drm/amd/display: Fix streamless connector update and ABM modesets
cover drm/amd/display: Fix streamless connector update and ABM modesets
Hi,
This series contains two independent fixes for connector property updates in amdgpu_dm.
The first patch fixes a reproduced NULL-stream dereference and is marked for stable.
The second avoids the visible panel blank caused by modesetting for every ABM level change. This is useful for desktop environments that want to adjust ABM dynamically for AC, battery and low-battery profiles without making the screen go black briefly on every profile transition.
It would be nice to have the ABM fix in stable kernels too, but I am not sure whether this behavior issue justifies a stable backport. I therefore kept its Fixes tag but did not explicitly Cc stable.
David
David Weber (2):
drm/amd/display: Skip connector updates without a stream drm/amd/display: Avoid modeset for ABM changes
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
Signature · 1 line
2.54.0
PATCH 1/2 drm/amd/display: Skip connector updates without a stream
Atomic DPMS can leave a connector assigned to an enabled but inactive CRTC after the driver has removed its DC stream. A later atomic commit can then change HDR_OUTPUT_METADATA or Broadcast RGB while leaving the CRTC mode, routing and active state unchanged.
That property-only commit does not require a modeset, so it reaches the connector stream-update path with a NULL stream. An HDR metadata change passes the update bailout and reaches dc_stream_get_status(NULL), while a Broadcast RGB change dereferences the stream earlier when comparing its output color space.
Skip stream updates until a stream exists. The connector state remains stored, and an enabling modeset applies it to the newly created stream.
The crash was reproduced on a Phoenix1 system with DCN 3.1.4 and Linux 7.2-rc5 using a one-shot AI-generated DRM atomic reproducer:
The reproducer first sets HDR_OUTPUT_METADATA on an active CRTC, then sets CRTC ACTIVE=0 to remove the DC stream while keeping the connector routed to the CRTC, and finally changes HDR_OUTPUT_METADATA from one non-NULL blob to another. The last commit does not require a modeset and reached dc_stream_get_status(NULL) from amdgpu_dm_atomic_commit_tail(). The same sequence completed without crashing with this fix applied.
--- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +3 -0
@@ -11519,6 +11519,9 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_commit *state) dm_new_crtc_state =to_dm_crtc_state (new_crtc_state); dm_old_crtc_state =to_dm_crtc_state (old_crtc_state); +/* DPMS-off leaves the connector routed to a streamless CRTC. */ +if (!dm_new_crtc_state->stream) +continue ; scaling_changed =is_scaling_state_different (dm_new_con_state, dm_old_con_state);
Signature · 1 line
2.54.0
[email protected]
AI reviewRe: [PATCH 1/2] drm/amd/display: Skip connector updates without a stream
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Pre-existing issues:
- [High] The function amdgpu_dm_atomic_commit_tail uses kzalloc with GFP_KERNEL and acquires a mutex, both of which can sleep, despite being an atomic commit path.
- [High] AMDGPU never updates the privacy screen hardware state because privacy screen updates force a modeset, but the commit tail skips privacy screen updates during modesets. --
--- Patch [1]: [PATCH 1/2] drm/amd/display: Skip connector updates without a stream --- commit 3a3a91e5bab6bdb82ebcebf89c73cd8b740ca904
Author: David Weber <[email protected]>
drm/amd/display: Skip connector updates without a stream
Atomic DPMS can leave a connector assigned to an enabled but inactive CRTC after the driver has removed its DC stream. A later atomic commit can then change HDR_OUTPUT_METADATA or Broadcast RGB while leaving the CRTC mode, routing and active state unchanged.
This patch skips stream updates until a stream exists, preventing null pointer dereferences when the connector stream-update path is reached with a NULL stream.
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_commit *state) · line 5052
Quoted 8 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +0
-0@@ -5050,6 +5050,9 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_commit *state) dm_new_crtc_state =to_dm_crtc_state (new_crtc_state); dm_old_crtc_state =to_dm_crtc_state (old_crtc_state);
[Severity: High]
This isn't a bug introduced by this patch, but does this skip logic prevent privacy screen updates?
