Re: [PATCH] mm/hugetlb: align cma on allocation order, not demotion order

[Frank van der Linden]

On Thu, May 2, 2024 at 6:15 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 30.04.24 18:14, Frank van der Linden wrote:
> > Align the CMA area for hugetlb gigantic pages to their size, not the
> > size that they can be demoted to. Otherwise there might be misaligned
> > sections at the start and end of the CMA area that will never be used
> > for hugetlb page allocations.
> >
> > Signed-off-by: Frank van der Linden <fvdl@xxxxxxxxxx>
> > Cc: Roman Gushchin <roman.gushchin@xxxxxxxxx>
> > Fixes: a01f43901cfb ("hugetlb: be sure to free demoted CMA pages to CMA")
> > ---
> >   mm/hugetlb.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index 5dc3f5ea3a2e..cfe7b025c576 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -7794,7 +7794,7 @@ void __init hugetlb_cma_reserve(int order)
> >                * huge page demotion.
> >                */
> >               res = cma_declare_contiguous_nid(0, size, 0,
> > -                                     PAGE_SIZE << HUGETLB_PAGE_ORDER,
> > +                                     PAGE_SIZE << order,
> >                                       HUGETLB_PAGE_ORDER, false, name,
> >                                       &hugetlb_cma[nid], nid);
> >               if (res) {
>
> I was wondering how that worked when reviewing your other patch.
> Wondering why we never got a BUG report, maybe we were always lucky
> about the alignment we actually got?

I think this issue was probably masked by the hugetlb allocator
falling back to direct alloc_contig_pages allocation if cma_alloc
fails. So if you're not under memory pressure, the failure to allocate
from the misaligned areas might not have been noticed.

I noticed it, because I was working with change I made: a flag that
prevents the fallback to straight alloc_contig_pages, as that behavior
may not be desired - you don't want to potentially eat in to
non-movable space that the kernel needs, it might be better to fail if
there's no CMA available.
>
> We round up size to PAGE_SIZE << order, so that's the alignment we need.
>
> Reviewed-by: David Hildenbrand <david@xxxxxxxxxx>

Thanks!