if (ilb_cpu >= nr_cpu_ids)
                return;
 
-       flags = atomic_fetch_or(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu));
-       if (flags & NOHZ_BALANCE_KICK)
+       flags = atomic_fetch_or(NOHZ_KICK_MASK, nohz_flags(ilb_cpu));
+       if (flags & NOHZ_KICK_MASK)
                return;
        /*
         * Use smp_send_reschedule() instead of resched_cpu().
        int need_serialize, need_decay = 0;
        u64 max_cost = 0;
 
-       update_blocked_averages(cpu);
-
        rcu_read_lock();
        for_each_domain(cpu, sd) {
                /*
  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
  * rebalancing for all the CPUs for whom scheduler ticks are stopped.
  */
-static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
+static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
 {
-       int this_cpu = this_rq->cpu;
-       struct rq *rq;
-       int balance_cpu;
        /* Earliest time when we have to do rebalance again */
        unsigned long next_balance = jiffies + 60*HZ;
        int update_next_balance = 0;
+       int this_cpu = this_rq->cpu;
+       unsigned int flags;
+       int balance_cpu;
+       struct rq *rq;
 
-       if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_BALANCE_KICK))
-               return;
+       if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
+               return false;
 
-       if (idle != CPU_IDLE)
-               goto end;
+       if (idle != CPU_IDLE) {
+               atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
+               return false;
+       }
+
+       flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
+
+       SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
 
        for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
                if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
                        cpu_load_update_idle(rq);
                        rq_unlock_irq(rq, &rf);
 
-                       rebalance_domains(rq, CPU_IDLE);
+                       update_blocked_averages(rq->cpu);
+                       if (flags & NOHZ_BALANCE_KICK)
+                               rebalance_domains(rq, CPU_IDLE);
                }
 
                if (time_after(next_balance, rq->next_balance)) {
                }
        }
 
+       update_blocked_averages(this_cpu);
+       if (flags & NOHZ_BALANCE_KICK)
+               rebalance_domains(this_rq, CPU_IDLE);
+
        /*
         * next_balance will be updated only when there is a need.
         * When the CPU is attached to null domain for ex, it will not be
         */
        if (likely(update_next_balance))
                nohz.next_balance = next_balance;
-end:
-       atomic_andnot(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
+
+       return true;
 }
 
 /*
        return kick;
 }
 #else
-static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
+static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
+{
+       return false;
+}
 #endif
 
 /*
         * load balance only within the local sched_domain hierarchy
         * and abort nohz_idle_balance altogether if we pull some load.
         */
-       nohz_idle_balance(this_rq, idle);
+       if (nohz_idle_balance(this_rq, idle))
+               return;
+
+       /* normal load balance */
+       update_blocked_averages(this_rq->cpu);
        rebalance_domains(this_rq, idle);
 }