On Mon, Jan 15, 2024 at 01:55:51PM -0400, Jason Gunthorpe wrote:
> On Wed, Jan 03, 2024 at 05:14:13PM +0800, peterx@xxxxxxxxxx wrote:
> > From: Peter Xu <peterx@xxxxxxxxxx>
> >
> > ARM defines pmd_thp_or_huge(), detecting either a THP or a huge PMD. It
> > can be a helpful helper if we want to merge more THP and hugetlb code
> > paths. Make it a generic default implementation, only exist when
> > CONFIG_MMU. Arch can overwrite it by defining its own version.
> >
> > For example, ARM's pgtable-2level.h defines it to always return false.
> >
> > Keep the macro declared with all config, it should be optimized to a false
> > anyway if !THP && !HUGETLB.
> >
> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> > ---
> > include/linux/pgtable.h | 4 ++++
> > mm/gup.c | 3 +--
> > 2 files changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
> > index 466cf477551a..2b42e95a4e3a 100644
> > --- a/include/linux/pgtable.h
> > +++ b/include/linux/pgtable.h
> > @@ -1362,6 +1362,10 @@ static inline int pmd_write(pmd_t pmd)
> > #endif /* pmd_write */
> > #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> >
> > +#ifndef pmd_thp_or_huge
> > +#define pmd_thp_or_huge(pmd) (pmd_huge(pmd) || pmd_trans_huge(pmd))
> > +#endif
>
> Why not just use pmd_leaf() ?
>
> This GUP case seems to me exactly like what pmd_leaf() should really
> do and be used for..
I think I mostly agree with you, and these APIs are indeed confusing. IMHO
the challenge is about the risk of breaking others on small changes in the
details where evil resides.
>
> eg x86 does:
>
> #define pmd_leaf pmd_large
> static inline int pmd_large(pmd_t pte)
> return pmd_flags(pte) & _PAGE_PSE;
>
> static inline int pmd_trans_huge(pmd_t pmd)
> return (pmd_val(pmd) & (_PAGE_PSE|_PAGE_DEVMAP)) == _PAGE_PSE;
>
> int pmd_huge(pmd_t pmd)
> return !pmd_none(pmd) &&
> (pmd_val(pmd) & (_PAGE_PRESENT|_PAGE_PSE)) != _PAGE_PRESENT;
For example, here I don't think it's strictly pmd_leaf()? As pmd_huge()
will return true if PRESENT=0 && PSE=0 (as long as none pte ruled out
first), while pmd_leaf() will return false; I think that came from
cbef8478bee5. I'm not sure whether that is the best solution, e.g., from a
1st glance it seems better to me to process swap entries separately
(including both migration and poisoned entries)..
Sparc has similar things there, which in that case I'm not sure whether a
direct replace is always safe.
Besides that, there're also other cases where it's not clear of such direct
replacement, not until further investigated. E.g., arm-3level has:
#define pmd_leaf(pmd) pmd_sect(pmd)
#define pmd_sect(pmd) ((pmd_val(pmd) & PMD_TYPE_MASK) == \
PMD_TYPE_SECT)
#define PMD_TYPE_SECT (_AT(pmdval_t, 1) << 0)
While pmd_huge() there relies on PMD_TABLE_BIT ()
int pmd_huge(pmd_t pmd)
{
return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
}
#define PMD_TABLE_BIT (_AT(pmdval_t, 1) << 1)
These are just the trivial details that I wanted to avoid to touch in this
series, so as to resolve the hugetlb issue separately from others.
The new pmd_huge_or_thp() is not ideal, but that easily isolates all these
trivial details / evils out of the picture, so that we can tackle them one
by one. It is strictly an OR or huge||thp, so it's hopefully safe to not
break anything yet from that regard.
>
> I spot checked a couple arches and it looks like it holds up.
>
> Further, it looks to me like this site in GUP is the only core code
> caller..
>
> So, I'd suggest a small series to go arch by arch and convert the arch
> to use pmd_huge() == pmd_leaf(). Then retire pmd_huge() as a public
> API.
>
> > diff --git a/mm/gup.c b/mm/gup.c
> > index df83182ec72d..eebae70d2465 100644
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -3004,8 +3004,7 @@ static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned lo
> > if (!pmd_present(pmd))
> > return 0;
> >
> > - if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
> > - pmd_devmap(pmd))) {
> > + if (unlikely(pmd_thp_or_huge(pmd) || pmd_devmap(pmd))) {
> > /* See gup_pte_range() */
> > if (pmd_protnone(pmd))
> > return 0;
>
> And the devmap thing here doesn't make any sense either. The arch
> should ensure that pmd_devmap() implies pmd_leaf(). Since devmap is a
> purely SW construct it almost certainly does already anyhow.
Yep, but only if pmd_leaf() is safe to be put here. A pmd devmap should
always imply as a pmd_leaf() indeed.
Thanks,
--
Peter Xu
