Intel Revs Linux Patches For Per-Client Engine Busyness - Allowing For Great GPU Insight

One of the set of patches for Intel's Linux kernel graphics driver that have been floating around for more than one year is about exposing per-client (process) statistics in how each application is making use of the GPU's render/blitter/video hardware and various insightful statistics related to that. The patches aren't queued for mainline yet but at least a new revision of the work was published.

Intel's Tvrtko Ursulin on Friday posted the latest series of his "per-client engine busyness" patches that ultimately expose enough information to user-space for creating a "top-like" tool for showing great detail over how individual system processes are making use of the GPU. Here is the textual/CLI visual of what could be exposed using this patch series:

It certainly would be quite useful and with the data being exposed over sysfs other user-space programs could be querying this information too. But it's unfortunate that it has taken more than one year and the work still isn't merged.

The new series adds support for providing the data in GuC mode and refactors some of the code. The infrastructure added to this kernel driver can also be re-used moving forward for improving the GPU scheduling decisions, cgroups controller, dynamic SSEU tuning, and other features mentioned by Ursulin.

It's becoming too late for seeing this in Linux 5.6 as the DRM-Next cut-off is upon us, but hopefully this work can get squared away in time for Linux 5.7.