Power Management

• Power management for embedded systems
• Application-driven framework
  • Framework components
  • Libraries and BSPs
  • Power manager
  • Power-managed objects
  • Power-aware applications
  • Persistent storage services
  • CPU power management

Power management for embedded systems

Traditional power management (PM) is aimed at conserving the power of computers that are usually left on. The general-purpose approach to PM for desktop PCs -- or even for "mobile" PCs, such as laptops -- doesn't take into account the specific demands of embedded systems, which can be off (or on standby) much of the time, yet must respond to external events in predictable ways.

The two main industry standards for PM -- APM and its successor ACPI -- deal with PCs, not embedded systems. The older APM standard was BIOS-oriented (how many embedded systems even have a BIOS?). The newer ACPI is more OS-focused and therefore more sophisticated, but the standard expressly concerns itself with desktops, laptops, and servers. An OS can never anticipate the application-specific scenarios of embedded systems and make appropriate power policy decisions.

Application-driven framework

Given the sheer variety of embedded systems (e.g hand-held medical instruments, network routers, in-car telematics systems, to name but a few), the application developer is in the best position to know the specific power-related needs of the system in question.

The QNX PM framework is precisely that -- a framework or set of mechanisms that developers can use to control and manage power in their embedded systems. The OS doesn't impose a power policy. Instead, you create your own policy based on the specific needs of your application.

Framework components

The PM framework consists of these main components:

Libraries and BSPs

These components are built using the following:

libpm

Client library that applications and device drivers use to interact with the power manager.

libps

Persistent storage library used to manipulate data managed by the persistent storage services.

libpmm

Server library used to build the power manager. This library implements the underlying protocol for communicating with device drivers and client applications.

BSPs

These provide reference implementations for the CPU power-mode support for reference hardware platforms (e.g. the Renesas Biscayne SH7760). You can customize a reference implementation for your particular embedded system.

The QNX PM framework facilitates application-specific designs.

Power manager

As the central component within the PM framework, the power manager is responsible for coordinating the power modes of each power-managed object according the system's PM policy.

The power manager is implemented as a user-written process that uses the libpmm library to manage the interactions with device drivers and power-aware applications.

The power manager consists of three main components:

• A resource manager interface used to represent power-managed objects. Device drivers and power-aware applications use this interface to manipulate the objects controlled by the power manager.

  The functionality for handling this communication is implemented entirely within the libpmm library.

• A policy that governs the behavior of power-managed objects in response to requests made via the resource manager interface.

  The policy is customizable. Your policy is informed whenever an operation is performed on an object. This allows your policy code to track the state of an object, and where appropriate, to allow or deny certain operations. For example, your policy may deny requests to change the power mode of an object if the new mode isn't compatible with the system's current power-mode state.

  The libpm library provides basic default behavior for these operations, so the system-specific policy needs to override only those operations that are required to implement the system's PM policy.

• Code to manage system power states. This user-written code defines the range of operational modes for the system along with their corresponding device and CPU power modes. This code is responsible for system initialization and actions that need to be performed based on system-specific events that are relevant to deciding the system's optimum power-mode state.

Resource manager interface

The most basic power manager simply initializes the resource manager interface to handle driver and client requests, and then starts a thread pool to service them:

#include <sys/procmgr.h>
#include <sys/pmm.h>

int
main()
{
    // initialise and start resource manager interface
    pmm_init(0, 0);
    pmm_start(0, 0);

    // become a daemon
    procmgr_daemon(0, 0);
    pthread_detach(pthread_self());
    pthread_exit(0);
    return 0;
}

This is sufficient to completely handle dynamic attachment by any drivers that start after the power manager; it implements a simple default policy that allows all power-mode changes to proceed as requested.

Power manager policy

The libpmm library handles the low-level details of interacting with device drivers and power-aware applications. However, the decisions about what actions are allowed to proceed are policy decisions that developers customize for their embedded system.

The library invokes this policy code whenever requests arrive that affect a power-managed object's state. If you don't supply you own policy function, an internal default is used.

The policy is informed (by invoking policy callback functions) when:

• An object is first created, and when it's deleted. This allows the policy to allocate its own specific data structures to be associated with the object.
• A client process attaches and detaches from an object. The policy can thus identify the client process and allow or reject access to the object. Since objects are accessed via a resource manager interface, basic authentication/access control is performed using the standard file access permissions of the object within the power manager namespace. Customizing this callback can allow additional authentication mechanisms.
• An object is assigned a name in the namespace. This allows the policy to determine how this object is related to other objects in the namespace, for example to manage power dependencies. As with the client attachment, basic permission checking is provided via file access permissions.
• An object is removed from the namespace. This allows the policy to remove any inter-object dependencies.
• A driver asks to manage an object. The policy can thus discover the supported device modes and can specify what initial mode the device should be placed in. For warm reboots from system standby modes, the policy can also instruct the driver to recover any data it saved in persistent storage.
• A driver detaches. This informs the policy that the object can no longer be power-managed.
• A client (or the driver itself) requests a power-mode change to an object. The policy can validate that the requested mode is compatible with the current system power state and allow or reject the request.
• A property is added to an object, or a property value is changed. This allows the policy to react to property changes that may affect the system's power-mode state.

