Audio Driver Streaming Interface

This document describes the audio streaming interface exposed by audio drivers in Zircon. It is meant to serve as a reference for both users and driver-authors, and to unambiguously define the interface contract which drivers must implement and users must follow.

Overview

Audio streams are device nodes published by driver services intended to be used by applications in order to capture or render audio on a Zircon device, or both. Each stream in the system (input or output) represents a stream of digital audio information which may be either received or transmitted by device. Streams are dynamic and may created or destroyed by the system at any time. Which streams exist at any given point in time, and what controls their lifecycles are considered to be issues of audio policy and codec management and are not discussed in this document. Additionally, the information present in audio outputs streams is exclusive to the application owner of the stream. Mixing of audio is not a service provided by the audio stream interface.

TODO(35523): The pre-FIDL serialization still in use as of 2020/02/04 is defined in audio_streaming_original.md, that file is a copy for reference of this file at 34d6afd26d0. Remove this paragraph once conversion of all users to FIDL is done.

TODO: extend this interface to support the concept of low-latency hardware mixers.

Definitions

Term Definition
Sample A representation of the sound rendered by a single speaker, or captured by a single microphone, at a single instant in time.
LPCM Linear pulse code modulation. The specific representation of audio samples present in all Zircon uncompressed audio streams. LPCM audio samples are representations of the amplitude of the audio signal at an instant in time where the numeric values of the encoded audio are linearly distributed across the amplitude levels of the rendering or capture device. This is in contrast to A-law and μ-law encodings which have non-linear mappings from numeric value to amplitude level.
Channel Within an audio stream, the subset of information which will be rendered by a single speaker, or which was captured by a single microphone in a stream.
Frame A set of audio samples for every channel of a audio stream captured/rendered at a single instant in time.
Frame Rate a.k.a. "Sample Rate". The rate (in Hz) at which audio frames are produced or consumed. Common sample rates include 44.1 KHz, 48 KHz, 96 KHz, and so on.
Client or User or Application These terms are used interchangeably in this document. They refer to modules that use these interfaces to communicate with an audio driver/device.

TODO: do we need to extend this interface to support non-linear audio sample encodings? This may be important for telephony oriented microphones which deliver μ-law encoded samples.

Basic Operation

Communication with an audio stream device is performed using messages sent over a channel. Applications open the device node for a stream and obtain a channel by issuing a FIDL request. After obtaining the channel, the device node may be closed. All subsequent communication with the stream occurs using channels.

The stream channel is used for most command and control tasks, including:

  • Capability interrogation
  • Format negotiation
  • Hardware gain control
  • Determining outboard latency
  • Plug detection notification
  • Access control capability detection and signalling

TODO: Should plug/unplug detection be done by sending notifications over the stream channel (as it is today), or by publishing/unpublishing the device nodes (and closing all channels in the case of unpublished channels)?

In order to actually send or receive audio information on the stream, the specific format to be used must first be set. The response to a successful CreateRingBuffer operation will contain a new "ring-buffer" channel. The ring-buffer channel may be used to request a shared buffer from the stream (delivered in the form of a VMO) which may be mapped into the address space of the application and used to send or receive audio data as appropriate. Generally, the operations conducted over the ring buffer channel include:

  • Requesting a shared buffer
  • Starting and Stopping stream playback and capture
  • Receiving notifications of playback and capture progress
  • Receiving clock recovery information in the case that the audio output clock is based on a different oscillator than the oscillator which backs ZX_CLOCK_MONOTONIC

Operational Details

Device nodes

Audio stream device nodes must be published by drivers using the protocol preprocessor symbol given in the table below. This will cause stream device nodes to be published in the locations given in the table. Applications can monitor these directories in order to discover new streams as they are published by the drivers.

Stream Type Protocol Location
Input ZX_PROTOCOL_AUDIO_INPUT /dev/class/audio-input-2
Output ZX_PROTOCOL_AUDIO_OUTPUT /dev/class/audio-output-2

Establishing the stream channel

After opening the device node, client applications may obtain a stream channel for subsequent communication using the fuchsia.hardware.audio.Device/GetChannel FIDL message.

