PCI Power Management

Copyright (c) 2010 Rafael J. Wysocki <rjw@sisk.pl>, Novell Inc.

An overview of concepts and the Linux kernel's interfaces related to PCI power
management.  Based on previous work by Patrick Mochel <mochel@transmeta.com>
(and others).

This document only covers the aspects of power management specific to PCI
devices.  For general description of the kernel's interfaces related to device
power management refer to Documentation/power/admin-guide/devices.rst and
Documentation/power/runtime_pm.txt.

---------------------------------------------------------------------------

1. Hardware and Platform Support for PCI Power Management
2. PCI Subsystem and Device Power Management
3. PCI Device Drivers and Power Management
4. Resources


1. Hardware and Platform Support for PCI Power Management
=========================================================

1.1. Native and Platform-Based Power Management
-----------------------------------------------
In general, power management is a feature allowing one to save energy by putting
devices into states in which they draw less power (low-power states) at the
price of reduced functionality or performance.

Usually, a device is put into a low-power state when it is underutilized or
completely inactive.  However, when it is necessary to use the device once
again, it has to be put back into the "fully functional" state (full-power
state).  This may happen when there are some data for the device to handle or
as a result of an external event requiring the device to be active, which may
be signaled by the device itself.

PCI devices may be put into low-power states in two ways, by using the device
capabilities introduced by the PCI Bus Power Management Interface Specification,
or with the help of platform firmware, such as an ACPI BIOS.  In the first
approach, that is referred to as the native PCI power management (native PCI PM)
in what follows, the device power state is changed as a result of writing a
specific value into one of its standard configuration registers.  The second
approach requires the platform firmware to provide special methods that may be
used by the kernel to change the device's power state.

Devices supporting the native PCI PM usually can generate wakeup signals called
Power Management Events (PMEs) to let the kernel know about external events
requiring the device to be active.  After receiving a PME the kernel is supposed
to put the device that sent it into the full-power state.  However, the PCI Bus
Power Management Interface Specification doesn't define any standard method of
delivering the PME from the device to the CPU and the operating system kernel.
It is assumed that the platform firmware will perform this task and therefore,
even though a PCI device is set up to generate PMEs, it also may be necessary to
prepare the platform firmware for notifying the CPU of the PMEs coming from the
device (e.g. by generating interrupts).

In turn, if the methods provided by the platform firmware are used for changing
the power state of a device, usually the platform also provides a method for
preparing the device to generate wakeup signals.  In that case, however, it
often also is necessary to prepare the device for generating PMEs using the
native PCI PM mechanism, because the method provided by the platform depends on
that.

Thus in many situations both the native and the platform-based power management
mechanisms have to be used simultaneously to obtain the desired result.

1.2. Native PCI Power Management
--------------------------------
The PCI Bus Power Management Interface Specification (PCI PM Spec) was
introduced between the PCI 2.1 and PCI 2.2 Specifications.  It defined a
standard interface for performing various operations related to power
management.

The implementation of the PCI PM Spec is optional for conventional PCI devices,
but it is mandatory for PCI Express devices.  If a device supports the PCI PM
Spec, it has an 8 byte power management capability field in its PCI
configuration space.  This field is used to describe and control the standard
features related to the native PCI power management.

The PCI PM Spec defines 4 operating states for devices (D0-D3) and for buses
(B0-B3).  The higher the number, the less power is drawn by the device or bus
in that state.  However, the higher the number, the longer the latency for
the device or bus to return to the full-power state (D0 or B0, respectively).

There are two variants of the D3 state defined by the specification.  The first
one is D3hot, referred to as the software accessible D3, because devices can be
programmed to go into it.  The second one, D3cold, is the state that PCI devices
are in when the supply voltage (Vcc) is removed from them.  It is not possible
to program a PCI device to go into D3cold, although there may be a programmable
interface for putting the bus the device is on into a state in which Vcc is
removed from all devices on the bus.

PCI bus power management, however, is not supported by the Linux kernel at the
time of this writing and therefore it is not covered by this document.

Note that every PCI device can be in the full-power state (D0) or in D3cold,
regardless of whether or not it implements the PCI PM Spec.  In addition to
that, if the PCI PM Spec is implemented by the device, it must support D3hot
as well as D0.  The support for the D1 and D2 power states is optional.

PCI devices supporting the PCI PM Spec can be programmed to go to any of the
supported low-power states (except for D3cold).  While in D1-D3hot the
standard configuration registers of the device must be accessible to software
(i.e. the device is required to respond to PCI configuration accesses), although
its I/O and memory spaces are then disabled.  This allows the device to be
programmatically put into D0.  Thus the kernel can switch the device back and
forth between D0 and the supported low-power states (except for D3cold) and the
possible power state transitions the device can undergo are the following:

+----------------------------+
| Current State | New State  |
+----------------------------+
| D0            | D1, D2, D3 |
+----------------------------+
| D1            | D2, D3     |
+----------------------------+
| D2            | D3         |
+----------------------------+
| D1, D2, D3    | D0         |
+----------------------------+

The transition from D3cold to D0 occurs when the supply voltage is provided to
the device (i.e. power is restored).  In that case the device returns to D0 with
a full power-on reset sequence and the power-on defaults are restored to the
device by hardware just as at initial power up.

PCI devices supporting the PCI PM Spec can be programmed to generate PMEs
while in a low-power state (D1-D3), but they are not required to be capable
of generating PMEs from all supported low-power states.  In particular, the
capability of generating PMEs from D3cold is optional and depends on the
presence of additional voltage (3.3Vaux) allowing the device to remain
sufficiently active to generate a wakeup signal.

1.3. ACPI Device Power Management
---------------------------------
The platform firmware support for the power management of PCI devices is
system-specific.  However, if the system in question is compliant with the
Advanced Configuration and Power Interface (ACPI) Specification, like the
majority of x86-based systems, it is supposed to implement device power
management interfaces defined by the ACPI standard.

For this purpose the ACPI BIOS provides special functions called "control
methods" that may be executed by the kernel to perform specific tasks, such as
putting a device into a low-power state.  These control methods are encoded
using special byte-code language called the ACPI Machine Language (AML) and
stored in the machine's BIOS.  The kernel loads them from the BIOS and executes
them as needed using an AML interpreter that translates the AML byte code into
computations and memory or I/O space accesses.  This way, in theory, a BIOS
writer can provide the kernel with a means to perform actions depending
on the system design in a system-specific fashion.

ACPI control methods may be divided into global control methods, that are not
associated with any particular devices, and device control methods, that have
to be defined separately for each device supposed to be handled with the help of
the platform.  This means, in particular, that ACPI device control methods can
only be used to handle devices that the BIOS writer knew about in advance.  The
ACPI methods used for device power management fall into that category.

The ACPI specification assumes that devices can be in one of four power states
labeled as D0, D1, D2, and D3 that roughly correspond to the native PCI PM
D0-D3 states (although the difference between D3hot and D3cold is not taken
into account by ACPI).  Moreover, for each power state of a device there is a
set of power resources that have to be enabled for the device to be put into
that state.  These power resources are controlled (i.e. enabled or disabled)
with the help of their own control methods, _ON and _OFF, that have to be
defined individually for each of them.

To put a device into the ACPI power state Dx (where x is a number between 0 and
3 inclusive) the kernel is supposed to (1) enable the power resources required
by the device in this state using their _ON control methods and (2) execute the
_PSx control method defined for the device.  In addition to that, if the device
is going to be put into a low-power state (D1-D3) and is supposed to generate
wakeup signals from that state, the _DSW (or _PSW, replaced with _DSW by ACPI
3.0) control method defined for it has to be executed before _PSx.  Power
resources that are not required by the device in the target power state and are
not required any more by any other device should be disabled (by executing their
_OFF control methods).  If the current power state of the device is D3, it can
only be put into D0 this way.

However, quite often the power states of devices are changed during a
system-wide transition into a sleep state or back into the working state.  ACPI
defines four system sleep states, S1, S2, S3, and S4, and denotes the system
working state as S0.  In general, the target system sleep (or working) state
determines the highest power (lowest number) state the device can be put
into and the kernel is supposed to obtain this information by executing the
device's _SxD control method (where x is a number between 0 and 4 inclusive).
If the device is required to wake up the system from the target sleep state, the
lowest power (highest number) state it can be put into is also determined by the
target state of the system.  The kernel is then supposed to use the device's
_SxW control method to obtain the number of that state.  It also is supposed to
use the device's _PRW control method to learn which power resources need to be
enabled for the device to be able to generate wakeup signals.

1.4. Wakeup Signaling
---------------------
Wakeup signals generated by PCI devices, either as native PCI PMEs, or as
a result of the execution of the _DSW (or _PSW) ACPI control method before
putting the device into a low-power state, have to be caught and handled as
appropriate.  If they are sent while the system is in the working state
(ACPI S0), they should be translated into interrupts so that the kernel can
put the devices generating them into the full-power state and take care of the
events that triggered them.  In turn, if they are sent while the system is
sleeping, they should cause the system's core logic to trigger wakeup.

On ACPI-based systems wakeup signals sent by conventional PCI devices are
converted into ACPI General-Purpose Events (GPEs) which are hardware signals
from the system core logic generated in response to various events that need to
be acted upon.  Every GPE is associated with one or more sources of potentially
interesting events.  In particular, a GPE may be associated with a PCI device
capable of signaling wakeup.  The information on the connections between GPEs
and event sources is recorded in the system's ACPI BIOS from where it can be
read by the kernel.

If a PCI device known to the system's ACPI BIOS signals wakeup, the GPE
associated with it (if there is one) is triggered.  The GPEs associated with PCI
bridges may also be triggered in response to a wakeup signal from one of the
devices below the bridge (this also is the case for root bridges) and, for
example, native PCI PMEs from devices unknown to the system's ACPI BIOS may be
handled this way.

A GPE may be triggered when the system is sleeping (i.e. when it is in one of
the ACPI S1-S4 states), in which case system wakeup is started by its core logic
(the device that was the source of the signal causing the system wakeup to occur
may be identified later).  The GPEs used in such situations are referred to as
wakeup GPEs.

Usually, however, GPEs are also triggered when the system is in the working
state (ACPI S0) and in that case the system's core logic generates a System
Control Interrupt (SCI) to notify the kernel of the event.  Then, the SCI
handler identifies the GPE that caused the interrupt to be generated which,
in turn, allows the kernel to identify the source of the event (that may be
a PCI device signaling wakeup).  The GPEs used for notifying the kernel of
events occurring while the system is in the working state are referred to as
runtime GPEs.

Unfortunately, there is no standard way of handling wakeup signals sent by
conventional PCI devices on systems that are not ACPI-based, but there is one
for PCI Express devices.  Namely, the PCI Express Base Specification introduced
a native mechanism for converting native PCI PMEs into interrupts generated by
root ports.  For conventional PCI devices native PMEs are out-of-band, so they
are routed separately and they need not pass through bridges (in principle they
may be routed directly to the system's core logic), but for PCI Express devices
they are in-band messages that have to pass through the PCI Express hierarchy,
including the root port on the path from the device to the Root Complex.  Thus
it was possible to introduce a mechanism by which a root port generates an
interrupt whenever it receives a PME message from one of the devices below it.
The PCI Express Requester ID of the device that sent the PME message is then
recorded in one of the root port's configuration registers from where it may be
read by the interrupt handler allowing the device to be identified.  [PME
messages sent by PCI Express endpoints integrated with the Root Complex don't
pass through root ports, but instead they cause a Root Complex Event Collector
(if there is one) to generate interrupts.]

