Re: [PATCH 1/6] mm: khugepaged: fix radix tree node leak in shmem collapse error path

[Jan Kara]

On Fri 11-11-16 13:59:21, Kirill A. Shutemov wrote:
> On Tue, Nov 08, 2016 at 11:12:45AM -0500, Johannes Weiner wrote:
> > On Tue, Nov 08, 2016 at 10:53:52AM +0100, Jan Kara wrote:
> > > On Mon 07-11-16 14:07:36, Johannes Weiner wrote:
> > > > The radix tree counts valid entries in each tree node. Entries stored
> > > > in the tree cannot be removed by simpling storing NULL in the slot or
> > > > the internal counters will be off and the node never gets freed again.
> > > > 
> > > > When collapsing a shmem page fails, restore the holes that were filled
> > > > with radix_tree_insert() with a proper radix tree deletion.
> > > > 
> > > > Fixes: f3f0e1d2150b ("khugepaged: add support of collapse for tmpfs/shmem pages")
> > > > Reported-by: Jan Kara <jack@xxxxxxx>
> > > > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> > > > ---
> > > >  mm/khugepaged.c | 3 ++-
> > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > > > index 728d7790dc2d..eac6f0580e26 100644
> > > > --- a/mm/khugepaged.c
> > > > +++ b/mm/khugepaged.c
> > > > @@ -1520,7 +1520,8 @@ static void collapse_shmem(struct mm_struct *mm,
> > > >  				if (!nr_none)
> > > >  					break;
> > > >  				/* Put holes back where they were */
> > > > -				radix_tree_replace_slot(slot, NULL);
> > > > +				radix_tree_delete(&mapping->page_tree,
> > > > +						  iter.index);
> > > 
> > > Hum, but this is inside radix_tree_for_each_slot() iteration. And
> > > radix_tree_delete() may end up freeing nodes resulting in invalidating
> > > current slot pointer and the iteration code will do use-after-free.
> > 
> > Good point, we need to do another tree lookup after the deletion.
> > 
> > But there are other instances in the code, where we drop the lock
> > temporarily and somebody else could delete the node from under us.
> > 
> > In the main collapse path, I *think* this is prevented by the fact
> > that when we drop the tree lock we still hold the page lock of the
> > regular page that's in the tree while we isolate and unmap it, thus
> > pin the node. Even so, it would seem a little hairy to rely on that.
> > 
> > Kirill?
> 
> [ sorry for delay ]
> 
> Yes, we make sure that locked page still belong to the radix tree and fall
> off if it's not. Locked page cannot be removed from radix-tree, so we
> should be fine.

Well, it cannot be removed from the radix tree but radix tree code is still
free to collapse / expand the tree nodes as it sees fit (currently the only
real case is when changing direct page pointer in the tree root to a node
pointer or vice versa but still...). So code should not really assume that
the node page is referenced from does not change once tree_lock is dropped.
It leads to subtle bugs...

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR