cpuidle: lpm-levels: Add support to parse LPM parameters

Add support for providing LPM prediction parameters in
device tree and parsing them.

Change-Id: I60fd7e8d32505e212100b886df29b72a9999ca6d
Signed-off-by: Raghavendra Kakarla <rkakarla@codeaurora.org>

Documentation/devicetree/bindings/arm/msm/lpm-levels.txt

@@ -28,6 +28,14 @@ Required properties:
 	mask of the cluster mode in the composite state ID used to define
 	cluster low power modes in PSCI.
 
+Optional properties:
+	- qcom,disable-prediction: This property is used to indicate the LPM
+	governor will not use LPM prediction for this cluster.
+	- qcom,clstr-tmr-add: This property is used as correction timer for
+	wrong prediction by lpm prediction algorithm for cluster predictions.
+	This value should be between 100 to 1500. Higher values would mean
+	longer time staying in shallower state before waking up to select a
+	deeper state in case of wrong prediction.
 	qcom,pm-cluster contains qcom,pm-cluster-level nodes which identify
 	the various low power modes that the cluster can enter. The
 	qcom,pm-cluster node should also include another cluster node or a cpu
@@ -77,8 +85,22 @@ that share the parameters.It contains the following properties.
 	- qcom,pm-cpu-levels: The different low power modes that a CPU could
 	enter. The following section explains the required properties of this
 	node.
-	-qcom,use-prediction: This optional property is used to indicate the
-	the LPM governor is to apply sleep prediction to this cluster.
+
+Optional properties:
+	- qcom,disable-prediction: This property is used to indicate the
+	LPM governor is to disable sleep prediction to this cpu.
+	- qcom,ref-stddev: This property is used as reference standard deviation
+	in lpm prediction algorithm. This value should be between 100 to 1000.
+	Higher value would result in more predictions and thereby resulting in
+	shallower low power modes.
+	- qcom,tmr-add: This property is used as correction timer for wrong
+	prediction by lpm prediction algorithm. This value should be between
+	100 to 1500. Higher values would mean longer time staying in shallower
+	state before waking up to select a deeper state in case of wrong prediction.
+	- qcom,ref-premature-cnt: This property is used as reference premature
+	count to predict next sleep state by the prediction algorithm. This value
+	should be between 1 to 5. Higher value for this parameter would result in
+	less predictions to disallow deeper low power modes.
 
 [Node bindings for qcom,pm-cpu-levels]
  Required properties:

drivers/cpuidle/lpm-levels-of.c

@@ -1,4 +1,4 @@
-/* Copyright (c) 2014-2017, The Linux Foundation. All rights reserved.
+/* Copyright (c) 2014-2018, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
@@ -178,8 +178,8 @@ static ssize_t lpm_latency_show(struct kobject *kobj,
 	struct kernel_param kp;
 	struct lpm_level_avail *avail = get_avail_ptr(kobj, attr);
 
-	if (!avail)
-		pr_info("Error\n");
+	if (WARN_ON(!avail))
+		return -EINVAL;
 
 	kp.arg = &avail->latency_us;
 
@@ -197,8 +197,15 @@ ssize_t lpm_enable_show(struct kobject *kobj, struct kobj_attribute *attr,
 {
 	int ret = 0;
 	struct kernel_param kp;
+	struct lpm_level_avail *avail = get_avail_ptr(kobj, attr);
+
+	if (WARN_ON(!avail))
+		return -EINVAL;
+
+	kp.arg = get_enabled_ptr(attr, avail);
+	if (WARN_ON(!kp.arg))
+		return -EINVAL;
 
-	kp.arg = get_enabled_ptr(attr, get_avail_ptr(kobj, attr));
 	ret = param_get_bool(buf, &kp);
 	if (ret > 0) {
 		strlcat(buf, "\n", PAGE_SIZE);
@@ -453,6 +460,17 @@ static int parse_cluster_params(struct device_node *node,
 	if (ret)
 		goto fail;
 
+	key = "qcom,disable-prediction";
+	c->lpm_prediction = !(of_property_read_bool(node, key));
+
+	if (c->lpm_prediction) {
+		key = "qcom,clstr-tmr-add";
+		ret = of_property_read_u32(node, key, &c->tmr_add);
+		if (ret || c->tmr_add < TIMER_ADD_LOW ||
+					c->tmr_add > TIMER_ADD_HIGH)
+			c->tmr_add = DEFAULT_TIMER_ADD;
+	}
+
 	/* Set default_level to 0 as default */
 	c->default_level = 0;
 
@@ -713,8 +731,28 @@ static int parse_cpu_levels(struct device_node *node, struct lpm_cluster *c)
 	if (ret)
 		goto failed_parse_params;
 
