Torvalds's AI-assisted Linux patches reach review, and a race condition surfaces

From an experimental branch to public review

Linus Torvalds’s AI-assisted Linux page-cache experiment has moved into a ten-patch public review, with an early correctness problem already reported. LWN’s October 14 account says reviewers identified a race condition, giving the discussion a concrete engineering question beyond whether using a language model is acceptable.

In his cover letter, dated October 13 in the mailing-list archive, Torvalds describes the rcu-free-v3 branch as a revision of the earlier noref-filemap-read-with-llm experiment. He says he used an LLM to help tackle a longstanding scalability problem affecting ordinary file reads. The new series delays freeing frequently accessed cached pages until an RCU grace period has elapsed, allowing a fast read path to avoid repeated reference-count updates when copying data does not trigger a page fault.

A striking number with a narrow meaning

Torvalds reports a 23-fold improvement on his Ryzen Threadripper in an artificial test that repeatedly reads the same small portion of a page across all CPUs. His displayed 64-thread comparison rises from 13.3 million to 304 million reads per second. He explicitly says real workloads will see much less benefit. Those are his measurements, not an independently reproduced result or a general Linux speedup.

The new event is the revised proposal and its review, following the earlier eight-patch branch. LWN’s race-condition report means further work is needed; the sources do not establish a completed merge or a fix already delivered to users. The English Hacker News discussion meanwhile debates expert supervision and code review, illustrating why successful benchmark runs and production readiness remain different questions.