cpuquiet: Update averaging of nr_runnables

Doing a Exponential moving average per nr_running++/-- does not
guarantee a fixed sample rate which induces errors if there are lots of
threads being enqueued/dequeued from the rq (Linpack mt). Instead of
keeping track of the avg, the scheduler now keeps track of the integral
of nr_running and allows the readers to perform filtering on top.

Implemented a proper exponential moving average for the runnables
governor and a straight 100ms average for the balanced governor. Tweaked
the thresholds for the runnables governor to minimize latency. Also,
decreased sample_rate for the runnables governor to the absolute minimum
of 10msecs.

Updated to K3.4

Change-Id: Ia25bf8baf2a1a015ba188b2c06e551e89b16c5f8
Signed-off-by: Sai Charan Gurrappadi <sgurrappadi@nvidia.com>
Signed-off-by: Peter De Schrijver <pdeschrijver@nvidia.com>
Reviewed-on: http://git-master/r/131147
Reviewed-by: Juha Tukkinen <jtukkinen@nvidia.com>

2 files changed, 147 insertions, 36 deletions

diff --git a/drivers/cpuquiet/governors/balanced.c b/drivers/cpuquiet/governors/balanced.c
index 187ac2e7f799..86a012a5ace5 100644
--- a/drivers/cpuquiet/governors/balanced.c
+++ b/drivers/cpuquiet/governors/balanced.c
@@ -204,6 +204,54 @@ static unsigned int *rt_profiles[] = {
 static unsigned int nr_run_hysteresis = 2;	/* 0.5 thread */
 static unsigned int nr_run_last;
 
+struct runnables_avg_sample {
+	u64 previous_integral;
+	unsigned int avg;
+	bool integral_sampled;
+	u64 prev_timestamp;
+};
+
+static DEFINE_PER_CPU(struct runnables_avg_sample, avg_nr_sample);
+
+static unsigned int get_avg_nr_runnables(void)
+{
+	unsigned int i, sum = 0;
+	struct runnables_avg_sample *sample;
+	u64 integral, old_integral, delta_integral, delta_time, cur_time;
+
+	for_each_online_cpu(i) {
+		sample = &per_cpu(avg_nr_sample, i);
+		integral = nr_running_integral(i);
+		old_integral = sample->previous_integral;
+		sample->previous_integral = integral;
+		cur_time = ktime_to_ns(ktime_get());
+		delta_time = cur_time - sample->prev_timestamp;
+		sample->prev_timestamp = cur_time;
+
+		if (!sample->integral_sampled) {
+			sample->integral_sampled = true;
+			/* First sample to initialize prev_integral, skip
+			 * avg calculation
+			 */
+			continue;
+		}
+
+		if (integral < old_integral) {
+			/* Overflow */
+			delta_integral = (ULLONG_MAX - old_integral) + integral;
+		} else {
+			delta_integral = integral - old_integral;
+		}
+
+		/* Calculate average for the previous sample window */
+		do_div(delta_integral, delta_time);
+		sample->avg = delta_integral;
+		sum += sample->avg;
+	}
+
+	return sum;
+}
+
 static CPU_SPEED_BALANCE balanced_speed_balance(void)
 {
 	unsigned long highest_speed = cpu_highest_speed();
@@ -211,7 +259,7 @@ static CPU_SPEED_BALANCE balanced_speed_balance(void)
 	unsigned long skewed_speed = balanced_speed / 2;
 	unsigned int nr_cpus = num_online_cpus();
 	unsigned int max_cpus = pm_qos_request(PM_QOS_MAX_ONLINE_CPUS) ? : 4;
-	unsigned int avg_nr_run = avg_nr_running();
+	unsigned int avg_nr_run = get_avg_nr_runnables();
 	unsigned int nr_run;
 	unsigned int *current_profile = rt_profiles[rt_profile_sel];
 
diff --git a/drivers/cpuquiet/governors/runnable_threads.c b/drivers/cpuquiet/governors/runnable_threads.c
index 44cf308befcf..8d75daff2228 100644
--- a/drivers/cpuquiet/governors/runnable_threads.c
+++ b/drivers/cpuquiet/governors/runnable_threads.c
@@ -33,14 +33,15 @@ typedef enum {
 	UP,
 } RUNNABLES_STATE;
 
