Merge "cpuidle: lpm-levels: Consider history during LPM selection"

drivers/cpuidle/lpm-levels-of.c

@@ -38,34 +38,138 @@ static const struct lpm_type_str lpm_types[] = {
 	{SUSPEND, "suspend_enabled"},
 };
 
+static DEFINE_PER_CPU(uint32_t *, max_residency);
+static DEFINE_PER_CPU(uint32_t *, min_residency);
 static struct lpm_level_avail *cpu_level_available[NR_CPUS];
 static struct platform_device *lpm_pdev;
 
-static void *get_avail_val(struct kobject *kobj, struct kobj_attribute *attr)
+static void *get_enabled_ptr(struct kobj_attribute *attr,
+					struct lpm_level_avail *avail)
 {
 	void *arg = NULL;
-	struct lpm_level_avail *avail = NULL;
 
-	if (!strcmp(attr->attr.name, lpm_types[IDLE].str)) {
+	if (!strcmp(attr->attr.name, lpm_types[IDLE].str))
-		avail = container_of(attr, struct lpm_level_avail,
-					idle_enabled_attr);
 		arg = (void *) &avail->idle_enabled;
-	} else if (!strcmp(attr->attr.name, lpm_types[SUSPEND].str)) {
+	else if (!strcmp(attr->attr.name, lpm_types[SUSPEND].str))
-		avail = container_of(attr, struct lpm_level_avail,
-					suspend_enabled_attr);
 		arg = (void *) &avail->suspend_enabled;
-	}
 
 	return arg;
 }
 
+static struct lpm_level_avail *get_avail_ptr(struct kobject *kobj,
+					struct kobj_attribute *attr)
+{
+	struct lpm_level_avail *avail = NULL;
+
+	if (!strcmp(attr->attr.name, lpm_types[IDLE].str))
+		avail = container_of(attr, struct lpm_level_avail,
+					idle_enabled_attr);
+	else if (!strcmp(attr->attr.name, lpm_types[SUSPEND].str))
+		avail = container_of(attr, struct lpm_level_avail,
+					suspend_enabled_attr);
+
+	return avail;
+}
+
+static void set_optimum_cpu_residency(struct lpm_cpu *cpu, int cpu_id,
+		bool probe_time)
+{
+	int i, j;
+	bool mode_avail;
+	uint32_t *maximum_residency = per_cpu(max_residency, cpu_id);
+	uint32_t *minimum_residency = per_cpu(min_residency, cpu_id);
+
+	for (i = 0; i < cpu->nlevels; i++) {
+		struct power_params *pwr = &cpu->levels[i].pwr;
+
+		mode_avail = probe_time ||
+			lpm_cpu_mode_allow(cpu_id, i, true);
+
+		if (!mode_avail) {
+			maximum_residency[i] = 0;
+			minimum_residency[i] = 0;
+			continue;
+		}
+
+		maximum_residency[i] = ~0;
+		for (j = i + 1; j < cpu->nlevels; j++) {
+			mode_avail = probe_time ||
+					lpm_cpu_mode_allow(cpu_id, j, true);
+
+			if (mode_avail &&
+				(maximum_residency[i] > pwr->residencies[j]) &&
+				(pwr->residencies[j] != 0))
+				maximum_residency[i] = pwr->residencies[j];
+		}
+
+		minimum_residency[i] = pwr->time_overhead_us;
+		for (j = i-1; j >= 0; j--) {
+			if (probe_time || lpm_cpu_mode_allow(cpu_id, j, true)) {
+				minimum_residency[i] = maximum_residency[j] + 1;
+				break;
+			}
+		}
+	}
+}
+
+static void set_optimum_cluster_residency(struct lpm_cluster *cluster,
+		bool probe_time)
+{
+	int i, j;
+	bool mode_avail;
+
+	for (i = 0; i < cluster->nlevels; i++) {
+		struct power_params *pwr = &cluster->levels[i].pwr;
+
+		mode_avail = probe_time ||
+			lpm_cluster_mode_allow(cluster, i,
+					true);
+
+		if (!mode_avail) {
+			pwr->max_residency = 0;
+			pwr->min_residency = 0;
+			continue;
+		}
+
+		pwr->max_residency = ~0;
+		for (j = i+1; j < cluster->nlevels; j++) {
+			mode_avail = probe_time ||
+					lpm_cluster_mode_allow(cluster, j,
+							true);
+			if (mode_avail &&
+				(pwr->max_residency > pwr->residencies[j]) &&
+				(pwr->residencies[j] != 0))
+				pwr->max_residency = pwr->residencies[j];
+		}
+
+		pwr->min_residency = pwr->time_overhead_us;
+		for (j = i-1;  j >= 0; j--) {
+			if (probe_time ||
+				lpm_cluster_mode_allow(cluster, j, true)) {
+				pwr->min_residency =
+				  cluster->levels[j].pwr.max_residency + 1;
+				break;
+			}
+		}
+	}
+}
+
+uint32_t *get_per_cpu_max_residency(int cpu)
+{
+	return per_cpu(max_residency, cpu);
+}
+
+uint32_t *get_per_cpu_min_residency(int cpu)
+{
+	return per_cpu(min_residency, cpu);
+}
 ssize_t lpm_enable_show(struct kobject *kobj, struct kobj_attribute *attr,
 				char *buf)
 {
 	int ret = 0;
 	struct kernel_param kp;
 