Client side termination of the stream channel

Clients may terminate the connection to the stream at any time simply by calling zx_handle_close(...) on the stream channel. Drivers must close any active ring-buffer channels established using this stream channel and must make every attempt to gracefully quiesce any on-going streaming operations in the process.

Sending and receiving messages on the stream and ring-buffer channels

All of the messages and message payloads which may be sent or received over stream and ring buffer channels are defined in stream.fidl and ring_buffer.fidl. Messages may be sent to the driver using the zx_channel_write(...) syscall. If a response is expected, it may be read using the zx_channel_read(...) syscall. Best practice, however, is to queue packets for your channel(s) port using the zx_port_queue(...) syscall, and use the zx_port_wait(...) syscall to determine when your set of channels have messages (either expected responses or asynchronous notifications) to be read. There are bindings for different languages to facilitate sending and receiving FIDL messages, and in particular for C++ drivers there is also a library SimpleAudioStream that facilitates the creation of drivers in C++, this library uses the LLCPP bindings to send and receive FIDL messages.

Format Negotiation

Sample Formats

The Format related protocol messages allow the driver to list its supported formats to the client. The supported formats may include multiple rates, bit per sample, etc. Each driver advertises what it can support and the client mandates what format is to be used for each driver.

To find out what formats are supported by a given driver, the client uses the GetSupportedFormats function. The driver replies with a vector of SupportedFormats, where each SupportedFormats includes a PcmSupportedFormats with:

  • A vector of number of channels. This lists the number of channels supported by the driver, for example <2,4,6,8>. A driver that supports either two or four channels would report a vector with two elements <2,4>. Must be in ascending order.
  • A vector of sample formats, e.g. PCM_SIGNED.
  • A vector of rates. Frame rates, for example 44100, 48000, and 96000. Must be in ascending order.
  • A number of bits per channel/slot/container. Number of bits in each channel allocated to hold a sample, e.g. 32 bits per channel. Must be in ascending order.
  • A vector of bits per sample. Sample widths, this could be smaller than the channel e.g. 24 bits per sample in a 32 bits channel. Must be in ascending order.

When not all combinations supported by the driver can be described with one PcmSupportedFormats, the driver returns more than one PcmSupportedFormats in the returned vector. For example, if one PcmSupportedFormats allows for 16 or 32 bits samples at 48KHz, and 16 bits samples at 96KHz, but not 32 bits samples at 96KHz, then the driver replies with 2 PcmSupportedFormats: <<16bits,32bits>,<48KHz>> and <<16bits>,<96KHz>>. For simplicity, this example ignores parameters other than rate and bits per sample. In the case where the driver supports either 16 or 32 bits samples at either 48 or 96KHz, the driver would reply with 1 PcmSupportedFormats: <<16bits,32bits>,<48KHz,96KHz>>.

Additionally, it is assumed that bits per sample is always smaller or equal to bits per channel. Hence, a driver can report <<16bits_per_channel,32bits_per_channel>,<16bits_per_sample,32bits_per_sample>> and this does not imply that it is reporting that 32 bits per sample on 16 bits samples is valid, it specifies only the 3 valid combinations:

  • 16 bits channels with 16 bits samples
  • 32 bits channels with 32 bits samples
  • 32 bits channels with 16 bits samples

The client specifies the format to use with the CreateRingBuffer function based on information that the driver provides in GetSupportedFormats reply, what is supported by the client, and any other requirements. This function takes a parameter that specifies:

  • A number of channels. This is the number of channels available in the buffer.
  • A bitmask of channels to use. These are the channels in the buffer to be used by the driver. For instance for stereo this must be a bitmask with 2 bits enabled 0x3, i.e. both channels 0 and 1 are used.
  • A sample format.
  • A frame rate.
  • A number of bits per channel.
  • A number of bits per sample.

