Re: [PATCH] proc: mm: export PTE sizes directly in smaps (v2)

[Catalin Marinas]

On Mon, Nov 28, 2016 at 01:39:49PM -0800, Dave Hansen wrote:
> On 11/28/2016 01:07 PM, Vlastimil Babka wrote:
> > On 11/28/2016 05:52 PM, Dave Hansen wrote:
> >> On 11/24/2016 06:22 AM, Vlastimil Babka wrote:
> >>> On 11/17/2016 01:28 AM, Dave Hansen wrote:
> >>>> @@ -702,11 +707,13 @@ static int smaps_hugetlb_range(pte_t *pt
> >>>>      }
> >>>>      if (page) {
> >>>>          int mapcount = page_mapcount(page);
> >>>> +        unsigned long hpage_size = huge_page_size(hstate_vma(vma));
> >>>>
> >>>> +        mss->rss_pud += hpage_size;
> >>>
> >>> This hardcoded pud doesn't look right, doesn't the pmd/pud depend on
> >>> hpage_size?
> >>
> >> Urg, nope.  Thanks for noticing that!  I think we'll need something
> >> along the lines of:
> >>
> >>                 if (hpage_size == PUD_SIZE)
> >>                         mss->rss_pud += PUD_SIZE;
> >>                 else if (hpage_size == PMD_SIZE)
> >>                         mss->rss_pmd += PMD_SIZE;
> > 
> > Sounds better, although I wonder whether there are some weird arches
> > supporting hugepage sizes that don't match page table levels. I recall
> > that e.g. MIPS could do arbitrary size, but dunno if the kernel supports
> > that...
> 
> arm64 seems to have pretty arbitrary sizes, and seems to be able to
> build them out of multiple hardware PTE sizes.  I think I can fix my
> code to handle those:
> 
>                 if (hpage_size >= PGD_SIZE)
>                         mss->rss_pgd += PGD_SIZE;
>                 else if (hpage_size >= PUD_SIZE)
>                         mss->rss_pud += PUD_SIZE;
>                 else if (hpage_size >= PMD_SIZE)
>                         mss->rss_pmd += PMD_SIZE;
>                 else
>                         mss->rss_pte += PAGE_SIZE;
> 
> But, I *think* that means that smaps_hugetlb_range() is *currently*
> broken for these intermediate arm64 sizes.  The code does:
> 
>                 if (mapcount >= 2)
>                         mss->shared_hugetlb += hpage_size;
>                 else
>                         mss->private_hugetlb += hpage_size;
> 
> So I *think* if we may count a hugetlbfs arm64 CONT_PTES page multiple
> times, and account hpage_size for *each* of the CONT_PTES.  That would
> artificially inflate the smaps output for those pages.

I don't think it would count them multiple times. As Vlastimil
mentioned, huge_page_size() would return (CONT_PTES * PAGE_SIZE) in such
case, so walk_hugetlb_range() skips the intermediate ptes. In general,
we try to keep the contiguous pte/pmd support visible only to the arm64
hugetlb code and hidden to the core code.

-- 
Catalin