[0/6] drm: Allow the damage helpers to handle buffer damage

  Hello,

This series is to fix an issue that surfaced after damage clipping was
enabled for the virtio-gpu by commit 01f05940a9a7 ("drm/virtio: Enable
fb damage clips property for the primary plane").

After that change, flickering artifacts was reported to be present with
both weston and wlroots wayland compositors when running in a virtual
machine. The cause was identified by Sima Vetter, who pointed out that
virtio-gpu does per-buffer uploads and for this reason it needs to do
a buffer damage handling, instead of frame damage handling.

Their suggestion was to extend the damage helpers to cover that case
and given that there's isn't a buffer damage accumulation algorithm
(e.g: buffer age), just do a full plane update if the framebuffer that
is attached to a plane changed since the last plane update (page-flip).

Patch #1 is just a refactoring to allow the logic of the frame damage
helpers to be shared by the buffer damage helpers.

Patch #2 adds the helpers that are needed for buffer damage handling.

Patch #3 fixes the virtio-gpu damage handling logic by using the
helper that is required by drivers that need to handle buffer damage.

Patch #4 fixes the vmwgfx similarly, since that driver also needs to
handle buffer damage and should have the same issue (although I have
not tested it due not having a VMWare setup).

Patch #5 adds to the KMS damage tracking kernel-doc some paragraphs
about damage tracking types and references to links that explain
frame damage vs buffer damage.

Finally patch #6 adds an item to the DRM/KMS todo, about the need to
implement some buffer damage accumulation algorithm instead of just
doing a full plane update in this case.

Because commit 01f05940a9a7 landed in v6.4, the first three patches
are marked as Fixes and Cc stable.

I've tested this on a VM with weston, was able to reproduce the issue
reported and the patches did fix the problem.

Please let me know what you think. Specially on the wording since could
made mistakes due just learning about these concepts yesterday thanks to
Sima, Simon and Pekka.

Best regards,
Javier


Javier Martinez Canillas (6):
  drm: Move drm_atomic_helper_damage_{iter_init,merged}() to helpers
  drm: Add drm_atomic_helper_buffer_damage_{iter_init,merged}() helpers
  drm/virtio: Use drm_atomic_helper_buffer_damage_merged() for buffer
    damage
  drm/vmwgfx: Use drm_atomic_helper_buffer_damage_iter_init() for buffer
    damage
  drm/plane: Extend damage tracking kernel-doc
  drm/todo: Add entry about implementing buffer age for damage tracking

 Documentation/gpu/todo.rst             |  20 +++
 drivers/gpu/drm/drm_damage_helper.c    | 166 +++++++++++++++++++------
 drivers/gpu/drm/drm_plane.c            |  22 +++-
 drivers/gpu/drm/virtio/virtgpu_plane.c |   2 +-
 drivers/gpu/drm/vmwgfx/vmwgfx_kms.c    |   2 +-
 include/drm/drm_damage_helper.h        |   7 ++
 6 files changed, 173 insertions(+), 46 deletions(-)
  

Hi Javier

Am 09.11.23 um 18:24 schrieb Javier Martinez Canillas:
> Hello,
> 
> This series is to fix an issue that surfaced after damage clipping was
> enabled for the virtio-gpu by commit 01f05940a9a7 ("drm/virtio: Enable
> fb damage clips property for the primary plane").
> 
> After that change, flickering artifacts was reported to be present with
> both weston and wlroots wayland compositors when running in a virtual
> machine. The cause was identified by Sima Vetter, who pointed out that
> virtio-gpu does per-buffer uploads and for this reason it needs to do
> a buffer damage handling, instead of frame damage handling.

I'm having problem understanding the types of damage. You never say what 
buffer damage is. I also don't know what a frame is in this context.

Regular damage handling marks parts of a plane as dirty/damaged. That is 
per-plane damage handling. The individual planes more or less 
independent from each other.

Buffer damage, I guess, marks the underlying buffer as dirty and 
requires synchronization of the buffer with some backing storage. The 
planes using that buffer are then updated more or less automatically.

Is that right?

And why does it flicker? Is there old data stored somewhere?

Best regards
Thomas

> 
> Their suggestion was to extend the damage helpers to cover that case
> and given that there's isn't a buffer damage accumulation algorithm
> (e.g: buffer age), just do a full plane update if the framebuffer that
> is attached to a plane changed since the last plane update (page-flip).
> 
> Patch #1 is just a refactoring to allow the logic of the frame damage
> helpers to be shared by the buffer damage helpers.
> 
> Patch #2 adds the helpers that are needed for buffer damage handling.
> 
> Patch #3 fixes the virtio-gpu damage handling logic by using the
> helper that is required by drivers that need to handle buffer damage.
> 
> Patch #4 fixes the vmwgfx similarly, since that driver also needs to
> handle buffer damage and should have the same issue (although I have
> not tested it due not having a VMWare setup).
> 
> Patch #5 adds to the KMS damage tracking kernel-doc some paragraphs
> about damage tracking types and references to links that explain
> frame damage vs buffer damage.
> 
> Finally patch #6 adds an item to the DRM/KMS todo, about the need to
> implement some buffer damage accumulation algorithm instead of just
> doing a full plane update in this case.
> 
> Because commit 01f05940a9a7 landed in v6.4, the first three patches
> are marked as Fixes and Cc stable.
> 
> I've tested this on a VM with weston, was able to reproduce the issue
> reported and the patches did fix the problem.
> 
> Please let me know what you think. Specially on the wording since could
> made mistakes due just learning about these concepts yesterday thanks to
> Sima, Simon and Pekka.
> 
> Best regards,
> Javier
> 
> 
> Javier Martinez Canillas (6):
>    drm: Move drm_atomic_helper_damage_{iter_init,merged}() to helpers
>    drm: Add drm_atomic_helper_buffer_damage_{iter_init,merged}() helpers
>    drm/virtio: Use drm_atomic_helper_buffer_damage_merged() for buffer
>      damage
>    drm/vmwgfx: Use drm_atomic_helper_buffer_damage_iter_init() for buffer
>      damage
>    drm/plane: Extend damage tracking kernel-doc
>    drm/todo: Add entry about implementing buffer age for damage tracking
> 
>   Documentation/gpu/todo.rst             |  20 +++
>   drivers/gpu/drm/drm_damage_helper.c    | 166 +++++++++++++++++++------
>   drivers/gpu/drm/drm_plane.c            |  22 +++-
>   drivers/gpu/drm/virtio/virtgpu_plane.c |   2 +-
>   drivers/gpu/drm/vmwgfx/vmwgfx_kms.c    |   2 +-
>   include/drm/drm_damage_helper.h        |   7 ++
>   6 files changed, 173 insertions(+), 46 deletions(-)
> 

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
  

Thomas Zimmermann <tzimmermann@suse.de> writes:

> Hi Javier
>
> Am 09.11.23 um 18:24 schrieb Javier Martinez Canillas:
>> Hello,
>> 
>> This series is to fix an issue that surfaced after damage clipping was
>> enabled for the virtio-gpu by commit 01f05940a9a7 ("drm/virtio: Enable
>> fb damage clips property for the primary plane").
>> 
>> After that change, flickering artifacts was reported to be present with
>> both weston and wlroots wayland compositors when running in a virtual
>> machine. The cause was identified by Sima Vetter, who pointed out that
>> virtio-gpu does per-buffer uploads and for this reason it needs to do
>> a buffer damage handling, instead of frame damage handling.
>
> I'm having problem understanding the types of damage. You never say what 
> buffer damage is. I also don't know what a frame is in this context.
>
> Regular damage handling marks parts of a plane as dirty/damaged. That is 
> per-plane damage handling. The individual planes more or less 
> independent from each other.
>
> Buffer damage, I guess, marks the underlying buffer as dirty and 
> requires synchronization of the buffer with some backing storage. The 
> planes using that buffer are then updated more or less automatically.
>
> Is that right?
>

In both cases the damage tracking information is the same, they mark
the damaged regions on the plane in framebuffer coordinates of the
framebuffer attached to the plane.

