value, assigning
>>>>>>> the COLOR_FILL_FORMAT property to a uint32_t value, and setting the
>>>>>>> framebuffer to NULL.
>>>>>>>
>>>>>>> Signed-off-by: Jessica Zhang <[e-mail]>
>>>>>>
>>>>>> Planes report supported formats using the drm_mode_getplane(). You'd
>>>>>> also need to tell userspace, which formats are supported for color fill.
>>>>>> I don't think one supports e.g. YV12.
>>>>>>
>>>>>> A bit of generic comment for the discussion (this is an RFC anyway).
>>>>>> Using color_fill/color_fill_format properties sounds simple, but this
>>>>>> might be not generic enough. Limiting color_fill to 32 bits would
>>>>>> prevent anybody from using floating point formats (e.g.
>>>>>> DRM_FORMAT_XRGB16161616F, 64-bit value). Yes, this can be solved with
>>>>>> e.g. using 64-bit for the color_fill value, but then this doesn't sound
>>>>>> extensible too much.
>>>>>>
>>>>>> So, a question for other hardware maintainers. Do we have hardware that
>>>>>> supports such 'color filled' planes? Do we want to support format
>>>>>> modifiers for filling color/data? Because what I have in mind is closer
>>>>>> to the blob structure, which can then be used for filling the plane:
>>>>>>
>>>>>> struct color_fill_blob {
>>>>>>        u32 pixel_format;
>>>>>>        u64 modifiers4];
>>>>>>        u32 pixel_data_size; // fixme: is this necessary?
>>>>>>        u8 pixel_data[];
>>>>>> };
>>>>>>
>>>>>> And then... This sounds a lot like a custom framebuffer.
>>>>>>
>>>>>> So, maybe what should we do instead is to add new DRM_MODE_FB_COLOR_FILL
>>>>>> flag to the framebuffers, which would e.g. mean that the FB gets stamped
>>>>>> all over the plane. This would also save us from changing if (!fb)
>>>>>> checks all over the drm core.
>>>>>>
>>>>>> Another approach might be using a format modifier instead of the FB flag.
>>>>>>
>>>>>> What do you think?
>>>>>
>>>>> First off, we only need to represent the value of a single pixel here. So I'm
>>>>> not quite following why we need format modifiers. Format modifiers describe how
>>>>> pixels are laid out in memory. Since there's a single pixel described, this
>>>>> is non-sensical to me, the format modifier is always LINEAR.
>>>
>>> Agreed.
>>>    
>>>>>
>>>>> Then, I can understand why putting the pixel_format in there is tempting to
>>>>> guarantee future extensibility, but it also adds complexity. For instance, how
>>>>> does user-space figure out which formats can be used for COLOR_FILL? Can
>>>>> user-space use any format supported by the plane? What does it mean for
>>>>> multi-planar formats? Do we really want the kernel to have conversion logic for
>>>>> all existing formats? Do we need to also add a new read-only blob prop to
>>>>> indicate supported COLOR_FILL formats?
>>
>> FWIW the formats supported by solid_fill wouldn't necessarily be all the
>> formats supported by the plane (ex. for msm/dpu, solid_fill only
>> supports all RGB color variants, though planes can normally support YUV
>> formats too).
>>
>> That being said, I'm ok with having the solid_fill take in only
>> RGBA32323232 format based on the comments below.
>>
>>>
>>> Right. This does not seem to require pixel formats at all.
>>>
>>> The point of pixel formats is to be able to feed large amounts of data
>>> as-is into hardware and avoid the CPU ever touching it. You do that
>>> with DRM FBs pointing to suitably allocated hardware buffers. But here
>>> we have exactly one pixel, which I imagine will always be read by the
>>> CPU so the driver will convert it into a hardware-specific format and
>>> program it; probably the driver will not create an internal DRM FB for
>>> it. >
>>> The above might also be a reason to not model this as a special-case
>>> DRM FB in UAPI. Or, at least you need a whole new ioctl to create such
>>> DRM FB to avoid the need to allocate e.g. a dumb buffer or a
>>> GPU-specific buffer. >
>>> What one does need is what Sebastian brought up: does it support alpha
>>> or not?
>> Hmm, the drm_plane struct already supports an alpha property so it seems
>> a bit redundant to also have a separate alpha value in the solid fill color.
> 
> Hi Jessica,
> 
> that's a good point! - Assuming that if hardware supports fill with
> alpha, it supports plane-alpha with real FBs as well.
> 
>> That being said, we could have it so that setting the alpha for the
>> solid_fill property will also change the value of the plane's alpha
>> property too.
> 
> No! Definitely not. That would be confusing.
> 
> One must not have properties that change the value of other
> non-immutable properties. It would become a real mess to handle in
> userspace and for backward compatibility. Just like the kernel must not
> spontaneously change the value of a non-immutable property. (Some
> mistakes exist already, and I think they cause userspace to need
> exceptional code for them.)

