smp: Fix error case handling in smp_call_function_*()

Commit 8053871d ("smp: Fix smp_call_function_single_async()
locking") fixed the locking for the asynchronous smp-call case, but in
the process of moving the lock handling around, one of the error cases
ended up not unlocking the call data at all.

This went unnoticed on x86, because this is a "caller is buggy" case,
where the caller is trying to call a non-existent CPU.  But apparently
ARM does that (at least under qemu-arm).  Bindly doing cross-cpu calls
to random CPU's that aren't even online seems a bit fishy, but the error
handling was clearly not correct.

Simply add the missing "csd_unlock()" to the error path.
Reported-and-tested-by: Guenter Roeck <linux@roeck-us.net>
Analyzed-by: Rabin Vincent <rabin@rab.in>
Acked-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

kernel/smp.c

@@ -159,8 +159,10 @@ static int generic_exec_single(int cpu, struct call_single_data *csd,
 	}
 
 
-	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu))
+	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
+		csd_unlock(csd);
 		return -ENXIO;
+	}
 
 	csd->func = func;
 	csd->info = info;