On 16/11/26 19:31, Hugo Sales via B4 Relay wrote:
A write to panel_power_savings stores the ABM level and sends a hotplug
uevent, but the level reaches the hardware only on the next atomic commit
that includes the eDP connector. Hyprland adds the connector to a commit
only on a modeset or when a connector property it sets changes, so on my
laptop the file read back the new level and the backlight never changed.
power-profiles-daemon writes this file on every profile, AC and battery
change, so none of its ABM changes applied

This shouldn't need a fix in every compositor. The file is for writers
that aren't the DRM master, like power-profiles-daemon, and they can't
commit anything themselves. Today they depend on the master taking a
generic hotplug uevent as a cue to commit the connector again. Nothing
documents that, and a compositor has no reason to read a hotplug event as
"a property you never touched has changed". A compositor that wants to own
ABM sets the 'adaptive backlight modulation' property, and from then on
the file returns -EBUSY. The file has to work without the master's help.
The backlight device already does: a brightness write goes straight to DC,
with no commit at all

This series makes the write commit the level itself:

   1. Render the existing kernel-doc, which no rst file included
   2. Add the CRTC's planes to a commit that changes the ABM level, which
      the forced modeset used to do. With VRR active, such a commit would
      otherwise carry only the primary plane
   3. Stop forcing a modeset when the connector's ABM level changes. DM
      already sends a new level as a stream update, and the commit that
      added the modeset flag also added the CRTC, which is all that update
      needs
   4. Move the store tests onto a full pipe the store can commit to
   5. Commit the level from the store through a duplicated connector
      state. The write returns the commit's error and no longer sends a
      uevent
   6. KUnit cases for a connector with the CRTC on, off and failing
      atomic check

Patch 3 also asks, below its cut, whether scaling can leave the same
check

 From my reading, other commits that add the CRTC but no planes, like a
change of the HDR metadata, can drop an overlay the same way while VRR is
active. That needs a fix in the CRTC check that pulls in only the primary
plane, which I'll send as a separate series

Hamza's v1 of the 2024 patch that added this file had no flush [1], and
Mario found that the level only took effect on the next modeset [2].
That patch's v2 changelog says "force an atomic commit" [3], but what
was merged, commit 63d0b87213a0 ("drm/amd/display: add
panel_power_savings sysfs entry to eDP connectors"), sends a hotplug
event instead

Commit 8d20a066fa9b ("drm/amd/display: change the panel power savings
level without a modeset") later applied the level outside atomic. It
broke igt abm_gradual, and commit 95aaa207e9ef ("Revert "drm/amd/display:
change the panel power savings level without a modeset"") reverted it.
The objections back then were to sysfs overriding the DRM master at all
[4], which the property above now covers. I didn't find any discussion
that rejects an internal atomic commit

A write now returns the commit's error instead of storing a level the
hardware may never get. power-profiles-daemon logs the error and tries
again on its next profile, AC or battery change. Nothing I found reads
the uevent for more than a cue to commit the connector, which the kernel
now does itself

The commit is blocking, so a write can hold the CRTC locks for up to a
frame while a compositor commit waits

Tested on my laptop (ASUS ROG Flow X13 GV301QH, DCN 2.1) with v3 on
7.3.4, under Hyprland with no modeset during the run:

   - 25 writes cycling levels 0 to 4 all succeeded and read back, with
     nothing in the kernel log. A write sends no uevent
   - I wrote 4, 0, 4, 0, 2 and 0 six seconds apart and read
     amdgpu_current_backlight_pwm and amdgpu_target_backlight_pwm every
     half second. After each 0, the backlight was back at full PWM by the
     next sample. After each 4 or 2, the target dropped, to 25% of full at
     4 and 52% at 2, and the backlight ramped down to it
   - This series on top of "drm/amd/display: Leave self-refresh before an
     ABM level update", with the panel in self-refresh before each write:
     all five writes that changed the level took the panel out of
     self-refresh for the update
   - Writes of the current level return without a commit, so they leave
     the panel in self-refresh

