Re: [PATCH 02/11] mm,migration: Do not try to migrate unmapped anonymous pages

[Mel Gorman <mel@csn.ul.ie>]

On Mon, Mar 15, 2010 at 02:34:20PM +0900, KAMEZAWA Hiroyuki wrote:
> On Mon, 15 Mar 2010 09:28:08 +0900
> Minchan Kim <minchan.kim@gmail.com> wrote:
> 
> > Hi, Mel.
> > On Sat, Mar 13, 2010 at 1:41 AM, Mel Gorman <mel@csn.ul.ie> wrote:
> > > rmap_walk_anon() was triggering errors in memory compaction that looks like
> > > use-after-free errors in anon_vma. The problem appears to be that between
> > > the page being isolated from the LRU and rcu_read_lock() being taken, the
> > > mapcount of the page dropped to 0 and the anon_vma was freed. This patch
> > > skips the migration of anon pages that are not mapped by anyone.
> > >
> > > Signed-off-by: Mel Gorman <mel@csn.ul.ie>
> > > Acked-by: Rik van Riel <riel@redhat.com>
> > > ---
> > >  mm/migrate.c |   10 ++++++++++
> > >  1 files changed, 10 insertions(+), 0 deletions(-)
> > >
> > > diff --git a/mm/migrate.c b/mm/migrate.c
> > > index 98eaaf2..3c491e3 100644
> > > --- a/mm/migrate.c
> > > +++ b/mm/migrate.c
> > > @@ -602,6 +602,16 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
> > >         * just care Anon page here.
> > >         */
> > >        if (PageAnon(page)) {
> > > +               /*
> > > +                * If the page has no mappings any more, just bail. An
> > > +                * unmapped anon page is likely to be freed soon but worse,
> > > +                * it's possible its anon_vma disappeared between when
> > > +                * the page was isolated and when we reached here while
> > > +                * the RCU lock was not held
> > > +                */
> > > +               if (!page_mapcount(page))
> > 
> > As looking code about mapcount of page, I got confused.
> > I think mapcount of page is protected by pte lock.
> > But I can't find pte lock in unmap_and_move.
>
> There is no pte_lock.
> 

Indeed. It is manipulated while some other lock is held but it can be read
without locks held. For example, when mapping a page either tha anon_vma
lock or i_mmap_lock is held but it is read without special locking in places
like page_referenced_ksm().

> > If I am right, what protects race between this condition check and
> > rcu_read_lock?
> > This patch makes race window very small but It can't remove race totally.
> > 
> > I think I am missing something.
> > Pz, point me out. :)
> > 
> 
> Hmm. This is my understanding of old story.
> 
> At migration.
>   1. we increase page_count().
>   2. isolate it from LRU.
>   3. call try_to_unmap() under rcu_read_lock(). Then, 
>   4. replace pte with swp_entry_t made by PFN. under pte_lock.
>   5. do migarate 
>   6. remap new pages. under pte_lock()>
>   7. release rcu_read_lock().
> 
> Here, we don't care whether page->mapping holds valid anon_vma or not.
> 
> Assume a racy threads which calls zap_pte_range() (or some other)
> 

I believe the race being hit is related to processes existing. A racy thread calling
zap_pte_range() while pages within were being migrated does appear to be the problem.

> a) When the thread finds valid pte under pte_lock and successfully call
>    page_remove_rmap().
>    In this case, migration thread finds try_to_unmap doesn't unmap any pte.
>    Then, at 6, remap pte will not work.
> b) When the thread finds migrateion PTE(as swap entry) in zap_page_range().
>    In this case, migration doesn't find migrateion PTE and remap fails.
> 
> Why rcu_read_lock() is necessary..
>  - When page_mapcount() goes to 0, we shouldn't trust page->mapping is valid.

I also believe this to be true.

>  - Possible cases are
> 	i) anon_vma (= page->mapping) is freed and used for other object.
>  	ii) anon_vma (= page->mapping) is freed
> 	iii) anon_vma (= page->mapping) is freed and used as anon_vma again.
> 
> Here, anon_vma_cachep is created  by SLAB_DESTROY_BY_RCU. Then, possible cases
> are only ii) and iii).

I believe it's (ii) that was being hit.

> While anon_vma is anon_vma, try_to_unmap and remap_page
> can work well because of the list of vmas and address check. IOW, remap routine
> just do nothing if anon_vma is freed.
> 
> I'm not sure by what logic "use-after-free anon_vma" is caught. But yes,
> there will be case, "anon_vma is touched after freed.", I think.
> 

The use after free looks like

1. page_mapcount(page) was zero so anon_vma was no longer reliable
2. rcu lock taken but the anon_vma at this point can already be garbage because the
   process exited
3. call try_to_unmap, looks up tha anon_vma and locks it. This causes problems

I thought the race would be closed but there is still a very tiny window there all
right. The following alternative should close it. What do you think?

        if (PageAnon(page)) {
		rcu_read_lock();

                /*
                 * If the page has no mappings any more, just bail. An
                 * unmapped anon page is likely to be freed soon but worse,
                 * it's possible its anon_vma disappeared between when
                 * the page was isolated and when we reached here while
                 * the RCU lock was not held
                 */
                if (!page_mapcount(page)) {
			rcu_read_unlock();
                        goto uncharge;
		}

                rcu_locked = 1;
                anon_vma = page_anon_vma(page);
                atomic_inc(&anon_vma->external_refcount);
        }

The rcu_unlock label is not used here because the reference counts were not taken in
the case where page_mapcount == 0.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab