Re: [PATCH] mm/hugetlb: avoid unnecessary check on pud and pmd entry in huge_pte_offset

[Jason Gunthorpe <jgg@mellanox.com>]

On Fri, Apr 24, 2020 at 12:07:50PM +0800, Li Xinhai wrote:
> On 2020-04-24 at 02:38 Jason Gunthorpe wrote:
> >On Thu, Apr 23, 2020 at 11:14:28AM -0700, Mike Kravetz wrote:
> >> Cc a few people who have looked at huge_pte_offset() recently.
> >>
> >> On 4/23/20 5:49 AM, Li Xinhai wrote:
> >> > When huge_pte_offset() is called, the parameter sz can only be PUD_SIZE
> >> > or PMD_SIZE.
> >> > If sz is PUD_SIZE and code can reach pud, then *pud must be none, or
> >> > normal hugetlb entry, or non-present (migration or hwpoisoned) hugetlb
> >> > entry, and we can directly return pud.
> >> > When sz is PMD_SIZE, pud must be none or present, and if code can reach
> >> > pmd, we can directly return pmd.
> >> >
> >> > So, after this patch, the code is simplified by first check on the
> >> > parameter sz, and avoid unnecessary checks in current code.
> >> >
> >> > Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
> >> > Cc: Mike Kravetz <mike.kravetz@oracle.com>
> >> > Cc: Andrew Morton <akpm@linux-foundation.org>
> >> >  mm/hugetlb.c | 24 +++++++++---------------
> >> >  1 file changed, 9 insertions(+), 15 deletions(-)
> >> >
> >> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> >> > index bcabbe0..e1424f5 100644
> >> > +++ b/mm/hugetlb.c
> >> > @@ -5365,8 +5365,8 @@ pte_t *huge_pte_offset(struct mm_struct *mm,
> >> >  {
> >> >  pgd_t *pgd;
> >> >  p4d_t *p4d;
> >> > -	pud_t *pud, pud_entry;
> >> > -	pmd_t *pmd, pmd_entry;
> >> > +	pud_t *pud;
> >> > +	pmd_t *pmd;
> >> > 
> >> >  pgd = pgd_offset(mm, addr);
> >> >  if (!pgd_present(*pgd))
> >> > @@ -5376,22 +5376,16 @@ pte_t *huge_pte_offset(struct mm_struct *mm,
> >> >  return NULL;
> >> > 
> >> >  pud = pud_offset(p4d, addr);
> >> > -	pud_entry = READ_ONCE(*pud);
> >> > -	if (sz != PUD_SIZE && pud_none(pud_entry))
> >> > -	return NULL;
> >> > -	/* hugepage or swap? */
> >> > -	if (pud_huge(pud_entry) || !pud_present(pud_entry))
> >> > +	if (sz == PUD_SIZE)
> >> > +	/* must be pud_huge or pud_none */
> >> >  return (pte_t *)pud;
> >> > -
> >> > -	pmd = pmd_offset(pud, addr);
> >> > -	pmd_entry = READ_ONCE(*pmd);
> >> > -	if (sz != PMD_SIZE && pmd_none(pmd_entry))
> >> > +	if (!pud_present(*pud))
> >> >  return NULL;
> >> > -	/* hugepage or swap? */
> >> > -	if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry))
> >> > -	return (pte_t *)pmd;
> >> > +	/* must have a valid entry and size to go further */
> >> > 
> >> > -	return NULL;
> >> > +	pmd = pmd_offset(pud, addr);
> >>
> >> Can we get here with sz = PMD_SIZE and pud_none(*pud)?  Would that be
> >> an issue for the pmd_offset() call?
> >
> >Certainly pmd_offset() must only be called if the PUD entry is
> >pointing at a pmd level.
> >
> >AFAIK this means it should not be called on pud_none(), pud_huge() or
> >!pud_present() cases. 
> 
> The test of !pud_present(*pud) also block pud_none(*pud)

Sure

> , so when sz == PMD_SIZE, pmd_offset() only called with a valid PUD
> entry which point to PMD page table.

But what prevents pud_huge?

This API seems kind of strange to be honest.. Should it be two
functions instead of a sz parameter?

huge_pud_offset() and huge_pmd_offset() ?

Jason