This patch series was initially motivated by a race condition (exposed
in PATCH 4/6) where we lacked synchronization for `job->file` access.
This led to use-after-free issues when a file descriptor was closed
while a job was still running.
However, beyond fixing this specific race, the series introduces
broader improvements to active job management and locking. While PATCH
1/6, 2/6, and 5/6 are primarily code refactors, PATCH 3/6 brings a
significant change to the locking scheme. Previously, all queues shared
the same spinlock, which caused unnecessary contention during high GPU
usage across different queues. PATCH 3/6 allows queues to operate more
independently.
Finally, PATCH 6/6 addresses a similar race condition to PATCH 4/6, but
this time, on the per-file descriptor reset counter.
Best Regards,
- Maíra
---
Maíra Canal (6):
drm/v3d: Store a pointer to `struct v3d_file_priv` inside each job
drm/v3d: Store the active job inside the queue's state
drm/v3d: Replace a global spinlock with a per-queue spinlock
drm/v3d: Address race-condition between per-fd GPU stats and fd release
drm/v3d: Synchronous operations can't timeout
drm/v3d: Protect per-fd reset counter against fd release
drivers/gpu/drm/v3d/v3d_drv.c | 14 ++++++-
drivers/gpu/drm/v3d/v3d_drv.h | 22 ++++-------
drivers/gpu/drm/v3d/v3d_fence.c | 11 +++---
drivers/gpu/drm/v3d/v3d_gem.c | 10 ++---
drivers/gpu/drm/v3d/v3d_irq.c | 68 +++++++++++++-------------------
drivers/gpu/drm/v3d/v3d_sched.c | 85 +++++++++++++++++++++++-----------------
drivers/gpu/drm/v3d/v3d_submit.c | 2 +-
7 files changed, 108 insertions(+), 104 deletions(-)
---
base-commit: ca2a6abdaee43808034cdb218428d2ed85fd3db8
change-id: 20250718-v3d-queue-lock-59babfb548bc