-static struct delayed_work runnables_work;
+static struct work_struct runnables_work;
 static struct kobject *runnables_kobject;
+static struct timer_list runnables_timer;
 
+static RUNNABLES_STATE runnables_state;
 /* configurable parameters */
-static unsigned int sample_rate = 20;		/* msec */
+static unsigned int sample_rate = 10;		/* msec */
 
 static RUNNABLES_STATE runnables_state;
-static struct workqueue_struct *runnables_wq;
 
 #define NR_FSHIFT_EXP	3
 #define NR_FSHIFT	(1 << NR_FSHIFT_EXP)
@@ -56,17 +57,93 @@ static unsigned int nr_run_thresholds[NR_CPUS];
 
 DEFINE_MUTEX(runnables_work_lock);
 
-static void update_runnables_state(void)
+struct runnables_avg_sample {
+	u64 previous_integral;
+	unsigned int avg;
+	bool integral_sampled;
+	u64 prev_timestamp;
+};
+
+static DEFINE_PER_CPU(struct runnables_avg_sample, avg_nr_sample);
+
+/* EXP = alpha in the exponential moving average.
+ * Alpha = e ^ (-sample_rate / window_size) * FIXED_1
+ * Calculated for sample_rate of 10ms, window size of 63.82ms
+ */
+#define EXP    1751
+
+static unsigned int get_avg_nr_runnables(void)
+{
+	unsigned int i, sum = 0;
+	static unsigned int avg;
+	struct runnables_avg_sample *sample;
+	u64 integral, old_integral, delta_integral, delta_time, cur_time;
+
+	for_each_online_cpu(i) {
+		sample = &per_cpu(avg_nr_sample, i);
+		integral = nr_running_integral(i);
+		old_integral = sample->previous_integral;
+		sample->previous_integral = integral;
+		cur_time = ktime_to_ns(ktime_get());
+		delta_time = cur_time - sample->prev_timestamp;
+		sample->prev_timestamp = cur_time;
+
+		if (!sample->integral_sampled) {
+			sample->integral_sampled = true;
+			/* First sample to initialize prev_integral, skip
+			 * avg calculation
+			 */
+			continue;
+		}
+
+		if (integral < old_integral) {
+			/* Overflow */
+			delta_integral = (ULLONG_MAX - old_integral) + integral;
+		} else {
+			delta_integral = integral - old_integral;
+		}
+
+		/* Calculate average for the previous sample window */
+		do_div(delta_integral, delta_time);
+		sample->avg = delta_integral;
+		sum += sample->avg;
+	}
+
+	/* Exponential moving average
+	 * Avgn = Avgn-1 * alpha + new_avg * (1 - alpha)
+	 */
+	avg *= EXP;
+	avg += sum * (FIXED_1 - EXP);
+	avg >>= FSHIFT;
+
+	return avg;
+}
+
+static void update_runnables_state(unsigned int nr_run)
 {
 	unsigned int nr_cpus = num_online_cpus();
 	int max_cpus = pm_qos_request(PM_QOS_MAX_ONLINE_CPUS) ? : 4;
 	int min_cpus = pm_qos_request(PM_QOS_MIN_ONLINE_CPUS);
-	unsigned int avg_nr_run = avg_nr_running();
-	unsigned int nr_run;
+
+	if ((nr_cpus > max_cpus || nr_run < nr_cpus) && nr_cpus >= min_cpus) {
+		runnables_state = DOWN;
+	} else if (nr_cpus < min_cpus || nr_run > nr_cpus) {
+		runnables_state =  UP;
+	} else {
+		runnables_state = IDLE;
+	}
+}
+
+static void runnables_avg_sampler(unsigned long data)
+{
+	unsigned int nr_run, avg_nr_run;
 
 	if (runnables_state == DISABLED)
 		return;
 
+	avg_nr_run = get_avg_nr_runnables();
+	mod_timer(&runnables_timer, jiffies + msecs_to_jiffies(sample_rate));
+
 	for (nr_run = 1; nr_run < ARRAY_SIZE(nr_run_thresholds); nr_run++) {
 		unsigned int nr_threshold = nr_run_thresholds[nr_run - 1];
 		if (nr_run_last <= nr_run)
@@ -74,15 +151,12 @@ static void update_runnables_state(void)
 		if (avg_nr_run <= (nr_threshold << (FSHIFT - NR_FSHIFT_EXP)))
 			break;
 	}
+
 	nr_run_last = nr_run;
+	update_runnables_state(nr_run);
 
