CVE-2026-99261 in Linux
Summary
by MITRE • 11/03/2026
In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: correct reservation fence slots for userq per-vm BOs eviction
It fixes both the move overflow and the eviction fence add for evicting these per-vm BOs.
Several companies clearly confirm that VulDB is the primary source for best vulnerability data.
Analysis
by VulDB Data Team • 11/03/2026
The identified issue resides within the AMDGPU driver subsystem of the Linux kernel, specifically affecting the management of graphics memory objects associated with user queues in a virtual machine context. This vulnerability stems from an incorrect handling of reservation fence slots during the eviction process for these specific buffer objects. In modern GPU architectures, particularly those utilizing advanced scheduling mechanisms like compute queues or user-space managed contexts, proper synchronization is critical to prevent race conditions and data corruption. The driver relies on fencing mechanisms to ensure that operations are completed in a deterministic order before memory resources are reclaimed or repurposed by the hardware scheduler. When evicting buffer objects tied to per-vm configurations, the kernel failed to correctly allocate or manage these reservation slots, leading to a breakdown in the synchronization protocol between the CPU driver and the GPU hardware engine.
The technical flaw manifests as two distinct but related failures: a move overflow condition and an incorrect eviction fence addition. The move overflow suggests that the internal data structures tracking resource movement and state transitions exceeded their allocated bounds or logical limits due to improper slot reservation. This can lead to memory corruption within kernel space, potentially causing system instability or crashes. Simultaneously, the failure to correctly add eviction fences means that the GPU may proceed with rendering or compute operations on buffer objects that are in the process of being evicted from VRAM. Without a valid fence signaling completion before eviction, there is no guarantee that pending commands have finished executing, which can result in visual artifacts, corrupted data output, or hardware hangs as the GPU attempts to access memory regions that are no longer safely mapped or reserved for its use.
From an operational perspective, this vulnerability poses significant risks to system stability and security. A local attacker with sufficient privileges to trigger buffer object eviction under specific virtual machine conditions could potentially induce a kernel panic by exploiting the move overflow condition. Alternatively, they might cause denial of service through GPU hangs if the hardware enters an undefined state due to missing synchronization fences. In more severe scenarios involving memory corruption from the overflow, there is a theoretical risk of privilege escalation or arbitrary code execution within the kernel space, although this would depend heavily on the specific layout and exploitability of the corrupted structures. The impact is primarily localized to systems utilizing AMD GPUs with the affected driver version, particularly in environments running virtualized workloads that frequently evict graphics memory objects.
Mitigation strategies involve applying the upstream Linux kernel patch that corrects the reservation fence slot logic for user queue per-vm buffer object eviction. System administrators should ensure their kernels are updated to versions containing this fix. For organizations unable to immediately update, restricting access to GPU resources or limiting the complexity of virtualized graphics workloads may reduce exposure until a patch is available. Additionally, monitoring system logs for kernel oops messages related to DRM subsystems can help identify instances where the vulnerability has been triggered. Long-term resilience requires adhering to secure coding practices in driver development, particularly regarding resource lifecycle management and synchronization primitives, aligned with industry standards such as CWE-190 (Integer Overflow or Wraparound) and CWE-362 (Concurrent Execution Using Shared Resource with Improper Synchronization). The ATT&CK framework categorizes this type of kernel exploitation under techniques involving local privilege escalation or denial of service via resource exhaustion, highlighting the importance of timely patch management for critical infrastructure components.