diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 0c57b20532be..21fb41c93caa 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1936,11 +1936,27 @@ out:
 void
 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
 {
+	/* Serialisation between user GTT access and our code depends upon
+	 * revoking the CPU's PTE whilst the mutex is held. The next user
+	 * pagefault then has to wait until we release the mutex.
+	 */
+	lockdep_assert_held(&obj->base.dev->struct_mutex);
+
 	if (!obj->fault_mappable)
 		return;
 
 	drm_vma_node_unmap(&obj->base.vma_node,
 			   obj->base.dev->anon_inode->i_mapping);
+
+	/* Ensure that the CPU's PTE are revoked and there are not outstanding
+	 * memory transactions from userspace before we return. The TLB
+	 * flushing implied above by changing the PTE above *should* be
+	 * sufficient, an extra barrier here just provides us with a bit
+	 * of paranoid documentation about our requirement to serialise
+	 * memory writes before touching registers / GSM.
+	 */
+	wmb();
+
 	obj->fault_mappable = false;
 }
 
@@ -3324,9 +3340,6 @@ static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
 	if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
 		return;
 
-	/* Wait for any direct GTT access to complete */
-	mb();
-
 	old_read_domains = obj->base.read_domains;
 	old_write_domain = obj->base.write_domain;