In principle the native PCI Express PME signaling may also be used on ACPI-based
systems along with the GPEs, but to use it the kernel has to ask the system's
ACPI BIOS to release control of root port configuration registers.  The ACPI
BIOS, however, is not required to allow the kernel to control these registers
and if it doesn't do that, the kernel must not modify their contents.  Of course
the native PCI Express PME signaling cannot be used by the kernel in that case.


2. PCI Subsystem and Device Power Management
============================================

2.1. Device Power Management Callbacks
--------------------------------------
The PCI Subsystem participates in the power management of PCI devices in a
number of ways.  First of all, it provides an intermediate code layer between
the device power management core (PM core) and PCI device drivers.
Specifically, the pm field of the PCI subsystem's struct bus_type object,
pci_bus_type, points to a struct dev_pm_ops object, pci_dev_pm_ops, containing
pointers to several device power management callbacks:

const struct dev_pm_ops pci_dev_pm_ops = {
	.prepare = pci_pm_prepare,
	.complete = pci_pm_complete,
	.suspend = pci_pm_suspend,
	.resume = pci_pm_resume,
	.freeze = pci_pm_freeze,
	.thaw = pci_pm_thaw,
	.poweroff = pci_pm_poweroff,
	.restore = pci_pm_restore,
	.suspend_noirq = pci_pm_suspend_noirq,
	.resume_noirq = pci_pm_resume_noirq,
	.freeze_noirq = pci_pm_freeze_noirq,
	.thaw_noirq = pci_pm_thaw_noirq,
	.poweroff_noirq = pci_pm_poweroff_noirq,
	.restore_noirq = pci_pm_restore_noirq,
	.runtime_suspend = pci_pm_runtime_suspend,
	.runtime_resume = pci_pm_runtime_resume,
	.runtime_idle = pci_pm_runtime_idle,
};

These callbacks are executed by the PM core in various situations related to
device power management and they, in turn, execute power management callbacks
provided by PCI device drivers.  They also perform power management operations
involving some standard configuration registers of PCI devices that device
drivers need not know or care about.

The structure representing a PCI device, struct pci_dev, contains several fields
that these callbacks operate on:

struct pci_dev {
	...
	pci_power_t     current_state;  /* Current operating state. */
	int		pm_cap;		/* PM capability offset in the
					   configuration space */
	unsigned int	pme_support:5;	/* Bitmask of states from which PME#
					   can be generated */
	unsigned int	pme_interrupt:1;/* Is native PCIe PME signaling used? */
	unsigned int	d1_support:1;	/* Low power state D1 is supported */
	unsigned int	d2_support:1;	/* Low power state D2 is supported */
	unsigned int	no_d1d2:1;	/* D1 and D2 are forbidden */
	unsigned int	wakeup_prepared:1;  /* Device prepared for wake up */
	unsigned int	d3_delay;	/* D3->D0 transition time in ms */
	...
};

They also indirectly use some fields of the struct device that is embedded in
struct pci_dev.

2.2. Device Initialization
--------------------------
The PCI subsystem's first task related to device power management is to
prepare the device for power management and initialize the fields of struct
pci_dev used for this purpose.  This happens in two functions defined in
drivers/pci/pci.c, pci_pm_init() and platform_pci_wakeup_init().

The first of these functions checks if the device supports native PCI PM
and if that's the case the offset of its power management capability structure
in the configuration space is stored in the pm_cap field of the device's struct
pci_dev object.  Next, the function checks which PCI low-power states are
supported by the device and from which low-power states the device can generate
native PCI PMEs.  The power management fields of the device's struct pci_dev and
the struct device embedded in it are updated accordingly and the generation of
PMEs by the device is disabled.

The second function checks if the device can be prepared to signal wakeup with
the help of the platform firmware, such as the ACPI BIOS.  If that is the case,
the function updates the wakeup fields in struct device embedded in the
device's struct pci_dev and uses the firmware-provided method to prevent the
device from signaling wakeup.

At this point the device is ready for power management.  For driverless devices,
however, this functionality is limited to a few basic operations carried out
during system-wide transitions to a sleep state and back to the working state.

2.3. Runtime Device Power Management
------------------------------------
The PCI subsystem plays a vital role in the runtime power management of PCI
devices.  For this purpose it uses the general runtime power management
(runtime PM) framework described in Documentation/power/runtime_pm.txt.
Namely, it provides subsystem-level callbacks:

	pci_pm_runtime_suspend()
	pci_pm_runtime_resume()
	pci_pm_runtime_idle()

that are executed by the core runtime PM routines.  It also implements the
entire mechanics necessary for handling runtime wakeup signals from PCI devices
in low-power states, which at the time of this writing works for both the native
PCI Express PME signaling and the ACPI GPE-based wakeup signaling described in
Section 1.

First, a PCI device is put into a low-power state, or suspended, with the help
of pm_schedule_suspend() or pm_runtime_suspend() which for PCI devices call
pci_pm_runtime_suspend() to do the actual job.  For this to work, the device's
driver has to provide a pm->runtime_suspend() callback (see below), which is
run by pci_pm_runtime_suspend() as the first action.  If the driver's callback
returns successfully, the device's standard configuration registers are saved,
the device is prepared to generate wakeup signals and, finally, it is put into
the target low-power state.