-	kp.arg = get_avail_val(kobj, attr);
+	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);
@@ -80,15 +184,25 @@ ssize_t lpm_enable_store(struct kobject *kobj, struct kobj_attribute *attr,
 {
 	int ret = 0;
 	struct kernel_param kp;
+	struct lpm_level_avail *avail;
 
-	kp.arg = get_avail_val(kobj, attr);
+	avail = get_avail_ptr(kobj, attr);
+	if (WARN_ON(!avail))
+		return -EINVAL;
+	kp.arg = get_enabled_ptr(attr, avail);
 	ret = param_set_bool(buf, &kp);
 
+	if (avail->cpu_node)
+		set_optimum_cpu_residency(avail->data, avail->idx, false);
+	else
+		set_optimum_cluster_residency(avail->data, false);
+
 	return ret ? ret : len;
 }
 
 static int create_lvl_avail_nodes(const char *name,
-			struct kobject *parent, struct lpm_level_avail *avail)
+			struct kobject *parent, struct lpm_level_avail *avail,
+			void *data, int index, bool cpu_node)
 {
 	struct attribute_group *attr_group = NULL;
 	struct attribute **attr = NULL;
@@ -139,6 +253,9 @@ static int create_lvl_avail_nodes(const char *name,
 	avail->idle_enabled = true;
 	avail->suspend_enabled = true;
 	avail->kobj = kobj;
+	avail->data = data;
+	avail->idx = index;
+	avail->cpu_node = cpu_node;
 
 	return ret;
 
@@ -181,7 +298,8 @@ static int create_cpu_lvl_nodes(struct lpm_cluster *p, struct kobject *parent)
 		for (i = 0; i < p->cpu->nlevels; i++) {
 
 			ret = create_lvl_avail_nodes(p->cpu->levels[i].name,
-					cpu_kobj[cpu_idx], &level_list[i]);
+					cpu_kobj[cpu_idx], &level_list[i],
+					(void *)p->cpu, cpu, true);
 			if (ret)
 				goto release_kobj;
 		}
@@ -215,7 +333,8 @@ int create_cluster_lvl_nodes(struct lpm_cluster *p, struct kobject *kobj)
 
 	for (i = 0; i < p->nlevels; i++) {
 		ret = create_lvl_avail_nodes(p->levels[i].level_name,
-				cluster_kobj, &p->levels[i].available);
+				cluster_kobj, &p->levels[i].available,
+				(void *)p, 0, false);
 		if (ret)
 			return ret;
 	}
@@ -421,6 +540,9 @@ static int parse_power_params(struct device_node *node,
 
 	key = "qcom,time-overhead";
 	ret = of_property_read_u32(node, key, &pwr->time_overhead_us);
+	if (ret)
+		goto fail;
+
 fail:
 	if (ret)
 		pr_err("%s(): %s Error reading %s\n", __func__, node->name,
@@ -615,11 +737,31 @@ static int get_cpumask_for_node(struct device_node *node, struct cpumask *mask)
 	return 0;
 }
 
+static int calculate_residency(struct power_params *base_pwr,
+					struct power_params *next_pwr)
+{
+	int32_t residency = (int32_t)(next_pwr->energy_overhead -
+						base_pwr->energy_overhead) -
+		((int32_t)(next_pwr->ss_power * next_pwr->time_overhead_us)
+		- (int32_t)(base_pwr->ss_power * base_pwr->time_overhead_us));
+
+	residency /= (int32_t)(base_pwr->ss_power  - next_pwr->ss_power);
+
+	if (residency < 0) {
+		__WARN_printf("%s: Incorrect power attributes for LPM\n",
+				__func__);
+		return next_pwr->time_overhead_us;
+	}
+
+	return residency < next_pwr->time_overhead_us ?
+				next_pwr->time_overhead_us : residency;
+}
+
 static int parse_cpu_levels(struct device_node *node, struct lpm_cluster *c)
 {
 	struct device_node *n;
 	int ret = -ENOMEM;
-	int i;
+	int i, j;
 	char *key;
 
 	c->cpu = devm_kzalloc(&lpm_pdev->dev, sizeof(*c->cpu), GFP_KERNEL);
@@ -676,6 +818,22 @@ static int parse_cpu_levels(struct device_node *node, struct lpm_cluster *c)
 		else if (ret)
 			goto failed;
 	}
+	for (i = 0; i < c->cpu->nlevels; i++) {
+		for (j = 0; j < c->cpu->nlevels; j++) {
+			if (i >= j) {
+				c->cpu->levels[i].pwr.residencies[j] = 0;
+				continue;
+			}
+
+			c->cpu->levels[i].pwr.residencies[j] =
+				calculate_residency(&c->cpu->levels[i].pwr,
+					&c->cpu->levels[j].pwr);
+
+			pr_err("%s: idx %d %u\n", __func__, j,
+					c->cpu->levels[i].pwr.residencies[j]);
+		}
+	}
+
 	return 0;
 failed:
 	for (i = 0; i < c->cpu->nlevels; i++) {
@@ -732,6 +890,7 @@ struct lpm_cluster *parse_cluster(struct device_node *node,
 	struct device_node *n;
 	char *key;
 	int ret = 0;
+	int i, j;
 
 	c = devm_kzalloc(&lpm_pdev->dev, sizeof(*c), GFP_KERNEL);
 	if (!c)
@@ -789,6 +948,22 @@ struct lpm_cluster *parse_cluster(struct device_node *node,
 				goto failed_parse_cluster;
 
 			c->aff_level = 1;
+
+			for_each_cpu(i, &c->child_cpus) {
+				per_cpu(max_residency, i) = devm_kzalloc(
+					&lpm_pdev->dev,
+					sizeof(uint32_t) * c->cpu->nlevels,
+					GFP_KERNEL);
+				if (!per_cpu(max_residency, i))
+					return ERR_PTR(-ENOMEM);
+				per_cpu(min_residency, i) = devm_kzalloc(
+					&lpm_pdev->dev,
+					sizeof(uint32_t) * c->cpu->nlevels,
+					GFP_KERNEL);
+				if (!per_cpu(min_residency, i))
+					return ERR_PTR(-ENOMEM);
+				set_optimum_cpu_residency(c->cpu, i, true);
+			}
 		}
 	}
 
@@ -797,6 +972,17 @@ struct lpm_cluster *parse_cluster(struct device_node *node,
 	else
 		c->last_level = c->nlevels-1;
 
