Re: [BUG] kmemleak on __radix_tree_preload

[Catalin Marinas <catalin.marinas@arm.com>]

On Thu, May 08, 2014 at 04:53:30PM +0100, Paul E. McKenney wrote:
> On Thu, May 08, 2014 at 04:29:48PM +0100, Catalin Marinas wrote:
> > BTW, is it safe to have a union overlapping node->parent and
> > node->rcu_head.next? I'm still staring at the radix-tree code but a
> > scenario I have in mind is that call_rcu() has been raised for a few
> > nodes, other CPU may have some reference to one of them and set
> > node->parent to NULL (e.g. concurrent calls to radix_tree_shrink()),
> > breaking the RCU linking. I can't confirm this theory yet ;)
> 
> If this were reproducible, I would suggest retrying with non-overlapping
> node->parent and node->rcu_head.next, but you knew that already.  ;-)

Reading the code, I'm less convinced about this scenario (though it's
worth checking without the union).

> But the usual practice would be to make node removal exclude shrinking.
> And the radix-tree code seems to delegate locking to the caller.
> 
> So, is the correct locking present in the page cache?  The radix-tree
> code seems to assume that all update operations for a given tree are
> protected by a lock global to that tree.

The calling code in mm/filemap.c holds mapping->tree_lock when deleting
radix-tree nodes, so no concurrent calls.

> Another diagnosis approach would be to build with
> CONFIG_DEBUG_OBJECTS_RCU_HEAD=y, which would complain about double
> call_rcu() invocations.  Rumor has it that is is necessary to turn off
> other kmem debugging for this to tell you anything -- I have seen cases
> where the kmem debugging obscures the debug-objects diagnostics.

Another test Jaegeuk could run (hopefully he has some time to look into
this).

Thanks for suggestions.

-- 
Catalin