Notes:

  • By default, multi-byte sample formats are assumed to use host-endianness.
  • The PCM_FLOAT encoding uses specifically the IEEE 754 floating point representation.
  • By default, non-floating point PCM encodings are assumed expressed using two's complement signed integers. eg. the bit values for a 16 bit PCM sample format would range from [0x8000, 0x7FFF] with 0x0000 representing zero speaker deflection. If the PCM_UNSIGNED sample format is used, the bit values would range from [0x0000, 0xFFFF] with 0x8000 representing zero deflection.
  • When encoding a smaller sample size in a larger channel (e.g. 20 or 24bit in 32), the most significant bits of the 32 bit container are used while the least significant bits will be ignored (left justified). e.g. a 20 bit sample would be mapped onto the range 12,31 of the 32 bit container.

Setting the desired stream format

In order to select a stream format, applications send a CreateRingBuffer message over the stream channel. In the message, the application specifies the format to be used.

The client specifies the new ring buffer channel over which streaming operations will be conducted. If a previous ring buffer channel had been established and was still active, the driver must close this channel and make every attempt to gracefully quiesce any on-going streaming operations in the process.

TODO: specify how compressed bitstream formats will be set

Determining external latency

The external latency of an audio stream is defined as the amount of time it takes outbound audio to travel from the system's interconnect to the speakers themselves, or inbound audio to travel from the microphone to the system's interconnect. As an example, consider an external codec connected to the system using a TDM interconnect: if this interconnect introduces a 4 frame delay between the reception of a TDM frame and the rendering of that frame at the speakers themselves, then the external delay of this audio path is the time duration equivalent to 4 audio frames.

External delay is reported in the external_delay field of a RingBufferProperties response to a GetProperties. Drivers should make their best attempt to accurately report the total of all of the sources of delay the driver knows about. Information about this delay can frequently be found in codec data sheets, dynamically reported as properties of codecs using protocols such as Intel HDA or the USB Audio specifications, or reported by down stream devices using mechanisms such as EDID when using HDMI or DisplayPort interconnects.

Hardware Gain Control

Hardware gain control capability reporting

In order to determine a stream's gain control capabilities, if it has not done so yet, an application sends a GetProperties message over the stream channel. No parameters need to be supplied with this message. The driver replies with a StreamProperties including gain capabilities among others. All stream drivers must respond to this message, regardless of whether or not the stream hardware is capable of any gain control. All gain values are expressed using 32 bit floating point numbers expressed in dB.

Drivers respond to this message with values which indicate the current stream's gain control capabilities. Current gain settings are expressed using a bool indicating whether the stream can be muted, a bool that indicates whether the stream can AGC, the minimum and maximum gain settings, and a gain_step_db. The gain_step_db indicates the smallest increment with which the gain can be controlled counting from the minimum gain value.

For example, an amplifier which has 5 gain steps of 7.5 dB each and a maximum 0 dB gain would indicate a range of (-30.0, 0.0) and a step size of 7.5. Amplifiers capable of functionally continuous gain control may encode their gain step size as 0.0.

Regardless of mute capabilities, drivers for fixed gain streams must report their min and max gain as (0.0, 0.0). gain_step_db is meaningless in this situation, but drivers should report it as 0.0.

Setting hardware gain control levels

In order to change a stream's current gain settings, applications send a SetGain message over the stream channel. This message include a parameter GainState indicating gain parameters to be configured including the dB gain which should be applied to the stream, muted and AGC enablement.

Presuming that the request is valid, drivers should round the request to the nearest supported gain step size. For example, if a stream can control its gain on the range from -60.0 to 0.0 dB, using a gain step size of 0.5 dB, then a request to set the gain to -33.3 dB should result in a gain of -33.5 being applied. A request to that same stream for a gain of -33.2 dB should result in a gain of -33.0 being applied.

Gain state notifications

Clients may request that streams send them asynchronous notifications of gain state changes by using the WatchGainState command. The driver will reply to the first |WatchGainState| sent by the client and will not respond to subsequent client |WatchGainState| calls until the gain state changes from what was most recently reported.

