inode_permission NULL pointer dereference in 3.13-rc1

[Al Viro]

On Fri, Nov 29, 2013 at 12:46:48PM +1100, Dave Chinner wrote:
> > 	* d_count(dentry) is -128
> 
> void lockref_mark_dead(struct lockref *lockref)
> {
>         assert_spin_locked(&lockref->lock);
>         lockref->count = -128;
> }

... done once refcount reaches zero and we decide to evict the sucker.
Which, for dentry that happens to be
	* pwd of at least one process
	* root of at least one process
	* root dentry of a filesystem that contains at least one
binary being executed
means that we have dropped several references too many.  Thus the
comment about unbalanced dput() somewhere; the question is _where_ had
that dput() been.