Linux kernel media framework
============================

This document describes the Linux kernel media framework, its data structures,
functions and their usage.


Introduction
------------

The media controller API is documented in DocBook format in
Documentation/DocBook/v4l/media-controller.xml. This document will focus on
the kernel-side implementation of the media framework.


Abstract media device model
---------------------------

Discovering a device internal topology, and configuring it at runtime, is one
of the goals of the media framework. To achieve this, hardware devices are
modeled as an oriented graph of building blocks called entities connected
through pads.

An entity is a basic media hardware building block. It can correspond to
a large variety of logical blocks such as physical hardware devices
(CMOS sensor for instance), logical hardware devices (a building block
in a System-on-Chip image processing pipeline), DMA channels or physical
connectors.

A pad is a connection endpoint through which an entity can interact with
other entities. Data (not restricted to video) produced by an entity
flows from the entity's output to one or more entity inputs. Pads should
not be confused with physical pins at chip boundaries.

A link is a point-to-point oriented connection between two pads, either
on the same entity or on different entities. Data flows from a source
pad to a sink pad.


Media device
------------

A media device is represented by a struct media_device instance, defined in
include/media/media-device.h. Allocation of the structure is handled by the
media device driver, usually by embedding the media_device instance in a
larger driver-specific structure.

Drivers register media device instances by calling

	media_device_register(struct media_device *mdev);

The caller is responsible for initializing the media_device structure before
registration. The following fields must be set:

 - dev must point to the parent device (usually a pci_dev, usb_interface or
   platform_device instance).

 - model must be filled with the device model name as a NUL-terminated UTF-8
   string. The device/model revision must not be stored in this field.

The following fields are optional:

 - serial is a unique serial number stored as a NUL-terminated ASCII string.
   The field is big enough to store a GUID in text form. If the hardware
   doesn't provide a unique serial number this field must be left empty.

 - bus_info represents the location of the device in the system as a
   NUL-terminated ASCII string. For PCI/PCIe devices bus_info must be set to
   "PCI:" (or "PCIe:") followed by the value of pci_name(). For USB devices,
   the usb_make_path() function must be used. This field is used by
   applications to distinguish between otherwise identical devices that don't
   provide a serial number.

 - hw_revision is the hardware device revision in a driver-specific format.
   When possible the revision should be formatted with the KERNEL_VERSION
   macro.

 - driver_version is formatted with the KERNEL_VERSION macro. The version
   minor must be incremented when new features are added to the userspace API
   without breaking binary compatibility. The version major must be
   incremented when binary compatibility is broken.

Upon successful registration a character device named media[0-9]+ is created.
The device major and minor numbers are dynamic. The model name is exported as
a sysfs attribute.

Drivers unregister media device instances by calling

	media_device_unregister(struct media_device *mdev);

Unregistering a media device that hasn't been registered is *NOT* safe.


Entities, pads and links
------------------------

- Entities

Entities are represented by a struct media_entity instance, defined in
include/media/media-entity.h. The structure is usually embedded into a
higher-level structure, such as a v4l2_subdev or video_device instance,
although drivers can allocate entities directly.

Drivers initialize entities by calling

	media_entity_init(struct media_entity *entity, u16 num_pads,
			  struct media_pad *pads, u16 extra_links);

The media_entity name, type, flags, revision and group_id fields can be
initialized before or after calling media_entity_init. Entities embedded in
higher-level standard structures can have some of those fields set by the
higher-level framework.

As the number of pads is known in advance, the pads array is not allocated
dynamically but is managed by the entity driver. Most drivers will embed the
pads array in a driver-specific structure, avoiding dynamic allocation.

Drivers must set the direction of every pad in the pads array before calling
media_entity_init. The function will initialize the other pads fields.

Unlike the number of pads, the total number of links isn't always known in
advance by the entity driver. As an initial estimate, media_entity_init
pre-allocates a number of links equal to the number of pads plus an optional
number of extra links. The links array will be reallocated if it grows beyond
the initial estimate.

