Back in October 2018 came the initial patches for providing per-process GPU usage reporting to be exposed to user-space for interesting metrics akin to the top command or other system monitoring utilities but for detailed GPU statistics. In October that interesting work finally saw a revision but went dark after that and didn't make it into the recent Linux 5.5 merge window. Now a new spin of that code has been sent out for review.

The per-client engine statistics reporting makes it possible for user-space utilities to report how busy a given system process is keeping the GPU. It's not just limited to a single metric but per-engine can report how much a particular game/application is utilizing the 3D hardware, blitter, video, and related engines. It opens interesting possibilities and here is an earlier patch illustrating a CLI-based Intel GPU TOP reporting example:

The metrics are to be exposed via sysfs so other existing/new applications could also scrape this data.

Sent out Thursday were the new patches that include the ability to track GPU time for clients that exit the GPU with work left running and other code improvements.

Given it should be quite useful for many different use-cases, hopefully we'll see this work finally land in Linux 5.6 or another near-term kernel release in 2020.