Re: linux-next: slab shrinkers: BUG at mm/list_lru.c:92

[Michal Hocko <mhocko@suse.cz>]

On Tue 18-06-13 10:31:05, Glauber Costa wrote:
> On Tue, Jun 18, 2013 at 12:46:23PM +1000, Dave Chinner wrote:
> > On Tue, Jun 18, 2013 at 02:30:05AM +0400, Glauber Costa wrote:
> > > On Mon, Jun 17, 2013 at 02:35:08PM -0700, Andrew Morton wrote:
> > > > On Mon, 17 Jun 2013 19:14:12 +0400 Glauber Costa <glommer@gmail.com> wrote:
> > > > 
> > > > > > I managed to trigger:
> > > > > > [ 1015.776029] kernel BUG at mm/list_lru.c:92!
> > > > > > [ 1015.776029] invalid opcode: 0000 [#1] SMP
> > > > > > with Linux next (next-20130607) with https://lkml.org/lkml/2013/6/17/203
> > > > > > on top. 
> > > > > > 
> > > > > > This is obviously BUG_ON(nlru->nr_items < 0) and 
> > > > > > ffffffff81122d0b:       48 85 c0                test   %rax,%rax
> > > > > > ffffffff81122d0e:       49 89 44 24 18          mov    %rax,0x18(%r12)
> > > > > > ffffffff81122d13:       0f 84 87 00 00 00       je     ffffffff81122da0 <list_lru_walk_node+0x110>
> > > > > > ffffffff81122d19:       49 83 7c 24 18 00       cmpq   $0x0,0x18(%r12)
> > > > > > ffffffff81122d1f:       78 7b                   js     ffffffff81122d9c <list_lru_walk_node+0x10c>
> > > > > > [...]
> > > > > > ffffffff81122d9c:       0f 0b                   ud2
> > > > > > 
> > > > > > RAX is -1UL.
> > > > > Yes, fearing those kind of imbalances, we decided to leave the counter as a signed quantity
> > > > > and BUG, instead of an unsigned quantity.
> > > > > 
> > > > > > 
> > > > > > I assume that the current backtrace is of no use and it would most
> > > > > > probably be some shrinker which doesn't behave.
> > > > > > 
> > > > > There are currently 3 users of list_lru in tree: dentries, inodes and xfs.
> > > > > Assuming you are not using xfs, we are left with dentries and inodes.
> > > > > 
> > > > > The first thing to do is to find which one of them is misbehaving. You can try finding
> > > > > this out by the address of the list_lru, and where it lays in the superblock.
> > > > > 
> > > > > Once we know each of them is misbehaving, then we'll have to figure out why.
> > > > 
> > > > The trace says shrink_slab_node->super_cache_scan->prune_icache_sb.  So
> > > > it's inodes?
> > > > 
> > > Assuming there is no memory corruption of any sort going on , let's check the code.
> > > nr_item is only manipulated in 3 places:
> > > 
> > > 1) list_lru_add, where it is increased
> > > 2) list_lru_del, where it is decreased in case the user have voluntarily removed the
> > >    element from the list
> > > 3) list_lru_walk_node, where an element is removing during shrink.
> > > 
> > > All three excerpts seem to be correctly locked, so something like this indicates an imbalance.
> > 
> > inode_lru_isolate() looks suspicious to me:
> > 
> >         WARN_ON(inode->i_state & I_NEW);
> >         inode->i_state |= I_FREEING;
> >         spin_unlock(&inode->i_lock);
> > 
> >         list_move(&inode->i_lru, freeable);
> >         this_cpu_dec(nr_unused);
> > 	return LRU_REMOVED;
> > }
> > 
> > All the other cases where I_FREEING is set and the inode is removed
> > from the LRU are completely done under the inode->i_lock. i.e. from
> > an external POV, the state change to I_FREEING and removal from LRU
> > are supposed to be atomic, but they are not here.
> > 
> > I'm not sure this is the source of the problem, but it definitely
> > needs fixing.
> > 
> Yes, I missed that yesterday, but that does look suspicious to me as well.
> 
> Michal, if you can manually move this one inside the lock as well and see
> if it fixes your problem as well... Otherwise I can send you a patch as well
> so we don't get lost on what is patched and what is not.

OK, I am testing with this now:
diff --git a/fs/inode.c b/fs/inode.c
index 604c15e..95e598c 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -733,9 +733,9 @@ inode_lru_isolate(struct list_head *item, spinlock_t *lru_lock, void *arg)
 
 	WARN_ON(inode->i_state & I_NEW);
 	inode->i_state |= I_FREEING;
+	list_move(&inode->i_lru, freeable);
 	spin_unlock(&inode->i_lock);
 
-	list_move(&inode->i_lru, freeable);
 	this_cpu_dec(nr_unused);
 	return LRU_REMOVED;
 }

> Let us at least know if this is the problem.
> 
> > > callers:
> > > iput_final, evict_inodes, invalidate_inodes.
> > > Both evict_inodes and invalidate_inodes will do the following pattern:
> > > 
> > >                 inode->i_state |= I_FREEING;                                            
> > >                 inode_lru_list_del(inode);
> > >                 spin_unlock(&inode->i_lock);
> > >                 list_add(&inode->i_lru, &dispose);
> > > 
> > > IOW, they will remove the element from the LRU, and add it to the dispose list.
> > > Both of them will also bail out if they see I_FREEING already set, so they are safe
> > > against each other - because the flag is manipulated inside the lock.
> > > 
> > > But how about iput_final? It seems to me that if we are calling iput_final at the
> > > same time as the other two, this *could* happen (maybe there is some extra protection
> > > that can be seen from Australia but not from here. Dave?)
> > 
> > If I_FREEING is set before we enter iput_final(), then something
> > else is screwed up. I_FREEING is only set once the last reference
> > has gone away and we are killing the inode. All the other callers
> > that set I_FREEING check that the reference count on the inode is
> > zero before they set I_FREEING. Hence I_FREEING cannot be set on the
> > transition of i_count from 1 to 0 when iput_final() is called. So
> > the patch won't do anything to avoid the problem being seen.
> > 
> Yes, but isn't things like evict_inodes and invalidate_inodes called at
> umount time, for instance?

JFYI No unmount is going on in my test case.
-- 
Michal Hocko
SUSE Labs