kref: prefer atomic_inc_not_zero to atomic_add_unless

On most platforms, there exists this ifdef:

 #define atomic_inc_not_zero(v) atomic_add_unless((v), 1, 0)

This makes this patch functionally useless. However, on PPC, there is
actually an explicit definition of atomic_inc_not_zero with its own
assembly that is slightly more optimized than atomic_add_unless. So,
this patch changes kref to use atomic_inc_not_zero instead, for PPC and
any future platforms that might provide an explicit implementation.

This also puts this usage of kref more in line with a verbatim reading
of the examples in Paul McKenney's paper [1] in the section titled "2.4
Atomic Counting With Check and Release Memory Barrier", which uses
atomic_inc_not_zero.

[1] http://open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2167.pdf

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

include/linux/kref.h

@@ -133,6 +133,6 @@ static inline int kref_put_mutex(struct kref *kref,
  */
 static inline int __must_check kref_get_unless_zero(struct kref *kref)
 {
-	return atomic_add_unless(&kref->refcount, 1, 0);
+	return atomic_inc_not_zero(&kref->refcount);
 }
 #endif /* _KREF_H_ */