Linus Torvalds Puts His Name on a Claude Code Kernel Patch for the First Time; Eight-Thread Cache Read Throughput Triples

On October 12, 2026, Linus Torvalds submitted the noref-filemap-read-with-llm patch series to the Linux kernel mailing list. The cover letter states that "the code, test programs, list review, and the initial draft of this message were all completed by an LLM coding assistant (Claude Code) under my guidance." Eight-thread 16-byte cache read throughput rose from about 10 million operations per second to 30–33 million operations per second, while CPU cycles fell from 2,950 to 725.

Factual reconstruction

The problem addressed by the patch centers on concurrent multithreaded reads of small chunks of the same file, where the kernel page cache must perform atomic reference-count updates on a folio; under high concurrency, these atomic operations contend for the same cache line. Torvalds's approach completes the read within an RCU read-side critical section, skips the reference-count operation, and defers the actual free until after an RCU grace period; when the number of pages queued for freeing per CPU exceeds 1,024, it triggers an accelerated grace period to prevent unbounded memory accumulation. The measurements come from Torvalds's own Core i7-8565U laptop; in the single-threaded unmapped-page case, cycles rose from about 4,150 to 6,550.

According to Phoronix, the series targets Linux 7.3-rc6 and had not yet been merged into the 7.4 merge window at the time of writing. In the cover letter, Torvalds also notes that the measurements were run only on his personal machine.

Mechanism breakdown

The original implementation updates an atomic reference count on every read; with eight concurrent threads, Load HITM events reached 30,538. The new approach uses RCU characteristics to remove the reference count, reducing HITM events to 101 and thereby improving throughput. In the single-threaded, no-concurrency case, added bookkeeping overhead raises cycles, illustrating the trade-off between concurrency optimization and the single-threaded path.

The patch submitter clearly distinguishes AI-generated content from human oversight: Torvalds himself ran the benchmarks and decided whether it was worth merging, a process that contrasts with the earlier case of vague AI-generated CVE reports being described as "nearly unmanageable."

Industry impact

The world's most closely watched kernel maintainer attributing AI collaboration in a formal patch for the first time signals that AI coding tools are moving from helper scripts and application-layer code into the lowest-level system software. In October 2024, Torvalds said 90% of AI discussion was marketing hype; in July 2026, he publicly said Linux is not an anti-AI project and that critics can fork or leave.

Google's open-source Sashiko AI code review system reportedly caught about 53% of kernel defects in a test set, echoing Torvalds personally testing and putting his name on this contribution. Faced with this signal, enterprises and infrastructure teams will find it difficult to treat AI tools as merely a marketing case.

Strategic assessment (analysis, not fact)

Judging from the shift in Torvalds's historical stance, whether AI tools can enter critical paths depends on whether developers retain final measurement and decision-making authority. The patch explicitly notes its limitations (single-machine benchmarks, single-threaded regression); this transparent approach may become a reference for other projects evaluating AI contributions. Over the long term, similar attribution may reduce the community's overall skepticism toward AI-generated code, provided contributors continue to supply reproducible hardware data rather than relying solely on model output.

If more kernel maintainers adopt the same model, the credibility of AI coding tools in safety-critical, low-tolerance codebases will rise further; conversely, without a human review step, low-quality AI-generated patches may still be rejected.