On 10/18/26 06:53, Qiliang Yuan wrote: > ksm_tests fills regions with identical data and checks KSM's counters > after merging them. allocate_memory() maps these regions without any > THP hint. > > With mTHP enabled the regions can be backed by large folios, which KSM > splits before merging. Splitting an anonymous folio maps its > zero-filled pages to the shared zeropage, so with use_zero_pages=0 > those pages never become KSM pages and "KSM zero page merging" fails > its count. "ksm_tests -Z -p 10 -z 0", 20 runs: > > THP setup before after > default 0/20 0/20 > every mTHP size "always" 20/20 0/20 > > Map the regions with MADV_NOHUGEPAGE, as ksm_functional_tests does. > ksm_merge_hugepages_time() maps its own THP-backed region and is not > affected.
That's actually something Nico intends to fix: https://lore.kernel.org/all/[email protected]/ > > Fixes: b1f202060afe ("mm: remap unused subpages to shared zeropage when > splitting isolated thp") > Signed-off-by: Qiliang Yuan <[email protected]> > --- > tools/testing/selftests/mm/ksm_tests.c | 10 ++++++++++ > 1 file changed, 10 insertions(+) > > diff --git a/tools/testing/selftests/mm/ksm_tests.c > b/tools/testing/selftests/mm/ksm_tests.c > index 5fd7792a0d479..140ea833353b6 100644 > --- a/tools/testing/selftests/mm/ksm_tests.c > +++ b/tools/testing/selftests/mm/ksm_tests.c > @@ -178,6 +178,16 @@ static void *allocate_memory(void *ptr, int prot, int > mapping, char data, size_ > ksft_perror("mmap"); > return NULL; > } > + /* > + * Splitting a large folio before merging maps its zero-filled pages > + * to the shared zeropage, so keep the pages small. Ignore EINVAL on > + * kernels without THP. > + */ > + if (madvise(map_ptr, map_size, MADV_NOHUGEPAGE) && errno != EINVAL) { > + ksft_perror("madvise"); > + munmap(map_ptr, map_size); > + return NULL; > + } > memset(map_ptr, data, map_size); > if (mprotect(map_ptr, map_size, prot)) { > ksft_perror("mprotect"); > > --- > base-commit: 90201351a862182bdabfc1ad1c97b1e1b6ed5931 > change-id: 20261016-bug-selftests-mm-ksm-tests-nohugepage-1414868db0ef > > Best regards, -- Cheers, David
