Fix assertions in crypto reference helpers

The assertions are racy and the use of `membar_exit()` did nothing to
fix that.

The helpers use atomic functions, so we cleverly get values from the
atomics that we can use to ensure that the assertions operate on the
correct values.

We also use `membar_producer()` prior to decrementing reference counts
so that operations that happened prior to a decrement to 0 will be
guaranteed to happen before the decrement on architectures that reorder
atomics.

This also slightly improves performance by eliminating unnecessary
reads, although I doubt it would be measurable in any benchmark.

Reviewed-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes #13880

module/icp/include/sys/crypto/sched_impl.h

@@ -73,9 +73,10 @@ typedef struct kcf_context {
  * context structure is freed along with the global context.
  */
 #define	KCF_CONTEXT_REFRELE(ictx) {				\
-	ASSERT((ictx)->kc_refcnt != 0);				\
-	membar_exit();						\
-	if (atomic_add_32_nv(&(ictx)->kc_refcnt, -1) == 0)	\
+	membar_producer();					\
+	int newval = atomic_add_32_nv(&(ictx)->kc_refcnt, -1);	\
+	ASSERT(newval != -1);					\
+	if (newval == 0)					\
 		kcf_free_context(ictx);				\
 }