Linux Kernel AMD GPU VRAM Leak Vulnerability via MMU Notifier Deadlock

A vulnerability in the Linux kernel's handling of AMD GPU memory can lead to a deadlock and a memory leak in the Video RAM (VRAM). This issue arises when the 'mmput' function is called from the Memory Management Unit (MMU) notifier callback during process termination or memory balancing operations. The premature release of the last memory reference can cause a deadlock, as the process fails to properly clean up, leaving the kernel's graphics processing unit (GPU) memory management in disarray. The deadlock can be observed in a task that becomes unresponsive for an extended period, blocked by the improper handling of memory references. As a result, the kernel's GPU virtual memory management leaks processes, exacerbating the VRAM waste.

Impact

The vulnerability causes a deadlock in the system, with a task being blocked for over 64,512 seconds. This blockage prevents the proper release of GPU memory, leading to a significant leak of VRAM resources.

Reproduction

The vulnerability can be reproduced by initiating a process that requires GPU memory allocation through the AMD GPU kernel driver. During the process, trigger a memory balancing operation or process termination. This sequence will cause the MMU notifier to invoke 'mmput', releasing the last memory reference prematurely and creating a deadlock. The blocked task can be observed, highlighting the issue.

Remediation

The vulnerability has been addressed in the Linux kernel. Users should upgrade to the latest version where this issue has been fixed.