Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit 7bd40ba9 authored by Mahesh Sivasubramanian's avatar Mahesh Sivasubramanian Committed by Archana Sathyakumar
Browse files

lpm-levels: Use residency instead of power and energy overhead values 



The driver looks through all the enabled modes and does energy
calculation runtime before it selects a low power mode to enter. With
residency values, the cpu/cluster can choose the low power mode as soon
as it comes across the low power mode which meets the residency
constraint.

Change-Id: I71235c212b26c3688a8ac5426942f2b88b681d47
Signed-off-by: default avatarMahesh Sivasubramanian <msivasub@codeaurora.org>
parent dc9078ea

drivers/cpuidle/lpm-levels-of.c

+148 −12
Original line number Diff line number Diff line
@@ -38,25 +38,87 @@ static const struct lpm_type_str lpm_types[] = { 
	{SUSPEND, "suspend_enabled"},

};



static DEFINE_PER_CPU(uint32_t *, max_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;



	if (!strcmp(attr->attr.name, lpm_types[IDLE].str))

		arg = (void *) &avail->idle_enabled;

	else if (!strcmp(attr->attr.name, lpm_types[SUSPEND].str))

		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)) {

	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 avail;

}



	return arg;

static void set_optimum_cpu_residency(struct lpm_cpu *cpu, int cpu_id,

		bool probe_time)

{

	int i, j;

	bool mode_avail;

	uint32_t *residency = per_cpu(max_residency, cpu_id);



	for (i = 0; i < cpu->nlevels; i++) {

		struct power_params *pwr = &cpu->levels[i].pwr;



		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 &&

				(residency[i] > pwr->residencies[j]) &&

				(pwr->residencies[j] != 0))

				residency[i] = pwr->residencies[j];

		}

	}

}



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;



		pwr->max_residency = ~0;

		for (j = 0; j < cluster->nlevels; j++) {

			if (i >= j)

				mode_avail = probe_time ||

					lpm_cluster_mode_allow(cluster, i,

							true);

			if (mode_avail &&

				(pwr->max_residency > pwr->residencies[j]) &&

				(pwr->residencies[j] != 0))

				pwr->max_residency = pwr->residencies[j];

		}

	}

}



uint32_t *get_per_cpu_max_residency(int cpu)

{

	return per_cpu(max_residency, cpu);

}



ssize_t lpm_enable_show(struct kobject *kobj, struct kobj_attribute *attr,
@@ -65,7 +127,7 @@ ssize_t lpm_enable_show(struct kobject *kobj, struct kobj_attribute *attr, 
	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 +142,23 @@ 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);

	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 +209,9 @@ 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 +254,8 @@ 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 +289,8 @@ 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 +496,9 @@ 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 +693,31 @@ 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 +774,22 @@ 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 +846,7 @@ 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 +904,16 @@ 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);

				set_optimum_cpu_residency(c->cpu, i, true);

			}

		}

	}
@@ -797,6 +922,17 @@ 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

+6 −41
Original line number Diff line number Diff line
@@ -433,18 +433,15 @@ 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;

	uint32_t lvl_latency_us = 0;

	uint32_t lvl_overhead_us = 0;

	uint32_t lvl_overhead_energy = 0;

	uint32_t *residency = get_per_cpu_max_residency(dev->cpu);



	if (!cpu)

		return -EINVAL;
@@ -468,12 +465,8 @@ 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)
@@ -484,32 +477,15 @@ static int cpu_power_select(struct cpuidle_device *dev, 
				next_wakeup_us = next_event_us - lvl_latency_us;

		}



		if (next_wakeup_us <= pwr_params->time_overhead_us)

			continue;



		/*

		 * If wakeup time greater than overhead by a factor of 1000

		 * assume that core steady state power dominates the power

		 * equation

		 */

		if ((next_wakeup_us >> 10) > lvl_overhead_us) {

			pwr = pwr_params->ss_power;

		} else {

			pwr = pwr_params->ss_power;

			pwr -= (lvl_overhead_us * pwr_params->ss_power) /

						next_wakeup_us;

			pwr += pwr_params->energy_overhead / next_wakeup_us;

		}



		if (best_level_pwr >= pwr) {

		if (next_wakeup_us <= residency[i]) {

			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;

			break;

		}

	}
@@ -567,8 +543,6 @@ 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;
@@ -620,18 +594,9 @@ 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) {

		if (sleep_us <= pwr_params->max_residency) {

			best_level = i;

			best_level_pwr = pwr;

			break;

		}

	}

drivers/cpuidle/lpm-levels.h

+6 −1
Original line number Diff line number Diff line
@@ -27,6 +27,8 @@ 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 max_residency;

};



struct lpm_cpu_level {
@@ -55,6 +57,9 @@ 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 +124,7 @@ 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);

extern struct lpm_cluster *lpm_root_node;



#ifdef CONFIG_SMP