Based on the description, it is inferred that the attack vector involves local kernel manipulation via GPU buffer object allocation and deallocation. A defect in the AMDGPU amdkfd GPU memory manager of the Linux kernel causes a buffer object (BO) to be unpinned twice during process teardown. The code removes the pin on a doorbell or MMIO BO before it confirms that the BO is still mapped to the GPU. When the BO remains mapped, the function returns an error but the BO has already been unpinned; a second unpin during final deallocation then triggers an underflow warning in ttm_bo_unpin(). This double‑unpin can corrupt the reference counter of the BO and may lead to a kernel panic or at least repeated warning messages, effectively providing a denial‑of‑service weakness for code that can manipulate AMDKFD GPU memory.

Based on the description, it is inferred that the attack vector requires local access to the GPU subsystem. All users of the Linux kernel that include the vulnerable AMDKFD code path, specifically those running the AMDGPU driver on AMD GPUs. The defect exists in kernel versions that predate commit 927c5b2defb9b09856444d94bebfd056a002bd75. Vendor information is limited to the Linux kernel itself with no version lattice supplied.

Based on the description, it is inferred that the attack vector is local to user‑space processes that allocate GPU buffers. The vulnerability requires entry into kernel space, so it is effectively a local escalation or a bug that can be triggered by any process that allocates and later frees GPU buffer objects. No public exploit has been documented and there is no KEV listing; EPSS data is not available. Nonetheless, the fault can cause kernel instability and potentially a crash under normal GPU workloads, presenting a high risk of denial of service. The fix reorders the unpin operation to prevent the double‑unpin but does not provide an alternative mitigation path beyond updating the kernel.