In amdgpu_dm_connector_atomic_check(), any change to the privacy screen software state forces a modeset:
new_crtc_state->mode_changed = true;
However, earlier in this loop in amdgpu_dm_atomic_commit_tail(), connectors undergoing a modeset are skipped:
/* Skip any modesets/resets */
if (!acrtc || drm_atomic_crtc_needs_modeset(new_crtc_state))
continue;Since the only call to update the privacy screen hardware is located at the very end of this loop:
drm_connector_update_privacy_screen(new_con_state);
Will this cause privacy screen updates to be permanently skipped because they force a modeset that bypasses the update code?
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_commit *state) · line 11527
Quoted 6 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +3
-0+/* DPMS-off leaves the connector routed to a streamless CRTC. */ +if (!dm_new_crtc_state->stream) +continue ; scaling_changed =is_scaling_state_different (dm_new_con_state, dm_old_con_state);
[Severity: High]
This is a pre-existing issue, but are there sleeping functions called from this atomic context?
Later in amdgpu_dm_atomic_commit_tail(), kzalloc is called with GFP_KERNEL:
dummy_updates = kzalloc(sizeof(struct dc_surface_update) * MAX_SURFACES, GFP_KERNEL);
And then a mutex is acquired:
mutex_lock(&dm->dc_lock);
Since drm_atomic_helper_commit_tail() and its variants execute in an atomic context where sleeping is forbidden, could this kzalloc(..., GFP_KERNEL) and mutex_lock() cause kernel warnings, system instability, or potential deadlocks?
Also, if the memory allocation fails, the subsequent hardware programming in dc_update_planes_and_stream() is silently skipped. Can this lead to an unrecoverable state mismatch since commit_tail cannot return error codes?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
PATCH 2/2 drm/amd/display: Avoid modeset for ABM changes
In reply to David WeberABM connector property changes currently set connectors_changed on the associated CRTC. DRM treats that flag as requiring a modeset, so each ABM change tears down and recreates the stream and visibly blanks the panel.
drm_atomic_get_crtc_state() already adds the affected CRTC to the atomic state. dm_update_crtc_state() then copies the connector ABM level to the CRTC state, allowing the commit tail to program it with the existing stream update path.
Stop marking the CRTC as connectors_changed for ABM changes. Continue doing so for scaling changes.
This makes runtime ABM changes practical for desktop power management, for example when switching between AC, battery and low-battery profiles.
Tested on a Phoenix1 system with DCN 3.1.4 and an internal eDP panel. Changing ABM through the panel_power_savings sysfs property was confirmed visually and by measuring the resulting change in system power consumption. It no longer caused the black-screen flicker from the forced modeset.
--- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +2 -3
@@ -13168,7 +13168,7 @@ static int amdgpu_dm_atomic_check(struct drm_device *dev,struct dm_connector_state *dm_old_con_state =to_dm_connector_state (old_con_state);struct dm_connector_state *dm_new_con_state =to_dm_connector_state (new_con_state); -/* Skip connectors that are disabled or part of modeset already. */ +/* Skip connectors without a CRTC. */ if (!new_con_state->crtc)continue ;
@@ -13179,8 +13179,7 @@ static int amdgpu_dm_atomic_check(struct drm_device *dev, goto fail; } -if (dm_old_con_state->abm_level != dm_new_con_state->abm_level || - dm_old_con_state->scaling != dm_new_con_state->scaling) +if (dm_old_con_state->scaling != dm_new_con_state->scaling) new_crtc_state->connectors_changed = true; }
Signature · 1 line
2.54.0
Re: [PATCH 2/2] drm/amd/display: Avoid modeset for ABM changes
Hi David,
Sorry this sat for so long. Hugo's series [1] carries the same change as its 2/5, together with a patch that commits panel_power_savings writes from the kernel, and it's going through our promotion testing this week, so it might be easiest to land it there. I've asked Hugo to credit you.
Without the forced modeset, an ABM change goes through the fast update while the panel may still be in PSR, and on DCN 2.1 that hangs DMUB. Hugo's other series [2] takes the panel out of self-refresh first, so this change has to land after it. For the same reason I'd keep it out of stable: the PSR events that series uses only exist since 3c108046e1d6.
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/[email protected]
Regards,
Chenyu