OpenCloudOS-Kernel/Documentation/power/devices.txt

Device Power Management

Copyright (c) 2010-2011 Rafael J. Wysocki <rjw@sisk.pl>, Novell Inc.
Copyright (c) 2010 Alan Stern <stern@rowland.harvard.edu>


Most of the code in Linux is device drivers, so most of the Linux power
management (PM) code is also driver-specific.  Most drivers will do very
little; others, especially for platforms with small batteries (like cell
phones), will do a lot.

This writeup gives an overview of how drivers interact with system-wide
power management goals, emphasizing the models and interfaces that are
shared by everything that hooks up to the driver model core.  Read it as
background for the domain-specific work you'd do with any specific driver.


Two Models for Device Power Management
======================================
Drivers will use one or both of these models to put devices into low-power
states:

    System Sleep model:
	Drivers can enter low-power states as part of entering system-wide
	low-power states like "suspend" (also known as "suspend-to-RAM"), or
	(mostly for systems with disks) "hibernation" (also known as
	"suspend-to-disk").

	This is something that device, bus, and class drivers collaborate on
	by implementing various role-specific suspend and resume methods to
	cleanly power down hardware and software subsystems, then reactivate
	them without loss of data.

	Some drivers can manage hardware wakeup events, which make the system
	leave the low-power state.  This feature may be enabled or disabled
	using the relevant /sys/devices/.../power/wakeup file (for Ethernet
	drivers the ioctl interface used by ethtool may also be used for this
	purpose); enabling it may cost some power usage, but let the whole
	system enter low-power states more often.

    Runtime Power Management model:
	Devices may also be put into low-power states while the system is
	running, independently of other power management activity in principle.
	However, devices are not generally independent of each other (for
	example, a parent device cannot be suspended unless all of its child
	devices have been suspended).  Moreover, depending on the bus type the
	device is on, it may be necessary to carry out some bus-specific
	operations on the device for this purpose.  Devices put into low power
	states at run time may require special handling during system-wide power
	transitions (suspend or hibernation).

	For these reasons not only the device driver itself, but also the
	appropriate subsystem (bus type, device type or device class) driver and
	the PM core are involved in runtime power management.  As in the system
	sleep power management case, they need to collaborate by implementing
	various role-specific suspend and resume methods, so that the hardware
	is cleanly powered down and reactivated without data or service loss.

There's not a lot to be said about those low-power states except that they are
very system-specific, and often device-specific.  Also, that if enough devices
have been put into low-power states (at runtime), the effect may be very similar
to entering some system-wide low-power state (system sleep) ... and that
synergies exist, so that several drivers using runtime PM might put the system
into a state where even deeper power saving options are available.

Most suspended devices will have quiesced all I/O: no more DMA or IRQs (except
for wakeup events), no more data read or written, and requests from upstream
drivers are no longer accepted.  A given bus or platform may have different
requirements though.

Examples of hardware wakeup events include an alarm from a real time clock,
network wake-on-LAN packets, keyboard or mouse activity, and media insertion
or removal (for PCMCIA, MMC/SD, USB, and so on).


Interfaces for Entering System Sleep States
===========================================
There are programming interfaces provided for subsystems (bus type, device type,
device class) and device drivers to allow them to participate in the power
management of devices they are concerned with.  These interfaces cover both
system sleep and runtime power management.


Device Power Management Operations
----------------------------------
Device power management operations, at the subsystem level as well as at the
device driver level, are implemented by defining and populating objects of type
struct dev_pm_ops:

struct dev_pm_ops {
	int (*prepare)(struct device *dev);
	void (*complete)(struct device *dev);
	int (*suspend)(struct device *dev);
	int (*resume)(struct device *dev);
	int (*freeze)(struct device *dev);
	int (*thaw)(struct device *dev);
	int (*poweroff)(struct device *dev);
	int (*restore)(struct device *dev);
	int (*suspend_noirq)(struct device *dev);
	int (*resume_noirq)(struct device *dev);
	int (*freeze_noirq)(struct device *dev);
	int (*thaw_noirq)(struct device *dev);
	int (*poweroff_noirq)(struct device *dev);
	int (*restore_noirq)(struct device *dev);
	int (*runtime_suspend)(struct device *dev);
	int (*runtime_resume)(struct device *dev);
	int (*runtime_idle)(struct device *dev);
};

This structure is defined in include/linux/pm.h and the methods included in it
are also described in that file.  Their roles will be explained in what follows.
For now, it should be sufficient to remember that the last three methods are
specific to runtime power management while the remaining ones are used during
system-wide power transitions.

