CVE-2026-68106: Linux Kernel AMDGPU DoS Vulnerability

CVE-2026-68106 Overview

CVE-2026-68106 is a division by zero vulnerability in the Linux kernel's AMD GPU driver (drm/amdgpu). The flaw resides in the Unified Video Decoder (UVD) H.264 decode paths, where invalid frame dimensions trigger a kernel-level arithmetic exception. When a caller supplies a width or height less than 16, the derived macroblock counts width_in_mb or height_in_mb evaluate to zero. The resulting fs_in_mb value of zero causes a division by zero when calculating num_dpb_buffer for both the standard H.264 and H.264 performance decode paths. The kernel fix adds input validation that rejects frames with dimensions below 16 before any dependent calculations execute.

Critical Impact
A local, authenticated user with access to the amdgpu device node can trigger a kernel division by zero, causing denial of service on affected Linux systems.

Affected Products

Linux kernel with the drm/amdgpu driver enabled
Systems using AMD GPUs with UVD hardware for H.264 decode
Distributions shipping the vulnerable kernel prior to the referenced stable commits

Discovery Timeline

2026-08-10 - CVE-2026-68106 published to NVD
2026-08-13 - Last updated in NVD database

Technical Details for CVE-2026-68106

Vulnerability Analysis

The vulnerability affects the AMD GPU UVD H.264 decode initialization logic in the Linux kernel's Direct Rendering Manager (DRM) subsystem. The decoder computes the frame size in macroblocks by dividing input width and height by 16. When either dimension falls below 16, integer truncation produces a zero macroblock count. The subsequent product fs_in_mb = width_in_mb * height_in_mb becomes zero. The driver then uses fs_in_mb as a divisor when computing num_dpb_buffer, the number of Decoded Picture Buffer entries required for H.264 decoding. This unchecked division triggers a kernel arithmetic exception.

A local user able to open the amdgpu render node and submit a UVD H.264 decode job with attacker-controlled dimensions can reach the vulnerable arithmetic. The result is a kernel oops or panic, depending on kernel configuration, disrupting availability on the affected host [CWE-369].

Root Cause

The root cause is missing input validation on user-supplied frame dimensions before macroblock arithmetic. The decode path assumed callers would provide dimensions of at least 16 pixels in both axes, matching the H.264 macroblock size, but did not enforce this constraint.

Attack Vector

Exploitation requires local access and privileges sufficient to open the amdgpu DRM device. The attacker submits a crafted UVD decode command with width or height less than 16. No user interaction is required beyond submitting the malformed job. The impact is limited to denial of service on the local host.

No verified exploit code has been published. Refer to the upstream fix commits for technical details: Kernel Git Commit 3e41d26 and related stable backports.

Detection Methods for CVE-2026-68106

Indicators of Compromise

Kernel oops or panic messages referencing divide error within amdgpu_uvd functions in dmesg or journalctl -k output
Unexpected reboots or GPU driver resets on hosts running vulnerable kernels with AMD GPUs
Repeated crashes correlated with unprivileged user processes invoking VA-API or VDPAU H.264 decode

Detection Strategies

Inventory hosts by running uname -r and comparing against fixed stable kernel versions referenced in the upstream commits
Monitor kernel ring buffer for divide_error exceptions originating in amdgpu symbols
Audit which local accounts and containers have access to /dev/dri/renderD* device nodes

Monitoring Recommendations

Forward kern.log and dmesg output to a central log platform and alert on divide error or RIP: .*amdgpu_uvd patterns
Track process lineage for user-space processes that submit UVD decode jobs prior to a crash event
Alert on abnormal frequencies of amdgpu driver resets within short time windows

How to Mitigate CVE-2026-68106

Immediate Actions Required

Apply the vendor-provided kernel update containing the upstream fix to all Linux systems using AMD GPUs
Restrict access to /dev/dri/renderD* nodes to trusted users and service accounts only
On multi-tenant hosts and container platforms, remove GPU device passthrough from untrusted workloads until patched

Patch Information

The fix is cherry-picked from upstream commit 3e41d26c70b0a459d041cc19482a226c4b7423cb, which adds validation to reject frames with width or height below 16 and uses warn_once to avoid log spam. Stable backports are tracked in the following commits: 0c01c81, 52f9a58, a00946b, be725ab, and ffb33d4. Track your distribution's security advisory for the corresponding backport version.

Workarounds

Remove or restrict permissions on amdgpu render nodes for untrusted local users using standard Linux group membership on the render and video groups
Disable UVD-based hardware H.264 decode in userspace media stacks (VA-API, VDPAU) where GPU decoding is not required
Isolate untrusted workloads that require GPU access into dedicated hosts until the kernel patch is deployed

bash

# Verify amdgpu render node permissions and group ownership
ls -l /dev/dri/renderD*
getent group render

# Confirm the running kernel version to compare against patched releases
uname -r