Ah, got it -- will have the value be RGB323232 instead.

> 
> 
>>>
>>> Userspace would also be interested in the supported precision of the
>>> values, but the hardware pixel component order is irrelevant because the
>>> driver will always convert the one pixel with CPU anyway.
>>>
>>> YUV vs. RGB is a another question. The KMS color pipeline is defined in
>>> terms of RGBA as far as I know, and alpha-blending YUV values makes no
>>> sense. So will there ever be any need to set an YUV fill? I have a hard
>>> time imagining it.
>>>
>>> If you do set an YUV fill, the KMS color pipeline most likely needs to
>>> convert it to RGBA for blending, and then you need the plane properties
>>> COLOR_ENCODING and COLOR_RANGE.
>>>
>>> But why bother when userspace can convert that one pixel to RGBA itself
>>> anyway?
>>
>> Noted, I think this is reasonable.
>>
>>>    
>>>>> We've recently-ish standardized a new Wayland protocol [1] which has the same
>>>>> purpose as this new kernel uAPI. The conclusion there was that using 32-bit
>>>>> values for each channel (R, G, B, A) would be enough for almost all use-cases.
>>>>> The driver can convert these high-precision values to what the hardware expects.
>>>>> The only concern was about sending values outside of the [0.0, 1.0] range,
>>>>> which may have HDR use-cases.
>>>
>>> This is what I would suggest, yes. This representation has enough
>>> precision to be future-proof, and the driver will be converting the
>>> value anyway.
>>>
>>> The question about values outside of the unit range is a good one, too.
>>> With Wayland, we can simply add another request to set a value in
>>> floating-point if that turns up necessary.
>>>
>>> Whether that will ever be necessary is connected to how the DRM KMS
>>> abstract color pipeline is modelled, and that you must define from the
>>> beginning:
>>>
>>> If DRM KMS gets color processing related plane properties like CTM,
>>> GAMMA or DEGAMMA (they already exist for CRTC, and these have been
>>> proposed for planes quite some time ago), does the fill color go
>>> through all these operations, or will the fill color skip all these
>>> operations and go straight to plane blending?
>>
>> The fill color would still go through color processing operations,
>> though FWIW the MSM driver doesn't support GAMMA/DEGAMMA.
> 
> That's ok. The important bit is to define what must happen *if* such
> plane properties are exposed by a driver. If they are not exposed, no
> problem.
> 
> Btw. I could easily expect disagreement between different hardware
> here, so I think this part will need many eyes to review.

Got it -- I'm not aware of any other HW outside of MSM devices that 
supports a similar color fill feature, but if there are any that have 
something similar I'm open to learning about how they've implemented 
this feature and adjusting accordingly.