+	for (i = 0; i < c->nlevels; i++) {
+		for (j = 0; j < c->nlevels; j++) {
+			if (i >= j) {
+				c->levels[i].pwr.residencies[j] = 0;
+				continue;
+			}
+			c->levels[i].pwr.residencies[j] = calculate_residency(
+				&c->levels[i].pwr, &c->levels[j].pwr);
+		}
+	}
+	set_optimum_cluster_residency(c, true);
 	return c;
 
 failed_parse_cluster:

drivers/cpuidle/lpm-levels.c

@@ -1,4 +1,6 @@
 /* Copyright (c) 2012-2016, The Linux Foundation. All rights reserved.
+ * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
+ * Copyright (C) 2009 Intel Corporation
  *
  * 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
@@ -83,9 +85,37 @@ struct lpm_debug {
 
 struct lpm_cluster *lpm_root_node;
 
+#define MAXSAMPLES 5
+
+static bool lpm_prediction;
+module_param_named(lpm_prediction,
+	lpm_prediction, bool, S_IRUGO | S_IWUSR | S_IWGRP);
+
+static uint32_t ref_stddev = 100;
+module_param_named(
+	ref_stddev, ref_stddev, uint, S_IRUGO | S_IWUSR | S_IWGRP
+);
+
+static uint32_t tmr_add = 100;
+module_param_named(
+	tmr_add, tmr_add, uint, S_IRUGO | S_IWUSR | S_IWGRP
+);
+
+struct lpm_history {
+	uint32_t resi[MAXSAMPLES];
+	int mode[MAXSAMPLES];
+	int nsamp;
+	uint32_t hptr;
+	uint32_t hinvalid;
+	uint32_t htmr_wkup;
+};
+
+static DEFINE_PER_CPU(struct lpm_history, hist);
+
 static DEFINE_PER_CPU(struct lpm_cluster*, cpu_cluster);
 static bool suspend_in_progress;
 static struct hrtimer lpm_hrtimer;
+static struct hrtimer histtimer;
 static struct lpm_debug *lpm_debug;
 static phys_addr_t lpm_debug_phys;
 static const int num_dbg_elements = 0x100;
@@ -327,10 +357,37 @@ static enum hrtimer_restart lpm_hrtimer_cb(struct hrtimer *h)
 	return HRTIMER_NORESTART;
 }
 
+static void histtimer_cancel(void)
+{
+	if (!lpm_prediction)
+		return;
+
+	hrtimer_try_to_cancel(&histtimer);
+}
+
+static enum hrtimer_restart histtimer_fn(struct hrtimer *h)
+{
+	int cpu = raw_smp_processor_id();
+	struct lpm_history *history = &per_cpu(hist, cpu);
+
+	history->hinvalid = 1;
+	return HRTIMER_NORESTART;
+}
+
+static void histtimer_start(uint32_t time_us)
+{
+	uint64_t time_ns = time_us * NSEC_PER_USEC;
+	ktime_t hist_ktime = ns_to_ktime(time_ns);
+
+	histtimer.function = histtimer_fn;
+	hrtimer_start(&histtimer, hist_ktime, HRTIMER_MODE_REL_PINNED);
+}
+
 static void msm_pm_set_timer(uint32_t modified_time_us)
 {
 	u64 modified_time_ns = modified_time_us * NSEC_PER_USEC;
 	ktime_t modified_ktime = ns_to_ktime(modified_time_ns);
+
 	lpm_hrtimer.function = lpm_hrtimer_cb;
 	hrtimer_start(&lpm_hrtimer, modified_ktime, HRTIMER_MODE_REL_PINNED);
 }
@@ -415,22 +472,160 @@ static int set_device_mode(struct lpm_cluster *cluster, int ndevice,
 		return -EINVAL;
 }
 
+static uint64_t lpm_cpuidle_predict(struct cpuidle_device *dev,
+		struct lpm_cpu *cpu, int *idx_restrict,
+		uint32_t *idx_restrict_time)
+{
+	int i, j, divisor;
+	uint64_t max, avg, stddev;
+	int64_t thresh = LLONG_MAX;
+	struct lpm_history *history = &per_cpu(hist, dev->cpu);
+	uint32_t *min_residency = get_per_cpu_min_residency(dev->cpu);
+
+	if (!lpm_prediction)
+		return 0;
+
+	/*
+	 * Samples are marked invalid when woken-up due to timer,
+	 * so donot predict.
+	 */
+	if (history->hinvalid) {
+		history->hinvalid = 0;
+		history->htmr_wkup = 1;
+		return 0;
+	}
+
+	/*
+	 * Predict only when all the samples are collected.
+	 */
+	if (history->nsamp < MAXSAMPLES)
+		return 0;
+
+	/*
+	 * Check if the samples are not much deviated, if so use the
+	 * average of those as predicted sleep time. Else if any
+	 * specific mode has more premature exits return the index of
+	 * that mode.
+	 */
+
+again:
+	max = avg = divisor = stddev = 0;
+	for (i = 0; i < MAXSAMPLES; i++) {
+		int64_t value = history->resi[i];
+
+		if (value <= thresh) {
+			avg += value;
+			divisor++;
+			if (value > max)
+				max = value;
+		}
+	}
+	do_div(avg, divisor);
+
+	for (i = 0; i < MAXSAMPLES; i++) {
+		int64_t value = history->resi[i];
+
+		if (value <= thresh) {
+			int64_t diff = value - avg;
+
+			stddev += diff * diff;
+		}
+	}
+	do_div(stddev, divisor);
+	stddev = int_sqrt(stddev);
+
+	/*
+	 * If the deviation is less, return the average, else
+	 * ignore one maximum sample and retry
+	 */
+	if (((avg > stddev * 6) && (divisor >= (MAXSAMPLES - 1)))
+					|| stddev <= ref_stddev) {
+		return avg;
+	} else if (divisor  > (MAXSAMPLES - 1)) {
+		thresh = max - 1;
+		goto again;
+	}
+
+	/*
+	 * Find the number of premature exits for each of the mode,
+	 * excluding clockgating mode, and they are more than fifty
+	 * percent restrict that and deeper modes.
+	 */
+	if (history->htmr_wkup != 1) {
+		for (j = 1; j < cpu->nlevels; j++) {
+			uint32_t failed = 0;
+			uint64_t total = 0;
+
+			for (i = 0; i < MAXSAMPLES; i++) {
+				if ((history->mode[i] == j) &&
+					(history->resi[i] < min_residency[j])) {
+					failed++;
+					total += history->resi[i];
+				}
+			}
+			if (failed > (MAXSAMPLES/2)) {
+				*idx_restrict = j;
+				do_div(total, failed);
+				*idx_restrict_time = total;
+				break;
+			}
+		}
+	}
+	return 0;
+}
+
+static inline void invalidate_predict_history(struct cpuidle_device *dev)
+{
+	struct lpm_history *history = &per_cpu(hist, dev->cpu);
+
+	if (!lpm_prediction)
+		return;
+
+	if (history->hinvalid) {
+		history->hinvalid = 0;
+		history->htmr_wkup = 1;
+	}
+}
+
+static void clear_predict_history(void)
+{
+	struct lpm_history *history;
+	int i;
+	unsigned int cpu;
+
+	if (!lpm_prediction)
+		return;
+
+	for_each_possible_cpu(cpu) {
+		history = &per_cpu(hist, cpu);
+		for (i = 0; i < MAXSAMPLES; i++) {
+			history->resi[i]  = 0;
+			history->mode[i] = -1;
+			history->hptr = 0;
+			history->nsamp = 0;
+		}
+	}
+}
+
+static void update_history(struct cpuidle_device *dev, int idx);
+
 static int cpu_power_select(struct cpuidle_device *dev,
 		struct lpm_cpu *cpu)
 {
 	int best_level = -1;
-	uint32_t best_level_pwr = ~0U;
 	uint32_t latency_us = pm_qos_request_for_cpu(PM_QOS_CPU_DMA_LATENCY,
 							dev->cpu);
 	uint32_t sleep_us =
 		(uint32_t)(ktime_to_us(tick_nohz_get_sleep_length()));
 	uint32_t modified_time_us = 0;
 	uint32_t next_event_us = 0;
-	uint32_t pwr;
+	int i, idx_restrict;
-	int i;
 	uint32_t lvl_latency_us = 0;
-	uint32_t lvl_overhead_us = 0;
+	uint64_t predicted = 0;
-	uint32_t lvl_overhead_energy = 0;
+	uint32_t htime = 0, idx_restrict_time = 0;
+	uint32_t next_wakeup_us = sleep_us;
+	uint32_t *min_residency = get_per_cpu_min_residency(dev->cpu);
+	uint32_t *max_residency = get_per_cpu_max_residency(dev->cpu);
 
