Intel's Habana Labs Continues Improving Their AI Accelerator Driver Stack With Linux 5.16

The "char/misc" changes have landed in Linux 5.16 as the catch-all area for code not fitting more appropriately within another subsystem or being a portion of the kernel that's too small for submitting pull requests directly to Linus Torvalds. For Linux 5.16, the IIO code has moved from being part of the staging area also maintained by Greg Kroah-Hartman to now being under the char/misc umbrella. Additionally, another big item is the Intel Habana Labs driver updates.

The Habana Labs kernel driver for supporting their Goya and Gaudi AI hardware continues to be part of the char/misc area with not yet having a formal "accelerator" subsystem short of the GPU/DRM area for the growing number of AI inference/training accelerators coming to market.

Most significant with the Intel-owned Habana Labs driver updates in Linux 5.16 is the adding of peer-to-peer support via DMA-BUF. That work was previously rejected by upstream DRM maintainers over at the time Habana Labs lacking any open-source user-space "client". That's now been addressed with Intel / Habana Labs having opened up more of their user-space code. Thus for Linux 5.16 there is a new API for this driver to export a DMA-BUF and is for peer-to-peer sharing over PCIe with a focus on sharing buffers directly between Gaudi training hardware and RDMA adapters.

Beyond the Habana Labs activity, there are routine IIO driver updates, counter driver updates, and other small changes. The overview of char/misc activity for Linux 5.16 can be found via this merge.