Plug detection

In addition to streams being published/unpublished in response to being connected or disconnected to/from their bus, streams may have the ability to be plugged or unplugged at any given point in time. For example, a set of USB headphones may publish a new output stream when connected to USB, but choose to be "hardwired" from a plug detection standpoint. A different USB audio adapter with a standard 3.5mm phono jack might publish an output stream when connected via USB, but choose to change its plugged/unplugged state as the user plugs and unplugs an analog device via the 3.5mm jack.

The ability to query the currently plugged or unplugged state of a stream, and to register for asynchonous notifications of plug state changes (if supported) is handled via plug detection messages.

Plug detect capabilities

In order to determine a stream's plug detection capabilities, if it has not done so yet, an application sends a GetProperties command over the stream channel. The driver replies with a StreamProperties including plug detect capabilities in plug_detect_capabilities among others fields.

Valid plug-detect capabilities flags currently defined are:

  • HARDWIRED is set when the stream hardware is considered to be "hardwired". In other words, the stream is considered to be connected as long as the device is published. Examples include a set of built-in speakers, a pair of USB headphones, or a pluggable audio device with no plug detection functionality.
  • CAN_ASYNC_NOTIFY is set when the stream hardware is capable of both asynchronously detecting that a device's plug state has changed, and sending a notification message if the client has requested these notifications.

Plug state notifications

Clients may request that streams send them asynchronous notifications of plug state changes by using the WatchPlugState command if the CAN_ASYNC_NOTIFY flag was sent by the driver in StreamProperties. I.e. drivers for streams which do not set the CAN_ASYNC_NOTIFY flag are free to ignore the WatchPlugState sent by applications. Driver with CAN_ASYNC_NOTIFY set will reply to the first |WatchPlugState| sent by the client and will not respond to subsequent client |WatchPlugState| calls until the plug state changes from what was most recently reported.

Access control capability detection and signaling

TODO: specify how this works. In particular, specify how drivers indicate to applications support for various digital access control mechanisms such as S/PDIF control words and HDCP.

Stream purpose and association

TODO: specify how drivers can indicate the general "purpose" of an audio stream in the system (if known), as well as its relationship to other streams (if known). For example, an embedded target like a phone or a tablet needs to indicate which output stream is the built-in speaker vs. which is the headset jack output. In addition, it needs to make clear which input stream is the microphone associated with the headset output vs. the builtin speaker.

Ring-Buffer Channels

Overview

Once an application has successfully set the format of a stream, it receives in the response a new channel representing its connection to the stream's ring-buffer. Clients use the ring-buffer channel to establish a shared memory buffer and start and stop playback and capture of audio stream data.

The ring buffer contents are produced by the client side (for playback) and the driver side (for recording). Hence, a client is a producer for playback and a consumer for recording and a driver is a producer for recording and a consumer for playback. The ring buffer contents may be directly consumed or produced by the audio hardware, or it may go through software processing of each sample done by the driver.

Ring buffer data production proceeds at the nominal rate from the point in time given in a successful response to the Start command. Note though that the ring-buffer will almost certainly have some form of FIFO buffer between the memory bus and the audio hardware which causes it to either read-ahead in the stream (in the case of playback), or potentially hold onto data (in the case of capturing). It is important for clients to query the size of this buffer before beginning operation so they know how far ahead/behind the stream's nominal inferred read/write position they need to stay in order to prevent audio glitching. Also note that because of the shared buffer nature of the system, and the fact that drivers are likely to be DMA-ing directly from this buffer to hardware, it is important for clients running on architectures which are not automatically cache coherent to be sure that they have properly written-back their cache after writing playback data to the buffer, or invalidated their cache before reading captured data.

Determining the FIFO depth

In order to determine a stream's fifo_depth, if it has not done so yet, an application sends a GetProperties command over the stream channel. The driver replies with a StreamProperties including fifo_depth among others fields. To ensure proper playback or capture of audio, applications and drivers must be careful to respect this value. Drivers must not read beyond the nominal playback position of the stream plus this number of bytes when playing audio stream data. Applications must stay this number of bytes behind the nominal capture point of the stream when capturing audio stream data.

The ring buffer data may be directly consumed/generated by hardware, in this case fifo_depth maps directly to the size of a hardware FIFO block, since the hardware FIFO block determines the amount of data read ahead or held back.