There also is a deprecated "old" or "legacy" interface for power management
operations available at least for some subsystems.  This approach does not use
struct dev_pm_ops objects and it is suitable only for implementing system sleep
power management methods.  Therefore it is not described in this document, so
please refer directly to the source code for more information about it.


Subsystem-Level Methods
-----------------------
The core methods to suspend and resume devices reside in struct dev_pm_ops
pointed to by the pm member of struct bus_type, struct device_type and
struct class.  They are mostly of interest to the people writing infrastructure
for buses, like PCI or USB, or device type and device class drivers.

Bus drivers implement these methods as appropriate for the hardware and the
drivers using it; PCI works differently from USB, and so on.  Not many people
write subsystem-level drivers; most driver code is a "device driver" that builds
on top of bus-specific framework code.

For more information on these driver calls, see the description later;
they are called in phases for every device, respecting the parent-child
sequencing in the driver model tree.


/sys/devices/.../power/wakeup files
-----------------------------------
All devices in the driver model have two flags to control handling of wakeup
events (hardware signals that can force the device and/or system out of a low
power state).  These flags are initialized by bus or device driver code using
device_set_wakeup_capable() and device_set_wakeup_enable(), defined in
include/linux/pm_wakeup.h.

The "can_wakeup" flag just records whether the device (and its driver) can
physically support wakeup events.  The device_set_wakeup_capable() routine
affects this flag.  The "should_wakeup" flag controls whether the device should
try to use its wakeup mechanism.  device_set_wakeup_enable() affects this flag;
for the most part drivers should not change its value.  The initial value of
should_wakeup is supposed to be false for the majority of devices; the major
exceptions are power buttons, keyboards, and Ethernet adapters whose WoL
(wake-on-LAN) feature has been set up with ethtool.

Whether or not a device is capable of issuing wakeup events is a hardware
matter, and the kernel is responsible for keeping track of it.  By contrast,
whether or not a wakeup-capable device should issue wakeup events is a policy
decision, and it is managed by user space through a sysfs attribute: the
power/wakeup file.  User space can write the strings "enabled" or "disabled" to
set or clear the "should_wakeup" flag, respectively.  This file is only present
for wakeup-capable devices (i.e. devices whose "can_wakeup" flags are set)
and is created (or removed) by device_set_wakeup_capable().  Reads from the
file will return the corresponding string.

The device_may_wakeup() routine returns true only if both flags are set.
This information is used by subsystems, like the PCI bus type code, to see
whether or not to enable the devices' wakeup mechanisms.  If device wakeup
mechanisms are enabled or disabled directly by drivers, they also should use
device_may_wakeup() to decide what to do during a system sleep transition.
However for runtime power management, wakeup events should be enabled whenever
the device and driver both support them, regardless of the should_wakeup flag.


/sys/devices/.../power/control files
------------------------------------
Each device in the driver model has a flag to control whether it is subject to
runtime power management.  This flag, called runtime_auto, is initialized by the
bus type (or generally subsystem) code using pm_runtime_allow() or
pm_runtime_forbid(); the default is to allow runtime power management.

The setting can be adjusted by user space by writing either "on" or "auto" to
the device's power/control sysfs file.  Writing "auto" calls pm_runtime_allow(),
setting the flag and allowing the device to be runtime power-managed by its
driver.  Writing "on" calls pm_runtime_forbid(), clearing the flag, returning
the device to full power if it was in a low-power state, and preventing the
device from being runtime power-managed.  User space can check the current value
of the runtime_auto flag by reading the file.

The device's runtime_auto flag has no effect on the handling of system-wide
power transitions.  In particular, the device can (and in the majority of cases
should and will) be put into a low-power state during a system-wide transition
to a sleep state even though its runtime_auto flag is clear.

For more information about the runtime power management framework, refer to
Documentation/power/runtime_pm.txt.


Calling Drivers to Enter and Leave System Sleep States
======================================================
When the system goes into a sleep state, each device's driver is asked to
suspend the device by putting it into a state compatible with the target
system state.  That's usually some version of "off", but the details are
system-specific.  Also, wakeup-enabled devices will usually stay partly
functional in order to wake the system.

