Re: [PATCH 1/3] mm: honor FOLL_GET flag in follow_hugetlb_page

[Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>]

On Tue, May 07, 2013 at 06:28:18PM -0400, Jerome Glisse wrote:
> On Tue, May 7, 2013 at 5:53 PM, Naoya Horiguchi
> <n-horiguchi@ah.jp.nec.com> wrote:
> > On Tue, May 07, 2013 at 04:45:54PM -0400, j.glisse@gmail.com wrote:
> >> From: Jerome Glisse <jglisse@redhat.com>
> >>
> >> Do not increase page count if FOLL_GET is not set.
> >>
> >> Signed-off-by: Jerome Glisse <jglisse@redhat.com>
> >> ---
> >>  mm/hugetlb.c | 4 +++-
> >>  1 file changed, 3 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> >> index 1a12f5b..5d1e46b 100644
> >> --- a/mm/hugetlb.c
> >> +++ b/mm/hugetlb.c
> >> @@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
> >>  same_page:
> >>               if (pages) {
> >>                       pages[i] = mem_map_offset(page, pfn_offset);
> >> -                     get_page(pages[i]);
> >> +                     if (flags & FOLL_GET) {
> >> +                             get_page_foll(pages[i]);
> >> +                     }
> >>               }
> >>
> >>               if (vmas)
> >> --
> >
> > Hi Jerome,
> >
> > I think that we need to be careful in handling tail pages, because
> > __get_page_tail_foll() uses page->_mapcount as refcount.
> > When you get refcount on a tail page and free the hugepage without
> > putting the *mapcount*, you will hit BUG_ON() in free_huge_page().
> > Yes, this is a very tricky workaround for thp, so to avoid making
> > things too complicated, I think either of the following is better:
> >  - to get refcount only for head pages, or
> >  - to introduce a hugetlbfs variant of get_page_foll().
> 
> Maybe a simpler variant is to just not take any refcount, ie like
> current code if FOLL_GET is set then take refcount on all page wether
> they are head/tail or not. I will resend with that.

Hmm, I think that FOLL_GET flag means "do get_page on page", so
the "not take any refcount" variant seems to make no sense for me.
Would it be better just call without FOLL_GET?

> > BTW, who do you expect is the caller of follow_hugetlb_page()
> > with FOLL_GET (I can't find your subsequent patches 2/3 or 3/3)?
> > I'm interested in this change because in my project it's necessary
> > to implement this for hugepage migration
> > (see https://lkml.org/lkml/2013/3/22/553).
> 
> I can not talk about the patchset yet (and it's not fully cook) but i
> need to be able to get the page without taking reference so without
> the FOLL_GET flag set but i need splitting, well no real splitting, i
> need pfn for each fake sub page of huge page (interested in physical
> address not in the page struct).

If the caller knows the vma is backed by hugetlbfs, we can get pfn of
the tail page by adding page offset (which can be calcurated by the virtual
address with proper huge page mask) to the head page's pfn.

Thanks,
Naoya Horiguchi

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>