-	if ((nr_cpus > max_cpus || nr_run < nr_cpus) && nr_cpus >= min_cpus) {
-		runnables_state = DOWN;
-	} else if (nr_cpus < min_cpus || nr_run > nr_cpus) {
-		runnables_state =  UP;
-	} else {
-		runnables_state = IDLE;
-	}
+	if (runnables_state != DISABLED && runnables_state != IDLE)
+		schedule_work(&runnables_work);
 }
 
 static unsigned int get_lightest_loaded_cpu_n(void)
@@ -92,8 +166,8 @@ static unsigned int get_lightest_loaded_cpu_n(void)
 	int i;
 
 	for_each_online_cpu(i) {
-		unsigned int nr_runnables = get_avg_nr_running(i);
-
+		struct runnables_avg_sample *s = &per_cpu(avg_nr_sample, i);
+		unsigned int nr_runnables = s->avg;
 		if (i > 0 && min_avg_runnables > nr_runnables) {
 			cpu = i;
 			min_avg_runnables = nr_runnables;
@@ -106,27 +180,23 @@ static unsigned int get_lightest_loaded_cpu_n(void)
 static void runnables_work_func(struct work_struct *work)
 {
 	bool up = false;
-	bool sample = false;
 	unsigned int cpu = nr_cpu_ids;
 
 	mutex_lock(&runnables_work_lock);
 
-	update_runnables_state();
+	/* Update state to avoid duplicate operations */
+	update_runnables_state(nr_run_last);
 
 	switch (runnables_state) {
 	case DISABLED:
-		break;
 	case IDLE:
-		sample = true;
 		break;
 	case UP:
 		cpu = cpumask_next_zero(0, cpu_online_mask);
 		up = true;
-		sample = true;
 		break;
 	case DOWN:
 		cpu = get_lightest_loaded_cpu_n();
-		sample = true;
 		break;
 	default:
 		pr_err("%s: invalid cpuquiet runnable governor state %d\n",
@@ -134,10 +204,6 @@ static void runnables_work_func(struct work_struct *work)
 		break;
 	}
 
-	if (sample)
-		queue_delayed_work(runnables_wq, &runnables_work,
-					msecs_to_jiffies(sample_rate));
-
 	if (cpu < nr_cpu_ids) {
 		if (up)
 			cpuquiet_wake_cpu(cpu);
@@ -190,7 +256,7 @@ static void runnables_device_busy(void)
 {
 	if (runnables_state != DISABLED) {
 		runnables_state = DISABLED;
-		cancel_delayed_work_sync(&runnables_work);
+		cancel_work_sync(&runnables_work);
 	}
 }
 
@@ -198,15 +264,14 @@ static void runnables_device_free(void)
 {
 	if (runnables_state == DISABLED) {
 		runnables_state = IDLE;
-		runnables_work_func(NULL);
+		mod_timer(&runnables_timer, jiffies + 1);
 	}
 }
 
 static void runnables_stop(void)
 {
 	runnables_state = DISABLED;
-	cancel_delayed_work_sync(&runnables_work);
-	destroy_workqueue(runnables_wq);
+	cancel_work_sync(&runnables_work);
 	kobject_put(runnables_kobject);
 }
 
@@ -218,12 +283,10 @@ static int runnables_start(void)
 	if (err)
 		return err;
 
-	runnables_wq = alloc_workqueue("cpuquiet-runnables",
-			WQ_UNBOUND | WQ_RESCUER | WQ_FREEZABLE, 1);
-	if (!runnables_wq)
-		return -ENOMEM;
+	INIT_WORK(&runnables_work, runnables_work_func);
 
-	INIT_DELAYED_WORK(&runnables_work, runnables_work_func);
+	init_timer(&runnables_timer);
+	runnables_timer.function = runnables_avg_sampler;
 
 	for(i = 0; i < ARRAY_SIZE(nr_run_thresholds); ++i) {
 		if (i < ARRAY_SIZE(default_thresholds))
@@ -236,7 +299,7 @@ static int runnables_start(void)
 	}
 
 	runnables_state = IDLE;
-	runnables_work_func(NULL);
+	runnables_avg_sampler(0);
 
 	return 0;
 }