Linus Torvalds' AI Code Improved: Performance Wins with Reworked Code

Linus Torvalds’ AI-assisted kernel idea has been reworked into a concrete patch series that eliminates folio reference taking for small, multi-threaded reads from the page cache. The change lets filemap_read() and shmem copy data directly from page cache pages without incrementing folio refcounts; instead, folios that such readers might access are only freed after an RCU grace period. That removes the frequent refcount bouncing that occurs when many CPUs repeatedly take and drop references on the same folio, reducing contention and overhead in workloads dominated by many small reads from a single file.

Meta engineer Kiryl Shutsemau tested the reworked code across EXT4, TMPFS, XFS and a raw block device using 64 threads on an AMD Ryzen Threadripper, with and without default CPU security mitigations enabled. The benchmarks show very large performance gains for the targeted workloads. The patch series is currently an RFC with some design details still open, but it demonstrates practical, measurable improvements over the existing approach and is posted for review on the kernel mailing lists as work toward upstream inclusion.