On Tue, May 17, 2022 at 10:31 PM Muchun Song <songmuchun@xxxxxxxxxxxxx> wrote:
>
> On Fri, May 13, 2022 at 12:17:05PM -0700, Yang Shi wrote:
> > IIUC PVMW checks if the vma is possibly huge PMD mapped by
> > transparent_hugepage_active() and "pvmw->nr_pages >= HPAGE_PMD_NR".
> >
> > Actually pvmw->nr_pages is returned by compound_nr() or
> > folio_nr_pages(), so the page should be THP as long as "pvmw->nr_pages
> > >= HPAGE_PMD_NR". And it is guaranteed THP is allocated for valid VMA
> > in the first place. But it may be not PMD mapped if the VMA is file
> > VMA and it is not properly aligned. The transhuge_vma_suitable()
> > is used to do such check, so replace transparent_hugepage_active() to
> > it, which is too heavy and overkilling.
> >
> > Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > Cc: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> > Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> > ---
> > v2: * Fixed build error for !CONFIG_TRANSPARENT_HUGEPAGE
> > * Removed fixes tag per Willy
> >
> > include/linux/huge_mm.h | 8 ++++++--
> > mm/page_vma_mapped.c | 2 +-
> > 2 files changed, 7 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> > index fbf36bb1be22..c2826b1f4069 100644
> > --- a/include/linux/huge_mm.h
> > +++ b/include/linux/huge_mm.h
> > @@ -117,8 +117,10 @@ extern struct kobj_attribute shmem_enabled_attr;
> > extern unsigned long transparent_hugepage_flags;
> >
> > static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
> > - unsigned long haddr)
> > + unsigned long addr)
> > {
> > + unsigned long haddr;
> > +
> > /* Don't have to check pgoff for anonymous vma */
> > if (!vma_is_anonymous(vma)) {
> > if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
> > @@ -126,6 +128,8 @@ static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
> > return false;
> > }
> >
> > + haddr = addr & HPAGE_PMD_MASK;
> > +
> > if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
> > return false;
> > return true;
> > @@ -328,7 +332,7 @@ static inline bool transparent_hugepage_active(struct vm_area_struct *vma)
> > }
> >
> > static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
> > - unsigned long haddr)
> > + unsigned long addr)
> > {
> > return false;
> > }
> > diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> > index c10f839fc410..e971a467fcdf 100644
> > --- a/mm/page_vma_mapped.c
> > +++ b/mm/page_vma_mapped.c
> > @@ -243,7 +243,7 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
> > * cleared *pmd but not decremented compound_mapcount().
> > */
> > if ((pvmw->flags & PVMW_SYNC) &&
> > - transparent_hugepage_active(vma) &&
> > + transhuge_vma_suitable(vma, pvmw->address) &&
>
> How about the following diff? Then we do not need to change
> transhuge_vma_suitable(). All the users of transhuge_vma_suitable()
> are already do the alignment by themselves.
Thanks for the suggestion. But TBH I don't think this is a better way.
I did think about this before proposing v2, but I don't prefer to
pollute the code with IS_ENABLED(CONFIG_xxx) since the definition of
transhuge_vma_suitable() is already protected by #ifdef. Rounding the
address in transhuge_vma_suitable() seems neater and more readable to
me IMHO.
Some callers of transhuge_vma_suitable() do round the address before
calling it, but the rounded address is used by other codes in the
callers too.
>
> Thanks.
>
> diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> index c10f839fc410..0aed5ca60c67 100644
> --- a/mm/page_vma_mapped.c
> +++ b/mm/page_vma_mapped.c
> @@ -243,7 +243,8 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
> * cleared *pmd but not decremented compound_mapcount().
> */
> if ((pvmw->flags & PVMW_SYNC) &&
> - transparent_hugepage_active(vma) &&
> + IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
> + transhuge_vma_suitable(vma, pvmw->address & HPAGE_PMD_MASK) &&
> (pvmw->nr_pages >= HPAGE_PMD_NR)) {
> spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
>
> > (pvmw->nr_pages >= HPAGE_PMD_NR)) {
> > spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
> >
> > --
> > 2.26.3
> >
> >