System power states

An embedded system may pass through a number of different power states. Each state defines the functionality and allowed power modes for the various power-managed objects in the system.

For example, an automotive device may define the following system states:

These system states can typically be modeled using a state machine, where system-specific events are used to cause transitions between states:

• Externally generated events can be implemented using properties (e.g. bus control messages associated with a vehicle bus, or discrete events such as key/switch status).
• Wakeup events are provided by the BSP-level IPL/startup code.

The policy callbacks interact with the state machine implementation to limit the allowed operations for the current system state, and to keep track of the status of each power-managed object so that they can be appropriately managed during state transitions.

Power-managed objects

Each entity that can be independently power-managed is represented by the power manager as a power-managed object.

The power manager's system PM policy determines the allowable power mode for each object, based on the system's current operational state. For example:

• System standby states require all power-managed objects to be placed in an appropriate standby mode.
• Some system states may operate with reduced power, providing a limited subset of the full system functionality. In this case, some devices are allowed to be powered up, while others must be powered down.

Power-managed objects are typically used to represent power-managed peripheral devices, allowing the power manager to control the device's power level. However, power-managed objects can also represent specialized applications. In this case, their operational characteristics are exposed to the power manager via its power mode; the power manager can control the application's behavior by changing its mode when the power manager changes the system's power-mode state.

Power manager namespace

The power manager presents all power-managed objects via a hierarchical namespace under /dev/pmm. This namespace provides a simple configuration database that can represent both physical and logical relationships between objects:

• The hierarchy can follow the hardware topology between buses and devices. Nonleaf nodes represent buses, whose children are devices or bridges to other buses. Leaf nodes represent individual devices.
• Logical relationships can be created where nonleaf nodes can represent system services or subsystems whose children are the power-managed objects managed by that subsystem.

Device drivers for power-managed devices register with the power manager by attaching to a node within the namespace or by creating a new node in the appropriate location. Once attached, the driver is responsible for responding to requests from the power manager to manage the device's power mode.

The framework allows for both statically and dynamically configured systems:

• A statically configured system contains a fixed hardware configuration, where all devices that are present are known to the power manager. In this case, the power manager can "pre-populate" the namespace with entries for each device, then start all device drivers during system initialization.
• A dynamically configured system contains devices that can be dynamically added to the system (e.g. devices on a USB bus or PCMCIA controller). In this case, drivers for these devices are often started after the power manager has initialized the system. When the drivers start up, they create nodes in the namespace to represent the new device.

How and when a driver is started makes no difference to the power manager; as soon as a driver registers with the power manager, its device will be managed according to the system's PM policy.

Power modes

The framework defines four generic power modes that are universally applied to all power-managed objects. These primarily define the operational state of the object, and by implication, its power level:

Note that all power-managed objects support the four operating modes, even if there's no specific IDLE or STANDBY mode.

If there's no IDLE mode, the object implements the IDLE mode using the same operational characteristics and hardware power level as its ACTIVE mode instead.

If there's no specific STANDBY, the object implements the STANDBY mode using the same physical power level as its OFF mode. In this case, the only difference is that entry into STANDBY may require the object to save context needed to reinitialize to an ACTIVE state.

These generic power modes provide a uniform way for the power manager (or power-aware applications) to control and reason about the power levels and functionality of a power-managed object, without requiring any device-specific knowledge.

The framework allows devices to implement an arbitrary number of additional device-specific modes that are sublevels of these four generic modes. For example, a driver may define:

• active modes that differ in their performance and power consumption
• idle modes that differ in their power consumption and the latency to return to an active mode
• standby modes that provide system wakeup functionality.

These device-specific modes allow finer-grained control over a device's power mode and operational characteristics where appropriate.

Properties

Each power-managed object may have arbitrary properties associated with it. Each property consists of an identifier,value pair:

identifier

An integer value used to the define the property type.

value

A data structure whose size and format is defined by the identifier.

The power manager's policy code is told whenever a property value is changed. This allows the policy code to respond to changes that are relevant to evaluating the optimum system power state.

The framework allows user-defined properties to be associated with any object, so you can define arbitrary PM-related properties or attributes that can be manipulated by the power manager or power-aware applications. For example, you can have properties defined that describe battery power levels, vehicle bus status, or other environmental conditions relevant to the embedded system.