 	if (!cpu)
 		return -EINVAL;
@@ -438,12 +633,13 @@ static int cpu_power_select(struct cpuidle_device *dev,
 	if (sleep_disabled)
 		return 0;
 
+	idx_restrict = cpu->nlevels + 1;
+
 	next_event_us = (uint32_t)(ktime_to_us(get_next_event_time(dev->cpu)));
 
 	for (i = 0; i < cpu->nlevels; i++) {
 		struct lpm_cpu_level *level = &cpu->levels[i];
 		struct power_params *pwr_params = &level->pwr;
-		uint32_t next_wakeup_us = sleep_us;
 		enum msm_pm_sleep_mode mode = level->mode;
 		bool allow;
 
@@ -454,56 +650,76 @@ static int cpu_power_select(struct cpuidle_device *dev,
 
 		lvl_latency_us = pwr_params->latency_us;
 
-		lvl_overhead_us = pwr_params->time_overhead_us;
-
-		lvl_overhead_energy = pwr_params->energy_overhead;
-
 		if (latency_us < lvl_latency_us)
-			continue;
+			break;
 
 		if (next_event_us) {
 			if (next_event_us < lvl_latency_us)
-				continue;
+				break;
 
 			if (((next_event_us - lvl_latency_us) < sleep_us) ||
 					(next_event_us < sleep_us))
 				next_wakeup_us = next_event_us - lvl_latency_us;
 		}
 
-		if (next_wakeup_us <= pwr_params->time_overhead_us)
+		if (!i) {
-			continue;
-
 			/*
-		 * If wakeup time greater than overhead by a factor of 1000
+			 * If the next_wake_us itself is not sufficient for
-		 * assume that core steady state power dominates the power
+			 * deeper low power modes than clock gating do not
-		 * equation
+			 * call prediction.
 			 */
-		if ((next_wakeup_us >> 10) > lvl_overhead_us) {
+			if (next_wakeup_us > max_residency[i]) {
-			pwr = pwr_params->ss_power;
+				predicted = lpm_cpuidle_predict(dev, cpu,
-		} else {
+					&idx_restrict, &idx_restrict_time);
-			pwr = pwr_params->ss_power;
+				if (predicted < min_residency[i])
-			pwr -= (lvl_overhead_us * pwr_params->ss_power) /
+					predicted = 0;
-						next_wakeup_us;
+			} else
-			pwr += pwr_params->energy_overhead / next_wakeup_us;
+				invalidate_predict_history(dev);
 		}
 
-		if (best_level_pwr >= pwr) {
+		if (i >= idx_restrict)
+			break;
+
 		best_level = i;
-			best_level_pwr = pwr;
+
 		if (next_event_us && next_event_us < sleep_us &&
 			(mode != MSM_PM_SLEEP_MODE_WAIT_FOR_INTERRUPT))
 			modified_time_us
 				= next_event_us - lvl_latency_us;
 		else
 			modified_time_us = 0;
-		}
+
+		if (predicted ? (predicted <= max_residency[i])
+			: (next_wakeup_us <= max_residency[i]))
+			break;
 	}
 
 	if (modified_time_us)
 		msm_pm_set_timer(modified_time_us);
 
+	/*
+	 * Start timer to avoid staying in shallower mode forever
+	 * incase of misprediciton
+	 */
+	if ((predicted || (idx_restrict != (cpu->nlevels + 1)))
+			&& ((best_level >= 0)
+			&& (best_level < (cpu->nlevels-1)))) {
+		htime = predicted + tmr_add;
+		if (htime == tmr_add)
+			htime = idx_restrict_time;
+		else if (htime > max_residency[best_level])
+			htime = max_residency[best_level];
+
+		if ((next_wakeup_us > htime) &&
+			((next_wakeup_us - htime) > max_residency[best_level]))
+			histtimer_start(htime);
+	}
+
 	trace_cpu_power_select(best_level, sleep_us, latency_us, next_event_us);
 
+	trace_cpu_pred_select(idx_restrict_time ? 2 : (predicted ? 1 : 0),
+			predicted, htime);
+
 	return best_level;
 }
 
@@ -554,8 +770,6 @@ static int cluster_select(struct lpm_cluster *cluster, bool from_idle)
 {
 	int best_level = -1;
 	int i;
-	uint32_t best_level_pwr = ~0U;
-	uint32_t pwr;
 	struct cpumask mask;
 	uint32_t latency_us = ~0U;
 	uint32_t sleep_us;
@@ -596,10 +810,10 @@ static int cluster_select(struct lpm_cluster *cluster, bool from_idle)
 			continue;
 
 		if (from_idle && latency_us < pwr_params->latency_us)
-			continue;
+			break;
 
 		if (sleep_us < pwr_params->time_overhead_us)
-			continue;
+			break;
 
 		if (suspend_in_progress && from_idle && level->notify_rpm)
 			continue;
@@ -607,19 +821,10 @@ static int cluster_select(struct lpm_cluster *cluster, bool from_idle)
 		if (level->notify_rpm && msm_rpm_waiting_for_ack())
 			continue;
 
-		if ((sleep_us >> 10) > pwr_params->time_overhead_us) {
-			pwr = pwr_params->ss_power;
-		} else {
-			pwr = pwr_params->ss_power;
-			pwr -= (pwr_params->time_overhead_us *
-					pwr_params->ss_power) / sleep_us;
-			pwr += pwr_params->energy_overhead / sleep_us;
-		}
-
-		if (best_level_pwr >= pwr) {
 		best_level = i;
-			best_level_pwr = pwr;
+
-		}
+		if (sleep_us <= pwr_params->max_residency)
+			break;
 	}
 
 	return best_level;
@@ -675,6 +880,7 @@ static int cluster_configure(struct lpm_cluster *cluster, int idx,
 		}
 
 		us = us + 1;
+		clear_predict_history();
 		do_div(us, USEC_PER_SEC/SCLK_HZ);
 		msm_mpm_enter_sleep(us, from_idle, cpumask);
 	}
@@ -1009,6 +1215,39 @@ static int lpm_cpuidle_select(struct cpuidle_driver *drv,
 	return idx;
 }
 
+static void update_history(struct cpuidle_device *dev, int idx)
+{
+	struct lpm_history *history = &per_cpu(hist, dev->cpu);
+	uint32_t tmr = 0;
+
+	if (!lpm_prediction)
+		return;
+
+	if (history->htmr_wkup) {
+		if (!history->hptr)
+			history->hptr = MAXSAMPLES-1;
+		else
+			history->hptr--;
+
+		history->resi[history->hptr] += dev->last_residency;
+		history->htmr_wkup = 0;
+		tmr = 1;
+	} else
+		history->resi[history->hptr] = dev->last_residency;
+
+	history->mode[history->hptr] = idx;
+
+	trace_cpu_pred_hist(history->mode[history->hptr],
+		history->resi[history->hptr], history->hptr, tmr);
+
+	if (history->nsamp < MAXSAMPLES)
+		history->nsamp++;
+
+	(history->hptr)++;
+	if (history->hptr >= MAXSAMPLES)
+		history->hptr = 0;
+}
+
 static int lpm_cpuidle_enter(struct cpuidle_device *dev,
 		struct cpuidle_driver *drv, int idx)
 {
@@ -1043,12 +1282,13 @@ exit:
 	cluster_unprepare(cluster, cpumask, idx, true, end_time);
 	cpu_unprepare(cluster, idx, true);
 	sched_set_cpu_cstate(smp_processor_id(), 0, 0, 0);
-
-	trace_cpu_idle_exit(idx, success);
 	end_time = ktime_to_ns(ktime_get()) - start_time;
-	dev->last_residency = do_div(end_time, 1000);
+	do_div(end_time, 1000);
+	dev->last_residency = end_time;
+	update_history(dev, idx);
+	trace_cpu_idle_exit(idx, success);
 	local_irq_enable();
-
+	histtimer_cancel();
 	return idx;
 }
 
