Linux kernel AMD XDNA driver: error path closes a live VMA and underflows its references

UnscoredCVE-2026-74721NVIDIA / GPU stackcurated

Impact

Two error paths in amdxdna_insert_pages() called vma->vm_ops->close(vma) before returning an error, dropping the shmem pages_pin_count and the GEM object reference that back the VMA while the mmap syscall has not even returned. The VMA stays alive, so kernel teardown calls close() a second time when the process unmaps the range, producing a reference count underflow and freeing the backing object early. The record does not establish controllable exploitation beyond that underflow, but a local user who can make the page insertion fail can reach it. Applies only to hosts with AMD XDNA NPU hardware and the amdxdna driver loaded.

Who can reach it

A local user or container with access to the amdxdna accel device node, mmapping a buffer object under conditions where page insertion hits one of the error paths. No elevated privileges beyond device access.

What to do

Update to a stable kernel with the deferred-fault fix and reboot. In the interim, gate access to the amdxdna accel device node with a device cgroup rule so only workloads that actually use the NPU can reach the path; unloaded drivers are not exposed.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-74721 https://git.kernel.org/stable/c/1501e4d07c6fee0d50531a0d1cb2be01a63e6e75 https://git.kernel.org/stable/c/8d51e0fd3e698919d2adeff71936377f0c0d4aa0

Related entries

All NVIDIA / GPU stack 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.