Executive brief
A race condition in the Linux kernel's AMD XDNA accelerator driver could cause the system to behave unexpectedly when processing commands. This occurs when a user application sends a command at the exact moment the hardware is automatically powering down to save energy. While primarily a stability issue, it could lead to failed operations or system instability on devices using these specific AMD accelerators.
Technical details
A race condition exists in the accel/amdxdna driver's power management logic. When autosuspend is triggered, the 'rpm_on' flag is set to indicate a state transition. If a userspace application submits a command during this window, amdxdna_pm_resume_get() may incorrectly skip the necessary resume operation because it sees the 'rpm_on' flag as already set. This results in hardware commands being dispatched to a device that has not fully resumed. The fix involves removing the 'rpm_on' flag and introducing aie2_pm_set_dpm() to explicitly manage device resumption before invoking Dynamic Power Management (DPM) changes.
Affected products
- Linux Linux Kernel 6.12, 6.13, 6.14
Timeline
- 2025-12-08: disclosed: Initial patch submission by AMD developers.
- 2026-02-26: patched: Patch merged into stable kernel branches.
- 2026-05-27: advisory: CVE-2025-71303 published.