The problem as far as I understand is whether the driver expects that
to determine the area that changed in the plane (and a plane flush is
enough) or the area that changed since that same buffer was last used.

> And why does it flicker? Is there old data stored somewhere?
>

It flickers because the framebuffer changed and so the damage tracking
is not used correctly to flush the damaged areas to the backing storage.

This is my understanding at least, please Sima or Simon correct me if I
got this wrong.

> Best regards
> Thomas
>
  

Hi

Am 14.11.23 um 17:28 schrieb Javier Martinez Canillas:
> Thomas Zimmermann <tzimmermann@suse.de> writes:
> 
>> Hi Javier
>>
>> Am 09.11.23 um 18:24 schrieb Javier Martinez Canillas:
>>> Hello,
>>>
>>> This series is to fix an issue that surfaced after damage clipping was
>>> enabled for the virtio-gpu by commit 01f05940a9a7 ("drm/virtio: Enable
>>> fb damage clips property for the primary plane").
>>>
>>> After that change, flickering artifacts was reported to be present with
>>> both weston and wlroots wayland compositors when running in a virtual
>>> machine. The cause was identified by Sima Vetter, who pointed out that
>>> virtio-gpu does per-buffer uploads and for this reason it needs to do
>>> a buffer damage handling, instead of frame damage handling.
>>
>> I'm having problem understanding the types of damage. You never say what
>> buffer damage is. I also don't know what a frame is in this context.
>>
>> Regular damage handling marks parts of a plane as dirty/damaged. That is
>> per-plane damage handling. The individual planes more or less
>> independent from each other.
>>
>> Buffer damage, I guess, marks the underlying buffer as dirty and
>> requires synchronization of the buffer with some backing storage. The
>> planes using that buffer are then updated more or less automatically.
>>
>> Is that right?
>>
> 
> In both cases the damage tracking information is the same, they mark
> the damaged regions on the plane in framebuffer coordinates of the
> framebuffer attached to the plane.
> 
> The problem as far as I understand is whether the driver expects that
> to determine the area that changed in the plane (and a plane flush is
> enough) or the area that changed since that same buffer was last used.
> 
>> And why does it flicker? Is there old data stored somewhere?
>>
> 
> It flickers because the framebuffer changed and so the damage tracking
> is not used correctly to flush the damaged areas to the backing storage.

I think I got it from the links in patch 5.  In out other drivers, 
there's a single backing storage for each plane (for example in the 
video memory). Here, there's a backing storage for each buffer. On page 
flips, the plane changes its backing storage.  Our GEM buffer is up to 
date, but the respective backing storage is missing all the intermediate 
changes.

If I'm not mistaken, an entirely different solution would be to 
implement a per-plane back storage in these drivers.

Best regards
Thomas

> 
> This is my understanding at least, please Sima or Simon correct me if I
> got this wrong.
> 
>> Best regards
>> Thomas
>>
> 

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
  

Thomas Zimmermann <tzimmermann@suse.de> writes:

Hello Thomas,

> Hi

[...]

>>> And why does it flicker? Is there old data stored somewhere?
>>>
>> 
>> It flickers because the framebuffer changed and so the damage tracking
>> is not used correctly to flush the damaged areas to the backing storage.
>
> I think I got it from the links in patch 5.  In out other drivers, 
> there's a single backing storage for each plane (for example in the 
> video memory). Here, there's a backing storage for each buffer. On page

Correct, that's what I understood too.

> flips, the plane changes its backing storage.  Our GEM buffer is up to 
> date, but the respective backing storage is missing all the intermediate 
> changes.
>
> If I'm not mistaken, an entirely different solution would be to 
> implement a per-plane back storage in these drivers.
>

I believe so but I'm not sure if that's possible since the virtio-gpu spec
defines that the VM should send a VIRTIO_GPU_CMD_RESOURCE_FLUSH to the VMM
in the host to do an update and the granularity for that is a framebuffer.

For that reason the only solution (other than forcing a full plane update
like this patch-set does) is to implement tracking suppor for buffer damage.

> Best regards
> Thomas
>