Re: [PATCH] smaps should deal with huge zero page exactly same as normal zero page

From: Fengwei Yin
Date: Fri Oct 10 2014 - 01:24:25 EST


On Thu, Oct 09, 2014 at 09:36:30AM -0700, Dave Hansen wrote:
> On 10/09/2014 02:19 AM, Fengwei Yin wrote:
> > diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> > index 80ca4fb..8550b27 100644
> > --- a/fs/proc/task_mmu.c
> > +++ b/fs/proc/task_mmu.c
> > @@ -476,7 +476,7 @@ static void smaps_pte_entry(pte_t ptent, unsigned long addr,
> > mss->nonlinear += ptent_size;
> > }
> >
> > - if (!page)
> > + if (!page || is_huge_zero_page(page))
> > return;
>
> This really seems like a bit of a hack. A normal (small) zero page
> won't make it to this point because of the vm_normal_page() check in
> smaps_pte_entry() which hits the _PAGE_SPECIAL bit in the pte.
>
> Is there a reason we can't set _PAGE_SPECIAL on the huge_zero_page ptes?
> If we did that, we wouldn't need a special case here.
>
> If we can't do that for some reason, can we at least teach
> vm_normal_page() about the huge_zero_page in some other way?
I suppose _PAGE_SPECIAL can't work. Two reasons:
1. Not all arch have HAVE_PTE_SPECIAL set. So always need another way to
handle the arch which has no PTE_SPECIAL.
2. _PAGE_SPECIAL is just for PTE now. If want to add it for huge page,
we need to introduce pmd_mkspecial() thing which I don't think it's
worth to do now (unless you want it. :)).

Yes. We could move the check to vm_normal_page(). But it still needs
export functions from huge_memory.c.

Please check the new patch.

>
> > if (PageAnon(page))
> > @@ -516,7 +516,8 @@ static int smaps_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
> > if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
> > smaps_pte_entry(*(pte_t *)pmd, addr, HPAGE_PMD_SIZE, walk);
> > spin_unlock(ptl);
> > - mss->anonymous_thp += HPAGE_PMD_SIZE;
> > + if (!is_huge_zero_pmd(*pmd))
> > + mss->anonymous_thp += HPAGE_PMD_SIZE;
> > return 0;
> > }
>
> How about we just move this hunk in to smaps_pte_entry()? Something
> along these lines:
>
> ...
> if (PageAnon(page)) {
> mss->anonymous += ptent_size;
> + if (PageTransHuge(page))
> + mss->anonymous_thp += ptent_size;
> }
Done.
>
> If we do that, plus teaching vm_normal_page() about huge_zero_pages, it
> will help keep the hacks and the extra code due to huge pages to a miniumum.
>
> > diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> > index 63579cb..758f569 100644
> > --- a/include/linux/huge_mm.h
> > +++ b/include/linux/huge_mm.h
> > @@ -34,6 +34,10 @@ extern int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> > unsigned long addr, pgprot_t newprot,
> > int prot_numa);
> >
> > +extern bool is_huge_zero_page(struct page *page);
> > +
> > +extern bool is_huge_zero_pmd(pmd_t pmd);
> > +
> > enum transparent_hugepage_flag {
> > TRANSPARENT_HUGEPAGE_FLAG,
> > TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index d9a21d06..bedc3ae 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -173,12 +173,12 @@ static int start_khugepaged(void)
> > static atomic_t huge_zero_refcount;
> > static struct page *huge_zero_page __read_mostly;
> >
> > -static inline bool is_huge_zero_page(struct page *page)
> > +bool is_huge_zero_page(struct page *page)
> > {
> > return ACCESS_ONCE(huge_zero_page) == page;
> > }
> >
> > -static inline bool is_huge_zero_pmd(pmd_t pmd)
> > +bool is_huge_zero_pmd(pmd_t pmd)
> > {
> > return is_huge_zero_page(pmd_page(pmd));
> > }
>
> ^^^ And all these exports.

A new function is_huge_zero_pfn() is added to mm/huge_memory.c
and exported.

Thanks.