[alsa-devel] [Intel-gfx] [RFC] set up an sync channel between audio and display driver (i.e. ALSA and DRM)

Daniel Vetter daniel at ffwll.ch
Tue May 20 12:02:04 CEST 2014

Adding Greg just as an fyi since we've chatted briefly about the avsink
bus. Comments below.

On Tue, May 20, 2014 at 02:52:19AM +0000, Lin, Mengdong wrote:
> This RFC is based on previous discussion to set up a generic communication channel between display and audio driver and
> an internal design of Intel MCG/VPG HDMI audio driver. It's still an initial draft and your advice would be appreciated
> to improve the design.
> The basic idea is to create a new avsink module and let both drm and alsa depend on it.
> This new module provides a framework and APIs for synchronization between the display and audio driver.
> 1. Display/Audio Client
> The avsink core provides APIs to create, register and lookup a display/audio client.
> A specific display driver (eg. i915) or audio driver (eg. HD-Audio driver) can create a client, add some resources
> objects (shared power wells, display outputs, and audio inputs, register ops) to the client, and then register this
> client to avisink core. The peer driver can look up a registered client by a name or type, or both. If a client gives
> a valid peer client name on registration, avsink core will bind the two clients as peer for each other. And we
> expect a display client and an audio client to be peers for each other in a system.
> int avsink_new_client ( const char *name,
>                             int type,   /* client type, display or audio */
>                             struct module *module,
>                             void *context,
>                             const char *peer_name,
>                             struct avsink_client **client_ret);
> int avsink_free_client (struct avsink_client *client);

Hm, my idea was to create a new avsink bus and let vga drivers register
devices on that thing and audio drivers register as drivers. There's a bit
more work involved in creating a full-blown bus, but it has a lot of
- Established infrastructure for matching drivers (i.e. audio drivers)
  against devices (i.e. avsinks exported by gfx drivers).
- Module refcounting.
- power domain handling and well-integrated into runtime pm.
- Allows integration into componentized device framework since we're
  dealing with a real struct device.
- Better decoupling between gfx and audio side since registration is done
  at runtime.
- We can attach drv private date which the audio driver needs.

> int avsink_register_client(struct avsink_client *client);
> int avisink_unregister_client(int client_handle);
> struct avsink_client *avsink_lookup_client(const char *name, int type);
> struct avsink_client {
>          const char *name;  /* client name */
>          int type; /* client type*/
>          void *context;
>          struct module *module;  /* top-level module for locking */
>          struct avsink_client *peer;      /* peer client */
>          /* shared power wells */
>          struct avsink_power_well *power_well;

We need to have an struct power_domain here so that we can do proper
runtime pm. But like I've said above I think we actually want a full blown
struct device.

>          int num_power_wells;
>          /* endpoints, display outputs or audio inputs */
>          struct avsink_endpoint * endpoint;
>          int num_endpints;
>          struct avsink_registers_ops *reg_ops; /* ops to access registers of a client */
>          void *private_data;
>          ...
> };

I think you're indeed implementing a full blown bus here ;-)

avsink->client = bus devices/childern
avsink->peer = driver for all this stuff
avsink->power_well = runtime pm support for the avsink bus
avsink->reg_ops = driver bind/unbind support

> On system boots, the avsink module is loaded before the display and audio driver module. And the display and audio
> driver may be loaded on parallel.
> * If a specific display driver (eg. i915) supports avsink, it can create a display client, add power wells and display
>   outputs to the client, and then register the display client to the avsink core. Then it may look up if there is any
>   audio client registered, by name or type, and may find an audio client registered by some audio driver.
> * If an audio driver supports avsink, it usually should look up a registered display client by name or type at first,
>   because it may need the shared power well in GPU and check the display outputs' name to bind the audio inputs. If
>   the display client is not registered yet, the audio driver can choose to wait (maybe in a work queue) or return
>   -EAGAIN for a deferred probe. After the display client is found, the audio driver can register an audio client with
>   the display client's name as the peer name, the avsink core will bind the display and audio clients to each other.
> Open question:
> If the display or audio driver is disabled by the black list, shall we introduce a time out to avoid waiting for the
> other client registered endlessly?

If the hdmi/dp side is a separate audio instance then we can just defer
forever I think. If that's not the case (i.e. other audio outputs are also
in the same alsa instance) then we need to be able to handle runtime
loading of the gfx driver.

Both cases would work easier I think if we have a real bus and
driver<->device matching.

