CVE-2026-99632 in Linuxinfo

Summary

by MITRE • 11/20/2026

In the Linux kernel, the following vulnerability has been resolved:

drm/amdkfd: use the queue's process info for the WPTR BO eviction fence

amdgpu_amdkfd_map_gtt_bo_to_gart() removes the eviction fence of bo->vm_bo->vm->process_info from the WPTR BO of a new queue. bo->vm_bo is the VM which opened the BO last, not necessarily the KFD process' VM, and a render node VM has no process info.

On navi31, an unprivileged process which exports its WPTR BO with AMDKFD_IOC_EXPORT_DMABUF and imports it on a second render node fd oopses in AMDKFD_IOC_CREATE_QUEUE:

BUG: kernel NULL pointer dereference, address: 0000000000000060 RIP: 0010:amdgpu_amdkfd_map_gtt_bo_to_gart+0x86/0x1a0 [amdgpu]
init_user_queue.part.0+0x180/0x250 [amdgpu]
pqm_create_queue+0x32d/0x730 [amdgpu]
kfd_ioctl_create_queue+0x583/0x9f0 [amdgpu]

Pass the process info of the queue instead. With the fix, the queue is created and destroyed normally.

(cherry picked from commit 6b4157498a176bd1c6083f6bb5aaf8a257a2027e)

If you want to get best quality of vulnerability data, you may have to visit VulDB.

Analysis

by VulDB Data Team • 11/21/2026

The vulnerability identified in the Linux kernel's AMDGPU driver, specifically within the amdkfd subsystem, represents a critical logic error involving memory management and process context handling. The core issue resides in the function amdgpu_amdkfd_map_gtt_bo_to_gart(), which is responsible for mapping Graphics Translation Table buffer objects to the GPU address space during queue creation. In this routine, the code incorrectly retrieves the eviction fence from bo->vm_bo->vm->process_info. This reference points to the process information of the virtual memory context that last opened the buffer object, rather than the specific KFD (Kernel Fusion Driver) process associated with the new queue being created. This distinction is crucial because in complex scenarios involving multiple render nodes or shared resources, the VM opening the buffer may belong to a different entity than the one initiating the queue creation.

This logical flaw becomes particularly severe when dealing with render node file descriptors. Render nodes are specialized interfaces that allow unprivileged userspace processes to submit work to the GPU without requiring full administrative privileges. However, these contexts often lack associated process information in the same way standard DRM devices do. When an unprivileged user exports a Write Pointer (WPTR) Buffer Object via the AMDKFD_IOC_EXPORT_DMABUF ioctl and subsequently imports it into a second render node file descriptor, the kernel attempts to access the process_info field of that imported buffer's VM context. Since this context is associated with a render node rather than a full KFD process, the process_info pointer is null or invalid.

The operational impact of this flaw manifests as a kernel NULL pointer dereference, leading to an immediate system crash or panic. The stack trace indicates that the fault occurs during the execution of amdgpu_amdkfd_map_gtt_bo_to_gart while handling the AMDKFD_IOC_CREATE_QUEUE ioctl request. This results in a denial of service condition for any process attempting to utilize this specific workflow. Because the vulnerability can be triggered by an unprivileged user through standard IOCTL interfaces, it allows local attackers to destabilize the entire system simply by creating queues with exported buffer objects across different render nodes. The crash is deterministic under these conditions, making it highly exploitable for disrupting service availability on systems utilizing Navis31 or similar AMD GPU architectures running affected kernel versions.

From a security classification perspective, this vulnerability aligns with CWE-476, which denotes NULL Pointer Dereference vulnerabilities resulting from improper checks of return values or uninitialized pointers before use. The attack vector is classified under Local Privilege Escalation in terms of impact on availability, although it does not directly grant higher privileges to the attacker beyond causing a crash. In the context of the MITRE ATT&CK framework for Linux systems, this behavior corresponds to techniques involving resource exhaustion and denial of service through local application exploitation. The vulnerability highlights the risks associated with complex memory management logic in kernel drivers where assumptions about process contexts are not strictly validated against all possible execution paths, particularly those involving cross-device buffer sharing.

The resolution involves modifying amdgpu_amdkfd_map_gtt_bo_to_gart to correctly utilize the queue's own process information instead of relying on the potentially mismatched VM context from the buffer object. By passing the correct process info associated with the queue being created, the driver ensures that eviction fences are managed accurately and safely, preventing access to invalid memory addresses. This fix restores normal operation for queue creation and destruction cycles involving exported DMA buffers. System administrators should ensure their kernels are updated to include this patch, which has been cherry-picked from upstream commit 6b4157498a176bd1c6083f6bb5aaf8a257a2027e. Regular kernel updates and the application of vendor-provided security patches are essential to mitigate this risk, as unpatched systems remain vulnerable to local denial-of-service attacks via GPU driver interactions.

Responsible

Linux

Reservation

09/25/2026

Disclosure

11/20/2026

Moderation

accepted

CPE

ready

EPSS

0.00167

KEV

no

Activities

very low

Sources

Want to know what is going to be exploited?

We predict KEV entries!