linux/drivers/powercap/dtpm.c

661 lines
16 KiB
C
Raw Normal View History

powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
// SPDX-License-Identifier: GPL-2.0-only
/*
* Copyright 2020 Linaro Limited
*
* Author: Daniel Lezcano <daniel.lezcano@linaro.org>
*
* The powercap based Dynamic Thermal Power Management framework
* provides to the userspace a consistent API to set the power limit
* on some devices.
*
* DTPM defines the functions to create a tree of constraints. Each
* parent node is a virtual description of the aggregation of the
* children. It propagates the constraints set at its level to its
* children and collect the children power information. The leaves of
* the tree are the real devices which have the ability to get their
* current power consumption and set their power limit.
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/dtpm.h>
#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/powercap.h>
#include <linux/slab.h>
#include <linux/mutex.h>
#include <linux/of.h>
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
#include "dtpm_subsys.h"
#define DTPM_POWER_LIMIT_FLAG 0
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
static const char *constraint_name[] = {
"Instantaneous",
};
static DEFINE_MUTEX(dtpm_lock);
static struct powercap_control_type *pct;
static struct dtpm *root;
static int get_time_window_us(struct powercap_zone *pcz, int cid, u64 *window)
{
return -ENOSYS;
}
static int set_time_window_us(struct powercap_zone *pcz, int cid, u64 window)
{
return -ENOSYS;
}
static int get_max_power_range_uw(struct powercap_zone *pcz, u64 *max_power_uw)
{
struct dtpm *dtpm = to_dtpm(pcz);
mutex_lock(&dtpm_lock);
*max_power_uw = dtpm->power_max - dtpm->power_min;
mutex_unlock(&dtpm_lock);
return 0;
}
static int __get_power_uw(struct dtpm *dtpm, u64 *power_uw)
{
struct dtpm *child;
u64 power;
int ret = 0;
if (dtpm->ops) {
*power_uw = dtpm->ops->get_power_uw(dtpm);
return 0;
}
*power_uw = 0;
list_for_each_entry(child, &dtpm->children, sibling) {
ret = __get_power_uw(child, &power);
if (ret)
break;
*power_uw += power;
}
return ret;
}
static int get_power_uw(struct powercap_zone *pcz, u64 *power_uw)
{
struct dtpm *dtpm = to_dtpm(pcz);
int ret;
mutex_lock(&dtpm_lock);
ret = __get_power_uw(dtpm, power_uw);
mutex_unlock(&dtpm_lock);
return ret;
}
static void __dtpm_rebalance_weight(struct dtpm *dtpm)
{
struct dtpm *child;
list_for_each_entry(child, &dtpm->children, sibling) {
pr_debug("Setting weight '%d' for '%s'\n",
child->weight, child->zone.name);
child->weight = DIV64_U64_ROUND_CLOSEST(
child->power_max * 1024, dtpm->power_max);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
__dtpm_rebalance_weight(child);
}
}
static void __dtpm_sub_power(struct dtpm *dtpm)
{
struct dtpm *parent = dtpm->parent;
while (parent) {
parent->power_min -= dtpm->power_min;
parent->power_max -= dtpm->power_max;
parent->power_limit -= dtpm->power_limit;
parent = parent->parent;
}
}
static void __dtpm_add_power(struct dtpm *dtpm)
{
struct dtpm *parent = dtpm->parent;
while (parent) {
parent->power_min += dtpm->power_min;
parent->power_max += dtpm->power_max;
parent->power_limit += dtpm->power_limit;
parent = parent->parent;
}
}
static int __dtpm_update_power(struct dtpm *dtpm)
{
int ret;
__dtpm_sub_power(dtpm);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
ret = dtpm->ops->update_power_uw(dtpm);
if (ret)
pr_err("Failed to update power for '%s': %d\n",
dtpm->zone.name, ret);
if (!test_bit(DTPM_POWER_LIMIT_FLAG, &dtpm->flags))
dtpm->power_limit = dtpm->power_max;
__dtpm_add_power(dtpm);
if (root)
__dtpm_rebalance_weight(root);
return ret;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
}
/**
* dtpm_update_power - Update the power on the dtpm
* @dtpm: a pointer to a dtpm structure to update
*
* Function to update the power values of the dtpm node specified in
* parameter. These new values will be propagated to the tree.
