bpf: Use rcu_trace_implies_rcu_gp() for program array freeing

To support both sleepable and normal uprobe bpf program, the freeing of
trace program array chains a RCU-tasks-trace grace period and a normal
RCU grace period one after the other.

With the introduction of rcu_trace_implies_rcu_gp(),
__bpf_prog_array_free_sleepable_cb() can check whether or not a normal
RCU grace period has also passed after a RCU-tasks-trace grace period
has passed. If it is true, it is safe to invoke kfree() directly.

Signed-off-by: Hou Tao <[email protected]>

Author: Hou Tao

kernel/bpf/core.c

@@ -2251,8 +2251,14 @@ static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
 {
 	struct bpf_prog_array *progs;
 
+	/* If RCU Tasks Trace grace period implies RCU grace period, there is
+	 * no need to call kfree_rcu(), just call kfree() directly.
+	 */
 	progs = container_of(rcu, struct bpf_prog_array, rcu);
-	kfree_rcu(progs, rcu);
+	if (rcu_trace_implies_rcu_gp())
+		kfree(progs);
+	else
+		kfree_rcu(progs, rcu);
 }
 
 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)