Linux kernel (drivers/infiniband/hw/mlx5): Memory-region deregistration hangs forever on the flagship AI-cluster NIC. A

Impact

Memory-region deregistration hangs forever on the flagship AI-cluster NIC. A reference taken on the parent region is never dropped, so the deregister call blocks in uninterruptible sleep - the task cannot be killed, it holds device references that prevent teardown, and the tenant's GPU job wedges without releasing its RDMA resources back to the node.

Who can reach it

A tenant container holding /dev/infiniband/uverbs* on mlx5 hardware with on-demand paging (implicit ODP) in use - deregistering a parent MR that still has implicit children is enough, which is normal application shutdown behaviour and therefore trivially repeatable. Only applies where ODP is enabled; non-ODP registrations do not reach this path.

What to do

Update to 6.12.18 / 6.13.6 or later. Interim: disable on-demand paging for tenant workloads (do not advertise ODP capability) so implicit MRs are never created, and drain jobs off nodes that have accumulated stuck deregistration tasks.

References