*
* Return: zero on success, -EINVAL if the values are inconsistent
*/
int dtpm_update_power(struct dtpm *dtpm)
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
{
int ret;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
mutex_lock(&dtpm_lock);
ret = __dtpm_update_power(dtpm);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
mutex_unlock(&dtpm_lock);
return ret;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
}
/**
* dtpm_release_zone - Cleanup when the node is released
* @pcz: a pointer to a powercap_zone structure
*
* Do some housecleaning and update the weight on the tree. The
* release will be denied if the node has children. This function must
* be called by the specific release callback of the different
* backends.
*
* Return: 0 on success, -EBUSY if there are children
*/
int dtpm_release_zone(struct powercap_zone *pcz)
{
struct dtpm *dtpm = to_dtpm(pcz);
struct dtpm *parent = dtpm->parent;
mutex_lock(&dtpm_lock);
if (!list_empty(&dtpm->children)) {
mutex_unlock(&dtpm_lock);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
return -EBUSY;
}
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
if (parent)
list_del(&dtpm->sibling);
__dtpm_sub_power(dtpm);
mutex_unlock(&dtpm_lock);
if (dtpm->ops)
dtpm->ops->release(dtpm);
if (root == dtpm)
root = NULL;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
kfree(dtpm);
return 0;
}
static int __get_power_limit_uw(struct dtpm *dtpm, int cid, u64 *power_limit)
{
*power_limit = dtpm->power_limit;
return 0;
}
static int get_power_limit_uw(struct powercap_zone *pcz,
int cid, u64 *power_limit)
{
struct dtpm *dtpm = to_dtpm(pcz);
int ret;
mutex_lock(&dtpm_lock);
ret = __get_power_limit_uw(dtpm, cid, power_limit);
mutex_unlock(&dtpm_lock);
return ret;
}
/*
* Set the power limit on the nodes, the power limit is distributed
* given the weight of the children.
*
* The dtpm node lock must be held when calling this function.
*/
static int __set_power_limit_uw(struct dtpm *dtpm, int cid, u64 power_limit)
{
struct dtpm *child;
int ret = 0;
u64 power;
/*
* A max power limitation means we remove the power limit,
* otherwise we set a constraint and flag the dtpm node.
*/
if (power_limit == dtpm->power_max) {
clear_bit(DTPM_POWER_LIMIT_FLAG, &dtpm->flags);
} else {
set_bit(DTPM_POWER_LIMIT_FLAG, &dtpm->flags);
}
pr_debug("Setting power limit for '%s': %llu uW\n",
dtpm->zone.name, power_limit);
/*
* Only leaves of the dtpm tree has ops to get/set the power
*/
if (dtpm->ops) {
dtpm->power_limit = dtpm->ops->set_power_uw(dtpm, power_limit);
} else {
dtpm->power_limit = 0;
list_for_each_entry(child, &dtpm->children, sibling) {
/*
* Integer division rounding will inevitably
* lead to a different min or max value when
* set several times. In order to restore the
* initial value, we force the child's min or
* max power every time if the constraint is
* at the boundaries.
*/
if (power_limit == dtpm->power_max) {
power = child->power_max;
} else if (power_limit == dtpm->power_min) {
power = child->power_min;
} else {
power = DIV_ROUND_CLOSEST_ULL(
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
power_limit * child->weight, 1024);
}
pr_debug("Setting power limit for '%s': %llu uW\n",
child->zone.name, power);
ret = __set_power_limit_uw(child, cid, power);
if (!ret)
ret = __get_power_limit_uw(child, cid, &power);
if (ret)
break;
dtpm->power_limit += power;
}
}
return ret;
}
static int set_power_limit_uw(struct powercap_zone *pcz,
int cid, u64 power_limit)
{
struct dtpm *dtpm = to_dtpm(pcz);
int ret;
mutex_lock(&dtpm_lock);
/*
* Don't allow values outside of the power range previously
* set when initializing the power numbers.
