CVE-2024-36949

What it actually does

Race condition in amd/amdkfd allows parallel device resets to skip full process eviction via kfd_suspend_all_processes, restoring processes that then trigger page faults on recovery.

Why it matters to the business

Systems with multiple AMD GPUs may experience instability, crashes, or failed GPU workload recovery during concurrent resets, disrupting compute or graphics operations.

How an attack unfolds

An attacker triggers simultaneous resets across multiple AMD KFD devices; the first device begins eviction while others proceed to recover without waiting, allowing unevicted processes to fault.

How to fix it

Apply the fix from commit b6f6626528fe724b512c34f3fb5946c36a135f58 in the Linux kernel stable tree; update to a kernel containing this commit or later.

How to confirm the fix worked

Verify the running kernel contains commit b6f6626528fe724b512c34f3fb5946c36a135f58 or equivalent patch; test multi-device reset scenarios for proper eviction ordering.

Typical attack path

1. Identify multi-device AMD GPU configuration supporting KFD
2. Trigger concurrent device reset paths in the kernel
3. Allow first device to initiate kfd_suspend_all_processes while others reset
4. Restore processes before eviction completes
5. Trigger page faults during recovery

How to detect it

Review kernel logs for KFD page faults occurring during or immediately after device reset events on systems with multiple AMD GPUs.

What your security team should do

• Prioritize kernel update on multi-GPU AMD systems within standard patching cycles
• Monitor for related stability issues in GPU reset logs
• Test reset/recovery behavior after applying the fix