GPU VulnDB

Database/NVIDIA / GPU stack

Linux kernel amdgpu: duplicate FENCE chunks in one submission leak a buffer-object reference per submit

CVE-2026-80539NVIDIA / GPU stackcurated

Impact

amdgpu_cs_pass1() dispatches on chunk_id without rejecting repeated ids, and p->uf_bo is a single slot. A submission carrying two AMDGPU_CHUNK_ID_FENCE chunks runs amdgpu_cs_p1_user_fence() twice, overwriting p->uf_bo with a newly referenced BO while the first reference is never dropped; amdgpu_cs_parser_fini() only unrefs the final one. The commit notes the leaked BO outlives handle close and process exit, so the memory is not reclaimed when the job ends. On a shared GPU node any tenant holding a render node can repeat this in a loop and pin VRAM or GTT until allocations for other tenants fail, and the only reliable recovery is a node reboot.

Who can reach it

Local user with access to an amdgpu render node (/dev/dri/renderD*) - which is exactly what a GPU container is given. No authentication beyond device access; submissions are ordinary ioctl calls, not privileged ones.

What to do

Apply the stable fix, which rejects duplicate FENCE chunks the same way commit fec5f8e8c6bc did for BO_HANDLES. Five stable commits are listed and the record names no fixed release version. Rollout is a patched kernel plus a per-node reboot; drain the node first. Leaked memory from prior exploitation is only recovered by rebooting.

References

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.