*/
power_limit = clamp_val(power_limit, dtpm->power_min, dtpm->power_max);
ret = __set_power_limit_uw(dtpm, cid, power_limit);
pr_debug("%s: power limit: %llu uW, power max: %llu uW\n",
dtpm->zone.name, dtpm->power_limit, dtpm->power_max);
mutex_unlock(&dtpm_lock);
return ret;
}
static const char *get_constraint_name(struct powercap_zone *pcz, int cid)
{
return constraint_name[cid];
}
static int get_max_power_uw(struct powercap_zone *pcz, int id, u64 *max_power)
{
struct dtpm *dtpm = to_dtpm(pcz);
mutex_lock(&dtpm_lock);
*max_power = dtpm->power_max;
mutex_unlock(&dtpm_lock);
return 0;
}
static struct powercap_zone_constraint_ops constraint_ops = {
.set_power_limit_uw = set_power_limit_uw,
.get_power_limit_uw = get_power_limit_uw,
.set_time_window_us = set_time_window_us,
.get_time_window_us = get_time_window_us,
.get_max_power_uw = get_max_power_uw,
.get_name = get_constraint_name,
};
static struct powercap_zone_ops zone_ops = {
.get_max_power_range_uw = get_max_power_range_uw,
.get_power_uw = get_power_uw,
.release = dtpm_release_zone,
};
/**
* dtpm_init - Allocate and initialize a dtpm struct
* @dtpm: The dtpm struct pointer to be initialized
* @ops: The dtpm device specific ops, NULL for a virtual node
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
*/
void dtpm_init(struct dtpm *dtpm, struct dtpm_ops *ops)
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
{
if (dtpm) {
INIT_LIST_HEAD(&dtpm->children);
INIT_LIST_HEAD(&dtpm->sibling);
dtpm->weight = 1024;
dtpm->ops = ops;
}
}
/**
* dtpm_unregister - Unregister a dtpm node from the hierarchy tree
* @dtpm: a pointer to a dtpm structure corresponding to the node to be removed
*
* Call the underlying powercap unregister function. That will call
* the release callback of the powercap zone.
*/
void dtpm_unregister(struct dtpm *dtpm)
{
powercap_unregister_zone(pct, &dtpm->zone);
pr_debug("Unregistered dtpm node '%s'\n", dtpm->zone.name);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
}
/**
* dtpm_register - Register a dtpm node in the hierarchy tree
* @name: a string specifying the name of the node
* @dtpm: a pointer to a dtpm structure corresponding to the new node
* @parent: a pointer to a dtpm structure corresponding to the parent node
*
* Create a dtpm node in the tree. If no parent is specified, the node
* is the root node of the hierarchy. If the root node already exists,
* then the registration will fail. The powercap controller must be
* initialized before calling this function.
*
* The dtpm structure must be initialized with the power numbers
* before calling this function.
*
* Return: zero on success, a negative value in case of error:
* -EAGAIN: the function is called before the framework is initialized.
