linux_dsm_epyc7002/Documentation/power/devices.txt

Most of the code in Linux is device drivers, so most of the Linux power
management 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-to-ram", or (mostly for systems with
	disks) "hibernate" (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 that low-power state.  This feature may be disabled using the
	relevant /sys/devices/.../power/wakeup file; enabling it may cost some
	power usage, but let the whole system enter low power states more often.

    Runtime Power Management model:
	Drivers may also enter low power states while the system is running,
	independently of other power management activity.  Upstream drivers
	will normally not know (or care) if the device is in some low power
	state when issuing requests; the driver will auto-resume anything
	that's needed when it gets a request.

	This doesn't have, or need much infrastructure; it's just something you
	should do when writing your drivers.  For example, clk_disable() unused
	clocks as part of minimizing power drain for currently-unused hardware.
	Of course, sometimes clusters of drivers will collaborate with each
	other, which could involve task-specific power management.

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
drivers put themselves into low power states (at "runtime"), the effect may be
the same as 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, 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
===========================================
Most of the programming interfaces a device driver needs to know about
relate to that first model:  entering a system-wide low power state,
rather than just minimizing power consumption by one device.


Bus Driver Methods
------------------
The core methods to suspend and resume devices reside in struct bus_type.
These are mostly of interest to people writing infrastructure for busses
like PCI or USB, or because they define the primitives that device drivers
may need to apply in domain-specific ways to their devices:

struct bus_type {
	...
	int  (*suspend)(struct device *dev, pm_message_t state);
	int  (*resume)(struct device *dev);
};

Bus drivers implement those methods as appropriate for the hardware and
the drivers using it; PCI works differently from USB, and so on.  Not many
people write bus 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.  Note that as this is being written,
only the suspend() and resume() are widely available; not many bus drivers
leverage all of those phases, or pass them down to lower driver levels.


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

The "can_wakeup" flag just records whether the device (and its driver) can
physically support wakeup events.  When that flag is clear, the sysfs
"wakeup" file is empty, and device_may_wakeup() returns false.

For devices that can issue wakeup events, a separate flag controls whether
that device should try to use its wakeup mechanism.  The initial value of
device_may_wakeup() will be true, so that the device's "wakeup" file holds
the value "enabled".  Userspace can change that to "disabled" so that
device_may_wakeup() returns false; or change it back to "enabled" (so that
it returns true again).


EXAMPLE:  PCI Device Driver Methods
-----------------------------------
PCI framework software calls these methods when the PCI device driver bound
to a device device has provided them:

struct pci_driver {
	...
	int  (*suspend)(struct pci_device *pdev, pm_message_t state);
	int  (*suspend_late)(struct pci_device *pdev, pm_message_t state);

	int  (*resume_early)(struct pci_device *pdev);
	int  (*resume)(struct pci_device *pdev);
};

Drivers will implement those methods, and call PCI-specific procedures
like pci_set_power_state(), pci_enable_wake(), pci_save_state(), and
pci_restore_state() to manage PCI-specific mechanisms.  (PCI config space
could be saved during driver probe, if it weren't for the fact that some
systems rely on userspace tweaking using setpci.)  Devices are suspended
before their bridges enter low power states, and likewise bridges resume
before their devices.


Upper Layers of Driver Stacks
-----------------------------
Device drivers generally have at least two interfaces, and the methods
sketched above are the ones which apply to the lower level (nearer PCI, USB,
or other bus hardware).  The network and block layers are examples of upper
level interfaces, as is a character device talking to userspace.

Power management requests normally need to flow through those upper levels,
which often use domain-oriented requests like "blank that screen".  In
some cases those upper levels will have power management intelligence that
relates to end-user activity, or other devices that work in cooperation.

When those interfaces are structured using class interfaces, there is a
standard way to have the upper layer stop issuing requests to a given
class device (and restart later):

struct class {
	...
	int  (*suspend)(struct device *dev, pm_message_t state);
	int  (*resume)(struct device *dev);
};

Those calls are issued in specific phases of the process by which the
system enters a low power "suspend" state, or resumes from it.


Calling Drivers to Enter System Sleep States
============================================
When the system enters a low power state, each device's driver is asked
to suspend the device by putting it into 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.  The suspend and resume operations always go together, and
both are multi-phase operations.

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

More power-aware drivers drivers will use more than one device low power
state, either at runtime or during system sleep states, and might trigger
system wakeup events.


Call Sequence Guarantees
------------------------
To ensure that bridges and similar links needed 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 (ie. 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.


