Torvalds Publishes Eight Page Cache Patches Written With Claude Code

Linus Torvalds pushed an eight-patch experimental series to his own kernel repository on Oct. 10, 2026, and the commit message says Claude Code wrote the code, the test programs and the first draft of the explanation. The branch is called noref-filemap-read-with-llm, sits on Linux 7.3-rc6 plus fixes, and is not in mainline.

The lead commit is titled "mm: read from the page cache without folio references." Linuxiac quotes its credit line: "The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction. The measurements were run on my machine." The kernel.org page for the commit returned an access error when The Terminal tried it, so the commit text here comes from that quotation and from It's FOSS, which also links the commit.

What the patches change

When several threads read the same cached file, each read takes and releases a reference on the cached memory, and every thread updates the same shared counter. The threads interfere with each other. According to It's FOSS, Matthew Wilcox flagged the issue in 2024, citing a customer's workload.

The series skips the counter. Memory leaving the page cache is not reused until a short delay passes, so a reader can copy data out without holding a reference to it.

The numbers, and what they were measured on

According to Linuxiac's reading of the commit message, the measurements ran on Torvalds's laptop, an Intel Core i7-8565U, using 16-byte reads from one cached file on ext4. Some kernel security mitigations were switched off for the test.

Workload	Before (reads per second)	After (reads per second)
8 threads, same file	about 10 million	about 30 to 33 million
4 threads, separate cores	about 9 to 12 million	about 25 to 26 million
1 thread	about 8 million	nearly 9 million

The single-thread row is the honest part of the table: the gain is a contention fix, and one thread has little contention to remove. These are one author's numbers on one laptop part, and the articles cited here report no independent run.

The costs the commit message admits

The patches add overhead in some cases, especially when reads hit missing pages, per Linuxiac. The commit message also lists unresolved issues: memory allocation behaviour, compatibility, possible regressions, and the price of delaying memory reuse, which leaves more memory idle for a short time. Neither article says when, or whether, the series will be merged, and the commit message carries no target release.

Torvalds on AI, in his own words

It's FOSS ties the branch to remarks Torvalds made at Open Source Summit Europe in Prague, in a conversation with Dirk Hohndel, head of Ericsson Software Technology, as reported by ZDNet's Steven Vaughan-Nichols. Torvalds said: "I actually really like using AI." He also said he uses AI "to do the things that I'm bad at," with a guitar pedal interface project as the example. In August, according to the same article, he used AI to help fix a bug in the Intel Xe graphics driver. In 2024 he had called the AI industry "90 percent marketing and 10 percent reality."

The contrast with other projects is direct. QGIS banned all AI-assisted code in the same week, while REA, an MCP server for reverse engineering, and a developer's claim that agents rebuilt a shooter's source show how widely agents are being pointed at existing code. Torvalds's branch is a different case: a maintainer using an agent under his direction, publishing the credit, and listing what he has not checked.

The next checkpoint is a reproduction. The commit sits in Torvalds's tree for anyone with a multi-core machine and an ext4 volume to run the same read test.