GPU VulnDB

Database/Control plane, storage & DevOps

Linux nfsd: async server-side COPY registers a stateid pointing into a reused request buffer

CVSS 9.8CVE-2026-89676Control plane, storage & DevOpscurated

Impact

nfsd4_copy() registered the copy stateid in the s2s_cp_stateids IDR before duplicating the fields, so the IDR slot pointed at memory in the per-rqstp COMPOUND buffer that the next request reuses. Any IDR walker - the laundromat, or manage_cpntf_state() via idr_find() - then dereferences recycled request memory and can be steered into refcount_dec(), idr_remove() and kfree() on attacker-influenced bytes. On a GPU cluster whose scratch or dataset tier is served by a Linux NFS server, a client that issues async COPY can corrupt server-side kernel state; the NFS server is a single point of failure for every node that mounts it, so a crash or compromise there stalls the whole fleet, not one tenant.

Who can reach it

Any NFSv4.2 client permitted to mount the export and issue an asynchronous server-side COPY. Standard NFS authentication only - no privilege on the server is needed.

What to do

Patch the NFS server kernel with one of the linked stable commits and reboot it; nfsd runs in-kernel, so no daemon restart is sufficient. Plan for the export to be unavailable during the reboot, or fail over to a second server first. Disabling NFSv4.2 server-side copy on the export removes the reachable path if a reboot cannot be scheduled.

References

Related entries

All Control plane, storage & DevOps 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.