scheduler: Re-compute time-average nr_running on read

Re-compute time-average nr_running when it is read. This would
prevent reading stalled average value if there were no run-queue
changes for a long time. New average value is returned to the reader,
but not stored to avoid concurrent writes. Light-weight sequential
counter synchronization is used to assure data consistency for
re-computing average.

Change-Id: I8e4ea1b28ea00b3ddaf6ef7cdcd27866f87d360b
Signed-off-by: Alex Frid <afrid@nvidia.com>
(cherry picked from commit 527a759d9b40bf57958eb002edd2bb82014dab99)
Reviewed-on: http://git-master/r/111637
Reviewed-by: Sai Gurrappadi <sgurrappadi@nvidia.com>
Tested-by: Sai Gurrappadi <sgurrappadi@nvidia.com>
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Peter Boonstoppel <pboonstoppel@nvidia.com>
Reviewed-by: Yu-Huan Hsu <yhsu@nvidia.com>

forward ported to Linux 3.4 for use on Mako

Signed-off-by: faux123 <reioux@gmail.com>
Signed-off-by: Seongmin Park <r_data@naver.com>

kernel/sched/core.c

@@ -2153,9 +2153,26 @@ unsigned long nr_iowait(void)
 unsigned long avg_nr_running(void)
 {
 	unsigned long i, sum = 0;
+	unsigned int seqcnt, ave_nr_running;
 
-	for_each_online_cpu(i)
-		sum += cpu_rq(i)->ave_nr_running;
+	for_each_online_cpu(i) {
+		struct rq *q = cpu_rq(i);
+
+		/*
+		 * Update average to avoid reading stalled value if there were
+		 * no run-queue changes for a long time. On the other hand if
+		 * the changes are happening right now, just read current value
+		 * directly.
+		 */
+		seqcnt = read_seqcount_begin(&q->ave_seqcnt);
+		ave_nr_running = do_avg_nr_running(q);
+		if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
+			read_seqcount_begin(&q->ave_seqcnt);
+			ave_nr_running = q->ave_nr_running;
+		}
+
+		sum += ave_nr_running;
+	}
 
 	return sum;
 }

kernel/sched/sched.h

@@ -366,6 +366,7 @@ struct rq {
 	/* time-based average load */
 	u64 nr_last_stamp;
 	unsigned int ave_nr_running;
+	seqcount_t ave_seqcnt;
 
 	/* capture load from *all* tasks on this cpu: */
 	struct load_weight load;
@@ -925,31 +926,39 @@ static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
 #define NR_AVE_PERIOD		(1 << NR_AVE_PERIOD_EXP)
 #define NR_AVE_DIV_PERIOD(x)	((x) >> NR_AVE_PERIOD_EXP)
 
-static inline void do_avg_nr_running(struct rq *rq)
+static inline unsigned int do_avg_nr_running(struct rq *rq)
 {
 	s64 nr, deltax;
+	unsigned int ave_nr_running = rq->ave_nr_running;
 
 	deltax = rq->clock_task - rq->nr_last_stamp;
-	rq->nr_last_stamp = rq->clock_task;
 	nr = NR_AVE_SCALE(rq->nr_running);
 
 	if (deltax > NR_AVE_PERIOD)
-		rq->ave_nr_running = nr;
+		ave_nr_running = nr;
 	else
-		rq->ave_nr_running +=
-			NR_AVE_DIV_PERIOD(deltax * (nr - rq->ave_nr_running));
+		ave_nr_running +=
+			NR_AVE_DIV_PERIOD(deltax * (nr - ave_nr_running));
+
+	return ave_nr_running;
 }
 
 static inline void inc_nr_running(struct rq *rq)
 {
-	do_avg_nr_running(rq);
+	write_seqcount_begin(&rq->ave_seqcnt);
+	rq->ave_nr_running = do_avg_nr_running(rq);
+	rq->nr_last_stamp = rq->clock_task;
 	rq->nr_running++;
+	write_seqcount_end(&rq->ave_seqcnt);
 }
 
 static inline void dec_nr_running(struct rq *rq)
 {
-	do_avg_nr_running(rq);
+	write_seqcount_begin(&rq->ave_seqcnt);
+	rq->ave_nr_running = do_avg_nr_running(rq);
+	rq->nr_last_stamp = rq->clock_task;
 	rq->nr_running--;
+	write_seqcount_end(&rq->ave_seqcnt);
 }
 
 extern void update_rq_clock(struct rq *rq);