linux_dsm_epyc7002/Documentation/power/pm_qos_interface.rst

===============================
PM Quality Of Service Interface
===============================

This interface provides a kernel and user mode interface for registering
performance expectations by drivers, subsystems and user space applications on
one of the parameters.

Two different PM QoS frameworks are available:
 * CPU latency QoS.
 * The per-device PM QoS framework provides the API to manage the
   per-device latency constraints and PM QoS flags.

The latency unit used in the PM QoS framework is the microsecond (usec).


1. PM QoS framework
===================

A global list of CPU latency QoS requests is maintained along with an aggregated
(effective) target value.  The aggregated target value is updated with changes
to the request list or elements of the list.  For CPU latency QoS, the
aggregated target value is simply the min of the request values held in the list
elements.

Note: the aggregated target value is implemented as an atomic variable so that
reading the aggregated value does not require any locking mechanism.

From kernel space the use of this interface is simple:

void cpu_latency_qos_add_request(handle, target_value):
  Will insert an element into the CPU latency QoS list with the target value.
  Upon change to this list the new target is recomputed and any registered
  notifiers are called only if the target value is now different.
  Clients of PM QoS need to save the returned handle for future use in other
  PM QoS API functions.

void cpu_latency_qos_update_request(handle, new_target_value):
  Will update the list element pointed to by the handle with the new target
  value and recompute the new aggregated target, calling the notification tree
  if the target is changed.

void cpu_latency_qos_remove_request(handle):
  Will remove the element.  After removal it will update the aggregate target
  and call the notification tree if the target was changed as a result of
  removing the request.

int cpu_latency_qos_limit():
  Returns the aggregated value for the CPU latency QoS.

int cpu_latency_qos_request_active(handle):
  Returns if the request is still active, i.e. it has not been removed from the
  CPU latency QoS list.

int cpu_latency_qos_add_notifier(notifier):
  Adds a notification callback function to the CPU latency QoS. The callback is
  called when the aggregated value for the CPU latency QoS is changed.

int cpu_latency_qos_remove_notifier(notifier):
  Removes the notification callback function from the CPU latency QoS.


From user space:

The infrastructure exposes one device node, /dev/cpu_dma_latency, for the CPU
latency QoS.

Only processes can register a PM QoS request.  To provide for automatic
cleanup of a process, the interface requires the process to register its
parameter requests as follows.

To register the default PM QoS target for the CPU latency QoS, the process must
open /dev/cpu_dma_latency.

As long as the device node is held open that process has a registered
request on the parameter.

To change the requested target value, the process needs to write an s32 value to
the open device node.  Alternatively, it can write a hex string for the value
using the 10 char long format e.g. "0x12345678".  This translates to a
cpu_latency_qos_update_request() call.

To remove the user mode request for a target value simply close the device
node.


2. PM QoS per-device latency and flags framework
================================================

For each device, there are three lists of PM QoS requests. Two of them are
maintained along with the aggregated targets of resume latency and active
state latency tolerance (in microseconds) and the third one is for PM QoS flags.
Values are updated in response to changes of the request list.

The target values of resume latency and active state latency tolerance are
simply the minimum of the request values held in the parameter list elements.
The PM QoS flags aggregate value is a gather (bitwise OR) of all list elements'
values.  One device PM QoS flag is defined currently: PM_QOS_FLAG_NO_POWER_OFF.

Note: The aggregated target values are implemented in such a way that reading
the aggregated value does not require any locking mechanism.


From kernel mode the use of this interface is the following:

int dev_pm_qos_add_request(device, handle, type, value):
  Will insert an element into the list for that identified device with the
  target value.  Upon change to this list the new target is recomputed and any
  registered notifiers are called only if the target value is now different.
  Clients of dev_pm_qos need to save the handle for future use in other
  dev_pm_qos API functions.

int dev_pm_qos_update_request(handle, new_value):
  Will update the list element pointed to by the handle with the new target
  value and recompute the new aggregated target, calling the notification
  trees if the target is changed.

int dev_pm_qos_remove_request(handle):
  Will remove the element.  After removal it will update the aggregate target
  and call the notification trees if the target was changed as a result of
  removing the request.

