GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel nvmet-tcp: over-long PDU writes past the 128-byte receive buffer

CVSS 9.8CVE-2026-89969Kernel, userspace & hypervisorcurated

Impact

A node exporting NVMe namespaces over TCP can have kernel memory corrupted by any host that can open a connection to the target port, before authentication. Once a header digest is negotiated, a second ICReq makes the target compute a 124-byte read into a 128-byte union at offset 8, writing 4 attacker-controlled bytes over the adjacent hdr_digest and data_digest fields; the duplicate ICReq is only rejected after the overflow has happened. On a GPU fleet this matters wherever storage nodes serve scratch or dataset volumes over NVMe/TCP on the same network tenants can reach - the blast radius is the storage kernel, not a single job, and a corrupted target takes every mounted namespace with it.

Who can reach it

Any host with TCP reach to the nvmet-tcp listener. No authentication required - the overflow occurs during connection setup.

What to do

Update to a kernel carrying the fix (the PDU length is now bounded against sizeof(queue->pdu) before the second recv) and reboot each affected storage node. Until then, restrict the NVMe/TCP listener to a trusted storage VLAN or an allowlist of initiator addresses. The record lists the stable commits but no distribution-fixed version.

References

Related entries

All Kernel, userspace & hypervisor 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.