Re: [PATCH 07/11] mm: shmem: use pgtable_has_pmd_leaves()

[Luiz Capitulino]

On 2025-12-16 21:03, Baolin Wang wrote:
> On 2025/12/16 21:47, Luiz Capitulino wrote:
> > On 2025-12-16 02:52, Baolin Wang wrote:
> > > On 2025/12/16 05:16, Luiz Capitulino wrote:
> > > > Shmem uses has_transparent_hugepage() to check if PMD-sized pages are
> > > > supported, use pgtable_has_pmd_leaves() instead.
> > > >
> > > > Signed-off-by: Luiz Capitulino <luizcap@xxxxxxxxxx>
> > > > ---
> > > >   mm/shmem.c | 7 ++++---
> > > >   1 file changed, 4 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/mm/shmem.c b/mm/shmem.c
> > > > index b329b5302c48..ad5825667b49 100644
> > > > --- a/mm/shmem.c
> > > > +++ b/mm/shmem.c
> > > > @@ -689,7 +689,8 @@ static int shmem_parse_huge(const char *str)
> > > >       else
> > > >           return -EINVAL;
> > > > -    if (!has_transparent_hugepage() &&
> > > > +    if (!(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
> > > > +          pgtable_has_pmd_leaves()) &&
> > > >           huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
> > > >           return -EINVAL;
> > > > @@ -4655,7 +4656,7 @@ static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
> > > >           ctx->huge = result.uint_32;
> > > >           if (ctx->huge != SHMEM_HUGE_NEVER &&
> > > >               !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
> > > > -              has_transparent_hugepage()))
> > > > +              pgtable_has_pmd_leaves()))
> > > >               goto unsupported_parameter;
> > > >           ctx->seen |= SHMEM_SEEN_HUGE;
> > > >           break;
> > > > @@ -5439,7 +5440,7 @@ void __init shmem_init(void)
> > > >   #endif
> > > >   #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > -    if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
> > > > +    if (pgtable_has_pmd_leaves() && shmem_huge > SHMEM_HUGE_DENY)
> > >
> > > Using pgtable_has_pmd_leaves() here is a bit confusing because the definition of pgtable_has_pmd_leaves() is: it returns true if the CPU supports PMD-sized pages and false otherwise.
> > >
> > > However, tmpfs and shmem already support other sizes of large folios, not just PMD-sized large folios.
> > >
> > > So, for me, using has_transparent_hugepage() to check would be at least clearer (even though it doesn't change the functionality).
> >
> > This is more of a naming issue, correct?
>
> Yes.
>
> > Would adding something like thp_has_pmd_support() which expands to:
> >
> >      return IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && pgtable_has_pmd_leaves();
> >
> > solve it for you? I suggested it in my RFC, but David advised not to do it.
>
> I agree with David. The thp_has_pmd_support() is not helpful too. What I mean is that the term 'pmd' shouldn't be used here. PMD-sized large folios aren't any more special than others.
>
> shmem also supports mTHP, and a better approach might be what you did in patch 10: even if PMD-sized pages are not supported on an architecture, shmem can still use other sizes of mTHP.

Oh, I get watch you mean now. I'll have to dig a little deeper into the
shmem code but I'll give your suggestion a try. Thanks for your
feedback.

> > Also, I'm not sure if the comparison with other folio sizes apply, as
> > PUD and PMD sizes are special.