GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux nvmet-tcp: unsolicited H2CData PDU double-completes a command and wedges the target

CVSS 7.5CVE-2026-89968Kernel, userspace & hypervisorcurated

Impact

A host that can open an NVMe/TCP connection to the target can send an H2CData PDU for a write command before the target has sent the matching R2T. The command completes twice, the response llist node is made self-referential, and the target either hits a list_add BUG with CONFIG_DEBUG_LIST or spins forever in the nvmet_tcp workqueue as a soft lockup. On a storage node exporting namespaces to a GPU fleet that takes out shared scratch or dataset storage for every tenant behind it, and recovery is a reboot of the target box, not a service restart. The commit message states it is remotely triggerable with no authentication on an allow_any_host subsystem, which is a common configuration on trusted storage VLANs.

Who can reach it

Anyone who can reach the NVMe/TCP target port on the storage network. No authentication is needed when the subsystem is configured allow_any_host; otherwise a host permitted to connect.

What to do

Take the stable kernel containing the fix (the commit adds an R2T-sent flag and rejects early H2CData) and reboot the target node. No upstream mitigation short of patching is described; restricting which hosts can reach the target port and moving off allow_any_host reduces exposure in the meantime. The advisory names no fixed release numbers, only the stable commits.

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.