@@ -1320,6 +1560,7 @@ static int lpm_probe(struct platform_device *pdev)
 	 */
 	suspend_set_ops(&lpm_suspend_ops);
 	hrtimer_init(&lpm_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+	hrtimer_init(&histtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 
 	ret = remote_spin_lock_init(&scm_handoff_lock, SCM_HANDOFF_LOCK_ID);
 	if (ret) {

drivers/cpuidle/lpm-levels.h

@@ -27,6 +27,9 @@ struct power_params {
 	uint32_t ss_power;		/* Steady state power */
 	uint32_t energy_overhead;	/* Enter + exit over head */
 	uint32_t time_overhead_us;	/* Enter + exit overhead */
+	uint32_t residencies[NR_LPM_LEVELS];
+	uint32_t min_residency;
+	uint32_t max_residency;
 };
 
 struct lpm_cpu_level {
@@ -55,6 +58,9 @@ struct lpm_level_avail {
 	struct kobject *kobj;
 	struct kobj_attribute idle_enabled_attr;
 	struct kobj_attribute suspend_enabled_attr;
+	void *data;
+	int idx;
+	bool cpu_node;
 };
 
 struct lpm_cluster_level {
@@ -119,7 +125,8 @@ bool lpm_cpu_mode_allow(unsigned int cpu,
 		unsigned int mode, bool from_idle);
 bool lpm_cluster_mode_allow(struct lpm_cluster *cluster,
 		unsigned int mode, bool from_idle);
-
+uint32_t *get_per_cpu_max_residency(int cpu);
+uint32_t *get_per_cpu_min_residency(int cpu);
 extern struct lpm_cluster *lpm_root_node;
 
 #ifdef CONFIG_SMP

include/trace/events/trace_msm_low_power.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2012, 2014-2015, The Linux Foundation. All rights reserved.
+/* Copyright (c) 2012, 2014-2016, 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
@@ -43,6 +43,54 @@ TRACE_EVENT(cpu_power_select,
 		__entry->next_event_us)
 );
 
+TRACE_EVENT(cpu_pred_select,
+
+	TP_PROTO(u32 predtype, u64 predicted, u32 tmr_time),
+
+	TP_ARGS(predtype, predicted, tmr_time),
+
+	TP_STRUCT__entry(
+		__field(u32, predtype)
+		__field(u64, predicted)
+		__field(u32, tmr_time)
+	),
+
+	TP_fast_assign(
+		__entry->predtype = predtype;
+		__entry->predicted = predicted;
+		__entry->tmr_time = tmr_time;
+	),
+
+	TP_printk("pred:%u time:%lu tmr_time:%u",
+		__entry->predtype, (unsigned long)__entry->predicted,
+		__entry->tmr_time)
+);
+
+TRACE_EVENT(cpu_pred_hist,
+
+	TP_PROTO(int idx, u32 resi, u32 sample, u32 tmr),
+
+	TP_ARGS(idx, resi, sample, tmr),
+
+	TP_STRUCT__entry(
+		__field(int, idx)
+		__field(u32, resi)
+		__field(u32, sample)
+		__field(u32, tmr)
+	),
+
+	TP_fast_assign(
+		__entry->idx = idx;
+		__entry->resi = resi;
+		__entry->sample = sample;
+		__entry->tmr = tmr;
+	),
+
+	TP_printk("idx:%d resi:%u sample:%u tmr:%u",
+		__entry->idx, __entry->resi,
+		__entry->sample, __entry->tmr)
+);
+
 TRACE_EVENT(cpu_idle_enter,
 
 	TP_PROTO(int index),