The low-power state to put the device into is the lowest-power (highest number)
state from which it can signal wakeup.  The exact method of signaling wakeup is
system-dependent and is determined by the PCI subsystem on the basis of the
reported capabilities of the device and the platform firmware.  To prepare the
device for signaling wakeup and put it into the selected low-power state, the
PCI subsystem can use the platform firmware as well as the device's native PCI
PM capabilities, if supported.

It is expected that the device driver's pm->runtime_suspend() callback will
not attempt to prepare the device for signaling wakeup or to put it into a
low-power state.  The driver ought to leave these tasks to the PCI subsystem
that has all of the information necessary to perform them.

A suspended device is brought back into the "active" state, or resumed,
with the help of pm_request_resume() or pm_runtime_resume() which both call
pci_pm_runtime_resume() for PCI devices.  Again, this only works if the device's
driver provides a pm->runtime_resume() callback (see below).  However, before
the driver's callback is executed, pci_pm_runtime_resume() brings the device
back into the full-power state, prevents it from signaling wakeup while in that
state and restores its standard configuration registers.  Thus the driver's
callback need not worry about the PCI-specific aspects of the device resume.

Note that generally pci_pm_runtime_resume() may be called in two different
situations.  First, it may be called at the request of the device's driver, for
example if there are some data for it to process.  Second, it may be called
as a result of a wakeup signal from the device itself (this sometimes is
referred to as "remote wakeup").  Of course, for this purpose the wakeup signal
is handled in one of the ways described in Section 1 and finally converted into
a notification for the PCI subsystem after the source device has been
identified.

The pci_pm_runtime_idle() function, called for PCI devices by pm_runtime_idle()
and pm_request_idle(), executes the device driver's pm->runtime_idle()
callback, if defined, and if that callback doesn't return error code (or is not
present at all), suspends the device with the help of pm_runtime_suspend().
Sometimes pci_pm_runtime_idle() is called automatically by the PM core (for
example, it is called right after the device has just been resumed), in which
cases it is expected to suspend the device if that makes sense.  Usually,
however, the PCI subsystem doesn't really know if the device really can be
suspended, so it lets the device's driver decide by running its
pm->runtime_idle() callback.

2.4. System-Wide Power Transitions
----------------------------------
There are a few different types of system-wide power transitions, described in
Documentation/power/admin-guide/devices.rst.  Each of them requires devices to be handled
in a specific way and the PM core executes subsystem-level power management
callbacks for this purpose.  They are executed in phases such that each phase
involves executing the same subsystem-level callback for every device belonging
to the given subsystem before the next phase begins.  These phases always run
after tasks have been frozen.

2.4.1. System Suspend

When the system is going into a sleep state in which the contents of memory will
be preserved, such as one of the ACPI sleep states S1-S3, the phases are:

	prepare, suspend, suspend_noirq.

The following PCI bus type's callbacks, respectively, are used in these phases:

	pci_pm_prepare()
	pci_pm_suspend()
	pci_pm_suspend_noirq()

The pci_pm_prepare() routine first puts the device into the "fully functional"
state with the help of pm_runtime_resume().  Then, it executes the device
driver's pm->prepare() callback if defined (i.e. if the driver's struct
dev_pm_ops object is present and the prepare pointer in that object is valid).

The pci_pm_suspend() routine first checks if the device's driver implements
legacy PCI suspend routines (see Section 3), in which case the driver's legacy
suspend callback is executed, if present, and its result is returned.  Next, if
the device's driver doesn't provide a struct dev_pm_ops object (containing
pointers to the driver's callbacks), pci_pm_default_suspend() is called, which
simply turns off the device's bus master capability and runs
pcibios_disable_device() to disable it, unless the device is a bridge (PCI
bridges are ignored by this routine).  Next, the device driver's pm->suspend()
callback is executed, if defined, and its result is returned if it fails.
Finally, pci_fixup_device() is called to apply hardware suspend quirks related
to the device if necessary.

Note that the suspend phase is carried out asynchronously for PCI devices, so
the pci_pm_suspend() callback may be executed in parallel for any pair of PCI
devices that don't depend on each other in a known way (i.e. none of the paths
in the device tree from the root bridge to a leaf device contains both of them).

The pci_pm_suspend_noirq() routine is executed after suspend_device_irqs() has
been called, which means that the device driver's interrupt handler won't be
invoked while this routine is running.  It first checks if the device's driver
implements legacy PCI suspends routines (Section 3), in which case the legacy
late suspend routine is called and its result is returned (the standard
configuration registers of the device are saved if the driver's callback hasn't
done that).  Second, if the device driver's struct dev_pm_ops object is not
present, the device's standard configuration registers are saved and the routine
returns success.  Otherwise the device driver's pm->suspend_noirq() callback is
executed, if present, and its result is returned if it fails.  Next, if the
device's standard configuration registers haven't been saved yet (one of the
device driver's callbacks executed before might do that), pci_pm_suspend_noirq()
saves them, prepares the device to signal wakeup (if necessary) and puts it into
a low-power state.

The low-power state to put the device into is the lowest-power (highest number)
state from which it can signal wakeup while the system is in the target sleep
state.  Just like in the runtime PM case described above, the mechanism of
signaling wakeup is system-dependent and determined by the PCI subsystem, which
is also responsible for preparing the device to signal wakeup from the system's
target sleep state as appropriate.