* -EBUSY: the root node is already inserted
* -EINVAL: * there is no root node yet and @parent is specified
* * no all ops are defined
* * parent have ops which are reserved for leaves
* Other negative values are reported back from the powercap framework
*/
int dtpm_register(const char *name, struct dtpm *dtpm, struct dtpm *parent)
{
struct powercap_zone *pcz;
if (!pct)
return -EAGAIN;
if (root && !parent)
return -EBUSY;
if (!root && parent)
return -EINVAL;
if (parent && parent->ops)
return -EINVAL;
if (!dtpm)
return -EINVAL;
if (dtpm->ops && !(dtpm->ops->set_power_uw &&
dtpm->ops->get_power_uw &&
dtpm->ops->update_power_uw &&
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
dtpm->ops->release))
return -EINVAL;
pcz = powercap_register_zone(&dtpm->zone, pct, name,
parent ? &parent->zone : NULL,
&zone_ops, MAX_DTPM_CONSTRAINTS,
&constraint_ops);
if (IS_ERR(pcz))
return PTR_ERR(pcz);
mutex_lock(&dtpm_lock);
if (parent) {
list_add_tail(&dtpm->sibling, &parent->children);
dtpm->parent = parent;
} else {
root = dtpm;
}
if (dtpm->ops && !dtpm->ops->update_power_uw(dtpm)) {
__dtpm_add_power(dtpm);
dtpm->power_limit = dtpm->power_max;
}
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
pr_debug("Registered dtpm node '%s' / %llu-%llu uW, \n",
dtpm->zone.name, dtpm->power_min, dtpm->power_max);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
mutex_unlock(&dtpm_lock);
return 0;
}
static struct dtpm *dtpm_setup_virtual(const struct dtpm_node *hierarchy,
struct dtpm *parent)
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
{
struct dtpm *dtpm;
int ret;
dtpm = kzalloc(sizeof(*dtpm), GFP_KERNEL);
if (!dtpm)
return ERR_PTR(-ENOMEM);
dtpm_init(dtpm, NULL);
ret = dtpm_register(hierarchy->name, dtpm, parent);
if (ret) {
pr_err("Failed to register dtpm node '%s': %d\n",
hierarchy->name, ret);
kfree(dtpm);
return ERR_PTR(ret);
}
return dtpm;
}
static struct dtpm *dtpm_setup_dt(const struct dtpm_node *hierarchy,
struct dtpm *parent)
{
struct device_node *np;
int i, ret;
np = of_find_node_by_path(hierarchy->name);
if (!np) {
pr_err("Failed to find '%s'\n", hierarchy->name);
return ERR_PTR(-ENXIO);
}
for (i = 0; i < ARRAY_SIZE(dtpm_subsys); i++) {
if (!dtpm_subsys[i]->setup)
continue;
ret = dtpm_subsys[i]->setup(parent, np);
if (ret) {
pr_err("Failed to setup '%s': %d\n", dtpm_subsys[i]->name, ret);
of_node_put(np);
return ERR_PTR(ret);
}
}
of_node_put(np);
/*
* By returning a NULL pointer, we let know the caller there
* is no child for us as we are a leaf of the tree
*/
return NULL;
}
typedef struct dtpm * (*dtpm_node_callback_t)(const struct dtpm_node *, struct dtpm *);
dtpm_node_callback_t dtpm_node_callback[] = {
[DTPM_NODE_VIRTUAL] = dtpm_setup_virtual,
[DTPM_NODE_DT] = dtpm_setup_dt,
};
static int dtpm_for_each_child(const struct dtpm_node *hierarchy,
const struct dtpm_node *it, struct dtpm *parent)
{
struct dtpm *dtpm;
int i, ret;
for (i = 0; hierarchy[i].name; i++) {
if (hierarchy[i].parent != it)
continue;
dtpm = dtpm_node_callback[hierarchy[i].type](&hierarchy[i], parent);
/*
* A NULL pointer means there is no children, hence we
* continue without going deeper in the recursivity.
*/
if (!dtpm)
continue;
/*
* There are multiple reasons why the callback could
* fail. The generic glue is abstracting the backend
* and therefore it is not possible to report back or
* take a decision based on the error. In any case,
* if this call fails, it is not critical in the
* hierarchy creation, we can assume the underlying
* service is not found, so we continue without this
* branch in the tree but with a warning to log the
* information the node was not created.
*/
if (IS_ERR(dtpm)) {
pr_warn("Failed to create '%s' in the hierarchy\n",
hierarchy[i].name);
continue;
}
ret = dtpm_for_each_child(hierarchy, &hierarchy[i], dtpm);
if (ret)
return ret;
}
return 0;
}
/**
* dtpm_create_hierarchy - Create the dtpm hierarchy
* @hierarchy: An array of struct dtpm_node describing the hierarchy
*
* The function is called by the platform specific code with the
* description of the different node in the hierarchy. It creates the
* tree in the sysfs filesystem under the powercap dtpm entry.
