From: Vitaly Prosyak <[email protected]>
This series fixes a lockdep false positive warning that occurs when memory
reclaim triggers MMU notifiers with amdgpu userptr buffers.
The root cause is that amdgpu_lockdep_init() was teaching lockdep the
incorrect lock ordering by calling fs_reclaim_acquire() while holding
notifier_lock. At runtime, memory reclaim establishes the opposite ordering,
creating a false circular dependency.
Patch 1 adds filelist_mutex to the lockdep hierarchy (preparation).
Patch 2 moves fs_reclaim registration before all locks to fix the false
positive.
This is a 7.2-cycle regression fix suitable for stable backport.
Testing: Mikhail Gavrilov's reproducer (IGT amd_lockdep test) runs 8 reclaim
cycles with no warnings after the fix. Before fix: warning on first cycle.
v3 changes (addressing Mikhail Gavrilov's technical review):
- Clarified that fs_reclaim_acquire/release pair only REGISTERS the
fs_reclaim lock class when called with no locks held (doesn't create
a static lockdep edge)
- Explained that the actual fs_reclaim -> notifier_lock dependency is
established at runtime when memory reclaim invokes MMU notifiers
- Added Cc: Arunpravin PaneerSelvam
v2 changes:
- Fixed author name: Michael -> Mikhail Gavrilov in all trailers
- Added Fixes: tag linking to original commit
- Added Tested-by: Mikhail Gavrilov
- Rewrote fs_reclaim comment to be technically accurate
- Noted this is a 7.2-cycle regression for stable routing
Vitaly Prosyak (2):
drm/amdgpu: add filelist_mutex to lockdep lock ordering annotations
drm/amdgpu: Fix lockdep false positive in amdgpu_lockdep_init
drivers/gpu/drm/amd/amdgpu/amdgpu_lockdep.c | 48 +++++++++++++++------
1 file changed, 35 insertions(+), 13 deletions(-)
--
2.54.0
