On Tue, May 03, 2022 at 03:00:05PM +0200, Sebastian Andrzej Siewior wrote: > On 2022-05-03 13:48:50 [+0100], Matthew Wilcox wrote: > > Thanks. Stephen already reported that; fix here: > > > > https://lore.kernel.org/all/Ym++SI1ftbRg+9zK@xxxxxxxxxxxxxxxxxxxx/ > > Stephen says "I applied the above patch to the mm tree merge today" and > I have here next-20220503. I don't have the BUILD_BUG() in > can_split_folio() anymore so I have this change. Ah! I didn't realise you were testing next; I thought you'd picked up these patches some other way. > My guess is that since THP_FILE_ALLOC is defined as BUILD_BUG() for > !CONFIG_TRANSPARENT_HUGEPAGE and there is nothing that removes that part > of the code, I end up in BUILD_BUG with CGROUP and no THP. > > PageTransHuge() used to "return false" for !CONFIG_TRANSPARENT_HUGEPAGE > which isn't the case for folio_test_large(). Indeed, indeed. I missed another case. This fixes it: diff --git a/mm/shmem.c b/mm/shmem.c index 5b161a92e6f1..019ad8bf0d21 100644 --- a/mm/shmem.c +++ b/mm/shmem.c @@ -716,7 +716,7 @@ static int shmem_add_to_page_cache(struct folio *folio, if (!folio_test_swapcache(folio)) { error = mem_cgroup_charge(folio, charge_mm, gfp); if (error) { - if (folio_test_large(folio)) { + if (folio_test_pmd_mappable(folio)) { count_vm_event(THP_FILE_FALLBACK); count_vm_event(THP_FILE_FALLBACK_CHARGE); }
