Linus Torvalds releases Linux 7.2 with AI-sourced patches and cache-aware scheduling

Linus Torvalds said the week leading up to the release was “the new normal”, with a large number of patches, many of which originated from scans by various AI tools. Contrary to what some might expect, the contributions are not new features written by models but security bugs that the tools detected. Torvalds clarified that bugs discovered by AI are “almost by definition not secret”, and handling them in private lists merely creates duplicate work and wastes time for everyone. He is not opposed to using the tools, but insists they should add value rather than generate artificial work.

The headline feature of the release is Cache Aware Scheduling, a series under development for several years. The idea is simple: instead of balancing load across cores without regard to shared memory, the scheduler groups tasks that share data in the same Last Level Cache (LLC) domain. The result is fewer cache bounces, fewer cache misses, and more efficient memory access. Guardrails prevent a domain from exceeding 25 % of CPU time or one third of its load, and threads are not migrated once a process is already consuming significant resources. Multi-threaded servers, AI/ML workloads, and some games stand to gain the most.

In the driver and I/O area, initial support for AMDGPU includes HDMI 2.1 Fixed Rate Link, enabling uncompressed 4K at 120 Hz and 8K. On the storage side, exFAT now uses the kernel’s iomap, delivering performance gains of up to 87 % on SD cards and USB-C flash drives; running a distribution from USB feels markedly faster. NTFS received bug fixes, and ext4 with XFS continue the I/O-map conversion work. Experimental folios up to 2 MB have been added, and support for Intel’s USB4STREAM protocol enables raw Host-to-Host streaming over USB4/Thunderbolt, which Torvalds called “super simple”.

Users who do not live on the bleeding edge will see the changes in fall releases such as Ubuntu 26.10. The full advantage will be felt on hardware with complex cache topologies, newer GPUs, mobile devices, and heavy storage workloads. The code is already in mainline and open for testing.