CVE-2023-54003

What it actually does

Reference leak in RDMA core GID table handling when create_ah fails, leaving sgid_attr unreleased and causing GID entry reference counts to be mishandled on table release.

Why it matters to the business

RDMA-dependent systems such as high-performance computing clusters or storage fabrics can experience gradual resource exhaustion, leading to performance degradation or instability in production workloads.

How an attack unfolds

An attacker with the ability to issue RDMA address handle creation requests that are crafted to fail can trigger the leak path repeatedly; over time this exhausts GID references and may surface as memory corruption or denial of service when the GID table is later released.

How to fix it

Apply the upstream fix by updating to Linux kernel commit 9c46c49ad3ffe84121715d392b5a0a94f9f10669 or later stable equivalents; rebuild and deploy the patched kernel across affected hosts.

How to confirm the fix worked

Confirm the fix commit is present in the running kernel source tree, then exercise AH creation failure scenarios while monitoring GID reference counts and table release behavior.

Typical attack path

1. Obtain local or RDMA-capable access to a target system
2. Issue repeated failing AH creation requests against an RDMA device
3. Trigger GID table cleanup paths
4. Observe reference count exhaustion or related kernel warnings

How to detect it

Review kernel logs for RDMA GID table warnings or reference leak messages; correlate with processes performing RDMA address handle operations around the time of anomalies.

What your security team should do

• Prioritize kernel upgrades on all RDMA-enabled hosts
• Restrict untrusted local access to RDMA interfaces until patched
• Include RDMA subsystems in regular kernel update and vulnerability scanning processes