On 18/02/2019 17:04, Mark Rutland wrote: > On Mon, Feb 18, 2019 at 03:30:38PM +0000, Steven Price wrote: >> On 18/02/2019 15:06, Peter Zijlstra wrote: >>> On Mon, Feb 18, 2019 at 02:29:52PM +0000, Mark Rutland wrote: >>>> I think that Peter means p?d_huge(x) should imply p?d_large(x), e.g. >>>> >>>> #define pmd_large(x) \ >>>> (pmd_sect(x) || pmd_huge(x) || pmd_trans_huge(x)) >>>> >>>> ... which should work regardless of CONFIG_HUGETLB_PAGE. >>> >>> Yep, that. >> >> I'm not aware of a situation where pmd_huge(x) is true but pmd_sect(x) >> isn't. Equally for pmd_huge(x) and pmd_trans_huge(x). >> >> What am I missing? > > Having dug for a bit, I think you're right in asserting that pmd_sect() > should cover those. > > I had worried that wouldn't cater for contiguous pmd entries, but those > have to be contiguous section entries, so they get picked up. > > That said, do we have any special handling for contiguous PTEs? We use > those in kernel mappings regardless of hugetlb support, and I didn't > spot a pte_large() helper.
There's no special handling for contiguous PTEs because the page walk code doesn't care - each PTE is valid individually even if it is part of a contiguous group. So the walker can descend all levels in this case. pte_large() if it existed would therefore always return 0. The pte_entry() callback obviously might go looking for the contiguous bit so that it can annotate the output correctly but that's different from a 'large' page. The code in arch/arm64/mm/dump.c simply looks for the PTE_CONT bit being set to do this annotation. Steve
