GPU VulnDB

Database/Firmware, BMC & network fabric

Linux kernel (drivers/infiniband/hw/mlx5): Memory-region deregistration self-deadlocks under memory pressure. An

CVE-2025-38373Firmware, BMC & network fabriccurated

Impact

Memory-region deregistration self-deadlocks under memory pressure. An allocation made while holding the page-mapping lock can drive reclaim, which calls back into the same driver's invalidation handler and blocks on the lock it already holds - the tenant's task hangs unkillably and the mlx5 registration path is blocked for everything else on the node.

Who can reach it

A tenant container holding /dev/infiniband/uverbs* on mlx5 hardware deregistering an on-demand-paging memory region while the node is under memory pressure. Both halves are tenant-influenceable: the tenant chooses when to deregister and can raise memory pressure itself, which is exactly the condition a busy shared GPU node is already in.

What to do

Update to 6.12.37 / 6.14 or later. Interim: disable on-demand paging for tenant workloads so ODP regions are never created, and keep node memory headroom high enough that reclaim is not entered during RDMA teardown.

References

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.