On Mon, 26 Oct 2026 10:38:50 +0100 Maxime Chevallier wrote: > On 10/23/26 19:52, Jakub Kicinski wrote: > > On Wed, 21 Oct 2026 11:08:51 +0200 Maxime Chevallier wrote: > >> The devmem selftest requires accessing /proc/sys/vm/nr_hugepages, > >> which is only available when hugetlbfs is supported. This isn't the > >> case on all CPU architectures, for example it isn't supported on ARM > >> 32-bits on cores that don't support LPAE. > >> > >> On these devices, the devmem kselftest doesn't skip gracefully > > > > We don't require graceful skipping. If anything IMO hiding the stack > > traces isn't great. > > > > If the check is extraneous and makes us fail some tests which don't > > actually require huge pages please make that clear in the commit > > message. Otherwise the net benefit of the changes is unclear. > > hugetlbfs isn't required for devmem AFAICS, but the blamed commit > moved the access to nr_hugepages before the actual "is devmem > actually supported" test. > > I don't expect the full test suite to gracefully run/skip for devmem > on these arm32 platforms TBH, it's likely that devmem is only something > that's been used on platforms that support hugetlbfs anyways, but > strictly speaking both features are unrelated.
For max clarity I should have simply asked - does it make any test which previously failed pass on any of your platforms? (incl. some WIP patches that'd add devmem support) That's basically the criteria I think we should use.
