diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
index 0e8e9a89553befd2cd10f35bc27f16ce161acd00..85abfb2bb9516e430dd50253f488d7899a6e204a 100644
--- a/net/sched/cls_api.c
+++ b/net/sched/cls_api.c
@@ -294,7 +294,22 @@ static void tcf_chain_put(struct tcf_chain *chain);
 static void tcf_proto_destroy(struct tcf_proto *tp, bool rtnl_held,
 			      bool sig_destroy, struct netlink_ext_ack *extack)
 {
-	tp->ops->destroy(tp, rtnl_held, extack);
+	/* A locked classifier's destroy callback (e.g. u32_destroy) uses
+	 * rtnl_dereference() and mutates shared structures (e.g. the
+	 * tc_u_common hash list) that are only safe under rtnl_lock. When an
+	 * unlocked classifier's request (e.g. flower on ingress) loses the
+	 * tcf_chain_tp_insert_unique() race and ends up dropping the last
+	 * reference on a locked classifier's proto, destroy() would run
+	 * without rtnl held. Take it here in that case.
+	 */
+	bool not_lockless = !rtnl_held &&
+		!(tp->ops->flags & TCF_PROTO_OPS_DOIT_UNLOCKED);
+
+	if (not_lockless)
+		rtnl_lock();
+	tp->ops->destroy(tp, rtnl_held || not_lockless, extack);
+	if (not_lockless)
+		rtnl_unlock();
 	if (sig_destroy)
 		tcf_proto_signal_destroyed(tp->chain, tp);
 	tcf_chain_put(tp->chain);