CVE-2026-53143 Overview
CVE-2026-53143 is a buffer overflow vulnerability in the Linux kernel's AMD Kernel Fusion Driver (amdkfd) affecting GFX11 (Navi3x) graphics hardware. The flaw resides in the v11 Memory Queue Descriptor (MQD) manager, which incorrectly assigned CP-compute variants of checkpoint_mqd and restore_mqd to KFD_MQD_TYPE_SDMA queues. These functions operate on 2048-byte v11_compute_mqd structures instead of the 512-byte v11_sdma_mqd structures, producing a 1536-byte overflow during Checkpoint/Restore in Userspace (CRIU) operations on System DMA (SDMA) queues.
Critical Impact
During CRIU checkpoint, 1536 bytes of adjacent Graphics Translation Table (GTT) kernel memory leak to userspace. During restore, the same region is overwritten, corrupting ring buffers or neighboring MQDs.
Affected Products
- Linux kernel drm/amdkfd driver on AMD GFX11 (Navi3x) GPUs
- Kernel builds using the v11 MQD manager with CRIU checkpoint/restore enabled
- Stable kernel branches referenced by commits 16dad1f, 2c5b66c, 352ea59, d02f05d, and d3efcad
Discovery Timeline
- 2026-06-25 - CVE-2026-53143 published to the National Vulnerability Database
- 2026-06-25 - Last updated in NVD database
Technical Details for CVE-2026-53143
Vulnerability Analysis
The defect is a copy-paste regression unique to the v11 MQD backend. All other Application-Specific Integrated Circuit (ASIC) backends — cik, vi, v9, v10, and v12 — correctly route SDMA queues to SDMA-specific checkpoint and restore handlers. The v11 backend instead reuses the compute-queue handlers, which assume a 2048-byte MQD layout.
When CRIU checkpoints an SDMA queue on Navi3x, checkpoint_mqd() reads 2048 bytes from a 512-byte SDMA MQD buffer. The extra 1536 bytes are copied from adjacent GTT memory directly into the userspace checkpoint image. On restore, restore_mqd() writes 2048 bytes back into the 512-byte SDMA MQD, overwriting whatever neighboring kernel objects occupy the trailing 1536 bytes, frequently the ring buffer or other MQDs.
Root Cause
The v11 MQD manager assigns the compute-MQD variants of checkpoint_mqd and restore_mqd to KFD_MQD_TYPE_SDMA queues. Both functions hardcode sizeof(struct v11_compute_mqd) as the copy length. SDMA queues allocate only sizeof(struct v11_sdma_mqd), yielding a 1536-byte out-of-bounds read on checkpoint and a 1536-byte out-of-bounds write on restore.
Attack Vector
Exploitation requires local access to a system with a Navi3x GPU and the ability to invoke CRIU checkpoint/restore on a process owning an SDMA queue. A local attacker can trigger the checkpoint path to exfiltrate kernel GTT memory to userspace, producing an information disclosure primitive [CWE-125]. The restore path provides a controlled out-of-bounds write [CWE-787] against adjacent kernel allocations, which can corrupt ring buffers or neighboring MQDs and may be leveraged for kernel memory corruption.
No public proof-of-concept code is available. The fix introduces dedicated checkpoint_mqd_sdma() and restore_mqd_sdma() functions that operate on the correct v11_sdma_mqd size, aligning v11 with the pattern used by other MQD managers. Refer to the Kernel Git Commit 6fa41db for the canonical patch.
Detection Methods for CVE-2026-53143
Indicators of Compromise
- Unexpected kernel ring buffer corruption messages from the amdgpu or amdkfd subsystems in dmesg after CRIU checkpoint or restore operations.
- CRIU image files containing larger-than-expected SDMA MQD blobs, indicating leaked adjacent GTT memory.
- GPU hangs, ring submission failures, or queue eviction errors on Navi3x hardware shortly after a restore operation.
Detection Strategies
- Audit running kernels for the presence of the patched commits (16dad1f, 2c5b66c, 352ea59, d02f05d, d3efcad) using uname -r and distribution kernel changelogs.
- Monitor process execution for criu dump and criu restore invocations against workloads using AMD compute queues (ROCm, HIP, OpenCL).
- Track loaded kernel modules and verify amdgpu/amdkfd versions correspond to a patched build.
Monitoring Recommendations
- Forward kernel logs to a centralized log platform and alert on GPU driver faults correlated with CRIU activity.
- Inventory hosts equipped with Navi3x (RX 7000 series, W7000 series) GPUs running unpatched kernels and prioritize them for remediation.
- Restrict the CAP_CHECKPOINT_RESTORE and CAP_SYS_ADMIN capabilities to trusted service accounts to reduce the attack surface.
How to Mitigate CVE-2026-53143
Immediate Actions Required
- Apply the upstream kernel patch introducing checkpoint_mqd_sdma() and restore_mqd_sdma() to all GFX11 hosts.
- Update to a distribution kernel that incorporates the referenced stable commits and reboot affected systems.
- Until patched, disable or restrict CRIU checkpoint/restore on workloads that allocate SDMA queues on Navi3x hardware.
Patch Information
The fix is published in the stable Linux kernel tree across multiple branches. Review the referenced commits: Kernel Git Commit 16dad1f, Kernel Git Commit 2c5b66c, Kernel Git Commit 352ea59, Kernel Git Commit d02f05d, and Kernel Git Commit d3efcad. The patch replaces the compute-MQD checkpoint/restore assignments for SDMA queues with SDMA-specific implementations that copy sizeof(struct v11_sdma_mqd) bytes.
Workarounds
- Remove or constrain CAP_CHECKPOINT_RESTORE for non-administrative users on hosts with Navi3x GPUs.
- Avoid running CRIU against ROCm, HIP, or OpenCL processes that allocate SDMA queues until the kernel is updated.
- Where feasible, schedule GPU compute workloads requiring checkpoint/restore on ASICs unaffected by the v11 backend.
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
