The client osc LRU (max_cached_mb, osc_lru_alloc(), osc_lru_del(), osc_lru_add_batch(), osc_lru_shrink() in lustre/osc/osc_page.c) accounts one slot per cl_page. With large folio support (LU-19895, https://review.whamcloud.com/65094) one cl_page can cover a folio of many base pages (cp_npgs), so a large folio takes a single LRU slot whatever its size. Cached base pages can then exceed max_cached_mb by up to the folio size per page, for written folios as well as readahead folios (readahead ra_cur_pages is already accounted in base pages and returns folio_nr_pages() on read completion).

The fix is to count LRU slots in base pages (cp_npgs) in all the osc LRU counters and in the LRU shrink/reclaim paths, with tests that fill max_cached_mb using large folios.

Found while reviewing change 65094 (thread on lustre/llite/rw.c, readahead folio accounting), where Andreas Dilger asked for this as a follow-on to that patch.