GPU VulnDB

Database/Firmware, BMC & network fabric

Linux kernel RDMA restrack: memory-region deletion fails to release the tracker, leaking task structs

CVSS 7.8CVE-2022-50822Firmware, BMC & network fabriccurated

Impact

Deleting an RDMA memory region did not release its restrack entry, so the associated task struct was never freed. A workload that repeatedly registers and deregisters memory regions leaks unreclaimable kernel memory until the node is under memory pressure or OOM. This is an availability problem rather than an escalation one, but MR register/deregister churn is exactly what NCCL, UCX and RDMA-backed storage clients do on every job, so a long-lived GPU node running back-to-back distributed jobs accumulates the leak without anyone issuing anything unusual. The failure mode is a node that degrades over days and has to be drained mid-job.

Who can reach it

Local, authenticated: any process with access to the RDMA verbs devices (/dev/infiniband/uverbs*) - in practice any tenant pod granted RDMA device access for InfiniBand or RoCE traffic. No privilege beyond that is needed, and the leak also accrues from ordinary legitimate use.

What to do

Move to a fixed stable kernel (commits in the record) and reboot. Fabric-side drivers live in the kernel image, so this is a drain-and-reboot per node. Short of that, the only mitigation is operational: watch kernel slab growth on nodes with heavy RDMA memory-region churn and recycle them before they hit memory pressure.

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.