Linux kernel NFS client: O_DIRECT writes can be silently corrupted on retransmission

Impact

The NFS client failed to rejoin subrequests into the head request before retransmitting, and did not put a synchronously written head back on the retransmission list for O_DIRECT. The result is silent data corruption on written files - no crash, no error returned to the application, wrong bytes on the server. For a GPU fleet this lands on the path that matters most: training checkpoints, dataset shards and model artifacts written over NFS, often with O_DIRECT, where a corrupted checkpoint is only discovered when a restart or an eval fails and there is no way to tell which files are affected. The CVSS vector is network-facing because the trigger is server-side write and retransmission behaviour rather than a local actor; integrity, not confidentiality, is what is lost.

Who can reach it

No authentication or attacker position is required in the usual sense - this is a correctness bug reached by normal NFS client I/O when writes are retransmitted (server-side synchronous writes, dropped or retried requests). Any node mounting NFS and doing O_DIRECT writes is exposed; an attacker able to influence server-side write and retransmit timing could aim it deliberately.

What to do

Update to a stable kernel carrying the fix on every NFS client node (five commits are linked in the record; no fixed version numbers are stated) and reboot - the NFS client is in-kernel, so there is no daemon to restart. Drain and reboot clients in batches. Given the failure mode is silent, treat checkpoints and datasets written over NFS from unpatched clients as suspect and re-verify checksums where you have them. Mitigation short of patching means avoiding O_DIRECT writes over NFS, which is usually not practical for training I/O.

References

Related entries

• QEMU (virtio): DMA reentrancy leads to double free across virtio devices - guest-to-host code execution in QEMUCVE-2024-3446 · QEMU (virtio)High
• Linux kernel (net/tls): A non-DATA record already copied out of the pending list could be merged with a second recordCVE-2024-58239 · Linux kernel (net/tls)High
• Intel ice driver (Ethernet 800 Series, Linux kernel mode): A missing check for an exceptional condition in theCVE-2025-20093 · Intel ice driver (Ethernet 800 Series, Linux kernel mode)High
• VMware ESXi: Arbitrary kernel write from the VMX process - sandbox escape completing the zero-day chainCVE-2025-22225 · VMware ESXiHigh
• KVM arm64: unvalidated FF-A offset lets the host kernel read and write past the hypervisor's memory-share bufferCVE-2025-40266 · Linux kernel KVM arm64 pKVM (FF-A memory share offset validation)High
• OpenStack Nova: crafted QCOW header on a Flat-backend disk lets a tenant destroy host data on resizeCVE-2026-24708 · OpenStack Nova (Flat image backend, qemu-img resize)High