Linux kernel (drivers/gpu/drm/scheduler): When a process is killed with GPU work still queued, the scheduler entity

Impact

When a process is killed with GPU work still queued, the scheduler entity tears down without releasing the dependency fences it holds and without taking a reference on the fence it last scheduled. A tenant that repeatedly starts and kills jobs grows kernel memory on the shared node and exercises an unreferenced fence in the common scheduler that amdgpu and xe both sit on.

Who can reach it

A tenant container holding /dev/dri/renderD* (or /dev/kfd on an AMD node) submits GPU work with outstanding dependencies and kills the process before it completes, in a loop. No privilege beyond the compute/render device node; the affected code is the shared DRM scheduler, so it applies on the mainstream datacenter GPU drivers rather than a niche one.

What to do

Boot a kernel carrying the drm/sched entity fence-reference fix below. Interim: enforce per-container kernel-memory limits so repeated kill cycles cannot exhaust the node, and watch for unexplained slab growth on nodes running short-lived GPU jobs.

References