Suspending Devices
------------------
Suspending a given device is done in several phases.  Suspending the
system always includes every phase, executing calls 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 phases are seen by driver notifications issued in this order:

   1	class.suspend(dev, message) is called after tasks are frozen, for
	devices associated with a class that has such a method.  This
	method may sleep.

	Since I/O activity usually comes from such higher layers, this is
	a good place to quiesce all drivers of a given type (and keep such
	code out of those drivers).

   2	bus.suspend(dev, message) is called next.  This method may sleep,
	and is often morphed into a device driver call with bus-specific
	parameters and/or rules.

	This call should handle parts of device suspend logic that require
	sleeping.  It probably does work to quiesce the device which hasn't
	been abstracted into class.suspend().

The pm_message_t parameter is currently used to refine those semantics
(described later).

At the end of those phases, drivers should normally 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 also use
clk_disable() to gate off one or more clock sources; sometimes they will
also switch off power supplies, or reduce voltages.  Drivers which have
runtime PM support may already have performed some or all of the steps
needed to prepare for the upcoming system sleep state.

When any driver sees that its device_can_wakeup(dev), it should make sure
to use the relevant hardware signals to trigger a system wakeup event.
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 PCI's PME# signal.

If a driver (or bus, or class) fails it suspend method, the system won't
enter the desired low power state; it will resume all the devices it's
suspended so far.

Note that drivers may need to perform different actions based on the target
system lowpower/sleep state.  At this writing, there are only platform
specific APIs through which drivers could determine those target states.


Device Low Power (suspend) States
---------------------------------
Device low-power states aren't very 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.


Meaning of pm_message_t.event
-----------------------------
Parameters to suspend calls include the device affected and a message of
type pm_message_t, which has one field:  the event.  If driver does not
recognize the event code, suspend calls may abort the request and return
a negative errno.  However, most drivers will be fine if they implement
PM_EVENT_SUSPEND semantics for all messages.

The event codes are used to refine the goal of suspending the device, and
mostly matter when creating or resuming system memory image snapshots, as
used with suspend-to-disk:

    PM_EVENT_SUSPEND -- quiesce the driver and put hardware into a low-power
	state.  When used with system sleep states like "suspend-to-RAM" or
	"standby", the upcoming resume() call will often be able to rely on
	state kept in hardware, or issue system wakeup events.

    PM_EVENT_HIBERNATE -- Put hardware into a low-power state and enable wakeup
	events as appropriate.  It is only used with hibernation
	(suspend-to-disk) and few devices are able to wake up the system from
	this state; most are completely powered off.

    PM_EVENT_FREEZE -- quiesce the driver, but don't necessarily change into
	any low power mode.  A system snapshot is about to be taken, often
	followed by a call to the driver's resume() method.  Neither wakeup
	events nor DMA are allowed.

    PM_EVENT_PRETHAW -- quiesce the driver, knowing that the upcoming resume()
	will restore a suspend-to-disk snapshot from a different kernel image.
	Drivers that are smart enough to look at their hardware state during
	resume() processing need that state to be correct ... a PRETHAW could
	be used to invalidate that state (by resetting the device), like a
	shutdown() invocation would before a kexec() or system halt.  Other
	drivers might handle this the same way as PM_EVENT_FREEZE.  Neither
	wakeup events nor DMA are allowed.

To enter "standby" (ACPI S1) or "Suspend to RAM" (STR, ACPI S3) states, or
the similarly named APM states, only PM_EVENT_SUSPEND is used; the other event
codes are used for hibernation ("Suspend to Disk", STD, ACPI S4).

There's also PM_EVENT_ON, a value which never appears as a suspend event
but is sometimes used to record the "not suspended" device state.


Resuming Devices
----------------
Resuming is done in multiple phases, much like suspending, with all
devices processing each phase's calls before the next phase begins.

The phases are seen by driver notifications issued in this order:

   1	bus.resume(dev) reverses the effects of bus.suspend().  This may
	be morphed into a device driver call with bus-specific parameters;
	implementations may sleep.

   2	class.resume(dev) is called for devices associated with a class
	that has such a method.  Implementations may sleep.

	This reverses the effects of class.suspend(), and would usually
	reactivate the device's I/O queue.

At the end of those phases, drivers should normally be as functional as
they were before suspending:  I/O can be performed using DMA and IRQs, and
the relevant clocks are gated on.  The device need not be "fully on"; it
might be in a runtime lowpower/suspend state that acts as if it were.

