On October 19, 2026, Phoronix reported that Meta engineer Kiryl Shutsemau had reworked an earlier Linux page-cache RFC by Linus Torvalds. The proposal aims to reduce contention when many CPUs make small reads from one file, by letting specific kernel read paths copy cached data without taking a folio reference.
How the proposed read path addresses contention
A folio is a kernel-managed unit of page-cache memory. When many CPUs make small reads from the same file, each read can take and drop a reference to the same folio, causing its reference count to bounce between CPUs. The proposal changes filemap_read() and shmem so they can copy from the page cache without taking that reference.
The cover-letter explanation reproduced by Phoronix says a folio that a reader may copy from is freed only after an RCU grace period. That delay is the proposed safeguard for the folio’s lifetime while the read path avoids taking a reference.
The rework builds on an earlier RFC by Linus Torvalds. NeoTeo previously covered Torvalds’ experimental page-cache work and its reported Claude Code assistance.
What the benchmark measures
Phoronix described the results as promising. The chart labels a specific workload: a 16-byte pread() at offset 0, with each thread using its own file descriptor. Its vertical axis measures millions of reads per second on a logarithmic scale.
That is a narrow small-read workload, so the chart describes performance for that operation rather than Linux file I/O in general. Phoronix reported tests across EXT4, TMPFS, XFS, and a raw block device, using 64 threads on an AMD Ryzen Threadripper CPU. The tests ran both with and without the default CPU security mitigations.
The RFC’s reported status
In its October 19, 2026 report, Phoronix described the series as an RFC with open design questions.
