Database/Firmware, BMC & network fabric
Linux kernel mana_ib: destroying an RSS QP leaves stale RX steering, causing RX completions on TX queues
Impact
mana_ib_destroy_qp_rss() tears down the RX WQ objects but never disables vPort RX steering in firmware, so the firmware keeps steering arriving traffic at objects that no longer exist. If a peer keeps transmitting and the VF is then brought up, completions can be delivered against stale CQ IDs that the ethernet driver has since reused for TX CQs - the kernel hits WARNs in mana_poll_tx_cq() and mana_gd_process_eq_events() and the queue state is no longer consistent. On a GPU node this is the RDMA path: the sequence is exactly what a DPDK or RDMA user-space process exiting and the interface being re-opened looks like, which is routine on a training node between jobs, and the damage lands in the shared NIC rather than in the exiting job. The record describes warnings and misdirected completions; it does not establish an information-disclosure or code-execution path, and this driver is only present on Azure-hosted guests.
Who can reach it
Local: any user who can open the RDMA device and create and destroy an RSS QP - in practice a tenant running DPDK or an RDMA application on an Azure VM - combined with continuing inbound traffic from a peer and a subsequent mana_open on the VF. No authentication beyond access to the RDMA device is needed.
What to do
Take the stable kernel fix, which disables vPort RX steering before the RX WQ objects are destroyed via a shared mana_disable_vport_rx() exported from mana_en; the stable commits are linked from the record. This is a kernel change, so it means installing the patched kernel and rebooting each affected node; there is no runtime mitigation short of not letting tenants drive RSS QPs on the VF.
References
Related entries
- Linux kernel (drivers/infiniband/hw/mana): The userspace ABI lets a tenant point several work queues at the sameCVE-2026-46117 · Linux kernel (drivers/infiniband/hw/mana)High
- Linux kernel (drivers/infiniband/hw/mana): The RSS hash-key length arrived from the userspace ABI structure and wentCVE-2026-46145 · Linux kernel (drivers/infiniband/hw/mana)High
- Linux kernel (drivers/infiniband/hw/mlx5): If the second of the two device-wide shared SRQs fails to allocate, theCVE-2026-46176 · Linux kernel (drivers/infiniband/hw/mlx5)High
- Linux kernel mlx4 InfiniBand driver: SRQ leaked on create failure, exhausting fabric resourcesCVE-2026-46178 · Linux kernel RDMA/mlx4 (mlx4_ib_create_srq error unwind)High
- Linux kernel RDMA core (ib_umem / IB_MR_REREG_ACCESS re-registration): An RDMA memory region registered read-only canCVE-2026-52908 · Linux kernel RDMA core (ib_umem / IB_MR_REREG_ACCESS re-registration)High
- Linux kernel RDS over InfiniBand (use outside the initial network namespace): The RDS/IB transport was never written toCVE-2026-53077 · Linux kernel RDS over InfiniBand (use outside the initial network namespace)High
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.