freenode
Kernel & Low-Level

Torvalds posts Linux page refcount scalability series

After years of incomplete fixes for buffered read bottlenecks, Linus Torvalds mails a new approach he refined with LLM help ahead of the maintainer summit.

Linus Torvalds has posted a series to improve page reference-count scalability in the Linux kernel, targeting a bottleneck that has limited concurrent buffered I/O for ordinary read and pread system calls.

Real users hit the problem years ago. Matthew Wilcox reported scalability limits in buffered I/O, which prompted a rough draft from Torvalds that he ran locally for more than a year until others showed it was badly broken. A later cleanup of that approach from Kirill Shutemov still had problems. Ilya Gladyshev then offered a different design, now at its sixth revision, that left known fundamental issues open.

With the topic still unresolved and LLM use in kernel development on the agenda for this year's maintainer summit, Torvalds set out to finish the job. "I decided that hey, I will try to actually solve this thing using an LLM to help with the issues these things have had," he wrote. He had already shown private versions to other developers and discussed the work at the summit before sending it to the list.

The series aims to make the common page-ref path scale without the breakage that sank earlier attempts. Torvalds said a separate, more routine fix for the assoc-array library would follow on its own.