Drivers register entities with a media device by calling

	media_device_register_entity(struct media_device *mdev,
				     struct media_entity *entity);

Entities are identified by a unique positive integer ID. Drivers can provide an
ID by filling the media_entity id field prior to registration, or request the
media controller framework to assign an ID automatically. Drivers that provide
IDs manually must ensure that all IDs are unique. IDs are not guaranteed to be
contiguous even when they are all assigned automatically by the framework.

Drivers unregister entities by calling

	media_device_unregister_entity(struct media_entity *entity);

Unregistering an entity will not change the IDs of the other entities, and the
ID will never be reused for a newly registered entity.

When a media device is unregistered, all its entities are unregistered
automatically. No manual entities unregistration is then required.

Drivers free resources associated with an entity by calling

	media_entity_cleanup(struct media_entity *entity);

This function must be called during the cleanup phase after unregistering the
entity. Note that the media_entity instance itself must be freed explicitly by
the driver if required.

Entities have flags that describe the entity capabilities and state.

	MEDIA_ENTITY_FLAG_DEFAULT indicates the default entity for a given
	type. This can be used to report the default audio and video devices
	or the default camera sensor.

Logical entity groups can be defined by setting the group ID of all member
entities to the same non-zero value. An entity group serves no purpose in the
kernel, but is reported to userspace during entities enumeration. The group_id
field belongs to the media device driver and must not by touched by entity
drivers.

Media device drivers should define groups if several entities are logically
bound together. Example usages include reporting

	- ALSA, VBI and video nodes that carry the same media stream
	- lens and flash controllers associated with a sensor

- Pads

Pads are represented by a struct media_pad instance, defined in
include/media/media-entity.h. Each entity stores its pads in a pads array
managed by the entity driver. Drivers usually embed the array in a
driver-specific structure.

Pads are identified by their entity and their 0-based index in the pads array.
Both information are stored in the media_pad structure, making the media_pad
pointer the canonical way to store and pass link references.

Pads have flags that describe the pad capabilities and state.

	MEDIA_PAD_FLAG_INPUT indicates that the pad supports sinking data.
	MEDIA_PAD_FLAG_OUTPUT indicates that the pad supports sourcing data.

One and only one of MEDIA_PAD_FLAG_INPUT and MEDIA_PAD_FLAG_OUTPUT must be set
for each pad.

- Links

Links are represented by a struct media_link instance, defined in
include/media/media-entity.h. Each entity stores all links originating at or
targetting any of its pads in a links array. A given link is thus stored
twice, once in the source entity and once in the target entity. The array is
pre-allocated and grows dynamically as needed.

Drivers create links by calling

	media_entity_create_link(struct media_entity *source, u16 source_pad,
				 struct media_entity *sink,   u16 sink_pad,
				 u32 flags);

An entry in the link array of each entity is allocated and stores pointers
to source and sink pads.

Links have flags that describe the link capabilities and state.

	MEDIA_LINK_FLAG_ENABLED indicates that the link is enabled and can be
	used to transfer media data. When two or more links target a sink pad,
	only one of them can be enabled at a time.
	MEDIA_LINK_FLAG_IMMUTABLE indicates that the link enabled state can't
	be modified at runtime. If MEDIA_LINK_FLAG_IMMUTABLE is set, then
	MEDIA_LINK_FLAG_ENABLED must also be set since an immutable link is
	always enabled.


Graph traversal
---------------

The media framework provides APIs to iterate over entities in a graph.

To iterate over all entities belonging to a media device, drivers can use the
media_device_for_each_entity macro, defined in include/media/media-device.h.

	struct media_entity *entity;

	media_device_for_each_entity(entity, mdev) {
		/* entity will point to each entity in turn */
		...
	}

Drivers might also need to iterate over all entities in a graph that can be
reached only through enabled links starting at a given entity. The media
framework provides a depth-first graph traversal API for that purpose.