> 
> If hardware is hard-wired to feed the fill color straight to blending,
> then if fill color UAPI is defined to go through per-plane color
> processing, the driver needs to apply that color processing on the CPU
> before programming the hardware.
> 
> If hardware allows processing the fill color through per-plane color
> processing, but fill color UAPI is defined to feed straight to blending,
> then the driver can simply ignore the per-plane color properties and
> program pass-through to the hardware.
> 
> For userspace, I think the choice makes little difference. Userspace
> can compensate for the choice the same way a driver would, except
> userspace can perhaps use more precise calculation methods. OTOH, if
> fill color is intended to match the color on a real FB on another
> plane, not going through the exact same computations might cause error.
> Whether that error is significant depends on the use case and is
> impossible to say here and now.
> 
> The important bit is to make that choice and document it.

Acked.

Thanks,

Jessica Zhang

> 
> ...
> 
>>>>> - Define a struct with a version and RGBA values:
>>>>>     struct color_fill_blob { u32 version; u32 rgba[4]; };
>>>>>     If we need to add more formats later, or new metadata:
>>>>>     struct color_fill_blob2 { u32 version; /* new fields */ };
>>>>>     where version must be set to 2.
>>>
>>> This could work.
>>
>> Leaning towards this option.
> 
> Yes, it seems the best to me too. Just rgb[3] rather than rgba[4],
> given the discussion about the plane alpha property. Or even u32 r;
> u32 g; u32 b; to avoid having to think about the index in code.
> 
> 
> Thanks,
> pq
> 
>>
>> Thanks,
>>
>> Jessica Zhang
>>
>>>    
>>>>> - Define a struct with a "pixel_format" prop, but force user-space to use a
>>>>>     fixed format for now. Later, if we need another format, add a new prop to
>>>>>     advertise supported formats.
>>>>> - More complicated solutions, e.g. advertise the list of supported formats from
>>>>>     the start.
>>>
>>> Feels more complicated than necessary.
>>>
>>> Anyway, the idea of creating a blob and then setting that into some KMS
>>> plane property sounds a very good mechanism.
>>>    
>>>>>
>>>>> [1]: https://gitlab.freedesktop.org/wayland/wayland-protocols/-/merge_requests/104
>>>>>      
>>>>
>>>> Agreeing with most of what you said here. However, what's the idea
>>>> behind a format anyway? The 4 values provided here are fed directly
>>>> into the color pipeline which seems to define the color channels it's
>>>> working on as RGBA (or doesn't define anything at all). The only
>>>> reason I can think of is that hardware might support only ingesting
>>>> values either in a format with high bit depth color channels and no
>>>> alpha or a format with low bit depth color but with alpha, so choosing
>>>> between the formats provides a real trade-off. Is that actually
>>>> something hardware might be restricted to or do they all just support
>>>> ingesting the color data with enough precision on every channel? >
>>> Right.
>>>
>>>
>>> Thanks,
>>> pq
>