PCI device drivers (that don't implement legacy power management callbacks) are
generally not expected to prepare devices for signaling wakeup or to put them
into low-power states.  However, if one of the driver's suspend callbacks
(pm->suspend() or pm->suspend_noirq()) saves the device's standard configuration
registers, pci_pm_suspend_noirq() will assume that the device has been prepared
to signal wakeup and put into a low-power state by the driver (the driver is
then assumed to have used the helper functions provided by the PCI subsystem for
this purpose).  PCI device drivers are not encouraged to do that, but in some
rare cases doing that in the driver may be the optimum approach.

2.4.2. System Resume

When the system is undergoing a transition from a sleep state in which the
contents of memory have been preserved, such as one of the ACPI sleep states
S1-S3, into the working state (ACPI S0), the phases are:

	resume_noirq, resume, complete.

The following PCI bus type's callbacks, respectively, are executed in these
phases:

	pci_pm_resume_noirq()
	pci_pm_resume()
	pci_pm_complete()

The pci_pm_resume_noirq() routine first puts the device into the full-power
state, restores its standard configuration registers and applies early resume
hardware quirks related to the device, if necessary.  This is done
unconditionally, regardless of whether or not the device's driver implements
legacy PCI power management callbacks (this way all PCI devices are in the
full-power state and their standard configuration registers have been restored
when their interrupt handlers are invoked for the first time during resume,
which allows the kernel to avoid problems with the handling of shared interrupts
by drivers whose devices are still suspended).  If legacy PCI power management
callbacks (see Section 3) are implemented by the device's driver, the legacy
early resume callback is executed and its result is returned.  Otherwise, the
device driver's pm->resume_noirq() callback is executed, if defined, and its
result is returned.

The pci_pm_resume() routine first checks if the device's standard configuration
registers have been restored and restores them if that's not the case (this
only is necessary in the error path during a failing suspend).  Next, resume
hardware quirks related to the device are applied, if necessary, and if the
device's driver implements legacy PCI power management callbacks (see
Section 3), the driver's legacy resume callback is executed and its result is
returned.  Otherwise, the device's wakeup signaling mechanisms are blocked and
its driver's pm->resume() callback is executed, if defined (the callback's
result is then returned).

The resume phase is carried out asynchronously for PCI devices, like the
suspend phase described above, which means that if two PCI devices don't depend
on each other in a known way, the pci_pm_resume() routine may be executed for
the both of them in parallel.

The pci_pm_complete() routine only executes the device driver's pm->complete()
callback, if defined.

2.4.3. System Hibernation

System hibernation is more complicated than system suspend, because it requires
a system image to be created and written into a persistent storage medium.  The
image is created atomically and all devices are quiesced, or frozen, before that
happens.

The freezing of devices is carried out after enough memory has been freed (at
the time of this writing the image creation requires at least 50% of system RAM
to be free) in the following three phases:

	prepare, freeze, freeze_noirq

that correspond to the PCI bus type's callbacks:

	pci_pm_prepare()
	pci_pm_freeze()
	pci_pm_freeze_noirq()

This means that the prepare phase is exactly the same as for system suspend.
The other two phases, however, are different.

The pci_pm_freeze() routine is quite similar to pci_pm_suspend(), but it runs
the device driver's pm->freeze() callback, if defined, instead of pm->suspend(),
and it doesn't apply the suspend-related hardware quirks.  It is executed
asynchronously for different PCI devices that don't depend on each other in a
known way.

The pci_pm_freeze_noirq() routine, in turn, is similar to
pci_pm_suspend_noirq(), but it calls the device driver's pm->freeze_noirq()
routine instead of pm->suspend_noirq().  It also doesn't attempt to prepare the
device for signaling wakeup and put it into a low-power state.  Still, it saves
the device's standard configuration registers if they haven't been saved by one
of the driver's callbacks.

Once the image has been created, it has to be saved.  However, at this point all
devices are frozen and they cannot handle I/O, while their ability to handle
I/O is obviously necessary for the image saving.  Thus they have to be brought
back to the fully functional state and this is done in the following phases:

	thaw_noirq, thaw, complete

using the following PCI bus type's callbacks:

	pci_pm_thaw_noirq()
	pci_pm_thaw()
	pci_pm_complete()

respectively.

The first of them, pci_pm_thaw_noirq(), is analogous to pci_pm_resume_noirq(),
but it doesn't put the device into the full power state and doesn't attempt to
restore its standard configuration registers.  It also executes the device
driver's pm->thaw_noirq() callback, if defined, instead of pm->resume_noirq().

The pci_pm_thaw() routine is similar to pci_pm_resume(), but it runs the device
driver's pm->thaw() callback instead of pm->resume().  It is executed
asynchronously for different PCI devices that don't depend on each other in a
known way.

The complete phase it the same as for system resume.

After saving the image, devices need to be powered down before the system can
enter the target sleep state (ACPI S4 for ACPI-based systems).  This is done in
three phases:

	prepare, poweroff, poweroff_noirq

where the prepare phase is exactly the same as for system suspend.  The other
two phases are analogous to the suspend and suspend_noirq phases, respectively.
The PCI subsystem-level callbacks they correspond to

	pci_pm_poweroff()
	pci_pm_poweroff_noirq()

work in analogy with pci_pm_suspend() and pci_pm_poweroff_noirq(), respectively,
although they don't attempt to save the device's standard configuration
registers.

2.4.4. System Restore

System restore requires a hibernation image to be loaded into memory and the
pre-hibernation memory contents to be restored before the pre-hibernation system
activity can be resumed.

As described in Documentation/power/admin-guide/devices.rst, the hibernation image is loaded
into memory by a fresh instance of the kernel, called the boot kernel, which in
turn is loaded and run by a boot loader in the usual way.  After the boot kernel
has loaded the image, it needs to replace its own code and data with the code
and data of the "hibernated" kernel stored within the image, called the image
kernel.  For this purpose all devices are frozen just like before creating
the image during hibernation, in the

	prepare, freeze, freeze_noirq

phases described above.  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 (that will only affect the devices having
drivers in the boot kernel), and then continue running normally.

