A set of patches to the Linux kernel's Direct Rendering Manager (DRM) scheduler that is shared among different kernel graphics drivers is showing the potential of delivering much lower job submission latency when the system is loaded with many runnable CPU processes.

Tvrtko Ursulin of Igalia sent out the request for comments (RFC) on Thursday to deal with cases where a system is overloaded with runnable CPU processes and in turn a high priority DRM client ends up seeing latency injected into the GPU submission path due to the DRM scheduler's use of kernel workqueues.

The proposed patches replace the use of workqueues with kthread_work as used for asynchronous task execution. They found a "striking improvement" with this DRM scheduler rework:

Ursulin added that, "this is obviously really good for preventing compositors from missing frames and similar."

This is very promising for the Panthor Arm Mali driver. Not all of the DRM scheduler using drivers have yet to be adapted to these RFC patches but so far just Intel Xe and Panthor. The AMDGPU driver and others that align the DRM scheduler as a hardware scheduler aren't expected to see a significant difference like the Panthor numbers above.

Those wanting to learn more about this DRM scheduler adaptation to replace workqueues with kthread_work can find the RFC patches on the dri-devel list.