Patch

diff --git a/drivers/gpu/drm/drm_atomic_uapi.c b/drivers/gpu/drm/drm_atomic_uapi.c
index 79730fa1dd8e..e1664463fca4 100644
--- a/drivers/gpu/drm/drm_atomic_uapi.c
+++ b/drivers/gpu/drm/drm_atomic_uapi.c
@@ -544,6 +544,10 @@  static int drm_atomic_plane_set_property(struct drm_plane *plane,
 		state->src_w = val;
 	} else if (property == config->prop_src_h) {
 		state->src_h = val;
+	} else if (property == plane->color_fill_format_property) {
+		state->color_fill_format = val;
+	} else if (property == plane->color_fill_property) {
+		state->color_fill = val;
 	} else if (property == plane->alpha_property) {
 		state->alpha = val;
 	} else if (property == plane->blend_mode_property) {
@@ -616,6 +620,10 @@  drm_atomic_plane_get_property(struct drm_plane *plane,
 		*val = state->src_w;
 	} else if (property == config->prop_src_h) {
 		*val = state->src_h;
+	} else if (property == plane->color_fill_format_property) {
+		*val = state->color_fill_format;
+	} else if (property == plane->color_fill_property) {
+		*val = state->color_fill;
 	} else if (property == plane->alpha_property) {
 		*val = state->alpha;
 	} else if (property == plane->blend_mode_property) {
diff --git a/drivers/gpu/drm/drm_blend.c b/drivers/gpu/drm/drm_blend.c
index b4c8cab7158c..b8c2b263fa51 100644
--- a/drivers/gpu/drm/drm_blend.c
+++ b/drivers/gpu/drm/drm_blend.c
@@ -616,3 +616,41 @@  int drm_plane_create_blend_mode_property(struct drm_plane *plane,
 	return 0;
 }
 EXPORT_SYMBOL(drm_plane_create_blend_mode_property);
+
+int drm_plane_create_color_fill_property(struct drm_plane *plane)
+{
+	struct drm_property *prop;
+
+	prop = drm_property_create_range(plane->dev, 0, "color_fill",
+					 0, 0xffffffff);
+	if (!prop)
+		return -ENOMEM;
+
+	drm_object_attach_property(&plane->base, prop, 0);
+	plane->color_fill_property = prop;
+
+	if (plane->state)
+		plane->state->color_fill = 0;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_plane_create_color_fill_property);
+
+int drm_plane_create_color_fill_format_property(struct drm_plane *plane)
+{
+	struct drm_property *prop;
+
+	prop = drm_property_create_range(plane->dev, 0, "color_fill_format",
+					 0, 0xffffffff);
+	if (!prop)
+		return -ENOMEM;
+
+	drm_object_attach_property(&plane->base, prop, 0);
+	plane->color_fill_format_property = prop;
+
+	if (plane->state)
+		plane->state->color_fill_format = 0;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_plane_create_color_fill_format_property);
diff --git a/include/drm/drm_blend.h b/include/drm/drm_blend.h
index 88bdfec3bd88..3e96f5e83cce 100644
--- a/include/drm/drm_blend.h
+++ b/include/drm/drm_blend.h
@@ -58,4 +58,6 @@  int drm_atomic_normalize_zpos(struct drm_device *dev,
 			      struct drm_atomic_state *state);
 int drm_plane_create_blend_mode_property(struct drm_plane *plane,
 					 unsigned int supported_modes);
+int drm_plane_create_color_fill_property(struct drm_plane *plane);
+int drm_plane_create_color_fill_format_property(struct drm_plane *plane);
 #endif
diff --git a/include/drm/drm_plane.h b/include/drm/drm_plane.h
index 89ea54652e87..dcbfdb0e1f71 100644
--- a/include/drm/drm_plane.h
+++ b/include/drm/drm_plane.h
@@ -116,6 +116,20 @@  struct drm_plane_state {
 	/** @src_h: height of visible portion of plane (in 16.16) */
 	uint32_t src_h, src_w;
 
+	/**
+	 * @color_fill_format:
+	 * Format of the color fill value.
+	 */
+	uint32_t color_fill_format;
+
+	/**
+	 * @color_fill:
+	 * Fill color of the plane with 0 as black and 0xffffffff as white.
+	 * Can be set by user by setting the COLOR_FILL property. See
+	 * drm_plane_create_color_fill_property() for more details.
+	 */
+	uint32_t color_fill;
+
 	/**
 	 * @alpha:
 	 * Opacity of the plane with 0 as completely transparent and 0xffff as
@@ -699,6 +713,20 @@  struct drm_plane {
 	 */
 	struct drm_plane_state *state;
 
+	/*
+	 * @color_fill_format_property:
+	 * Optional color fill format property for this plane. See
+	 * drm_plane_create_color_fill_format_property().
+	 */
+	struct drm_property *color_fill_format_property;
+
+	/*
+	 * @color_fill_property:
+	 * Optional color fill property for this plane. See
+	 * drm_plane_create_color_fill_property().
+	 */
+	struct drm_property *color_fill_property;
+
 	/**
 	 * @alpha_property:
 	 * Optional alpha property for this plane. See