If the pre-hibernation memory contents are restored successfully, which is the
usual situation, control is passed to the image kernel, which then becomes
responsible for bringing the system back to the working state.  To achieve this,
it must restore the devices' pre-hibernation functionality, which is done much
like waking up from the memory sleep state, although it involves different
phases:

	restore_noirq, restore, complete

The first two of these are analogous to the resume_noirq and resume phases
described above, respectively, and correspond to the following PCI subsystem
callbacks:

	pci_pm_restore_noirq()
	pci_pm_restore()

These callbacks work in analogy with pci_pm_resume_noirq() and pci_pm_resume(),
respectively, but they execute the device driver's pm->restore_noirq() and
pm->restore() callbacks, if available.

The complete phase is carried out in exactly the same way as during system
resume.


3. PCI Device Drivers and Power Management
==========================================

3.1. Power Management Callbacks
-------------------------------
PCI device drivers participate in power management by providing callbacks to be
executed by the PCI subsystem's power management routines described above and by
controlling the runtime power management of their devices.

At the time of this writing there are two ways to define power management
callbacks for a PCI device driver, the recommended one, based on using a
dev_pm_ops structure described in Documentation/power/admin-guide/devices.rst, and the
"legacy" one, in which the .suspend(), .suspend_late(), .resume_early(), and
.resume() callbacks from struct pci_driver are used.  The legacy approach,
however, doesn't allow one to define runtime power management callbacks and is
not really suitable for any new drivers.  Therefore it is not covered by this
document (refer to the source code to learn more about it).

It is recommended that all PCI device drivers define a struct dev_pm_ops object
containing pointers to power management (PM) callbacks that will be executed by
the PCI subsystem's PM routines in various circumstances.  A pointer to the
driver's struct dev_pm_ops object has to be assigned to the driver.pm field in
its struct pci_driver object.  Once that has happened, the "legacy" PM callbacks
in struct pci_driver are ignored (even if they are not NULL).

The PM callbacks in struct dev_pm_ops are not mandatory and if they are not
defined (i.e. the respective fields of struct dev_pm_ops are unset) the PCI
subsystem will handle the device in a simplified default manner.  If they are
defined, though, they are expected to behave as described in the following
subsections.

3.1.1. prepare()

The prepare() callback is executed during system suspend, during hibernation
(when a hibernation image is about to be created), during power-off after
saving a hibernation image and during system restore, when a hibernation image
has just been loaded into memory.

This callback is only necessary if the driver's device has children that in
general may be registered at any time.  In that case the role of the prepare()
callback is to prevent new children of the device from being registered until
one of the resume_noirq(), thaw_noirq(), or restore_noirq() callbacks is run.

In addition to that the prepare() callback may carry out some operations
preparing the device to be suspended, although it should not allocate memory
(if additional memory is required to suspend the device, it has to be
preallocated earlier, for example in a suspend/hibernate notifier as described
in Documentation/driver-api/pm/notifiers.rst).

3.1.2. suspend()

The suspend() callback is only executed during system suspend, after prepare()
callbacks have been executed for all devices in the system.

This callback is expected to quiesce the device and prepare it to be put into a
low-power state by the PCI subsystem.  It is not required (in fact it even is
not recommended) that a PCI driver's suspend() callback save the standard
configuration registers of the device, prepare it for waking up the system, or
put it into a low-power state.  All of these operations can very well be taken
care of by the PCI subsystem, without the driver's participation.

However, in some rare case it is convenient to carry out these operations in
a PCI driver.  Then, pci_save_state(), pci_prepare_to_sleep(), and
pci_set_power_state() should be used to save the device's standard configuration
registers, to prepare it for system wakeup (if necessary), and to put it into a
low-power state, respectively.  Moreover, if the driver calls pci_save_state(),
the PCI subsystem will not execute either pci_prepare_to_sleep(), or
pci_set_power_state() for its device, so the driver is then responsible for
handling the device as appropriate.

While the suspend() callback is being executed, the driver's interrupt handler
can be invoked to handle an interrupt from the device, so all suspend-related
operations relying on the driver's ability to handle interrupts should be
carried out in this callback.

3.1.3. suspend_noirq()

The suspend_noirq() callback is only executed during system suspend, after
suspend() callbacks have been executed for all devices in the system and
after device interrupts have been disabled by the PM core.

The difference between suspend_noirq() and suspend() is that the driver's
interrupt handler will not be invoked while suspend_noirq() is running.  Thus
suspend_noirq() can carry out operations that would cause race conditions to
arise if they were performed in suspend().

3.1.4. freeze()

The freeze() callback is hibernation-specific and is executed in two situations,
during hibernation, after prepare() callbacks have been executed for all devices
in preparation for the creation of a system image, and during restore,
after a system image has been loaded into memory from persistent storage and the
prepare() callbacks have been executed for all devices.

The role of this callback is analogous to the role of the suspend() callback
described above.  In fact, they only need to be different in the rare cases when
the driver takes the responsibility for putting the device into a low-power
state.

In that cases the freeze() callback should not prepare the device system wakeup
or put it into a low-power state.  Still, either it or freeze_noirq() should
save the device's standard configuration registers using pci_save_state().

3.1.5. freeze_noirq()

The freeze_noirq() callback is hibernation-specific.  It is executed during
hibernation, after prepare() and freeze() callbacks have been executed for all
devices in preparation for the creation of a system image, and during restore,
after a system image has been loaded into memory and after prepare() and
freeze() callbacks have been executed for all devices.  It is always executed
after device interrupts have been disabled by the PM core.

