cpu/hotplug: Fix smpboot thread ordering

Commit 931ef163309e moved the smpboot thread park/unpark invocation to the
state machine. The move of the unpark invocation was premature as it depends
on work in progress patches.

As a result cpu down can fail, because rcu synchronization in takedown_cpu()
eventually requires a functional softirq thread. I never encountered the
problem in testing, but 0day testing managed to provide a reliable reproducer.

Remove the smpboot_threads_park() call from the state machine for now and put
it back into the original place after the rcu synchronization.

I'm embarrassed as I knew about the dependency and still managed to get it
wrong. Hotplug induced brain melt seems to be the only sensible explanation
for that.

Fixes: 931ef163309e "cpu/hotplug: Unpark smpboot threads from the state machine"
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>

kernel/cpu.c

@@ -706,8 +706,9 @@ static int takedown_cpu(unsigned int cpu)
 	else
 		synchronize_rcu();
 
-	/* Park the hotplug thread */
+	/* Park the smpboot threads */
 	kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
+	smpboot_park_threads(cpu);
 
 	/*
 	 * Prevent irq alloc/free while the dying cpu reorganizes the
@@ -1206,7 +1207,7 @@ static struct cpuhp_step cpuhp_ap_states[] = {
 	[CPUHP_AP_SMPBOOT_THREADS] = {
 		.name			= "smpboot:threads",
 		.startup		= smpboot_unpark_threads,
-		.teardown		= smpboot_park_threads,
+		.teardown		= NULL,
 	},
 	[CPUHP_AP_NOTIFY_ONLINE] = {
 		.name			= "notify:online",