refcounts have a generic implementation and an asm optimized one.  The
generic version has extra debugging to make sure that once a refcount
goes to zero, refcount_inc won't increase it.

The btrfs delayed inode code wasn't expecting this, and we're tripping
over the warnings when the generic refcounts are used.  We ended up with
this race:

Process A                                         Process B
                                                  btrfs_get_delayed_node()
						  spin_lock(root->inode_lock)
						  radix_tree_lookup()
__btrfs_release_delayed_node()
refcount_dec_and_test(&delayed_node->refs)
our refcount is now zero
						  refcount_add(2) <---
						  warning here, refcount
                                                  unchanged

spin_lock(root->inode_lock)
radix_tree_delete()

With the generic refcounts, we actually warn again when process B above
tries to release his refcount because refcount_add() turned into a
no-op.

We saw this in production on older kernels without the asm optimized
refcounts.

The fix used here is to use refcount_inc_not_zero() to detect when the
object is in the middle of being freed and return NULL.  This is almost
always the right answer anyway, since we usually end up pitching the
delayed_node if it didn't have fresh data in it.

This also changes __btrfs_release_delayed_node() to remove the extra
check for zero refcounts before radix tree deletion.
btrfs_get_delayed_node() was the only path that was allowing refcounts
to go from zero to one.

Fixes: 6de5f18e ("btrfs: fix refcount_t usage when deleting btrfs_delayed_node")
CC: <stable@vger.kernel.org> # 4.12+
Signed-off-by: Chris Mason <clm@fb.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>