The role of this callback is analogous to the role of the suspend_noirq()
callback described above and it very rarely is necessary to define
freeze_noirq().

The difference between freeze_noirq() and freeze() is analogous to the
difference between suspend_noirq() and suspend().

3.1.6. poweroff()

The poweroff() callback is hibernation-specific.  It is executed when the system
is about to be powered off after saving a hibernation image to a persistent
storage.  prepare() callbacks are executed for all devices before poweroff() is
called.

The role of this callback is analogous to the role of the suspend() and freeze()
callbacks described above, although it does not need to save the contents of
the device's registers.  In particular, if the driver wants to put the device
into a low-power state itself instead of allowing the PCI subsystem to do that,
the poweroff() callback should use pci_prepare_to_sleep() and
pci_set_power_state() to prepare the device for system wakeup and to put it
into a low-power state, respectively, but it need not save the device's standard
configuration registers.

3.1.7. poweroff_noirq()

The poweroff_noirq() callback is hibernation-specific.  It is executed after
poweroff() callbacks have been executed for all devices in the system.

The role of this callback is analogous to the role of the suspend_noirq() and
freeze_noirq() callbacks described above, but it does not need to save the
contents of the device's registers.

The difference between poweroff_noirq() and poweroff() is analogous to the
difference between suspend_noirq() and suspend().

3.1.8. resume_noirq()

The resume_noirq() callback is only executed during system resume, after the
PM core has enabled the non-boot CPUs.  The driver's interrupt handler will not
be invoked while resume_noirq() is running, so this callback can carry out
operations that might race with the interrupt handler.

Since the PCI subsystem unconditionally puts all devices into the full power
state in the resume_noirq phase of system resume and restores their standard
configuration registers, resume_noirq() is usually not necessary.  In general
it should only be used for performing operations that would lead to race
conditions if carried out by resume().

3.1.9. resume()

The resume() callback is only executed during system resume, after
resume_noirq() callbacks have been executed for all devices in the system and
device interrupts have been enabled by the PM core.

This callback is responsible for restoring the pre-suspend configuration of the
device and bringing it back to the fully functional state.  The device should be
able to process I/O in a usual way after resume() has returned.

3.1.10. thaw_noirq()

The thaw_noirq() callback is hibernation-specific.  It is executed after a
system image has been created and the non-boot CPUs have been enabled by the PM
core, in the thaw_noirq phase of hibernation.  It also may be executed if the
loading of a hibernation image fails during system restore (it is then executed
after enabling the non-boot CPUs).  The driver's interrupt handler will not be
invoked while thaw_noirq() is running.

The role of this callback is analogous to the role of resume_noirq().  The
difference between these two callbacks is that thaw_noirq() is executed after
freeze() and freeze_noirq(), so in general it does not need to modify the
contents of the device's registers.

3.1.11. thaw()

The thaw() callback is hibernation-specific.  It is executed after thaw_noirq()
callbacks have been executed for all devices in the system and after device
interrupts have been enabled by the PM core.

This callback is responsible for restoring the pre-freeze configuration of
the device, so that it will work in a usual way after thaw() has returned.

3.1.12. restore_noirq()

The restore_noirq() callback is hibernation-specific.  It is executed in the
restore_noirq phase of hibernation, when the boot kernel has passed control to
the image kernel and the non-boot CPUs have been enabled by the image kernel's
PM core.

This callback is analogous to resume_noirq() with the exception that it cannot
make any assumption on the previous state of the device, even if the BIOS (or
generally the platform firmware) is known to preserve that state over a
suspend-resume cycle.

For the vast majority of PCI device drivers there is no difference between
resume_noirq() and restore_noirq().

3.1.13. restore()

The restore() callback is hibernation-specific.  It is executed after
restore_noirq() callbacks have been executed for all devices in the system and
after the PM core has enabled device drivers' interrupt handlers to be invoked.

This callback is analogous to resume(), just like restore_noirq() is analogous
to resume_noirq().  Consequently, the difference between restore_noirq() and
restore() is analogous to the difference between resume_noirq() and resume().

For the vast majority of PCI device drivers there is no difference between
resume() and restore().

3.1.14. complete()

The complete() callback is executed in the following situations:
  - during system resume, after resume() callbacks have been executed for all
    devices,
  - during hibernation, before saving the system image, after thaw() callbacks
    have been executed for all devices,
  - during system restore, when the system is going back to its pre-hibernation
    state, after restore() callbacks have been executed for all devices.
It also may be executed if the loading of a hibernation image into memory fails
(in that case it is run after thaw() callbacks have been executed for all
devices that have drivers in the boot kernel).

This callback is entirely optional, although it may be necessary if the
prepare() callback performs operations that need to be reversed.

3.1.15. runtime_suspend()

The runtime_suspend() callback is specific to device runtime power management
(runtime PM).  It is executed by the PM core's runtime PM framework when the
device is about to be suspended (i.e. quiesced and put into a low-power state)
at run time.

This callback is responsible for freezing the device and preparing it to be
put into a low-power state, but it must allow the PCI subsystem to perform all
of the PCI-specific actions necessary for suspending the device.

3.1.16. runtime_resume()

The runtime_resume() callback is specific to device runtime PM.  It is executed
by the PM core's runtime PM framework when the device is about to be resumed
(i.e. put into the full-power state and programmed to process I/O normally) at
run time.

This callback is responsible for restoring the normal functionality of the
device after it has been put into the full-power state by the PCI subsystem.
The device is expected to be able to process I/O in the usual way after
runtime_resume() has returned.

