CVE-2025-68174: 
Linux Kernel vulnerability analysis and mitigation

CVE-2025-68174 is a race condition vulnerability in the Linux kernel's AMD GPU driver (amdkfd) that can lead to a divide-by-zero kernel crash. The flaw exists in the partition switching logic, where the kfd_processes_table emptiness check is insufficient to prevent concurrent access between a workqueue releasing a KFD process and a partition switch tearing down KFD nodes. It was published on December 16, 2025, and affects the Linux kernel's amdgpu driver. The EPSS score is approximately 0.018% (Low), indicating limited exploitation probability at this time (Feedly).

The root cause is a time-of-check/time-of-use (TOCTOU) race condition (CWE-362) in the AMD KFD (Kernel Fusion Driver) partition switching code. When Process A (a workqueue) executes kfd_process_wq_release and accesses kfd_node members, and Process B concurrently triggers a partition switch via amdgpu_xcp_pre_partition_switch → amdgpu_amdkfd_device_fini_sw → KFD node teardown, the node can be destroyed while still in use. The existing guard only checks whether kfd_processes_table is empty, but the table entry is removed in kfd_process_notifier_release — before the actual process teardown completes in kfd_process_wq_release — leaving a window where the node is accessed after being freed. The fix introduces an atomic counter kfd_processes_count per kfd_dev that is incremented on KFD process creation and decremented only after kfd_process_wq_release completes, closing the race (Feedly).

Successful exploitation of this race condition results in a divide-by-zero kernel panic (divide error: 0000 [#1] SMP NOPTI) triggered in kfd_get_num_sdma_engines, causing a system crash and denial of service. The crash occurs in kernel context (workqueue kfd_process_wq), meaning it affects the entire host system rather than a single process. There is no evidence of confidentiality or integrity impact; the primary consequence is availability loss through a kernel crash on systems using AMD GPUs with KFD support (Feedly).

• Logs: Kernel dmesg output containing divide error: 0000 [#1] SMP NOPTI with a stack trace referencing kfd_get_num_sdma_engines in the amdgpu module and workqueue kfd_process_wq kfd_process_wq_release.
• Process: Kernel crash (oops/panic) originating from kworker threads associated with kfd_process_wq on systems with AMD GPU hardware.
• System: Unexpected system reboots or kernel panics on hosts running AMD GPU workloads (e.g., ROCm/HPC environments) coinciding with GPU partition switching operations.

The fix has been merged into the Linux kernel stable tree via commits 45da20e00d5da842e17dfc633072b127504f0d0e and 536d80f660ec12058e461f4db387ea42bee9250d. Administrators should update to a patched kernel version that includes these commits. As a workaround, avoiding concurrent KFD process teardown and GPU partition switching operations can reduce exposure. Distributions including SUSE have begun shipping kernel updates addressing this issue (Feedly, Linux Security).