GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel sched/cache: use-after-free on mm_struct when a task switches address space during account_mm_sched()

UnscoredCVE-2026-98385Kernel, userspace & hypervisorcurated

Impact

The LLC cache-aware scheduling statistics lived inside mm_struct, so their lifetime was tied to the address space. account_mm_sched() on one CPU could dereference task->mm->sc_stat while the task was switching mm on another CPU, reaching a freed mm - a use-after-free found with KASAN. The fix extracts the stats into a separately refcounted sched_cache_group. This is core scheduler code, so it is reachable on any host running an affected kernel with cache-aware scheduling, including GPU nodes, and does not depend on a device driver. Exposure is local only: it needs the ability to run processes on the node, which on a GPU fleet means any tenant with shell or container execution. The practical risk is a kernel crash taking the whole node and its in-flight training jobs with it; memory corruption from a UAF race is harder to steer but not excluded by the record.

Who can reach it

Local - any unprivileged user or container that can run processes and trigger address-space switches on the host. No special device access and no network reachability required. Not exploitable remotely.

What to do

Take a stable kernel containing commits 28f9c0e0a0b9 / 852ecb6bc0bc and reboot each node on the affected series - the fix is a structural change to mm_struct and the scheduler, not a tunable, so there is no runtime mitigation and no way to patch this without a kernel replacement. Plan it as a rolling drain-and-reboot across the fleet; for GPU nodes that means evacuating long-running training jobs first. The kernel git record does not name a distribution release.

References

Related entries

All Kernel, userspace & hypervisor entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.