GPU VulnDB

Database/Firmware, BMC & network fabric

Linux kernel RDMA/core: deleting an IB sub-device leaks a device reference on the -EOPNOTSUPP path

CVSS 7.8CVE-2025-71157Firmware, BMC & network fabriccurated

Impact

The netlink handler that deletes an IB sub-device takes a device reference with ib_device_get_by_index() before calling ib_del_sub_device_and_put(), but that function returned -EOPNOTSUPP without dropping the reference. Each rejected delete therefore pins the IB device structure forever. On a GPU fleet the IB devices are the InfiniBand or RoCE HBAs that carry collective traffic, and a leaked device reference means the driver can no longer be unbound or unloaded and the device cannot be cleanly torn down - so a node that needs an HCA driver reload for firmware or a fabric change has to be rebooted instead. The exposure is narrow: the path is the RDMA netlink sub-device management interface, not anything a tenant job touches.

Who can reach it

Local user holding the privilege to issue RDMA netlink device-management commands (CAP_NET_ADMIN in the relevant namespace) - in practice the host administrator or a fabric-management agent, not a tenant with a GPU pod. No remote or unauthenticated path in the record.

What to do

Update to a stable kernel carrying the fix and reboot the node; the leak is in ib_core, and once a reference has leaked the module cannot be unloaded, so a reboot is the only way to clear an already-leaked state. Until patched, avoid repeated RDMA netlink sub-device delete attempts against devices that do not support them. Only the upstream commits are in the record; no vendor advisory with a fixed version is present.

References

Related entries

All Firmware, BMC & network fabric 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.