struct drm_sched_entity *entity)
 {
        struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
-       unsigned idx = centity->sequence & (amdgpu_sched_jobs - 1);
-       struct dma_fence *other = centity->fences[idx];
+       struct dma_fence *other;
+       unsigned idx;
+       long r;
 
-       if (other) {
-               signed long r;
-               r = dma_fence_wait(other, true);
-               if (r < 0) {
-                       if (r != -ERESTARTSYS)
-                               DRM_ERROR("Error (%ld) waiting for fence!\n", r);
+       spin_lock(&ctx->ring_lock);
+       idx = centity->sequence & (amdgpu_sched_jobs - 1);
+       other = dma_fence_get(centity->fences[idx]);
+       spin_unlock(&ctx->ring_lock);
 
-                       return r;
-               }
-       }
+       if (!other)
+               return 0;
 
-       return 0;
+       r = dma_fence_wait(other, true);
+       if (r < 0 && r != -ERESTARTSYS)
+               DRM_ERROR("Error (%ld) waiting for fence!\n", r);
+
+       dma_fence_put(other);
+       return r;
 }
 
 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)