diff --git a/freedreno/freedreno-symbol-check b/freedreno/freedreno-symbol-check
index 6da9d667..002a3989 100755
--- a/freedreno/freedreno-symbol-check
+++ b/freedreno/freedreno-symbol-check
@@ -49,6 +49,7 @@ fd_ringbuffer_flush
 fd_ringbuffer_grow
 fd_ringbuffer_new
 fd_ringbuffer_new_object
+fd_ringbuffer_ref
 fd_ringbuffer_reloc
 fd_ringbuffer_reloc2
 fd_ringbuffer_reset
diff --git a/freedreno/freedreno_ringbuffer.c b/freedreno/freedreno_ringbuffer.c
index 1fa33b7a..80af736f 100644
--- a/freedreno/freedreno_ringbuffer.c
+++ b/freedreno/freedreno_ringbuffer.c
@@ -71,11 +71,21 @@ fd_ringbuffer_new_object(struct fd_pipe *pipe, uint32_t size)
 
 drm_public void fd_ringbuffer_del(struct fd_ringbuffer *ring)
 {
-	if (!(ring->flags & FD_RINGBUFFER_OBJECT))
-		fd_ringbuffer_reset(ring);
+	if (!atomic_dec_and_test(&ring->refcnt))
+		return;
+
+	fd_ringbuffer_reset(ring);
 	ring->funcs->destroy(ring);
 }
 
+drm_public struct fd_ringbuffer *
+fd_ringbuffer_ref(struct fd_ringbuffer *ring)
+{
+	STATIC_ASSERT(sizeof(ring->refcnt) <= sizeof(ring->__pad));
+	atomic_inc(&ring->refcnt);
+	return ring;
+}
+
 /* ringbuffers which are IB targets should set the toplevel rb (ie.
  * the IB source) as it's parent before emitting reloc's, to ensure
  * the bookkeeping works out properly.
diff --git a/freedreno/freedreno_ringbuffer.h b/freedreno/freedreno_ringbuffer.h
index 69e7ed99..b2e8024d 100644
--- a/freedreno/freedreno_ringbuffer.h
+++ b/freedreno/freedreno_ringbuffer.h
@@ -53,6 +53,19 @@ struct fd_ringbuffer {
 	void *user;
 
 	uint32_t flags;
+
+	/* This is a bit gross, but we can't use atomic_t in exported
+	 * headers.  OTOH, we don't need the refcnt to be publicly
+	 * visible.  The only reason that this struct is exported is
+	 * because fd_ringbuffer_emit needs to be something that can
+	 * be inlined for performance reasons.
+	 */
+	union {
+#ifdef HAS_ATOMIC_OPS
+		atomic_t refcnt;
+#endif
+		uint64_t __pad;
+	};
 };
 
 struct fd_ringbuffer * fd_ringbuffer_new(struct fd_pipe *pipe,
@@ -60,6 +73,7 @@ struct fd_ringbuffer * fd_ringbuffer_new(struct fd_pipe *pipe,
 struct fd_ringbuffer * fd_ringbuffer_new_object(struct fd_pipe *pipe,
 		uint32_t size);
 
+struct fd_ringbuffer *fd_ringbuffer_ref(struct fd_ringbuffer *ring);
 void fd_ringbuffer_del(struct fd_ringbuffer *ring);
 void fd_ringbuffer_set_parent(struct fd_ringbuffer *ring,
 		struct fd_ringbuffer *parent);
diff --git a/freedreno/kgsl/kgsl_ringbuffer.c b/freedreno/kgsl/kgsl_ringbuffer.c
index e4fdf342..7361f7d5 100644
--- a/freedreno/kgsl/kgsl_ringbuffer.c
+++ b/freedreno/kgsl/kgsl_ringbuffer.c
@@ -216,6 +216,8 @@ drm_private struct fd_ringbuffer * kgsl_ringbuffer_new(struct fd_pipe *pipe,
 	}
 
 	ring = &kgsl_ring->base;
+	atomic_set(&ring->refcnt, 1);
+
 	ring->funcs = &funcs;
 	ring->size = size;
 
diff --git a/freedreno/msm/msm_ringbuffer.c b/freedreno/msm/msm_ringbuffer.c
index 5a088818..a102ca35 100644
--- a/freedreno/msm/msm_ringbuffer.c
+++ b/freedreno/msm/msm_ringbuffer.c
@@ -29,6 +29,7 @@
 #include <assert.h>
 #include <inttypes.h>
 