-	key = "qcom,use-prediction";
-	cpu->lpm_prediction = of_property_read_bool(node, key);
+	key = "qcom,disable-prediction";
+	cpu->lpm_prediction = !(of_property_read_bool(node, key));
+
+	if (cpu->lpm_prediction) {
+		key = "qcom,ref-stddev";
+		ret = of_property_read_u32(node, key, &cpu->ref_stddev);
+		if (ret || cpu->ref_stddev < STDDEV_LOW ||
+					cpu->ref_stddev > STDDEV_HIGH)
+			cpu->ref_stddev = DEFAULT_STDDEV;
+
+		key = "qcom,tmr-add";
+		ret = of_property_read_u32(node, key, &cpu->tmr_add);
+		if (ret || cpu->tmr_add < TIMER_ADD_LOW ||
+					cpu->tmr_add > TIMER_ADD_HIGH)
+			cpu->tmr_add = DEFAULT_TIMER_ADD;
+
+		key = "qcom,ref-premature-cnt";
+		ret = of_property_read_u32(node, key, &cpu->ref_premature_cnt);
+		if (ret || cpu->ref_premature_cnt < PREMATURE_CNT_LOW ||
+				cpu->ref_premature_cnt > PREMATURE_CNT_HIGH)
+			cpu->ref_premature_cnt = DEFAULT_PREMATURE_CNT;
+	}
 
 	key = "parse_cpu";
 	ret = parse_cpu(node, cpu);

drivers/cpuidle/lpm-levels.c

@@ -92,15 +92,6 @@ struct lpm_cluster *lpm_root_node;
 static bool lpm_prediction = true;
 module_param_named(lpm_prediction, lpm_prediction, bool, 0664);
 
-static uint32_t ref_stddev = 500;
-module_param_named(ref_stddev, ref_stddev, uint, 0664);
-
-static uint32_t tmr_add = 1000;
-module_param_named(tmr_add, tmr_add, uint, 0664);
-
-static uint32_t ref_premature_cnt = 1;
-module_param_named(ref_premature_cnt, ref_premature_cnt, uint, 0664);
-
 static uint32_t bias_hyst;
 module_param_named(bias_hyst, bias_hyst, uint, 0664);
 
@@ -496,7 +487,7 @@ again:
 	 * ignore one maximum sample and retry
 	 */
 	if (((avg > stddev * 6) && (divisor >= (MAXSAMPLES - 1)))
-					|| stddev <= ref_stddev) {
+					|| stddev <= cpu->ref_stddev) {
 		history->stime = ktime_to_us(ktime_get()) + avg;
 		return avg;
 	} else if (divisor  > (MAXSAMPLES - 1)) {
@@ -521,7 +512,7 @@ again:
 					total += history->resi[i];
 				}
 			}
-			if (failed >= ref_premature_cnt) {
+			if (failed >= cpu->ref_premature_cnt) {
 				*idx_restrict = j;
 				do_div(total, failed);
 				for (i = 0; i < j; i++) {
@@ -545,8 +536,9 @@ again:
 static inline void invalidate_predict_history(struct cpuidle_device *dev)
 {
 	struct lpm_history *history = &per_cpu(hist, dev->cpu);
+	struct lpm_cpu *lpm_cpu = per_cpu(cpu_lpm, dev->cpu);
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !lpm_cpu->lpm_prediction)
 		return;
 
 	if (history->hinvalid) {
@@ -561,8 +553,9 @@ static void clear_predict_history(void)
 	struct lpm_history *history;
 	int i;
 	unsigned int cpu;
+	struct lpm_cpu *lpm_cpu = per_cpu(cpu_lpm, raw_smp_processor_id());
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !lpm_cpu->lpm_prediction)
 		return;
 
 	for_each_possible_cpu(cpu) {
@@ -681,8 +674,8 @@ static int cpu_power_select(struct cpuidle_device *dev,
 	if ((predicted || (idx_restrict != (cpu->nlevels + 1)))
 			&& ((best_level >= 0)
 			&& (best_level < (cpu->nlevels-1)))) {
-		htime = predicted + tmr_add;
-		if (htime == tmr_add)
+		htime = predicted + cpu->tmr_add;
+		if (htime == cpu->tmr_add)
 			htime = idx_restrict_time;
 		else if (htime > max_residency[best_level])
 			htime = max_residency[best_level];
@@ -746,14 +739,14 @@ static uint64_t get_cluster_sleep_time(struct lpm_cluster *cluster,
 			next_event.tv64 = next_event_c->tv64;
 		}
 
-		if (from_idle && lpm_prediction) {
+		if (from_idle && lpm_prediction && cluster->lpm_prediction) {
 			history = &per_cpu(hist, cpu);
 			if (history->stime && (history->stime < prediction))
 				prediction = history->stime;
 		}
 	}
 