Power-aware applications

Power-aware applications communicate with the power manager using an API implemented by the libpm library. The API allows the application to manipulate power-managed objects by:

• querying an object's power-mode status
• modifying an object's power-mode status (if allowed by the policy)
• receiving notification (via a SIGEVENT) when the power-mode status changes
• querying or modifying properties associated with an object.

Power-aware applications fall into two main categories:

• system-monitoring applications -- these monitor aspects of the system that may be relevant to the PM policy (e.g. control messages on a vehicle bus, discrete events such as battery-level indications, ignition key insertion/removal, transducer readings, etc.).

  This information can be represented using power manager properties. The monitoring application can update the property value to inform the power manager's PM policy. The policy can then use this information to evaluate the appropriate system power state.

• power-sensitive applications -- these register with the power manager for notification when the state of a power-managed object changes. The application can then change its behavior as required (e.g. by enabling or disabling features based on available devices).

  Power-sensitive applications can also request to change the power mode of an object, which can be used by the power manager's PM policy to intelligently decide which devices are eligible to be powered down at any given time.

Persistent storage services

Most CPUs support a low-power mode, where the CPU and peripherals are powered down with DRAM maintained in self-refresh mode. But for embedded systems that spend most of their time in a standby state for prolonged periods of time, this is unacceptable, because the power used for self-refresh over prolonged periods will exceed the reserve capacity provided by limited battery sources.

For these systems, the typical standby state will disable DRAM, so that only the limited hardware necessary to wake the system remains powered. When the system is woken up, a warm reboot is performed; the power manager then needs to reinitialize the system according to the power-mode state it decides the system should be in. This may require drivers or power-aware applications to save volatile configuration information in persistent storage so that this information can be recovered when they are restarted.

Services for persistent storage are provided by a persistent storage manager that controls one or more persistent stores. These persistent stores can make use of a number of different media:

⁺

For example, SRAM or a limited set of SDRAM banks that remain in self-refresh mode.

The persistent storage manager can manage multiple storage media, with one of those selected as the currently "active store." The underlying storage is hidden from clients, who simply request to save or restore their data.

Your power manager controls the choice of media used. Given your custom set of available stores for your embedded system, your power manager policy sets the "active store" based on the nature of the standby state it is entering before powering down the system.

CPU power management

Many CPUs offer a variety of available operating modes that provide different levels of power consumption:

• Frequency or voltage scaling can reduce the power consumption of the CPU while maintaining normal instruction execution.

  The system as a whole is fully operational; these different modes simply provide a tradeoff between performance and power consumption.

• The CPU can be halted, with RAM and peripherals remaining powered. The CPU will be powered up when an interrupt occurs; the latency to return to normal operation is very short, typically only a few CPU cycles.

  The system as a whole is fully operational; this mode simply reduces the power consumption due to CPU activity. However, the device will still draw power for peripherals and RAM refresh.

  This mode is generally used by the kernel's idle loop to reduce power consumption when there are no runnable threads, without impacting the system's realtime performance in responding to peripheral interrupts.

• The CPU can be placed in a low-power sleep or suspend mode where RAM is placed into self-refresh mode. The CPU will be powered up by some configured wakeup source (e.g. an RTC, peripheral interrupt, or GPIO line).

  For some system-on-chip processors, this mode may also power down on-chip peripherals, further reducing the overall power consumption.

  The latency to return to normal operation is relatively long because the memory controller and on-chip peripherals may need to be reinitialized. However, the system state remains preserved in RAM -- this can be a way to significantly reduce system power consumption while still providing a relatively low-latency to return to a fully active system.

• The CPU can be placed in a sleep or suspend mode where RAM is powered down, along with all on-chip peripherals. The CPU will be powered up by an appropriately configured wakeup source (e.g. an RTC or GPIO line).

  Returning to normal operation is essentially a complete system reboot, because all system state information preserved in RAM is lost. However, this provides the lowest power consumption of all, requiring power for only those components required to wake up the CPU.

The power manager's PM policy determines which CPU modes are appropriate for the various system states required for your device. Remember that this policy is user-written -- as the embedded system designer, you are in the best position to decide on the optimum balance between power consumption and performance or latency to return to an operational mode.

These CPU modes are highly processor-specific. Also, the nature of the wakeup sources depends on the system hardware implementation. For this reason, the implementation is provided by the power() callout within the BSP-specific startup program, and may also require support from the BSP-specific IPL (if the CPU wakeup causes a warm reboot via the reset vector).

Reference implementations provide the basic CPU-level support required for the modes supported by a particular CPU. You may then customize this as required to handle board-specific wakeup configuration, if necessary.