*
* The expected tree has the format:
*
* struct dtpm_node hierarchy[] = {
* [0] { .name = "topmost", type = DTPM_NODE_VIRTUAL },
* [1] { .name = "package", .type = DTPM_NODE_VIRTUAL, .parent = &hierarchy[0] },
* [2] { .name = "/cpus/cpu0", .type = DTPM_NODE_DT, .parent = &hierarchy[1] },
* [3] { .name = "/cpus/cpu1", .type = DTPM_NODE_DT, .parent = &hierarchy[1] },
* [4] { .name = "/cpus/cpu2", .type = DTPM_NODE_DT, .parent = &hierarchy[1] },
* [5] { .name = "/cpus/cpu3", .type = DTPM_NODE_DT, .parent = &hierarchy[1] },
* [6] { }
* };
*
* The last element is always an empty one and marks the end of the
* array.
*
* Return: zero on success, a negative value in case of error. Errors
* are reported back from the underlying functions.
*/
int dtpm_create_hierarchy(struct of_device_id *dtpm_match_table)
{
const struct of_device_id *match;
const struct dtpm_node *hierarchy;
struct device_node *np;
int i, ret;
if (pct)
return -EBUSY;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
pct = powercap_register_control_type(NULL, "dtpm", NULL);
if (IS_ERR(pct)) {
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
pr_err("Failed to register control type\n");
ret = PTR_ERR(pct);
goto out_pct;
}
ret = -ENODEV;
np = of_find_node_by_path("/");
if (!np)
goto out_err;
match = of_match_node(dtpm_match_table, np);
of_node_put(np);
if (!match)
goto out_err;
hierarchy = match->data;
if (!hierarchy) {
ret = -EFAULT;
goto out_err;
}
ret = dtpm_for_each_child(hierarchy, NULL, NULL);
if (ret)
goto out_err;
for (i = 0; i < ARRAY_SIZE(dtpm_subsys); i++) {
if (!dtpm_subsys[i]->init)
continue;
ret = dtpm_subsys[i]->init();
if (ret)
pr_info("Failed to initialze '%s': %d",
dtpm_subsys[i]->name, ret);
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
}
return 0;
out_err:
powercap_unregister_control_type(pct);
out_pct:
pct = NULL;
return ret;
powercap/drivers/dtpm: Add API for dynamic thermal power management On the embedded world, the complexity of the SoC leads to an increasing number of hotspots which need to be monitored and mitigated as a whole in order to prevent the temperature to go above the normative and legally stated 'skin temperature'. Another aspect is to sustain the performance for a given power budget, for example virtual reality where the user can feel dizziness if the GPU performance is capped while a big CPU is processing something else. Or reduce the battery charging because the dissipated power is too high compared with the power consumed by other devices. The userspace is the most adequate place to dynamically act on the different devices by limiting their power given an application profile: it has the knowledge of the platform. These userspace daemons are in charge of the Dynamic Thermal Power Management (DTPM). Nowadays, the dtpm daemons are abusing the thermal framework as they act on the cooling device state to force a specific and arbitrary state without taking care of the governor decisions. Given the closed loop of some governors that can confuse the logic or directly enter in a decision conflict. As the number of cooling device support is limited today to the CPU and the GPU, the dtpm daemons have little control on the power dissipation of the system. The out of tree solutions are hacking around here and there in the drivers, in the frameworks to have control on the devices. The common solution is to declare them as cooling devices. There is no unification of the power limitation unit, opaque states are used. This patch provides a way to create a hierarchy of constraints using the powercap framework. The devices which are registered as power limit-able devices are represented in this hierarchy as a tree. They are linked together with intermediate nodes which are just there to propagate the constraint to the children. The leaves of the tree are the real devices, the intermediate nodes are virtual, aggregating the children constraints and power characteristics. Each node have a weight on a 2^10 basis, in order to reflect the percentage of power distribution of the children's node. This percentage is used to dispatch the power limit to the children. The weight is computed against the max power of the siblings. This simple approach allows to do a fair distribution of the power limit. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> Tested-by: Lukasz Luba <lukasz.luba@arm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2020-12-08 16:41:44 +00:00
}
EXPORT_SYMBOL_GPL(dtpm_create_hierarchy);