Skip to contentSkip to conversation

drm/atomic: Clear plane pipeline for legacy clients

PATCH In review Robert Mader 4 participants 4 messages Jul 9 – Jul 30

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/atomic: Clear plane pipeline for legacy clients

Clients that do not enable DRM_CLIENT_CAP_PLANE_COLOR_PIPELINE and try to apply a new plane state currently have no way to unset/reset an existing non-bypass color pipeline, resulting in unexpected behavior.

As we can be sure that such clients do not expect plane pipelines to be set, let's reset the later for them, ensuring clean bypass-only pipeline states.

This is a less comprehensive but more backward compatible approach compared to a DRM_MODE_ATOMIC_RESET flag (link) that is currently being discussed. Considering the simplicity and clearly defined behavior of this case, this change could serve as an additional stop-gap solution to smoothen the transition to the new APIs, resulting in less breakage on systems with existing clients.

---

If an approach like this was previously rejected or clearly goes against API rules (or is unacceptable for other reason), please excuse me (and just let me know).

The main motivation for the patch came up during testing the Weston implementation (link) for the fixed-matrix color-ops. The test there checks both the color pipeline and legacy property ways of offloading YCbCr buffers and the current reset-less situation results in the legacy one failing once the color pipeline test ran before.

---
 drivers/gpu/drm/drm_atomic.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/drm_atomic.c +3 -0
index 735ab7badc2e..581fde90dbd8 100644
@@ -656,6 +656,9 @@ drm_atomic_get_plane_state(struct drm_atomic_commit *state,
 	state->planes[index].new_state = plane_state;
 	plane_state->state = state;
 
+	if (!state->plane_color_pipeline)
+		plane_state->color_pipeline = NULL;
+
 	drm_dbg_atomic(plane->dev, "Added [PLANE:%d:%s] %p state to %p\n",
 		       plane->base.id, plane->name, plane_state, state);
 
Signature · 1 line
2.55.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [High] Implicitly clearing `color_pipeline` in `drm_atomic_get_plane_state` without adding the affected…

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] Implicitly clearing color_pipeline in drm_atomic_get_plane_state without adding the affected colorops causes hardware state desynchronization.
- [Medium] Piecemeal reset of color pipelines for legacy clients leads to visually inconsistent planes on the same CRTC.
--

--- Patch [1]: [PATCH v1] drm/atomic: Clear plane pipeline for legacy clients --- commit 58ecff67ea2c28b7e6969be3a82eb9da9e435470
Author: Robert Mader <[email protected]>

drm/atomic: Clear plane pipeline for legacy clients

This commit clears the color_pipeline pointer in drm_atomic_get_plane_state() for legacy clients that do not enable DRM_CLIENT_CAP_PLANE_COLOR_PIPELINE. This ensures non-bypass color pipelines are reset to clean bypass-only states.

drivers/gpu/drm/drm_atomic.c drm_atomic_get_plane_state(struct drm_atomic_commit *state, · line 660

Quoted 10 lines
drivers/gpu/drm/drm_atomic.c +2 -0
index 735ab7badc2ec..581fde90dbd8d 100644
@@ -656,6 +656,9 @@ drm_atomic_get_plane_state(struct drm_atomic_commit *state,
 	state->planes[index].new_state = plane_state;
 	plane_state->state = state;
 
+	if (!state->plane_color_pipeline)
+		plane_state->color_pipeline = NULL;

[Severity: High]
Does clearing the color_pipeline here prevent the old colorops from being added to the atomic transaction?

If a legacy client takes over the display after a modern compositor configured a color pipeline, setting plane_state->color_pipeline to NULL means callers will skip adding the affected colorops.

This appears to orphan the old active colorops from the atomic commit, meaning driver hardware programming will not see the transition to disable them, leaving the hardware out of sync with software state.

[Severity: Medium]
Will this piecemeal reset lead to visually inconsistent planes on the same CRTC?

Because this reset is tied to plane state retrieval, a legacy client performing a partial update (like DRM_IOCTL_MODE_CURSOR) will only add the specific target plane to the transaction.

Other active planes on the same CRTC would not have drm_atomic_get_plane_state() called for them, so their complex color pipelines would remain fully active while the updated plane is forced to bypass.

