[PATCH 0/3] drm/amd/display: Fix FRL status polling work lifetime

From: Eren Özen via B4 Relay

Date: Mon Nov 09 2026 - 23:31:41 EST


The HDMI FRL status polling work and its workqueue, added in commit
5c9b8b27a883 ("drm/amd/display: Tie FRL support into amdgpu_dm"), have
three lifetime problems:

1. The delayed work is initialized in amdgpu_dm_connector_init_helper(),
which also runs at runtime for every new MST connector. A hotplug
while FRL polling is armed re-initializes a pending delayed work.

2. amdgpu_dm_fini() never cancels the self-rearming work and never
destroys the workqueue, so the workqueue and its rescuer thread
leak on every unbind, and the work can run after dc_destroy().

3. If the workqueue allocation fails, amdgpu_dm_atomic_commit_tail()
still passes the NULL workqueue to queue_delayed_work().

Patch 1 moves the initialization to amdgpu_dm_init(). Patch 2 tears
the work and workqueue down in amdgpu_dm_fini(), and depends on
patch 1. Patch 3 skips queueing when the workqueue does not exist.

These issues were pointed out by an automated review of the
workqueue rename patch earlier this year.

This series should combine cleanly with Timur's pending "Check HDMI
FRL support before creating polling workqueue". With that patch, the
workqueue is NULL on hardware without FRL, and the NULL checks in
patches 2 and 3 handle that case. If his patch lands first, patch 1
needs a trivial rebase because both touch the same context.

Testing:
- Built for x86_64 with GCC 16, no new warnings with W=1 on the
changed files.
- amdgpu_dm KUnit suite (UML): 1910/1912 pass both with and without
this series. The two failures are pre-existing
(dm_test_detect_hides_secondary_tile and
dm_test_dmub_sw_init_gtt_only_asic).
- Booted on a Cezanne (DCN 2.1) laptop with
CONFIG_DEBUG_OBJECTS_WORK and CONFIG_DEBUG_OBJECTS_TIMERS: one
polling workqueue thread, no ODEBUG warnings, and a clean s2idle
suspend/resume.
- Unbinding amdgpu through sysfs on the same laptop, then counting the
hdmi_frl_status_polling rescuer threads:
without this series (Fedora 7.2.15): 1 before, 1 after (leaked)
with this series: 1 before, 0 after
No ODEBUG warnings or splats during the unbind with this series.
- Not tested: MST hotplug with an active HDMI FRL stream, since I
have no FRL sink or MST hub. Cezanne has no HDMI FRL, so the
polling work itself never runs on my machine.

Signed-off-by: Eren Özen <erenozen@xxxxxxxxxxxxxx>
---
Eren Özen (3):
drm/amd/display: Initialize FRL status polling work once at DM init
drm/amd/display: Destroy FRL status polling workqueue in amdgpu_dm_fini()
drm/amd/display: Don't queue FRL status polling on a NULL workqueue

drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 12 +++++++++++-
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_connector.c | 2 --
2 files changed, 11 insertions(+), 3 deletions(-)
---
base-commit: 47df73aa49d1988c8f684b7caf8be6570e966844
change-id: 20261110-frl-polling-wq-teardown-2b3336cce9e6

Best regards,
--
Eren Özen <erenozen@xxxxxxxxxxxxxx>