Commit f183324133ea ("percpu: implement partial chunk depopulation") placed pcpu_sidelined_slot at __pcpu_size_to_slot(pcpu_unit_size) + 1 and moved pcpu_free_slot to pcpu_sidelined_slot + 1 so pcpu_alloc() checks sidelined chunks before free chunks. pcpu_balance_populated() also loops from pcpu_size_to_slot(PAGE_SIZE) through pcpu_free_slot, but was not updated to skip pcpu_sidelined_slot. When pcpu_nr_empty_pop_pages drops below PCPU_EMPTY_POP_PAGES_HIGH and active chunks have no unpopulated pages, pcpu_balance_populated() walks into pcpu_sidelined_slot and repopulates chunks that were just sidelined to save memory. Because chunk->isolated is still true on sidelined chunks, pcpu_update_empty_pages() does not count those newly populated pages toward pcpu_nr_empty_pop_pages. That leaves the global empty page count below the threshold, so subsequent balance passes keep refilling every sidelined chunk back to 100% before reaching pcpu_free_slot. Skip pcpu_sidelined_slot in pcpu_balance_populated() so sidelined chunks stay depopulated until pcpu_alloc() actually needs them. Benchmark results (4-CPU arm64, 4 sidelined chunks with 60 reclaimed pages after triggering pcpu_balance_populated()): Metric Before After Delta -------------------------------------------------------------------- Sidelined pages repopulated 60 0 -100% Sidelined backing pages (kB) 960 0 -960 kB Fixes: f183324133ea ("percpu: implement partial chunk depopulation") Cc: stable@vger.kernel.org Signed-off-by: Chris Wiseman --- mm/percpu.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/mm/percpu.c b/mm/percpu.c index 3eff382e565a..ca83773a6728 100644 --- a/mm/percpu.c +++ b/mm/percpu.c @@ -2045,6 +2045,9 @@ static void pcpu_balance_populated(void) if (!nr_to_pop) break; + if (slot == pcpu_sidelined_slot) + continue; + list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) { nr_unpop = chunk->nr_pages - chunk->nr_populated; if (nr_unpop) -- 2.47.0