GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel NVMe: SUBMIT_IO skips the command permission check, allowing writes through read-only handles

CVSS 7.8CVE-2026-90227Kernel, userspace & hypervisorcurated

Impact

Unlike NVME_IOCTL_IO_CMD and IO64_CMD, NVME_IOCTL_SUBMIT_IO never called nvme_cmd_allowed(), so an unprivileged caller holding a device handle could issue I/O against a partition device or write through a file descriptor opened read-only. That is a direct storage integrity and confidentiality break: the check that normally confines partition-scoped and read-only handles is simply absent on that path. On a GPU node this is the local NVMe that holds scratch datasets, container images and model checkpoints, often carved into partitions shared between tenants or system roles; a tenant with access to one partition handle can read or write outside its intended scope. Remediation is a kernel update, which means draining the node. NVD scores it local, low complexity, high on confidentiality, integrity and availability.

Who can reach it

Local user with an open file descriptor on an NVMe block device or partition, including a read-only one. No administrative privilege is needed beyond that handle - which is exactly the case this fix closes.

What to do

Apply the stable kernel update that passes flags and open_for_write through to nvme_cmd_allowed() and returns -EACCES for disallowed commands (commits 736b7c6adc7a, 95b0591d51db, b116e9dae433, b53d495c7f0d, c35cd98eed1a), then drain and reboot each node. As an interim control, tighten who can open NVMe block devices and do not pass raw NVMe devices or partitions into tenant containers. The record names no fixed release numbers.

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.