CVE-2026-98510

In the Linux kernel, the following vulnerability has been resolved:

drm/xe/vm: nuke PTs only after unlinking contested VMAs

In xe vm close and put(), external-BO VMAs are queued on the contested list for deferred destruction via xe vma destroy unlocked(). However, xe vm pt destroy() was previously invoked before processing contested VMAs, destroying vm->pt root while those VMAs were still linked to their respective buffer objects (vm bo->list.gpuva).

If a concurrent thread evicts one of those shared buffer objects, xe bo trigger rebind() holding only bo->resv walks the BO's VMAs and, in fault mode, calls xe vm invalidate vma() -> xe pt zap ptes(). Because vm->pt root[tile->id] is already NULL, dereferencing pt->level causes a NULL ptr deref.

Fix this by deferring xe vm free scratch() and xe vm pt destroy() until after all contested VMAs have been unlinked and destroyed.

User is reporting hitting a NULL ptr deref in xe pt zap ptes(), which could be explained by this race.

(cherry picked from commit c2863648959489767f08892fd6e90577d2ea0b6a)