On Fri, 7 Dec 2018, Vlastimil Babka wrote:
> > This reverts commit 89c83fb539f95491be80cdd5158e6f0ce329e317.
> >
> > There are a couple of issues with 89c83fb539f9 independent of its partial
> > revert in 2f0799a0ffc0 ("mm, thp: restore node-local hugepage
> > allocations"):
> >
> > Firstly, the interaction between alloc_hugepage_direct_gfpmask() and
> > alloc_pages_vma() is racy wrt __GFP_THISNODE and MPOL_BIND.
> > alloc_hugepage_direct_gfpmask() makes sure not to set __GFP_THISNODE for
> > an MPOL_BIND policy but the policy used in alloc_pages_vma() may not be
> > the same for shared vma policies, triggering the WARN_ON_ONCE() in
> > policy_node().
>
> AFAICS 2f0799a0ffc0 removed the policy check in
> alloc_hugepage_direct_gfpmask() comlpetely, so it's not racy and the
> warning will always trigger for a MPOL_BIND policy right now?
>
Yup, looks like you hit it on the head. This revert should have been done
alongside 2f0799a0ffc0 ("mm, thp: restore node-local hugepage
allocations"), I didn't appreciate how invasive the consolidation patch
was.
I noticed the race in 89c83fb539f9 ("mm, thp: consolidate THP gfp handling
into alloc_hugepage_direct_gfpmask") that is fixed by the revert, but as
you noted it didn't cleanup the second part which is the balancing act for
gfp flags between alloc_hugepage_direct_gfpmask() and alloc_pages_vma().
Syzbot found this to trigger the WARN_ON_ONCE() you mention.
So we certainly need this patch for 4.20 as a follow-up to 2f0799a0ffc0.
It's likely better to regroup and discuss NUMA aspects of all thp
allocations separately with a stable 4.20.