The ring buffer data may instead be consumed/generated by audio driver software that is conceptually situated between the ring buffer and the audio hardware. In this case, for playback the fifo_depth read ahead amount is set large enough such that the driver guarantees no undetected underruns, this assuming the client is generating the data as determined by the CreateRingBuffer and Start commands. For capture, the fifo_depth held back amount is set large enough such that the driver guarantees no undetected underruns when generating the data as determined by the CreateRingBuffer and Start commands. The driver must set fifo_depth big enough such that the potential delays added by any software interfacing with the audio hardware do not occur under most scenarios, and must detect and report underruns. How an underrun is reported is not defined here.

Once the format of a stream is set and a ring-buffer channel has been opened, the driver must not change this value. From an application's point of view, it is a constant property of the ring-buffer channel.

Obtaining a shared buffer

To send or receive audio, the application must first establish a shared memory buffer. This is done by sending an CreateRingBuffer request over the ring-buffer channel. This may only be done while the ring-buffer is stopped.

If the channel created via CreateRingBuffer is closed by the driver for instance because a buffer has already been established and the ring-buffer has already been started, it must not either stop the ring-buffer, or discard the existing shared memory. If the application requests a new buffer after having already established a buffer while the ring buffer is stopped, it must consider the existing buffer ii has to be invalid, the old buffer is now gone.

Applications must specify two parameters when requesting a ring buffer: min_frames and clock_recovery_notifications_per_ring.

min_frames

The minimum number of frames of audio the client needs allocated for the ring buffer. Drivers may make this buffer larger to meet hardware requirements. Clients must use the returned VMOs size (in bytes) to determine the actual size of the ring buffer. Clients must not assume that the size of the buffer (as determined by the driver) is exactly the size they requested. Drivers must ensure that the size of the ring buffer is an integral number of audio frames.

TODO : Is it reasonable to require that drivers produce buffers which are an integral number of audio frames in length? It certainly makes the audio client's life easier (client code never needs to split or re-assemble a frame before processing), but it might make it difficult for some audio hardware to meet its requirements without making the buffer significantly larger than the client asked for.

clock_recovery_notifications_per_ring

Optional number of position update notifications the client would like the driver to send per cycle through the ring buffer, these notifications are meant to be used for clock recovery. Drivers must only send these as a reply to a WatchClockRecoveryPositionInfo request. Drivers should attempt to space notifications uniformly throughout the ring; however clients must not rely on perfectly uniform spacing of the update notifications.

ring_buffer

If the request succeeds, the driver must return a handle to a VMO with permissions which allow applications to map the VMO into their address space using zx_vmar_map, and to read/write data in the buffer in the case of playback, or simply to read the data in the buffer in the case of capture.

num_frames

If the request succeeds, the driver will also return the actual number of frames of audio it will use in the buffer. The size of the VMO returned (as reported by zx_vmo_get_size()) must not be larger than this number of frames (when converted to bytes). This number may be larger than the min_frames request from the client but must not be smaller than this number.

Starting and Stopping the ring-buffer

Clients may request that a ring-buffer start or stop using the Start and Stop commands. Attempting to start a stream which is already started must be considered a failure. Attempting to stop a stream which is already stopped should be considered a success. Ring-buffers cannot be either stopped or started until after a shared buffer has been established using the CreateRingBuffer operation.

Upon successfully starting a stream, drivers must provide their best estimate of the time at which their hardware began to transmit or capture the stream in the start_time field of the response. This time stamp must be taken from the clock exposed via the ZX_CLOCK_MONOTONIC syscall. Along with the FIFO depth property of the ring buffer, this timestamp allows applications to send or receive stream data without the need for periodic position updates from the driver. Along with the outboard latency estimate provided by the stream channel, this timestamp allows applications to synchronize presentation of audio information across multiple streams, or even multiple devices (provided that an external time synchronization protocol is used to synchronize the ZX_CLOCK_MONOTONIC timelines across the cohort of synchronized devices).