When the system leaves that low-power state, the device's driver is asked to
resume it by returning it to full power.  The suspend and resume operations
always go together, and both are multi-phase operations.

For simple drivers, suspend might quiesce the device using class code
and then turn its hardware as "off" as possible during suspend_noirq.  The
matching resume calls would then completely reinitialize the hardware
before reactivating its class I/O queues.

More power-aware drivers might prepare the devices for triggering system wakeup
events.


Call Sequence Guarantees
------------------------
To ensure that bridges and similar links needing to talk to a device are
available when the device is suspended or resumed, the device tree is
walked in a bottom-up order to suspend devices.  A top-down order is
used to resume those devices.

The ordering of the device tree is defined by the order in which devices
get registered:  a child can never be registered, probed or resumed before
its parent; and can't be removed or suspended after that parent.

The policy is that the device tree should match hardware bus topology.
(Or at least the control bus, for devices which use multiple busses.)
In particular, this means that a device registration may fail if the parent of
the device is suspending (i.e. has been chosen by the PM core as the next
device to suspend) or has already suspended, as well as after all of the other
devices have been suspended.  Device drivers must be prepared to cope with such
situations.


System Power Management Phases
------------------------------
Suspending or resuming the system is done in several phases.  Different phases
are used for standby or memory sleep states ("suspend-to-RAM") and the
hibernation state ("suspend-to-disk").  Each phase involves executing callbacks
for every device before the next phase begins.  Not all busses or classes
support all these callbacks and not all drivers use all the callbacks.  The
various phases always run after tasks have been frozen and before they are
unfrozen.  Furthermore, the *_noirq phases run at a time when IRQ handlers have
been disabled (except for those marked with the IRQ_WAKEUP flag).

All phases use bus, type, or class callbacks (that is, methods defined in
dev->bus->pm, dev->type->pm, or dev->class->pm).  These callbacks are mutually
exclusive, so if the device type provides a struct dev_pm_ops object pointed to
by its pm field (i.e. both dev->type and dev->type->pm are defined), the
callbacks included in that object (i.e. dev->type->pm) will be used.  Otherwise,
if the class provides a struct dev_pm_ops object pointed to by its pm field
(i.e. both dev->class and dev->class->pm are defined), the PM core will use the
callbacks from that object (i.e. dev->class->pm).  Finally, if the pm fields of
both the device type and class objects are NULL (or those objects do not exist),
the callbacks provided by the bus (that is, the callbacks from dev->bus->pm)
will be used (this allows device types to override callbacks provided by bus
types or classes if necessary).

These callbacks may in turn invoke device- or driver-specific methods stored in
dev->driver->pm, but they don't have to.


Entering System Suspend
-----------------------
When the system goes into the standby or memory sleep state, the phases are:

		prepare, suspend, suspend_noirq.

    1.	The prepare phase is meant to prevent races by preventing new devices
	from being registered; the PM core would never know that all the
	children of a device had been suspended if new children could be
	registered at will.  (By contrast, devices may be unregistered at any
	time.)  Unlike the other suspend-related phases, during the prepare
	phase the device tree is traversed top-down.

	In addition to that, if device drivers need to allocate additional
	memory to be able to hadle device suspend correctly, that should be
	done in the prepare phase.

	After the prepare callback method returns, no new children may be
	registered below the device.  The method may also prepare the device or
	driver in some way for the upcoming system power transition (for
	example, by allocating additional memory required for this purpose), but
	it should not put the device into a low-power state.

    2.	The suspend methods should quiesce the device to stop it from performing
	I/O.  They also may save the device registers and put it into the
	appropriate low-power state, depending on the bus type the device is on,
	and they may enable wakeup events.

    3.	The suspend_noirq phase occurs after IRQ handlers have been disabled,
	which means that the driver's interrupt handler will not be called while
	the callback method is running.  The methods should save the values of
	the device's registers that weren't saved previously and finally put the
	device into the appropriate low-power state.

	The majority of subsystems and device drivers need not implement this
	callback.  However, bus types allowing devices to share interrupt
	vectors, like PCI, generally need it; otherwise a driver might encounter
	an error during the suspend phase by fielding a shared interrupt
	generated by some other device after its own device had been set to low
	power.

At the end of these phases, drivers should have stopped all I/O transactions
(DMA, IRQs), saved enough state that they can re-initialize or restore previous
state (as needed by the hardware), and placed the device into a low-power state.
On many platforms they will gate off one or more clock sources; sometimes they
will also switch off power supplies or reduce voltages.  (Drivers supporting
runtime PM may already have performed some or all of these steps.)

