The vulnerability arises in the Linux kernel's AMDGPU kernel driver (amdkfd). A cached address_space for the first open of /dev/kfd is used by the GPU reset path to unmap user‑space mappings. When multiple containers use SRIOV, the cache may refer to an inode owned by the first container that later exits. The freed inode leads to a use‑after‑free when the GPU reset attempts to unmap, causing a rwsem deadlock and a hard hang observable as a non‑interruptible NMI wait. In another scenario, only the first container's mappings are cleared, leaving later containers’ doorbell and MMIO pages dangling, which can produce PCIe transaction timeouts and NMI panics. The impact is a denial of service that can lock the entire system or kernel.

This issue affects systems running the Linux kernel that use the amdgpu kernel module in environments employing SRIOV with multiple containers. The fix concerns the amdkfd module and is applied at kernel version containing commit 1128b4a52de1572e87431de837fd9850cb99542c. No specific affected kernel releases are enumerated in the advisory; users should verify whether their kernel includes the stated commit.

The vulnerability requires privileged kernel execution in a SRIOV context and access to /dev/kfd. Exploitation would trigger only during a GPU reset, which is typically performed by the driver during fault recovery or power management, rather than arbitrary code execution. The EPSS score is not available, and it is not listed in the CISA KEV catalog. Despite the lack of a quantified risk metric, the potential for a hard kernel hang and uninterruptible wait makes the severity high for environments that rely on GPU stability. The attack vector is likely local, i.e., an attacker with control over container lifecycle or GPU reset operations.