Quoted 4 lines
drivers/gpu/drm/drm_atomic.c +1 -0
+
 	drm_dbg_atomic(plane->dev, "Added [PLANE:%d:%s] %p state to %p\n",
 		       plane->base.id, plane->name, plane_state, state);
 
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Robert Mader
Quoted 41 lines · Robert Mader

Clients that do not enable DRM_CLIENT_CAP_PLANE_COLOR_PIPELINE and try to apply a new plane state currently have no way to unset/reset an existing non-bypass color pipeline, resulting in unexpected behavior.

As we can be sure that such clients do not expect plane pipelines to be set, let's reset the later for them, ensuring clean bypass-only pipeline states.

This is a less comprehensive but more backward compatible approach compared to a DRM_MODE_ATOMIC_RESET flag (link) that is currently being discussed. Considering the simplicity and clearly defined behavior of this case, this change could serve as an additional stop-gap solution to smoothen the transition to the new APIs, resulting in less breakage on systems with existing clients.

---

If an approach like this was previously rejected or clearly goes against API rules (or is unacceptable for other reason), please excuse me (and just let me know).

The main motivation for the patch came up during testing the Weston implementation (link) for the fixed-matrix color-ops. The test there checks both the color pipeline and legacy property ways of offloading YCbCr buffers and the current reset-less situation results in the legacy one failing once the color pipeline test ran before.

---
 drivers/gpu/drm/drm_atomic.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/drm_atomic.c +0 -0
index 735ab7badc2e..581fde90dbd8 100644

drivers/gpu/drm/drm_atomic.c drm_atomic_get_plane_state(struct drm_atomic_commit *state, · line 664

Quoted 9 lines
drivers/gpu/drm/drm_atomic.c +3 -0
@@ -656,6 +656,9 @@ drm_atomic_get_plane_state(struct drm_atomic_commit *state,
 	state->planes[index].new_state = plane_state;
 	plane_state->state = state;
 
+	if (!state->plane_color_pipeline)
+		plane_state->color_pipeline = NULL;
+
 	drm_dbg_atomic(plane->dev, "Added [PLANE:%d:%s] %p state to %p\n",
 		       plane->base.id, plane->name, plane_state, state);
 

Is drm_atomic_get_plane_state always called for all planes which might have a color pipeline enabled?

If yes, this looks reasonable.

Signature · 2 lines
Earthling Michel Dänzer       \        GNOME / Xwayland / Mesa developer
https://redhat.com             \               Libre software enthusiast
Acked-by on v1 1/1
#3 lore raw
In reply to Robert Mader
Quoted 31 lines · On Thu, 9 Jul 2026 13:58:55 +0200 Robert Mader <[email protected]>…

Clients that do not enable DRM_CLIENT_CAP_PLANE_COLOR_PIPELINE and try to apply a new plane state currently have no way to unset/reset an existing non-bypass color pipeline, resulting in unexpected behavior.

As we can be sure that such clients do not expect plane pipelines to be set, let's reset the later for them, ensuring clean bypass-only pipeline states.

This is a less comprehensive but more backward compatible approach compared to a DRM_MODE_ATOMIC_RESET flag (link) that is currently being discussed. Considering the simplicity and clearly defined behavior of this case, this change could serve as an additional stop-gap solution to smoothen the transition to the new APIs, resulting in less breakage on systems with existing clients.

---

If an approach like this was previously rejected or clearly goes against API rules (or is unacceptable for other reason), please excuse me (and just let me know).

The main motivation for the patch came up during testing the Weston implementation (link) for the fixed-matrix color-ops. The test there checks both the color pipeline and legacy property ways of offloading YCbCr buffers and the current reset-less situation results in the legacy one failing once the color pipeline test ran before.

Quoted 4 lines
---
 drivers/gpu/drm/drm_atomic.c | 3 +++
 1 file changed, 3 insertions(+)

Hi Robert,

this sounds reasonable indeed.

Thanks,
pq

Quoted 14 lines
drivers/gpu/drm/drm_atomic.c +3 -0
index 735ab7badc2e..581fde90dbd8 100644
@@ -656,6 +656,9 @@ drm_atomic_get_plane_state(struct drm_atomic_commit *state,
 	state->planes[index].new_state = plane_state;
 	plane_state->state = state;
 
+	if (!state->plane_color_pipeline)
+		plane_state->color_pipeline = NULL;
+
 	drm_dbg_atomic(plane->dev, "Added [PLANE:%d:%s] %p state to %p\n",
 		       plane->base.id, plane->name, plane_state, state);