Linus Torvalds submits his first AI-assisted patch to tackle Linux page cache bottlenecks

Linux creator Linus Torvalds has published his first AI-assisted patch series to address an old memory bottleneck that occurs when multiple processor cores read the same file. Developing the eight-patch series on his noref-filemap-read-with-llm branch with Anthropic's Claude Code, Torvalds targeted severe cache contention in the Linux page cache. Torvalds reported that a synthetic test on an eight-thread Intel laptop pushed read throughput from roughly 10 million reads per second to over 30 million, and Tom's Hardware noted synthetic benchmarks showed up to a 23x performance leap. [1][2]

The experimental patches alter page cache behavior by deferring memory reuse until a read-copy-update (RCU) grace period passes, allowing threads to copy file data without updating shared reference counters. However, Torvalds emphasized that real-world workloads will not see those synthetic gains, noting that deferred memory reclamation remains the least solved part of the proposal. The code also introduces a performance penalty when reading unmapped user pages, leaving the patches outside mainline Linux 7.3-rc6 pending further review. [1][2]

The experiment underscores a broader shift in kernel governance toward embracing generative tools. The Linux team has stopped requiring developers to disclose AI assistance in their patches, with Torvalds comparing large language models to standard editors and compilers. Meanwhile, AI tools have introduced new operational hurdles for the project, flooding security mailing lists and driving fix rates to roughly 2,000 vulnerabilities per release. [1]