Re: [PATCH 1/7] mm/thp: fix __split_huge_pmd_locked() on shmem migration entry

[Hugh Dickins]

On Thu, 3 Jun 2021, Yang Shi wrote:
> On Tue, Jun 1, 2021 at 2:05 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> >
> > Are there more places that need to be careful about pmd migration entries?
> > None hit in practice, but several of those is_huge_zero_pmd() tests were
> > done without checking pmd_present() first: I believe a pmd migration entry
> > could end up satisfying that test.  Ah, the inversion of swap offset, to
> > protect against L1TF, makes that impossible on x86; but other arches need
> > the pmd_present() check, and even x86 ought not to apply pmd_page() to a
> > swap-like pmd.  Fix those instances; __split_huge_pmd_locked() was not
> > wrong to be checking with pmd_trans_huge() instead, but I think it's
> > clearer to use pmd_present() in each instance.
...
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 63ed6b25deaa..9fb7b47da87e 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -1676,7 +1676,7 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
> >                 spin_unlock(ptl);
> >                 if (is_huge_zero_pmd(orig_pmd))
> >                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
> > -       } else if (is_huge_zero_pmd(orig_pmd)) {
> > +       } else if (pmd_present(orig_pmd) && is_huge_zero_pmd(orig_pmd)) {
> 
> If it is a huge zero migration entry, the code would fallback to the
> "else". But IIUC the "else" case doesn't handle the huge zero page
> correctly. It may mess up the rss counter.

A huge zero migration entry?  I hope that's not something special
that I've missed.

Do we ever migrate a huge zero page - and how do we find where it's
mapped, to insert the migration entries?  But if we do, I thought it
would use the usual kind of pmd migration entry; and the first check
in is_pmd_migration_entry() is !pmd_present(pmd).

(I have to be rather careful to check such details, after getting
burnt once by pmd_present(): which includes the "huge" bit even when
not otherwise present, to permit races with pmdp_invalidate().
I mentioned in private mail that I'd dropped one of my "fixes" because
it was harmless but mistaken: I had misunderstood pmd_present().)

The point here (see commit message above) is that some unrelated pmd
migration entry could pass the is_huge_zero_pmd() test, which rushes
off to use pmd_page() without even checking pmd_present() first.  And
most of its users have, one way or another, checked pmd_present() first;
but this place and a couple of others had not.

I'm just verifying that it's really a a huge zero pmd before handling
its case; the "else" still does not need to handle the huge zero page.

Hugh