The percpu-counter-sum code does a for_each_online_cpu() protected
by a spinlock, which makes it look like it needs to use
get/put_online_cpus_atomic(), going forward. However, the code has
adequate synchronization with CPU hotplug, via a hotplug callback
and the fbc->lock.

So use for_each_online_cpu_nocheck() to avoid false-positive warnings
from the hotplug locking validator. And add a comment justifying the
same.

Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Srivatsa S. Bhat <redacted>
---

 lib/percpu_counter.c |    9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/lib/percpu_counter.c b/lib/percpu_counter.c
index ba6085d..2d80e8a 100644
--- a/lib/percpu_counter.c
+++ b/lib/percpu_counter.c

@@ -98,9 +98,16 @@ s64 __percpu_counter_sum(struct percpu_counter *fbc)
 	s64 ret;
 	int cpu;
 
+	/*
+	 * CPU hotplug synchronization is explicitly handled via the
+	 * hotplug callback, which synchronizes through fbc->lock.
+	 * So it is safe to use the _nocheck() version of
+	 * for_each_online_cpu() here (to avoid false-positive warnings
+	 * from the CPU hotplug debug code).
+	 */
 	raw_spin_lock(&fbc->lock);
 	ret = fbc->count;
-	for_each_online_cpu(cpu) {
+	for_each_online_cpu_nocheck(cpu) {
 		s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
 		ret += *pcount;
 	}