> 2. Shared power wells (optional)
> The audio and display devices, maybe only part of them, may share a common power well (e.g. for Intel Haswell and
> Broadwell). If so, the driver that controls the power well should define a power well object, implement the get/put ops,
> and add it to its avsink client before registering the client to avsink core. Then the peer client can look up this
> power well by its name, and get/put this power well as a user.
> A client can have multiple power well objects.
> struct avsink_power_well {
>          const char *name; /* name of the power well */
>          void *context;   /* parameter of get/put ops, maybe device pointer for this power well */
>          struct avsink_power_well_ops *ops
> };
> struct avsink_power_well_ops {
>          int (*get)(void *context);
>          int (*put)(void *context);
> };
> API:
> int avsink_new_power(struct avsink_client *client,
>                    const char *power_name,
>                    void * power_context,
>                    struct avsink_power_well_ops *ops,
>                    struct avsink_power_well **power_ret);
> struct avsink_power_well *avisnk_lookup_power(const char *name);
> int avsink_get_power(struct avsink_power_well *power);  /* Reqesut the power */
> int avsink_put_power(struct avsink_power_well *power);    /* Release the power */
> For example, the i915 display driver can create a device for the shared power well in Haswell GPU, implement its PM
> functions, and use the device pointer as the context when creating the power well object, like below
> struct avsink_power_well_ops i915_power_well_ops = {
>          .get = pm_runtime_get_sync;
>          .put = pm_runtime_put_sync;
> };
> ...
> avsink_new_power ( display_client,
>                    "i915_display_power_well",
>                    pdev,  /* pointer of the power well device */
>                    &i915_power_well_ops,
>                    ...)
> Power domain is not used here since a single device seems enough to represent a power well.

Imo the point of the avsink stuff is _not_ to reinvent the wheel again. A
real struct device per endpoint + runtime pm should be able to do
everything we want.

> 3. Display output and audio input endpoints
> A display client should register the display output endpoints and its audio peer client should register the audio input
> endpoints. A client can have multiple endpoints. The avsink core will bind an audio input and a display output as peer
> to each other. This is to allow the audio and display driver to synchronize with each other for each display pipeline.
> All endpoints should be added to a client before the client is registered to avsink core. Dynamic endpoints are not
> supported now.
> A display out here represents a physical HDMI/DP output port. And as long as it's usable in the system (i.e. physically
> connected to the HDMP/DP port on the machine board), the display output should be registered not matter the port is
> connected to an external display device or not. And if HW and display driver can support DP1.2 daisy chain (multiple DP
> display devices can be connected to a single port), multiple static displays outputs should be defined for the DP port
> according to the HW capability. The port & display device number can be indicated by the name (e.g. "i915_DDI_B",
> "i915_DDI_B_DEV0", "i915_DDI_B_DEV1", or "i915_DDI_B_DEV2"), defined by the display driver.
> The audio driver can check the endpoints of its peer display client and use an display endpoint's name, or a presumed
> display endpoint name, as peer name when registering an audio endpoint, thus the avsink core will bind the two display
> and audio endpoints as peers.
> struct avsink_endpoint {
>          const char *name;  /*name of the endpoint */
>          int type;            /* DISPLAY_OUTPUT or AUDIO_INPUT */
>          void *context;           /* private data, used as parameter of the ops */
>          struct avsink_endpoint_ops *ops;
>          struct avsink_endpoint *peer; /* peer endpoint */
> };
> struct avsink_endpoint_ops {
>          int (*get_caps) (enum had_caps_list query_element,
>                             void *capabilities,
>                             void *context);
>          int (*set_caps) (enum had_caps_list set_element,
>                             void *capabilities,
>                             void *context);
>          int (*event_handler) (enum avsink_event_type event_type, void *context);
> };

Ok, this is confusing since get/set_caps are implemented by the gfx side.
The event handler otoh is implemented by the audio side. This needs to be
split up.

With a full device model the set/get stuff would be attached to the device
while the event handler would be part of the driver.

> API:
> int avsink_new_endpoint (struct avsink_client *client,
>                             const char *name,
>                             int type, /* DISPLAY_OUTPUT or AUDIO_INPUT*/
>                             void *context,
>                             const char *peer_name, /* can be NULL if no clue */
>                             avsink_endpoint_ops *ops,
>                             struct avsink_endpoint **endpoint_ret);
> int avsink_endpoint_get_caps(struct avsink_endpoint *endpoint,
>                             enum avsink_caps_list t get_element,
>                             void *capabilities);
> int avsink_endpoint_set_caps(struct avsink_endpoint *endpoint,
>                             enum had_caps_list set_element,
>                             void *capabilities);
> int avsink_endpoint_post_event(struct avsink_endpoint *endpoint,
>                             enum avsink_event_type event_type);
> 4. Get/Set caps on an endpoint
> The display or audio driver can get or set capabilities on an endpoint. Depending on the capability ID, the avsink core
> will call get_caps/set_caps ops of this endpoint, or call get_caps/set_caps ops of its peer endpoint and return the
> result to the caller.
> enum avsink_caps_list {
>          /* capabilities for display output endpoints */
>          AVSINK_GET_DISPLAY_ELD = 1,
>          AVSINK_GET_DISPLAY_TYPE, /* HDMI or DisplayPort */
>          AVSINK_GET_DISPLAY_NAME, /* Hope to use display device name under /sys/class/drm, like "card0-DP-1", for user
>                                      * space to figure out which HDMI/DP output on the drm side corresponds to which audio
>                                        * stream device on the alsa side */
>          AVSINK_GET_DISPLAY_SAMPLING_FREQ,      /* HDMI TMDS clock or DP link symbol clock, for audio driver to
>                                                          * program N value
>                                                         */
>          AVSINK_GET_DISPLAY_AUDIO_STATUS, /* Whether audio is enabled */
>          AVSINK_SET_DISPLAY_ENABLE_AUDIO, /* Enable audio */
>          AVSINK_SET_DISPLAY_DISABLE_AUDIO,         /* Disable audio */
>          AVSINK_SET_DISPLAY_ENABLE_AUDIO_INT, /* Enable audio interrupt */
>          AVSINK_SET_DISPLAY_DISABLE_AUDIO_INT,         /* Disable audio interrupt */
>          /* capabilities for audio input endpoints */
>          AVSINK_GET_AUDIO_IS_BUSY,  /* Whether there is an active audio streaming */
>          OTHERS_TBD,
> };