+#include "xf86atomic.h"
 #include "freedreno_ringbuffer.h"
 #include "msm_priv.h"
 
@@ -50,8 +51,6 @@ struct msm_cmd {
 struct msm_ringbuffer {
 	struct fd_ringbuffer base;
 
-	atomic_t refcnt;
-
 	/* submit ioctl related tables:
 	 * Note that bos and cmds are tracked by the parent ringbuffer, since
 	 * that is global to the submit ioctl call.  The reloc's table is tracked
@@ -97,9 +96,6 @@ static inline struct msm_ringbuffer * to_msm_ringbuffer(struct fd_ringbuffer *x)
 	return (struct msm_ringbuffer *)x;
 }
 
-static void msm_ringbuffer_unref(struct fd_ringbuffer *ring);
-static void msm_ringbuffer_ref(struct fd_ringbuffer *ring);
-
 #define INIT_SIZE 0x1000
 
 static pthread_mutex_t idx_lock = PTHREAD_MUTEX_INITIALIZER;
@@ -454,7 +450,7 @@ static int msm_ringbuffer_flush(struct fd_ringbuffer *ring, uint32_t *last_start
 		if (msm_cmd->ring->flags & FD_RINGBUFFER_OBJECT) {
 			/* we could have dropped last reference: */
 			msm_ring->cmds[i] = NULL;
-			msm_ringbuffer_unref(msm_cmd->ring);
+			fd_ringbuffer_del(msm_cmd->ring);
 			free(U642VOID(cmd->relocs));
 		}
 	}
@@ -568,7 +564,7 @@ static uint32_t msm_ringbuffer_emit_reloc_ring(struct fd_ringbuffer *ring,
 	 * destroyed after emitted but before flush, so we must hold a ref:
 	 */
 	if (added_cmd && (target->flags & FD_RINGBUFFER_OBJECT)) {
-		msm_ringbuffer_ref(target);
+		fd_ringbuffer_ref(target);
 	}
 
 	return size;
@@ -579,13 +575,10 @@ static uint32_t msm_ringbuffer_cmd_count(struct fd_ringbuffer *ring)
 	return to_msm_ringbuffer(ring)->cmd_count;
 }
 
-static void msm_ringbuffer_unref(struct fd_ringbuffer *ring)
+static void msm_ringbuffer_destroy(struct fd_ringbuffer *ring)
 {
 	struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring);
 
-	if (!atomic_dec_and_test(&msm_ring->refcnt))
-		return;
-
 	flush_reset(ring);
 	delete_cmds(msm_ring);
 
@@ -596,12 +589,6 @@ static void msm_ringbuffer_unref(struct fd_ringbuffer *ring)
 	free(msm_ring);
 }
 
-static void msm_ringbuffer_ref(struct fd_ringbuffer *ring)
-{
-	struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring);
-	atomic_inc(&msm_ring->refcnt);
-}
-
 static const struct fd_ringbuffer_funcs funcs = {
 		.hostptr = msm_ringbuffer_hostptr,
 		.flush = msm_ringbuffer_flush,
@@ -610,7 +597,7 @@ static const struct fd_ringbuffer_funcs funcs = {
 		.emit_reloc = msm_ringbuffer_emit_reloc,
 		.emit_reloc_ring = msm_ringbuffer_emit_reloc_ring,
 		.cmd_count = msm_ringbuffer_cmd_count,
-		.destroy = msm_ringbuffer_unref,
+		.destroy = msm_ringbuffer_destroy,
 };
 
 drm_private struct fd_ringbuffer * msm_ringbuffer_new(struct fd_pipe *pipe,
@@ -633,9 +620,10 @@ drm_private struct fd_ringbuffer * msm_ringbuffer_new(struct fd_pipe *pipe,
 
 	list_inithead(&msm_ring->cmd_list);
 	msm_ring->seqno = ++to_msm_device(pipe->dev)->ring_cnt;
-	atomic_set(&msm_ring->refcnt, 1);
 
 	ring = &msm_ring->base;
+	atomic_set(&ring->refcnt, 1);
+
 	ring->funcs = &funcs;
 	ring->size = size;
 	ring->pipe = pipe;   /* needed in ring_cmd_new() */