Rewrite Linux page-cache reads to drop folio reference counting

I want to improve the performance of [a hot path in this codebase, e.g. the function that handles small reads from the page cache] under contention from many threads. First read the current implementation and explain where the contention comes from. Then propose a change that avoids [the shared counter or lock] on the fast path, implement it as a series of small commits, and write a test program that reproduces the contention so we can measure before and after. Run the benchmark and report the numbers. Draft the cover letter for the series, stating plainly which parts you wrote, and stop before anything is sent.