-	if (from_idle && lpm_prediction) {
+	if (from_idle && lpm_prediction && cluster->lpm_prediction) {
 		if (prediction > ktime_to_us(ktime_get()))
 			*pred_time = prediction - ktime_to_us(ktime_get());
 	}
@@ -772,7 +765,7 @@ static int cluster_predict(struct lpm_cluster *cluster,
 	struct cluster_history *history = &cluster->history;
 	int64_t cur_time = ktime_to_us(ktime_get());
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !cluster->lpm_prediction)
 		return 0;
 
 	if (history->hinvalid) {
@@ -847,7 +840,7 @@ static void update_cluster_history(struct cluster_history *history, int idx)
 	struct lpm_cluster *cluster =
 			container_of(history, struct lpm_cluster, history);
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !cluster->lpm_prediction)
 		return;
 
 	if ((history->entry_idx == -1) || (history->entry_idx == idx)) {
@@ -908,7 +901,7 @@ static void clear_cl_predict_history(void)
 	struct lpm_cluster *cluster = lpm_root_node;
 	struct list_head *list;
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !cluster->lpm_prediction)
 		return;
 
 	clear_cl_history_each(&cluster->history);
@@ -1034,7 +1027,7 @@ static int cluster_configure(struct lpm_cluster *cluster, int idx,
 			cluster->child_cpus.bits[0], from_idle);
 		lpm_stats_cluster_enter(cluster->stats, idx);
 
-		if (from_idle && lpm_prediction)
+		if (from_idle && lpm_prediction && cluster->lpm_prediction)
 			update_cluster_history_time(&cluster->history, idx,
 						ktime_to_us(ktime_get()));
 	}
@@ -1066,7 +1059,8 @@ static int cluster_configure(struct lpm_cluster *cluster, int idx,
 	if (predicted && (idx < (cluster->nlevels - 1))) {
 		struct power_params *pwr_params = &cluster->levels[idx].pwr;
 
-		clusttimer_start(cluster, pwr_params->max_residency + tmr_add);
+		clusttimer_start(cluster, pwr_params->max_residency +
+							cluster->tmr_add);
 	}
 
 	return 0;
@@ -1120,7 +1114,8 @@ static void cluster_prepare(struct lpm_cluster *cluster,
 						&cluster->levels[0].pwr;
 
 			clusttimer_start(cluster,
-					pwr_params->max_residency + tmr_add);
+					pwr_params->max_residency +
+					cluster->tmr_add);
 
 			goto failed;
 		}
@@ -1335,8 +1330,9 @@ static void update_history(struct cpuidle_device *dev, int idx)
 {
 	struct lpm_history *history = &per_cpu(hist, dev->cpu);
 	uint32_t tmr = 0;
+	struct lpm_cpu *lpm_cpu = per_cpu(cpu_lpm, dev->cpu);
 
-	if (!lpm_prediction)
+	if (!lpm_prediction || !lpm_cpu->lpm_prediction)
 		return;
 
 	if (history->htmr_wkup) {
@@ -1394,7 +1390,7 @@ exit:
 	update_history(dev, idx);
 	trace_cpu_idle_exit(idx, success);
 	local_irq_enable();
-	if (lpm_prediction) {
+	if (lpm_prediction && cpu->lpm_prediction) {
 		histtimer_cancel();
 		clusttimer_cancel();
 	}

drivers/cpuidle/lpm-levels.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2014-2017, The Linux Foundation. All rights reserved.
+/* Copyright (c) 2014-2018, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
@@ -16,6 +16,15 @@
 #define NR_LPM_LEVELS 8
 #define MAXSAMPLES 5
 #define CLUST_SMPL_INVLD_TIME 40000
+#define DEFAULT_PREMATURE_CNT 3
+#define DEFAULT_STDDEV 100
+#define DEFAULT_TIMER_ADD 100
+#define TIMER_ADD_LOW 100
+#define TIMER_ADD_HIGH 1500
+#define STDDEV_LOW 100
+#define STDDEV_HIGH 1000
+#define PREMATURE_CNT_LOW 1
+#define PREMATURE_CNT_HIGH 5
 
 struct power_params {
 	uint32_t latency_us;		/* Enter + Exit latency */
@@ -43,6 +52,9 @@ struct lpm_cpu {
 	int nlevels;
 	unsigned int psci_mode_shift;
 	unsigned int psci_mode_mask;
+	uint32_t ref_stddev;
+	uint32_t ref_premature_cnt;
+	uint32_t tmr_add;
 	bool lpm_prediction;
 	struct cpuidle_driver *drv;
 	struct lpm_cluster *parent;
@@ -97,6 +109,8 @@ struct lpm_cluster {
 	int min_child_level;
 	int default_level;
 	int last_level;
+	uint32_t tmr_add;
+	bool lpm_prediction;
 	struct list_head cpu;
 	spinlock_t sync_lock;
 	struct cpumask child_cpus;