damon_pa_pageout() uses damon_get_folio(), which rejects hugetlb folios because they are not on the LRU. The loop then advances by PAGE_SIZE and looks up every base page of a hugetlb folio that cannot be reclaimed. Use damon_get_monitor_folio() and skip a hugetlb folio as a whole by advancing to its end. Do this before the filters, the accounting and the last_applied check, so that adjacent regions within the same hugetlb folio are not walked by base page either. Test the LRU flag first to keep the hugetlb check off the path of ordinary LRU folios. In a pageout microbenchmark in an x86-64 KVM guest, traversal time dropped by about 99.7% for regions covering whole 2 MiB hugetlb folios. As a side note, there is a small overhead of about 1 ns per region when each region is a single 4 KiB page. Signed-off-by: Aveline Noir --- Tested in x86-64 KVM guests. KUnit passes. The damos_quota.py selftest can also fail without this patch; I will send a separate fix for it. --- mm/damon/paddr.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/mm/damon/paddr.c b/mm/damon/paddr.c index 5b773012a876..aca48bf752c8 100644 --- a/mm/damon/paddr.c +++ b/mm/damon/paddr.c @@ -268,7 +268,9 @@ static unsigned long damon_pa_pageout(struct damon_region *r, addr = damon_pa_phys_addr(r->ar.start, addr_unit); while (addr < damon_pa_phys_addr(r->ar.end, addr_unit)) { - folio = damon_get_folio(PHYS_PFN(addr)); + folio = damon_get_monitor_folio(PHYS_PFN(addr)); + if (folio && !folio_test_lru(folio) && folio_test_hugetlb(folio)) + goto put_folio; if (damon_pa_invalid_damos_folio(folio, s)) { addr += PAGE_SIZE; continue; --- base-commit: 67f3b04b277373438f17300e056306adbd9edf7b change-id: 20261117-fix-damon-hugetlb-pageout-20261117-a4e03d9d5a5b Best regards, -- Aveline Noir