Torvalds is testing page-cache reads that skip folio references

A branch named noref-filemap-read-with-llm showed up in Linus Torvalds’ kernel tree on 11 October 2026. He does not usually park feature work on his own repository. The name is literal. The draft, the tests, and the first pass of the cover text were produced with Claude Code, under his direction. He says the measurements were run on his own machine.

The problem is contention. When many threads perform small reads of one file, they pile up on folio reference counting in the memory-management code. The branch tries to read from the page cache without taking those references, so the counter is no longer the choke point for that pattern.

This is not a merged feature. It sits on a side branch while 7.3 is still in release-candidate fixes. Torvalds has said he expects to see whether it belongs in the 7.4 merge window. That is the right pace. A page-cache change that skips a reference count can be faster and also wrong in ways that only show up under reclaim, truncation, or a file that is being written while it is read.

The AI part is narrower than the headlines suggest. Torvalds has used models for debugging and for code outside the kernel before, and he has described them as a tool. Here the tool drafted a subsystem change he is willing to measure himself. The kernel process in front of that draft is unchanged: list review, benchmarks other people can reproduce, and a merge decision that is still open.

Anyone running a workload with many threads reading small pieces of the same file has a reason to watch the branch. Nobody should treat it as the 7.3 fix set.