Intel Still Hasn't Gotten To Landing Per-Client Engine Busyness Reporting For Linux

One of the patch series I have been looking forward to see land going back to 2018 has been the per-process GPU load statistics or as it's known officially the "per client engine busyness" series. The work didn't land for Linux 5.14 but at least this week the latest revision was posted.

The newest "per-client engine busyness" patches are pretty much the same as what has been volleyed out now for about three years. This functionality is not only about reporting how busy the render/3D, blitter, video engines, and other blocks are but for being able to report such information on a per-process basis. Thus quite interesting to performance-oriented folks like myself for having this per-process and per-hardware-block reporting.

While this newest patch series is basically re-basing this long standing patch series against the latest code, new this time around is adding documentation on the format for reporting this information to user-space. The hope is this DRM documentation can become a common specification so user-space tools can implement reading of this information but ideally seeing other Linux DRM graphics drivers also export their per-process reporting in a compatible format. This is a much welcomed addition but remains to be seen if/when other drivers will jump on this bandwagon, which would likely motivate more user-space tooling to utilize this per-process GPU information.

This latest Intel Linux graphics driver patch series can be seen on the driver mailing list. This patch has the new documentation outlining what is exposed to user-space.

Here's to hoping it doesn't take another year before these patches are finally merged... This functionality would also be useful if an Intel Linux graphics control panel / UI materializes.