Skip to contentSkip to conversation

drm/amd/display: Copy crc_skip_count when duplicating CRTC state

PATCH In review Rodrigo Siqueira 3 participants 3 messages Jan 18, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amd/display: Copy crc_skip_count when duplicating CRTC state

From: Leo Li <[email protected]>

[Why]
crc_skip_count is used to track how many frames to skip to allow the OTG CRC engine to "warm up" before it outputs correct CRC values. Experimentally, this seems to be 2 frames.

When duplicating CRTC states, this value was not copied to the duplicated state. Therefore, when this state is committed, we will needlessly wait 2 frames before outputing CRC values. Even if the CRC engine is already warmed up.

[How]
Copy the crc_skip_count as part of dm_crtc_duplicate_state.

Cc Mark Yacoub <[email protected]>
Cc Hayden Goodfellow <[email protected]>
Cc Harry Wentland <[email protected]>
Cc Nicholas Choi <[email protected]>
---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 1 +
 1 file changed, 1 insertion(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +1 -0
index 87299e62fe12..5482b0925396 100644
@@ -6568,6 +6568,7 @@ dm_crtc_duplicate_state(struct drm_crtc *crtc)
 	state->freesync_config = cur->freesync_config;
 	state->cm_has_degamma = cur->cm_has_degamma;
 	state->cm_is_degamma_srgb = cur->cm_is_degamma_srgb;
+	state->crc_skip_count = cur->crc_skip_count;
 	state->force_dpms_off = cur->force_dpms_off;
 	/* TODO Duplicate dc_stream after objects are stream object is flattened */
 
Signature · 1 line
2.25.1
#1 lore raw
Quoted 7 lines · Rodrigo Siqueira

From: Leo Li <[email protected]>

[Why]
crc_skip_count is used to track how many frames to skip to allow the OTG CRC engine to "warm up" before it outputs correct CRC values. Experimentally, this seems to be 2 frames.

Quoted 6 lines

When duplicating CRTC states, this value was not copied to the duplicated state. Therefore, when this state is committed, we will needlessly wait 2 frames before outputing CRC values. Even if the CRC engine is already warmed up. >
[How]
Copy the crc_skip_count as part of dm_crtc_duplicate_state.

This likely introduces regressions.

Here's an example case where it can take two frames even after the CRTC is enabled:

1. VUPDATE is before line 0, in the front porch, counter=0
2. Flip arrives before VUPDATE is signaled, but does not finish programming until after VUPDATE point, counter=0.
3. Vblank counter increments, counter=1.
4. Flip programming finishes, counter=1.
5. OS delay happens, cursor programming is delayed, counter=1.
6. Cursor programming starts, counter=1.
7. VUPDATE fires, updating frame but missing cursor, counter=1.
8. Cursor programming finishes, counter=2.
9. Cursor programming pending for counter=2.

This is a little contrived, but I've seen something similar happen during IGT testing before.

This is because cursor update happens independent of the rest of plane programming and is tied to a separate lock. That lock part can't change due to potential for stuttering, but the first part could be fixed.

Regards,
Nicholas Kazlauskas

Quoted 24 lines
Cc Mark Yacoub <[email protected]>
Cc Hayden Goodfellow <[email protected]>
Cc Harry Wentland <[email protected]>
Cc Nicholas Choi <[email protected]>
---
  drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 1 +
  1 file changed, 1 insertion(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +1 -0
index 87299e62fe12..5482b0925396 100644
@@ -6568,6 +6568,7 @@ dm_crtc_duplicate_state(struct drm_crtc *crtc)
  	state->freesync_config = cur->freesync_config;
  	state->cm_has_degamma = cur->cm_has_degamma;
  	state->cm_is_degamma_srgb = cur->cm_is_degamma_srgb;
+	state->crc_skip_count = cur->crc_skip_count;
  	state->force_dpms_off = cur->force_dpms_off;
  	/* TODO Duplicate dc_stream after objects are stream object is flattened */
 
#2 lore raw
Quoted 29 lines · Nicholas
Quoted 7 lines · Rodrigo Siqueira

From: Leo Li <[email protected]>

[Why]
crc_skip_count is used to track how many frames to skip to allow the OTG CRC engine to "warm up" before it outputs correct CRC values. Experimentally, this seems to be 2 frames.

Quoted 6 lines

When duplicating CRTC states, this value was not copied to the duplicated state. Therefore, when this state is committed, we will needlessly wait 2 frames before outputing CRC values. Even if the CRC engine is already warmed up. >
[How]
Copy the crc_skip_count as part of dm_crtc_duplicate_state.

This likely introduces regressions.

Here's an example case where it can take two frames even after the CRTC is enabled:

1. VUPDATE is before line 0, in the front porch, counter=0
2. Flip arrives before VUPDATE is signaled, but does not finish programming until after VUPDATE point, counter=0.
3. Vblank counter increments, counter=1.
4. Flip programming finishes, counter=1.
5. OS delay happens, cursor programming is delayed, counter=1.
6. Cursor programming starts, counter=1.
7. VUPDATE fires, updating frame but missing cursor, counter=1.
8. Cursor programming finishes, counter=2.
9. Cursor programming pending for counter=2.

Quoted 1 line

This is a little contrived, but I've seen something similar happen during IGT testing before.

Hi Nick,

I'm wondering if we have a test that can reproduce the issue that you described. For this specific patch, I ran it through our pre-submission CI, and everything worked as expected, except the subtest pipe-a-primary-size-* in the kms_plane_cursor. However, I submitted the below IGT patch that fixed the issue:

https://patchwork.freedesktop.org/patch/469919/?series=98994&rev=1
(Note that the IGT fix is not related to the scenario you described.)

In summary, I do not see regressions in other tests, whereas this patch fixes multiple CRC mismatches when running IGT (e.g., kms_plane and plane_scaling).

Do you recommend running some specific test to reproduce the issue? Maybe some particular set of steps that enable me to see the issue?

Ps.: Some IGT tests already compensate for the two extra frames (see link [1]), this behavior also happens on other drivers such as i915, and we still fail in those tests because we are skipping four frames without this patch.
1. https://gitlab.freedesktop.org/drm/igt-gpu-tools/-/blob/master/tests/kms_plane.c#L612

Thanks
Siqueira

Quoted 29 lines

This is because cursor update happens independent of the rest of plane programming and is tied to a separate lock. That lock part can't change due to potential for stuttering, but the first part could be fixed.

Regards,
Nicholas Kazlauskas

Quoted 23 lines
Cc Mark Yacoub <[email protected]>
Cc Hayden Goodfellow <[email protected]>
Cc Harry Wentland <[email protected]>
Cc Nicholas Choi <[email protected]>

---
  drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 1 +

  1 file changed, 1 insertion(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +1 -0
index 87299e62fe12..5482b0925396 100644
@@ -6568,6 +6568,7 @@ dm_crtc_duplicate_state(struct drm_crtc *crtc)
       state->freesync_config = cur->freesync_config;
       state->cm_has_degamma = cur->cm_has_degamma;
       state->cm_is_degamma_srgb = cur->cm_is_degamma_srgb;
+    state->crc_skip_count = cur->crc_skip_count;
       state->force_dpms_off = cur->force_dpms_off;
       /* TODO Duplicate dc_stream after objects are stream object is flattened */