Linus Torvalds posts his first patch written with an LLM

The maintainer of the Linux kernel wrote a patch series with help from a language model, posted it to the kernel mailing list for review, and said so out loud. LWN reported on 14 October that Linus Torvalds has posted his first LLM-assisted patch, a short series of ten patches that improves how the kernel looks up folios in memory. The series had already been reviewed by a human and a race condition had been found before LWN went to press.

What the series actually does

The ten patches touch the memory management code. They let the kernel notice when a folio, a group of memory pages, is being read often. For those hot folios the code delays releasing the folio until after a read-side critical section finishes, which is a standard pattern in the kernel's read-copy-update (RCU) subsystem. The payoff is that the kernel can read the folio's contents under RCU as well as find it under RCU, with no reference count taken on the fast path when the user copy does not cause a page fault.

Torvalds quoted a one-benchmark result of an order-of-magnitude speedup on read-heavy workloads. LWN prints the quote in full and gives no second benchmark, so one number against one workload is what the series stands on so far.

The review found a bug the LLM missed

The LWN brief links to a review comment on the mailing list that identifies a race condition in the code. LWN says the race is real and that further work will be needed before the series can land. The piece does not describe the race in detail.

That is the part worth reading twice. The patch series moved faster than a hand-written one would have, and the review still had to find a concurrency bug the model did not notice. On the previous day of the Maintainers Summit, which LWN links from the same brief, Torvalds had said he had been using "a large language model" on the problem. The LWN brief does not name a vendor. A top comment on the HN thread observes that Torvalds has done management rather than direct kernel coding for years, so the series is also his first new code in the kernel in a while.

What it tells a team deciding whether to try this

Any team arguing about whether a senior developer should be writing with a model's help now has a public worked example to read. The series is small enough to read in full, the result is quoted with a benchmark, and the review process on the kernel mailing list is open: anyone can look up the comment that caught the race condition and see what the human reader did that the model did not. The question for the team is how the review stays honest when the author is faster than before.

Source

Linus Torvalds's first LLM-assisted patch series, corbet, LWN.