If device_may_wakeup(dev) returns true, the device should be prepared for
generating hardware wakeup signals to trigger a system wakeup event when the
system is in the sleep state.  For example, enable_irq_wake() might identify
GPIO signals hooked up to a switch or other external hardware, and
pci_enable_wake() does something similar for the PCI PME signal.

If any of these callbacks returns an error, the system won't enter the desired
low-power state.  Instead the PM core will unwind its actions by resuming all
the devices that were suspended.


Leaving System Suspend
----------------------
When resuming from standby or memory sleep, the phases are:

		resume_noirq, resume, complete.

    1.	The resume_noirq callback methods should perform any actions needed
	before the driver's interrupt handlers are invoked.  This generally
	means undoing the actions of the suspend_noirq phase.  If the bus type
	permits devices to share interrupt vectors, like PCI, the method should
	bring the device and its driver into a state in which the driver can
	recognize if the device is the source of incoming interrupts, if any,
	and handle them correctly.

	For example, the PCI bus type's ->pm.resume_noirq() puts the device into
	the full-power state (D0 in the PCI terminology) and restores the
	standard configuration registers of the device.  Then it calls the
	device driver's ->pm.resume_noirq() method to perform device-specific
	actions.

    2.	The resume methods should bring the the device back to its operating
	state, so that it can perform normal I/O.  This generally involves
	undoing the actions of the suspend phase.

    3.	The complete phase uses only a bus callback.  The method should undo the
	actions of the prepare phase.  Note, however, that new children may be
	registered below the device as soon as the resume callbacks occur; it's
	not necessary to wait until the complete phase.

At the end of these phases, drivers should be as functional as they were before
suspending: I/O can be performed using DMA and IRQs, and the relevant clocks are
gated on.  Even if the device was in a low-power state before the system sleep
because of runtime power management, afterwards it should be back in its
full-power state.  There are multiple reasons why it's best to do this; they are
discussed in more detail in Documentation/power/runtime_pm.txt.

However, the details here may again be platform-specific.  For example,
some systems support multiple "run" states, and the mode in effect at
the end of resume might not be the one which preceded suspension.
That means availability of certain clocks or power supplies changed,
which could easily affect how a driver works.

Drivers need to be able to handle hardware which has been reset since the
suspend methods were called, for example by complete reinitialization.
This may be the hardest part, and the one most protected by NDA'd documents
and chip errata.  It's simplest if the hardware state hasn't changed since
the suspend was carried out, but that can't be guaranteed (in fact, it usually
is not the case).

Drivers must also be prepared to notice that the device has been removed
while the system was powered down, whenever that's physically possible.
PCMCIA, MMC, USB, Firewire, SCSI, and even IDE are common examples of busses
where common Linux platforms will see such removal.  Details of how drivers
will notice and handle such removals are currently bus-specific, and often
involve a separate thread.

These callbacks may return an error value, but the PM core will ignore such
errors since there's nothing it can do about them other than printing them in
the system log.


Entering Hibernation
--------------------
Hibernating the system is more complicated than putting it into the standby or
memory sleep state, because it involves creating and saving a system image.
Therefore there are more phases for hibernation, with a different set of
callbacks.  These phases always run after tasks have been frozen and memory has
been freed.

The general procedure for hibernation is to quiesce all devices (freeze), create
an image of the system memory while everything is stable, reactivate all
devices (thaw), write the image to permanent storage, and finally shut down the
system (poweroff).  The phases used to accomplish this are:

	prepare, freeze, freeze_noirq, thaw_noirq, thaw, complete,
	prepare, poweroff, poweroff_noirq

    1.	The prepare phase is discussed in the "Entering System Suspend" section
	above.

    2.	The freeze methods should quiesce the device so that it doesn't generate
	IRQs or DMA, and they may need to save the values of device registers.
	However the device does not have to be put in a low-power state, and to
	save time it's best not to do so.  Also, the device should not be
	prepared to generate wakeup events.

    3.	The freeze_noirq phase is analogous to the suspend_noirq phase discussed
	above, except again that the device should not be put in a low-power
	state and should not be allowed to generate wakeup events.

At this point the system image is created.  All devices should be inactive and
the contents of memory should remain undisturbed while this happens, so that the
image forms an atomic snapshot of the system state.

