On Mon, Jun 24, 2013 at 10:55:35AM -0700, Tejun Heo wrote:

quoted

@@ -105,6 +106,7 @@ s64 __percpu_counter_sum(struct percpu_counter *fbc)
 		ret += *pcount;
 	}
 	raw_spin_unlock(&fbc->lock);
+	put_online_cpus_atomic();

I don't think this is necessary.  CPU on/offlining is explicitly
handled via the hotplug callback which synchronizes through fbc->lock.
__percpu_counter_sum() racing with actual on/offlining doesn't affect
correctness and adding superflous get_online_cpus_atomic() around it
can be misleading.

Ah, okay, so you added a debug feature which triggers warning if
online mask is accessed without synchronization.  Yeah, that makes
sense and while the above is not strictly necessary, it probably is
better to just add it rather than suppressing the warning in a
different way.  Can you please at least add a comment explaining that?

Thanks.

-- 
tejun