Note that graphs with cycles (whether directed or undirected) are *NOT*
supported by the graph traversal API. To prevent infinite loops, the graph
traversal code limits the maximum depth to MEDIA_ENTITY_ENUM_MAX_DEPTH,
currently defined as 16.

Drivers initiate a graph traversal by calling

	media_entity_graph_walk_start(struct media_entity_graph *graph,
				      struct media_entity *entity);

The graph structure, provided by the caller, is initialized to start graph
traversal at the given entity.

Drivers can then retrieve the next entity by calling

	media_entity_graph_walk_next(struct media_entity_graph *graph);

When the graph traversal is complete the function will return NULL.

Graph traversal can be interrupted at any moment. No cleanup function call is
required and the graph structure can be freed normally.

Helper functions can be used to find a link between two given pads, or a pad
connected to another pad through an enabled link

	media_entity_find_link(struct media_pad *source,
			       struct media_pad *sink);

	media_entity_remote_source(struct media_pad *pad);

Refer to the kerneldoc documentation for more information.


Use count and power handling
----------------------------

Due to the wide differences between drivers regarding power management needs,
the media controller does not implement power management. However, the
media_entity structure includes a use_count field that media drivers can use to
track the number of users of every entity for power management needs.

The use_count field is owned by media drivers and must not be touched by entity
drivers. Access to the field must be protected by the media device graph_mutex
lock.


Links setup
-----------

Link properties can be modified at runtime by calling

	media_entity_setup_link(struct media_link *link, u32 flags);

The flags argument contains the requested new link flags.

The only configurable property is the ENABLED link flag to enable/disable a
link. Links marked with the IMMUTABLE link flag can not be enabled or disabled.

When a link is enabled or disabled, the media framework calls the
link_setup operation for the two entities at the source and sink of the link,
in that order. If the second link_setup call fails, another link_setup call is
made on the first entity to restore the original link flags.

Media device drivers can be notified of link setup operations by setting the
media_device::link_notify pointer to a callback function. If provided, the
notification callback will be called before enabling and after disabling
links.

Entity drivers must implement the link_setup operation if any of their links
is non-immutable. The operation must either configure the hardware or store
the configuration information to be applied later.

Link configuration must not have any side effect on other links. If an enabled
link at a sink pad prevents another link at the same pad from being disabled,
the link_setup operation must return -EBUSY and can't implicitly disable the
first enabled link.


Pipelines and media streams
---------------------------

When starting streaming, drivers must notify all entities in the pipeline to
prevent link states from being modified during streaming by calling

	media_entity_pipeline_start(struct media_entity *entity,
				    struct media_pipeline *pipe);

The function will mark all entities connected to the given entity through
enabled links, either directly or indirectly, as streaming.

The media_pipeline instance pointed to by the pipe argument will be stored in
every entity in the pipeline. Drivers should embed the media_pipeline structure
in higher-level pipeline structures and can then access the pipeline through
the media_entity pipe field.

Calls to media_entity_pipeline_start() can be nested. The pipeline pointer must
be identical for all nested calls to the function.

When stopping the stream, drivers must notify the entities with

	media_entity_pipeline_stop(struct media_entity *entity);

If multiple calls to media_entity_pipeline_start() have been made the same
number of media_entity_pipeline_stop() calls are required to stop streaming. The
media_entity pipe field is reset to NULL on the last nested stop call.

Link configuration will fail with -EBUSY by default if either end of the link is
a streaming entity. Links that can be modified while streaming must be marked
with the MEDIA_LINK_FLAG_DYNAMIC flag.

If other operations need to be disallowed on streaming entities (such as
changing entities configuration parameters) drivers can explictly check the
media_entity stream_count field to find out if an entity is streaming. This
operation must be done with the media_device graph_mutex held.