Linux kernel (drivers/gpu/drm/scheduler): When one tenant's scheduler entity is killed, its scheduled fences are not

Impact

When one tenant's scheduler entity is killed, its scheduled fences are not signalled, so any other tenant whose job depends on those fences waits forever. One workload dying - or one attacker deliberately killing processes in a loop - permanently hangs unrelated jobs belonging to different tenants on the same GPU, which is a cross-tenant availability break, not just a local crash.

Who can reach it

An unprivileged process in a container with /dev/dri/renderD* creates cross-process fence dependencies (shared dma-buf / sync_file / syncobj, which is how compositors, media pipelines and multi-process GPU workloads normally work) and then exits or is killed. Lives in the shared drm/scheduler layer, so every scheduler-based driver on the node is affected.

What to do

Update to a kernel containing the fix commits below. Interim: avoid sharing fences/syncobjs across tenant boundaries, and be prepared to reset the GPU (node drain plus device reset) to clear hung dependent jobs.

References