TODO: Redefine start_time to allow it to be an arbitrary 'audio stream clock' instead of the ZX_CLOCK_MONOTONIC clock. If the stream clock is made to count in audio frames since start, then this start_time can be replaced with the terms for a segment of a piecewise linear transformation which can be subsequently updated via notifications sent by the driver in the case that the audio hardware clock is rooted in a different oscillator from the system's tick counter. Clients can then use this transformation either to control the rate of consumption of input streams, or to determine where to sample in the input stream to effect clock correction.

Upon successfully starting a stream, drivers must guarantee that no position notifications will be sent before the start response has been enqueued into the ring-buffer channel.

Upon successfully stopping a stream, drivers must guarantee that no position notifications will be enqueued into the ring-buffer channel after the stop response has been enqueued.

Position notifications

If requested by the client via a non-zero clock_recovery_notifications_per_ring in the CreateRingBuffer operation, the driver will periodically send updates to the client informing it of its current production or consumption position in the buffer. This position is expressed in bytes in the position field of the RingBufferPositionInfo struct sent on a reply to the WatchClockRecoveryPositionInfo message. The message also includes a timestamp field that contains the time (as zx::time) that this byte position was valid. WatchClockRecoveryPositionInfo replies must only be sent while the ring-buffer is started. Note, these position notifications indicate where in the buffer the driver has consumed or produced data, not the nominal playback or capture position (sometimes called the "write cursor" or "read cursor" respectively). The timing of their arrival is not guaranteed to be perfectly uniform and should not be used to effect clock recovery. However, the correspondence pair (timestamp, position) values themselves ARE intended to be used to recover the clock for the audio stream. If a client discovers that a driver has consumed past the point in the ring buffer where that client has written playback data, audio presentation is undefined. Clients should increase their clock lead time and be certain to stay ahead of this point in the stream in the future. Likewise, clients which capture audio should not attempt to read beyond the point in the ring buffer indicated by the most recent position notification sent by the driver.

Driver playback and capture position must always begin at ring buffer byte 0, immediately following a successful Start command. When the ring buffer position reaches the end of the VMO (as indicated by zx_vmo_get_size(...)), the ring buffer position wraps back to zero. Drivers are not required to consume or produce data in integral numbers of audio frames. Clients whose notion of stream position depends on position notifications should take care to request that a sufficient number of notifications per ring be sent (minimum 2) and to process them quickly enough that aliasing does not occur.

Clock recovery and synchronization

Upon receiving AUDIO_STREAM_CMD_GET_CLOCK_DOMAIN message, the driver must respond with the identifier of the clock domain containing that device. If the audio device is locked to the local system monotonic clock and does not expose a mechanism by which its rate would be fine-tuned, then it should return the value 0 to represent the local CLOCK_MONOTONIC domain. A client may use this information (in addition to AUDIO_RB_POSITION_NOTIFY messages) to simplify the process of recovering the audio device's clock.

TODO: extend this section to include how clock recovery occurs, and how this is exposed to clients. Also, detail how slewable oscillators are discovered and controlled. We may need rate-change notifications to clients of slewable clocks.

Previous content: TODO: define a way that clock recovery information can be sent to clients in the case that the audio output oscillator is not derived from the ZX_CLOCK_MONOTONIC oscillator. In addition, if the oscillator is slew-able in hardware, provide the ability to discover this capability and control the slew rate. Given the fact that this oscillator is likely to be shared by multiple streams, it might be best to return some form of system wide clock identifier and provide the ability to obtain a channel on which clock recovery notifications can be delivered to clients and hardware slewing command can be sent from clients to the clock.

Error notifications

TODO: define these and what driver behavior should be, if/when they occur.

Unexpected client termination

If the client side of a ring buffer control channel is closed for any reason, drivers must immediately close the control channel and shut down the ring buffer, such that no further audio is emitted nor captured. While drivers are encouraged to do so in a way which produces a graceful transition to silence, they must ensure that the audio stream goes silent instead of looping. Once the transition to silence is complete, resources associated with playback or capture may be released and reused by the driver.

This way, if a playback client teminates unexpectedly, the system will close the client channels, causing audio playback to stop instead of continuing to loop.