On a text console, igt amd_abm passed all five subtests. I also ran an
IGT test that puts an overlay plane over the primary plane and turns VRR
on, so the CRTC idles at the panel's 40 Hz minimum. On v2, a write took
the overlay off the screen until the next flip. With v3, the overlay
stayed. With VRR off, it stayed on both

[1] 
https://lore.kernel.org/amd-gfx/[email protected]/
[2] 
https://lore.kernel.org/amd-gfx/[email protected]/
[3] https://lore.kernel.org/amd-gfx/[email protected]/
[4] 
https://lore.kernel.org/amd-gfx/[email protected]/

Changes in v3:
- New patch 2: add the CRTC's planes to a commit that changes the
   level. Without the modeset, such a commit with VRR active carried only
   the primary plane, and DC dropped an active overlay until the next
   flip. Sashiko's finding, reproduced on v2
- Patch 3: credit David Weber, who sent the same condition change in
   August (Chenyu)
- Patch 5: say that a compositor commit waits for the write instead of
   failing with -EBUSY, in the kernel-doc and below the cut (Chenyu)
- Link to v2: 
https://lore.kernel.org/r/[email protected]

Changes in v2:
- Patch 3 (v2's 2): name commit b05e2c5e81f9 ("drm/amd/display: Update
   ABM crtc state on non-modeset") as the one that lets the non-modeset
   path carry the level, say the KMS property loses its modeset too, and
   say "drm/amd/display: Leave self-refresh before an ABM level update"
   has to go first
- Patch 4 (v2's 3): set level 2 before the test that writes 0, so the
   write changes the level
- Patch 5 (v2's 4): return without a commit when the level doesn't
   change (Harry). A write that changes nothing, or that the DRM master
   forbids, no longer wakes the GPU. A write that a signal interrupts
   restarts instead of failing with -EINTR
- Patch 6 (v2's 5): test a write of the current level, check that it and
   a forbidden write leave a runtime-suspended GPU suspended, and fold
   the CRTC on and off cases into one parameterised test
- Link to v1: 
https://lore.kernel.org/r/[email protected]

---
Hugo Sales (6):
       drm/amd/display: Render the panel power savings kernel-doc
       drm/amd/display: Add the CRTC's planes when the ABM level changes
       drm/amd/display: Don't force a modeset when the ABM level changes
       drm/amd/display: Give the panel_power_savings store tests a full pipe
       drm/amd/display: Commit panel_power_savings writes from the kernel
       drm/amd/display: Test panel_power_savings commits

  .../gpu/amdgpu/display/display-manager.rst         |   3 +
  drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  |  17 +-
  .../amd/display/amdgpu_dm/amdgpu_dm_backlight.c    | 118 ++++++++-
  .../amdgpu_dm/tests/amdgpu_dm_backlight_test.c     | 271 ++++++++++++++++++++-
  4 files changed, 390 insertions(+), 19 deletions(-)
---
base-commit: ba496bb837882eb1af58f41572fc289e46c77b8c
change-id: 20261101-abm-sysfs-commit-96a38de1ac66
prerequisite-change-id: 20261113-amdgpu-psr-abm-update-next-de20bc23b12c:v2
prerequisite-patch-id: c6cbeb4671914753ab7526369c2612039ae16504
prerequisite-patch-id: efc64b5a0a1bc2d46494d4797b0d25447b174076
prerequisite-patch-id: 496307c9687a947cfbfdcf1b6d0ad6ba1b6d4089

Best regards,
--
Hugo Sales <[email protected]>


Hi Hugo,

Tested this v3 series on top of your self-refresh v2 series on Rembrandt
(DCN 3.1.2), running 7.1.0-hugo-abm. All five amd_abm subtests, including
abm_gradual, passed on A/C.

My panel reports Sink support: no and Driver support: no [0xffffffff],
so I couldn't exercise the PSR-active case Chenyu requested.

Tested-by: Shivam Kalra <[email protected]>

Reply via email to