On Mon, Jul 25, 2016 at 03:12:24PM -0700, bsegall@google.com wrote:

quoted

@@ -2690,15 +2707,15 @@ static inline int update_cfs_rq_load_avg
 
 	if (atomic_long_read(&cfs_rq->removed_load_avg)) {
 		s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
-		sa->load_avg = max_t(long, sa->load_avg - r, 0);
-		sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
+		sub_positive(&sa->load_avg, r);
+		sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
 		removed = 1;
 	}
 
 	if (atomic_long_read(&cfs_rq->removed_util_avg)) {
 		long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
-		sa->util_avg = max_t(long, sa->util_avg - r, 0);
-		sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
+		sub_positive(&sa->util_avg, r);
+		sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
 	}
 
 	decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,

I missed this the first time around, and I have no problem with this
backport, but it's not remotely obvious that update_cfs_rq_h_load() is
intended to be something approaching racing-safe when not under
rq->lock. (And given only somewhat-adversarial compilers I agree that it
probably won't do any worse than skip updates, though I certainly won't
swear to it)

Right, this is somewhat 'new' because we now use the avg load value for
the load-balancer, which is entirely unserialized.