I really don't like caps based apis. It's imo much better to have specific
set/get functions. Also a lot of this could be passed to more specific
event handlers directly (like the eld or the sampling freq).

If you have a void* somewhere in your interface you're throwing out an
awful lot of safety checks gcc provides. Which is not good.

> For example, the audio driver can query ELD info on an audio input endpoint by using caps AVSINK_GET_DISPLAY_ELD, and
> avsink core will call get_caps() on the peer display output endpoint and return the ELD info to the audio driver.
> Some audio driver may only use part of these caps. E.g. HD-Audio driver can use bus commands instead of the ops to
> control the audio on gfx side, so it doesn't use caps like ENABLE/DISABLE_AUDIO or ENABLE/DISABLE_AUDIO.
> When the display driver want to disable a display pipeline for hot-plug, mode change or power saving, it can use caps
> AVSINK_GET_AUDIO_IS_BUSY to check if the audio input is busy (active streaming) on this display pipeline. And if audio
> is busy, the display driver can choose to wait or go ahead to disable display pipeline anyway. For the latter case, the
> audio input endpoint will be notified by an event and should abort audio streaming.
> 5. Event handling of endpoints
> A driver can post events on an endpoint. Depending on the event type, the avsink core will call the endpoint's event
> handler or pass the event to its peer endpoint and trigger the peer's event handler function if defined.
> int avsink_endpoint_post_event(struct avsink_endpoint *endpoint,
>                                      enum avsink_event_type event_type);
> Now we only defined event types which should be handled by the audio input endpoints. The event types can be extended
> in the future.
> enum avsink_event_type {
>         AVSINK_EVENT_DISPLAY_DISABLE = 1,  /* The display pipeline is disabled for hot-plug, mode change or
>                                                          * suspend. Audio driver should stop any active streaming.
>                                                         */
>         AVSINK_EVENT_DISPLAY_ENABLE,                   /* The display pipeline is enabled after hot-plug, mode change or
>                                                          * resume. Audio driver can restore previously interrupted streaming
>                                                          */
>         AVSINK_EVENT_DISPLAY_MODE_CHANGE,   /* Display mode change event. At this time, the new display mode is
>                                                          * configured but the display pipeline is not enabled yet. Audio driver
>                                                         * can do some configurations such as programing N value */
>         AVSINK_EVENT_DISPLAY_AUDIO_BUFFER_DONE,         /* Audio Buffer done interrupts. Only for audio drivers if DMA and
>                                                          * interrupt are handled by GPU
>                                                         */
>         AVSINK_EVENT_DISPLAY_AUDIO_BUFFER_UNDERRUN,       /* Audio Buffer under run interrupts. Only for audio drivers if
>                                                                   * DMA and interrupt are handled by GPU
>                                                                  */
> };
> So for a display driver, it can post an event on a display output endpoint and get processed by the peer audio input
> endpoint. Or it can also directly post an event on a peer audio input endpoint, by using the 'peer' pointer on a
> display output endpoint.

Again I don't like the enumeration, much better to have a bunch of
specific callbacks. They can also supply interesting information to the
driver directly if instead of audio driver needing to jump through a few
get/set hooks.

> 6. Display register operation (optional)
> Some audio driver needs to access GPU audio registers. The register ops are provided by the peer display client.
> struct avsink_registers_ops {
>          int (*read_register) (uint32_t reg_addr, uint32_t *data, void *context);
>          int (*write_register) (uint32_t reg_addr, uint32_t data, void *context);
>          int (*read_modify_register) (uint32_t reg_addr, uint32_t data, uint32_t mask, void *context);
> int avsink_define_reg_ops (struct avsink_client *client, struct avsink_registers_ops *ops);
> And avsink core provides API for the audio driver to access the display registers:
> int avsink_read_display_register(struct avsink_client *client , uint32_t offset, uint32_t *data);
> int avsink_write_display_register(struct avsink_client *client , uint32_t offset, uint32_t data);
> int avsink_read_modify_display_register(struct avsink_client *client, uint32_t offset, uint32_t data, uint32_t mask);
> If the client is an audio client, the avsink core will find it peer display client and call its register ops;
> and if the client is a display client, the avsink core will just call its own register ops.

Oh dear. Where do we need this? Imo this is really horrible, but if we
indeed need this then a struct device is better - we can attach mmio
resources to devices and let the audio side remap them as best as they see
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

More information about the Alsa-devel mailing list