    4.	The thaw_noirq phase is analogous to the resume_noirq phase discussed
	above.  The main difference is that its methods can assume the device is
	in the same state as at the end of the freeze_noirq phase.

    5.	The thaw phase is analogous to the resume phase discussed above.  Its
	methods should bring the device back to an operating state, so that it
	can be used for saving the image if necessary.

    6.	The complete phase is discussed in the "Leaving System Suspend" section
	above.

At this point the system image is saved, and the devices then need to be
prepared for the upcoming system shutdown.  This is much like suspending them
before putting the system into the standby or memory sleep state, and the phases
are similar.

    7.	The prepare phase is discussed above.

    8.	The poweroff phase is analogous to the suspend phase.

    9.	The poweroff_noirq phase is analogous to the suspend_noirq phase.

The poweroff and poweroff_noirq callbacks should do essentially the same things
as the suspend and suspend_noirq callbacks.  The only notable difference is that
they need not store the device register values, because the registers should
already have been stored during the freeze or freeze_noirq phases.


Leaving Hibernation
-------------------
Resuming from hibernation is, again, more complicated than resuming from a sleep
state in which the contents of main memory are preserved, because it requires
a system image to be loaded into memory and the pre-hibernation memory contents
to be restored before control can be passed back to the image kernel.

Although in principle, the image might be loaded into memory and the
pre-hibernation memory contents restored by the boot loader, in practice this
can't be done because boot loaders aren't smart enough and there is no
established protocol for passing the necessary information.  So instead, the
boot loader loads a fresh instance of the kernel, called the boot kernel, into
memory and passes control to it in the usual way.  Then the boot kernel reads
the system image, restores the pre-hibernation memory contents, and passes
control to the image kernel.  Thus two different kernels are involved in
resuming from hibernation.  In fact, the boot kernel may be completely different
from the image kernel: a different configuration and even a different version.
This has important consequences for device drivers and their subsystems.

To be able to load the system image into memory, the boot kernel needs to
include at least a subset of device drivers allowing it to access the storage
medium containing the image, although it doesn't need to include all of the
drivers present in the image kernel.  After the image has been loaded, the
devices managed by the boot kernel need to be prepared for passing control back
to the image kernel.  This is very similar to the initial steps involved in
creating a system image, and it is accomplished in the same way, using prepare,
freeze, and freeze_noirq phases.  However the devices affected by these phases
are only those having drivers in the boot kernel; other devices will still be in
whatever state the boot loader left them.

Should the restoration of the pre-hibernation memory contents fail, the boot
kernel would go through the "thawing" procedure described above, using the
thaw_noirq, thaw, and complete phases, and then continue running normally.  This
happens only rarely.  Most often the pre-hibernation memory contents are
restored successfully and control is passed to the image kernel, which then
becomes responsible for bringing the system back to the working state.

To achieve this, the image kernel must restore the devices' pre-hibernation
functionality.  The operation is much like waking up from the memory sleep
state, although it involves different phases:

	restore_noirq, restore, complete

    1.	The restore_noirq phase is analogous to the resume_noirq phase.

    2.	The restore phase is analogous to the resume phase.

    3.	The complete phase is discussed above.

The main difference from resume[_noirq] is that restore[_noirq] must assume the
device has been accessed and reconfigured by the boot loader or the boot kernel.
Consequently the state of the device may be different from the state remembered
from the freeze and freeze_noirq phases.  The device may even need to be reset
and completely re-initialized.  In many cases this difference doesn't matter, so
the resume[_noirq] and restore[_norq] method pointers can be set to the same
routines.  Nevertheless, different callback pointers are used in case there is a
situation where it actually matters.


Device Power Domains
--------------------
Sometimes devices share reference clocks or other power resources.  In those
cases it generally is not possible to put devices into low-power states
individually.  Instead, a set of devices sharing a power resource can be put
into a low-power state together at the same time by turning off the shared
power resource.  Of course, they also need to be put into the full-power state
together, by turning the shared power resource on.  A set of devices with this
property is often referred to as a power domain.