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 called, but that can't always be guaranteed.

Drivers must also be prepared to notice that the device has been removed
while the system was powered off, 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.


Note that the bus-specific runtime PM wakeup mechanism can exist, and might
be defined to share some of the same driver code as for system wakeup.  For
example, a bus-specific device driver's resume() method might be used there,
so it wouldn't only be called from bus.resume() during system-wide wakeup.
See bus-specific information about how runtime wakeup events are handled.


System Devices
--------------
System devices follow a slightly different API, which can be found in

	include/linux/sysdev.h
	drivers/base/sys.c

System devices will only be suspended with interrupts disabled, and after
all other devices have been suspended.  On resume, they will be resumed
before any other devices, and also with interrupts disabled.

That is, IRQs are disabled, the suspend_late() phase begins, then the
sysdev_driver.suspend() phase, and the system enters a sleep state.  Then
the sysdev_driver.resume() phase begins, followed by the resume_early()
phase, after which IRQs are enabled.

Code to actually enter and exit the system-wide low power state sometimes
involves hardware details that are only known to the boot firmware, and
may leave a CPU running software (from SRAM or flash memory) that monitors
the system and manages its wakeup sequence.


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 a bus the device uses, and will
usually include hardware states that are also used in system sleep states.

However, note that if a driver puts a device into a runtime low power state
and the system then goes into a system-wide sleep state, it normally ought
to resume into that runtime low power state rather than "full on".  Such
distinctions would be part of the driver-internal state machine for that
hardware; the whole point of runtime power management is to be sure that
drivers are decoupled in that way from the state machine governing phases
of the system-wide power/sleep state transitions.


Power Saving Techniques
-----------------------
Normally runtime power management is handled by the drivers without specific
userspace or kernel intervention, by device-aware use of techniques like:

    Using information provided by other system layers
	- stay deeply "off" except between open() and close()
	- if transceiver/PHY indicates "nobody connected", stay "off"
	- application protocols may include power commands or hints

    Using fewer CPU cycles
	- using DMA instead of PIO
	- removing timers, or making them lower frequency
	- shortening "hot" code paths
	- eliminating cache misses
	- (sometimes) offloading work to device firmware

    Reducing other resource costs
	- gating off unused clocks in software (or hardware)
	- switching off unused power supplies
	- eliminating (or delaying/merging) IRQs
	- tuning DMA to use word and/or burst modes

    Using device-specific low power states
	- using lower voltages
	- avoiding needless DMA transfers

Read your hardware documentation carefully to see the opportunities that
may be available.  If you can, measure the actual power usage and check
it against the budget established for your project.


Examples:  USB hosts, system timer, system CPU
----------------------------------------------
USB host controllers make interesting, if complex, examples.  In many cases
these have no work to do:  no USB devices are connected, or all of them are
in the USB "suspend" state.  Linux host controller drivers can then disable
periodic DMA transfers that would otherwise be a constant power drain on the
memory subsystem, and enter a suspend state.  In power-aware controllers,
entering that suspend state may disable the clock used with USB signaling,
saving a certain amount of power.

The controller will be woken from that state (with an IRQ) by changes to the
signal state on the data lines of a given port, for example by an existing
peripheral requesting "remote wakeup" or by plugging a new peripheral.  The
same wakeup mechanism usually works from "standby" sleep states, and on some
systems also from "suspend to RAM" (or even "suspend to disk") states.
(Except that ACPI may be involved instead of normal IRQs, on some hardware.)

System devices like timers and CPUs may have special roles in the platform
power management scheme.  For example, system timers using a "dynamic tick"
approach don't just save CPU cycles (by eliminating needless timer IRQs),
but they may also open the door to using lower power CPU "idle" states that
cost more than a jiffie to enter and exit.  On x86 systems these are states
like "C3"; note that periodic DMA transfers from a USB host controller will
also prevent entry to a C3 state, much like a periodic timer IRQ.

That kind of runtime mechanism interaction is common.  "System On Chip" (SOC)
processors often have low power idle modes that can't be entered unless
certain medium-speed clocks (often 12 or 48 MHz) are gated off.  When the
drivers gate those clocks effectively, then the system idle task may be able
to use the lower power idle modes and thereby increase battery life.

If the CPU can have a "cpufreq" driver, there also may be opportunities
to shift to lower voltage settings and reduce the power cost of executing
a given number of instructions.  (Without voltage adjustment, it's rare
for cpufreq to save much power; the cost-per-instruction must go down.)