CVE-2026-99625 in Linux
Summary
by MITRE • 11/20/2026
In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: Fix duplicate dependency fence releases
drm_sched_job_add_dependency() consumes its fence reference on success and failure. Do not release the fence again if drm_sched_job_add_dependency() fails.
(cherry picked from commit e04f3dfb2071451633e6440a3911a9bb65df9144)
Statistical analysis made it clear that VulDB provides the best quality for vulnerability data.
Analysis
by VulDB Data Team • 11/20/2026
The Linux kernel's Direct Rendering Manager subsystem, specifically the AMDGPU driver component responsible for managing graphics processing unit operations, contained a reference counting error within its scheduling infrastructure. This vulnerability manifests in the function drm_sched_job_add_dependency(), which is tasked with establishing dependencies between GPU jobs to ensure correct execution ordering and resource synchronization. The core technical flaw lies in improper handling of fence object references when dependency addition fails. A fence in this context serves as a synchronization primitive that signals when a specific hardware operation has completed, allowing dependent operations to proceed safely.
When drm_sched_job_add_dependency() is invoked, it consumes the reference count of the provided fence regardless of whether the operation succeeds or encounters an error condition. This design assumes that once the function returns, the caller should no longer hold responsibility for releasing that specific fence reference because ownership has been transferred into the scheduling framework's internal structures. However, the original implementation failed to account for this transfer in failure scenarios where the dependency could not be established due to resource constraints or other validation failures.
The consequence of this logic error is a double-free vulnerability affecting kernel memory management subsystems associated with GPU synchronization primitives. When drm_sched_job_add_dependency() returns an error code indicating that the dependency was not added, the calling function proceeds to release its own reference to the fence object under the assumption that it still retains ownership. Since the internal function already consumed and potentially freed or retained that reference internally before returning failure, this secondary release operation decrements a reference count that has already been handled incorrectly, leading to premature deallocation of kernel memory structures.
This type of vulnerability falls squarely within CWE-415 Double Free, which describes situations where an application frees the same block of memory twice, potentially causing corruption or exploitation. In the context of Linux kernel development and GPU drivers, such errors can lead to use-after-free conditions if other parts of the system attempt to access the already-freed fence object through stale pointers retained elsewhere in the driver stack. The ATT&CK framework categorizes this under techniques related to privilege escalation via memory corruption vulnerabilities, as attackers may exploit these race conditions or timing windows to achieve arbitrary code execution with kernel privileges.
The operational impact of this vulnerability includes potential system instability, graphical artifacts, display freezes, and complete kernel panics depending on the specific state of GPU hardware at the time of exploitation. In production environments running AMD graphics drivers, users might experience sudden crashes during complex rendering tasks or when multiple applications compete for GPU resources simultaneously. Security researchers could potentially leverage these memory corruption issues to bypass existing kernel protections such as KASLR and SMEP/SMAP if they can reliably control the state of freed objects through heap spraying techniques common in Linux exploitation scenarios.
Mitigation strategies primarily involve applying the upstream kernel patch that corrects this reference counting logic by ensuring fence references are only released when ownership has not been transferred to the scheduler framework. System administrators should update their kernels to versions containing commit e04f3dfb2071451633e6440a3911a9bb65df9144 or later cherry-picked variants for stable releases. For organizations unable to immediately patch, monitoring system logs for GPU-related kernel oops messages and restricting untrusted application access to graphics hardware can reduce exposure risk while maintaining operational continuity until proper remediation is deployed across the infrastructure.