Support for power domains is provided through the pwr_domain field of struct
device.  This field is a pointer to an object of type struct dev_power_domain,
defined in include/linux/pm.h, providing a set of power management callbacks
analogous to the subsystem-level and device driver callbacks that are executed
for the given device during all power transitions, in addition to the respective
subsystem-level callbacks.  Specifically, the power domain "suspend" callbacks
(i.e. ->runtime_suspend(), ->suspend(), ->freeze(), ->poweroff(), etc.) are
executed after the analogous subsystem-level callbacks, while the power domain
"resume" callbacks (i.e. ->runtime_resume(), ->resume(), ->thaw(), ->restore,
etc.) are executed before the analogous subsystem-level callbacks.  Error codes
returned by the "suspend" and "resume" power domain callbacks are ignored.

Power domain ->runtime_idle() callback is executed before the subsystem-level
->runtime_idle() callback and the result returned by it is not ignored.  Namely,
if it returns error code, the subsystem-level ->runtime_idle() callback will not
be called and the helper function rpm_idle() executing it will return error
code.  This mechanism is intended to help platforms where saving device state
is a time consuming operation and should only be carried out if all devices
in the power domain are idle, before turning off the shared power resource(s).
Namely, the power domain ->runtime_idle() callback may return error code until
the pm_runtime_idle() helper (or its asychronous version) has been called for
all devices in the power domain (it is recommended that the returned error code
be -EBUSY in those cases), preventing the subsystem-level ->runtime_idle()
callback from being run prematurely.

The support for device power domains is only relevant to platforms needing to
use the same subsystem-level (e.g. platform bus type) and device driver power
management callbacks in many different power domain configurations and wanting
to avoid incorporating the support for power domains into the subsystem-level
callbacks.  The other platforms need not implement it or take it into account
in any way.


Device Low Power (suspend) States
---------------------------------
Device low-power states aren't standard.  One device might only handle
"on" and "off, while another might support a dozen different versions of
"on" (how many engines are active?), plus a state that gets back to "on"
faster than from a full "off".

Some busses define rules about what different suspend states mean.  PCI
gives one example:  after the suspend sequence completes, a non-legacy
PCI device may not perform DMA or issue IRQs, and any wakeup events it
issues would be issued through the PME# bus signal.  Plus, there are
several PCI-standard device states, some of which are optional.

In contrast, integrated system-on-chip processors often use IRQs as the
wakeup event sources (so drivers would call enable_irq_wake) and might
be able to treat DMA completion as a wakeup event (sometimes DMA can stay
active too, it'd only be the CPU and some peripherals that sleep).

Some details here may be platform-specific.  Systems may have devices that
can be fully active in certain sleep states, such as an LCD display that's
refreshed using DMA while most of the system is sleeping lightly ... and
its frame buffer might even be updated by a DSP or other non-Linux CPU while
the Linux control processor stays idle.

Moreover, the specific actions taken may depend on the target system state.
One target system state might allow a given device to be very operational;
another might require a hard shut down with re-initialization on resume.
And two different target systems might use the same device in different
ways; the aforementioned LCD might be active in one product's "standby",
but a different product using the same SOC might work differently.


Power Management Notifiers
--------------------------
There are some operations that cannot be carried out by the power management
callbacks discussed above, because the callbacks occur too late or too early.
To handle these cases, subsystems and device drivers may register power
management notifiers that are called before tasks are frozen and after they have
been thawed.  Generally speaking, the PM notifiers are suitable for performing
actions that either require user space to be available, or at least won't
interfere with user space.

For details refer to Documentation/power/notifiers.txt.


Runtime Power Management
========================
Many devices are able to dynamically power down while the system is still
running. This feature is useful for devices that are not being used, and
can offer significant power savings on a running system.  These devices
often support a range of runtime power states, which might use names such
as "off", "sleep", "idle", "active", and so on.  Those states will in some
cases (like PCI) be partially constrained by the bus the device uses, and will
usually include hardware states that are also used in system sleep states.

A system-wide power transition can be started while some devices are in low
power states due to runtime power management.  The system sleep PM callbacks
should recognize such situations and react to them appropriately, but the
necessary actions are subsystem-specific.

In some cases the decision may be made at the subsystem level while in other
cases the device driver may be left to decide.  In some cases it may be
desirable to leave a suspended device in that state during a system-wide power
transition, but in other cases the device must be put back into the full-power
state temporarily, for example so that its system wakeup capability can be
disabled.  This all depends on the hardware and the design of the subsystem and
device driver in question.

During system-wide resume from a sleep state it's best to put devices into the
full-power state, as explained in Documentation/power/runtime_pm.txt.  Refer to
that document for more information regarding this particular issue as well as
for information on the device runtime power management framework in general.