s32 dev_pm_qos_read_value(device, type):
  Returns the aggregated value for a given device's constraints list.

enum pm_qos_flags_status dev_pm_qos_flags(device, mask)
  Check PM QoS flags of the given device against the given mask of flags.
  The meaning of the return values is as follows:

	PM_QOS_FLAGS_ALL:
		All flags from the mask are set
	PM_QOS_FLAGS_SOME:
		Some flags from the mask are set
	PM_QOS_FLAGS_NONE:
		No flags from the mask are set
	PM_QOS_FLAGS_UNDEFINED:
		The device's PM QoS structure has not been initialized
		or the list of requests is empty.

int dev_pm_qos_add_ancestor_request(dev, handle, type, value)
  Add a PM QoS request for the first direct ancestor of the given device whose
  power.ignore_children flag is unset (for DEV_PM_QOS_RESUME_LATENCY requests)
  or whose power.set_latency_tolerance callback pointer is not NULL (for
  DEV_PM_QOS_LATENCY_TOLERANCE requests).

int dev_pm_qos_expose_latency_limit(device, value)
  Add a request to the device's PM QoS list of resume latency constraints and
  create a sysfs attribute pm_qos_resume_latency_us under the device's power
  directory allowing user space to manipulate that request.

void dev_pm_qos_hide_latency_limit(device)
  Drop the request added by dev_pm_qos_expose_latency_limit() from the device's
  PM QoS list of resume latency constraints and remove sysfs attribute
  pm_qos_resume_latency_us from the device's power directory.

int dev_pm_qos_expose_flags(device, value)
  Add a request to the device's PM QoS list of flags and create sysfs attribute
  pm_qos_no_power_off under the device's power directory allowing user space to
  change the value of the PM_QOS_FLAG_NO_POWER_OFF flag.

void dev_pm_qos_hide_flags(device)
  Drop the request added by dev_pm_qos_expose_flags() from the device's PM QoS
  list of flags and remove sysfs attribute pm_qos_no_power_off from the device's
  power directory.

Notification mechanisms:

The per-device PM QoS framework has a per-device notification tree.

int dev_pm_qos_add_notifier(device, notifier, type):
  Adds a notification callback function for the device for a particular request
  type.

  The callback is called when the aggregated value of the device constraints
  list is changed.

int dev_pm_qos_remove_notifier(device, notifier, type):
  Removes the notification callback function for the device.


Active state latency tolerance
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

This device PM QoS type is used to support systems in which hardware may switch
to energy-saving operation modes on the fly.  In those systems, if the operation
mode chosen by the hardware attempts to save energy in an overly aggressive way,
it may cause excess latencies to be visible to software, causing it to miss
certain protocol requirements or target frame or sample rates etc.

If there is a latency tolerance control mechanism for a given device available
to software, the .set_latency_tolerance callback in that device's dev_pm_info
structure should be populated.  The routine pointed to by it is should implement
whatever is necessary to transfer the effective requirement value to the
hardware.

Whenever the effective latency tolerance changes for the device, its
.set_latency_tolerance() callback will be executed and the effective value will
be passed to it.  If that value is negative, which means that the list of
latency tolerance requirements for the device is empty, the callback is expected
to switch the underlying hardware latency tolerance control mechanism to an
autonomous mode if available.  If that value is PM_QOS_LATENCY_ANY, in turn, and
the hardware supports a special "no requirement" setting, the callback is
expected to use it.  That allows software to prevent the hardware from
automatically updating the device's latency tolerance in response to its power
state changes (e.g. during transitions from D3cold to D0), which generally may
be done in the autonomous latency tolerance control mode.

If .set_latency_tolerance() is present for the device, sysfs attribute
pm_qos_latency_tolerance_us will be present in the devivce's power directory.
Then, user space can use that attribute to specify its latency tolerance
requirement for the device, if any.  Writing "any" to it means "no requirement,
but do not let the hardware control latency tolerance" and writing "auto" to it
allows the hardware to be switched to the autonomous mode if there are no other
requirements from the kernel side in the device's list.

Kernel code can use the functions described above along with the
DEV_PM_QOS_LATENCY_TOLERANCE device PM QoS type to add, remove and update
latency tolerance requirements for devices.