3.1.17. runtime_idle()

The runtime_idle() callback is specific to device runtime PM.  It is executed
by the PM core's runtime PM framework whenever it may be desirable to suspend
the device according to the PM core's information.  In particular, it is
automatically executed right after runtime_resume() has returned in case the
resume of the device has happened as a result of a spurious event.

This callback is optional, but if it is not implemented or if it returns 0, the
PCI subsystem will call pm_runtime_suspend() for the device, which in turn will
cause the driver's runtime_suspend() callback to be executed.

3.1.18. Pointing Multiple Callback Pointers to One Routine

Although in principle each of the callbacks described in the previous
subsections can be defined as a separate function, it often is convenient to
point two or more members of struct dev_pm_ops to the same routine.  There are
a few convenience macros that can be used for this purpose.

The SIMPLE_DEV_PM_OPS macro declares a struct dev_pm_ops object with one
suspend routine pointed to by the .suspend(), .freeze(), and .poweroff()
members and one resume routine pointed to by the .resume(), .thaw(), and
.restore() members.  The other function pointers in this struct dev_pm_ops are
unset.

The UNIVERSAL_DEV_PM_OPS macro is similar to SIMPLE_DEV_PM_OPS, but it
additionally sets the .runtime_resume() pointer to the same value as
.resume() (and .thaw(), and .restore()) and the .runtime_suspend() pointer to
the same value as .suspend() (and .freeze() and .poweroff()).

The SET_SYSTEM_SLEEP_PM_OPS can be used inside of a declaration of struct
dev_pm_ops to indicate that one suspend routine is to be pointed to by the
.suspend(), .freeze(), and .poweroff() members and one resume routine is to
be pointed to by the .resume(), .thaw(), and .restore() members.

3.2. Device Runtime Power Management
------------------------------------
In addition to providing device power management callbacks PCI device drivers
are responsible for controlling the runtime power management (runtime PM) of
their devices.

The PCI device runtime PM is optional, but it is recommended that PCI device
drivers implement it at least in the cases where there is a reliable way of
verifying that the device is not used (like when the network cable is detached
from an Ethernet adapter or there are no devices attached to a USB controller).

To support the PCI runtime PM the driver first needs to implement the
runtime_suspend() and runtime_resume() callbacks.  It also may need to implement
the runtime_idle() callback to prevent the device from being suspended again
every time right after the runtime_resume() callback has returned
(alternatively, the runtime_suspend() callback will have to check if the
device should really be suspended and return -EAGAIN if that is not the case).

The runtime PM of PCI devices is enabled by default by the PCI core.  PCI
device drivers do not need to enable it and should not attempt to do so.
However, it is blocked by pci_pm_init() that runs the pm_runtime_forbid()
helper function.  In addition to that, the runtime PM usage counter of
each PCI device is incremented by local_pci_probe() before executing the
probe callback provided by the device's driver.

If a PCI driver implements the runtime PM callbacks and intends to use the
runtime PM framework provided by the PM core and the PCI subsystem, it needs
to decrement the device's runtime PM usage counter in its probe callback
function.  If it doesn't do that, the counter will always be different from
zero for the device and it will never be runtime-suspended.  The simplest
way to do that is by calling pm_runtime_put_noidle(), but if the driver
wants to schedule an autosuspend right away, for example, it may call
pm_runtime_put_autosuspend() instead for this purpose.  Generally, it
just needs to call a function that decrements the devices usage counter
from its probe routine to make runtime PM work for the device.

It is important to remember that the driver's runtime_suspend() callback
may be executed right after the usage counter has been decremented, because
user space may already have caused the pm_runtime_allow() helper function
unblocking the runtime PM of the device to run via sysfs, so the driver must
be prepared to cope with that.

The driver itself should not call pm_runtime_allow(), though.  Instead, it
should let user space or some platform-specific code do that (user space can
do it via sysfs as stated above), but it must be prepared to handle the
runtime PM of the device correctly as soon as pm_runtime_allow() is called
(which may happen at any time, even before the driver is loaded).

When the driver's remove callback runs, it has to balance the decrementation
of the device's runtime PM usage counter at the probe time.  For this reason,
if it has decremented the counter in its probe callback, it must run
pm_runtime_get_noresume() in its remove callback.  [Since the core carries
out a runtime resume of the device and bumps up the device's usage counter
before running the driver's remove callback, the runtime PM of the device
is effectively disabled for the duration of the remove execution and all
runtime PM helper functions incrementing the device's usage counter are
then effectively equivalent to pm_runtime_get_noresume().]

The runtime PM framework works by processing requests to suspend or resume
devices, or to check if they are idle (in which cases it is reasonable to
subsequently request that they be suspended).  These requests are represented
by work items put into the power management workqueue, pm_wq.  Although there
are a few situations in which power management requests are automatically
queued by the PM core (for example, after processing a request to resume a
device the PM core automatically queues a request to check if the device is
idle), device drivers are generally responsible for queuing power management
requests for their devices.  For this purpose they should use the runtime PM
helper functions provided by the PM core, discussed in
Documentation/power/runtime_pm.txt.

Devices can also be suspended and resumed synchronously, without placing a
request into pm_wq.  In the majority of cases this also is done by their
drivers that use helper functions provided by the PM core for this purpose.

For more information on the runtime PM of devices refer to
Documentation/power/runtime_pm.txt.


4. Resources
============

PCI Local Bus Specification, Rev. 3.0
PCI Bus Power Management Interface Specification, Rev. 1.2
Advanced Configuration and Power Interface (ACPI) Specification, Rev. 3.0b
PCI Express Base Specification, Rev. 2.0
Documentation/power/admin-guide/devices.rst
Documentation/power/runtime_pm.txt