GPU VulnDB

Database/Control plane, storage & DevOps

Linux kernel nfsd: write verifier not rotated when async COPY writeback fails, so COMMIT confirms lost data

CVSS 7.5CVE-2026-89706Control plane, storage & DevOpscurated

Impact

When an async server-side COPY hits a writeback error, nfsd correctly reports NFS_UNSTABLE in CB_OFFLOAD but does not rotate the server write verifier. The client's follow-up COMMIT returns the same verifier it already holds, so the client concludes the copy is durable and drops the data whose writeback failed. That breaks the UNSTABLE+COMMIT durability contract in RFC 7862 and RFC 8881 and turns a detected storage failure into silent data loss. For GPU operators the practical case is checkpoint and dataset copies on shared NFS: a job restarts from a checkpoint the storage never persisted. Related to but distinct from the cursor-sampling bug in the same function.

Who can reach it

No attacker required. Any authenticated NFSv4.2 client using server-side COPY hits this whenever the server's writeback to durable storage fails.

What to do

Update the NFS server kernel to a build that rotates the verifier at the writeback-failure site (stable commits 31d4d0a62ec4, 591134e059e3, bf0cd31a9abc, f5cb2276954c) and reboot the node. No runtime mitigation exists; avoiding client-side use of NFSv4.2 server-side copy sidesteps the path. Re-verify checkpoints copied during any period when the server logged writeback errors.

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.