Opening and closing an audio device

First the application must call the sio_open() function to obtain a handle to the device; later it will be passed as the hdl argument of most other functions. The name parameter gives the device string discussed in sndio(7). In most cases it should be set to SIO_DEVANY to allow the user to select it using the AUDIODEVICE environment variable.

The following values of the mode parameter are supported:

SIO_PLAY

Play-only mode: data written will be played by the device.

SIO_REC

Record-only mode: samples are recorded by the device and must be read.

SIO_PLAY | SIO_REC

The device plays and records synchronously; this means that the n-th recorded sample was physically sampled exactly when the n-th played sample was actually played.

If the nbio_flag argument is true (i.e. non-zero), then the sio_read() and sio_write() functions (see below) will be non-blocking.

The sio_close() function stops the device as if sio_stop() is called and frees the handle. Thus, no samples submitted with sio_write() are discarded.

Negotiating audio parameters

Audio samples are interleaved. A frame consists of one sample for each channel. For example, a 16-bit stereo encoding has two samples per frame and, two bytes per sample (thus 4 bytes per frame).

The set of parameters of the device that can be controlled is given by the following structure:

struct sio_par {
	unsigned int bits;	/* bits per sample */
	unsigned int bps;	/* bytes per sample */
	unsigned int sig;	/* 1 = signed, 0 = unsigned int */
	unsigned int le;	/* 1 = LE, 0 = BE byte order */
	unsigned int msb;	/* 1 = MSB, 0 = LSB aligned */
	unsigned int rchan;	/* number channels for recording */
	unsigned int pchan;	/* number channels for playback */
	unsigned int rate;	/* frames per second */
	unsigned int appbufsz;	/* minimum buffer size without xruns */
	unsigned int bufsz;	/* end-to-end buffer size (read-only) */
	unsigned int round;	/* optimal buffer size divisor */
#define SIO_IGNORE	0	/* pause during xrun */
#define SIO_SYNC	1	/* resync after xrun */
#define SIO_ERROR	2	/* terminate on xrun */
	unsigned int xrun;	/* what to do on overrun/underrun */
};

The parameters are as follows:

bits

Number of bits per sample: must be between 1 and 32.

bps

Bytes per samples; if specified, it must be large enough to hold all bits. By default it's set to the smallest power of two large enough to hold bits.

sig

If set (i.e. non-zero) then the samples are signed, else unsigned.

le

If set, then the byte order is little endian, else big endian; it's meaningful only if bps > 1.

msb

If set, then the bits are aligned in the packet to the most significant bit (i.e. lower bits are padded), else to the least significant bit (i.e. higher bits are padded); it's meaningful only if bits < bps * 8.

rchan

The number of recorded channels; meaningful only if SIO_REC mode was selected.

pchan

The number of played channels; meaningful only if SIO_PLAY mode was selected.

rate

The sampling frequency in Hz.

bufsz

The maximum number of frames that may be buffered. This parameter takes into account any buffers, and can be used for latency calculations. It is read-only.

appbufsz

Size of the buffer in frames the application must maintain non-empty (on the play end) or non-full (on the record end) by calling sio_write() or sio_read() fast enough to avoid overrun or underrun conditions. The audio subsystem may use additional buffering, thus this parameter cannot be used for latency calculations.

round

Optimal number of frames that the application buffers should be a multiple of, to get best performance. Applications can use this parameter to round their block size.

xrun

The action when the client doesn't accept recorded data or doesn't provide data to play fast enough; it can be set to one of the SIO_IGNORE, SIO_SYNC, or SIO_ERROR constants.

The following approach is recommended to negotiate device parameters:

• Initialize a sio_par structure using sio_initpar() and fill it with the desired parameters. Then call sio_setpar() to request the device to use them. Parameters left unset in the sio_par structure will be set to device-specific defaults.
• Call sio_getpar() to retrieve the actual parameters of the device and check that they are usable. If they are not, then fail or set up a conversion layer. Sometimes the rate set can be slightly different to what was requested. A difference of about 0.5% is not audible and should be ignored.

Parameters cannot be changed after sio_start() has been called, sio_stop() or sio_flush() must be called before parameters can be changed.

If the device is exposed by the sndiod(8) server, which is the default configuration, a transparent emulation layer will automatically be set up, and in this case any combination of rate, encoding and numbers of channels is supported.

To ease filling the sio_par structure, the following macros can be used:

SIO_BPS(bits)

Return the smallest value for bps that is a power of two and that is large enough to hold bits.

SIO_LE_NATIVE

Can be used to set the le parameter when native byte order is required. It is 1 if the native byte order is little endian or 0 otherwise.

Getting device capabilities

There's no way to get an exhaustive list of all parameter combinations the device supports. Applications that need to have a set of working parameter combinations in advance can use the sio_getcap() function. However, for most new applications it's generally not recommended to use sio_getcap(). Instead, follow the recommendations for negotiating device parameters (see above).

The sio_cap structure contains the list of parameter configurations. Each configuration contains multiple parameter sets. The application must examine all configurations, and choose its parameter set from one of the configurations. Parameters of different configurations are not usable together.

struct sio_cap {
	struct sio_enc {		/* allowed encodings */
		unsigned int bits;
		unsigned int bps;
		unsigned int sig;
		unsigned int le;
		unsigned int msb;
	} enc[SIO_NENC];
	unsigned int rchan[SIO_NCHAN];	/* allowed rchans */
	unsigned int pchan[SIO_NCHAN];	/* allowed pchans */
	unsigned int rate[SIO_NRATE];	/* allowed rates */
	unsigned int nconf;		/* num. of confs[] */
	struct sio_conf {
		unsigned int enc;	/* bitmask of enc[] indexes */
		unsigned int rchan;	/* bitmask of rchan[] indexes */
		unsigned int pchan;	/* bitmask of pchan[] indexes */
		unsigned int rate;	/* bitmask of rate[] indexes */
	} confs[SIO_NCONF];
};

The parameters are as follows:

enc[SIO_NENC]

Array of supported encodings. The tuple of bits, bps, sig, le, and msb parameters are usable in the corresponding parameters of the sio_par structure.

rchan[SIO_NCHAN]

Array of supported channel numbers for recording usable in the sio_par structure.

pchan[SIO_NCHAN]

Array of supported channel numbers for playback usable in the sio_par structure.

rate[SIO_NRATE]

Array of supported sample rates usable in the sio_par structure.

nconf

Number of different configurations available, i.e. number of filled elements of the confs[] array.

confs[SIO_NCONF]

Array of available configurations. Each configuration contains bitmasks indicating which elements of the above parameter arrays are valid for the given configuration. For instance, if the second bit of rate is set, in the sio_conf structure, then the second element of the rate[SIO_NRATE] array of the sio_cap structure is valid for this configuration. As such, when reading the array elements in the sio_cap structure, the corresponding sio_conf bitmasks should always be used.

Starting and stopping the device

The sio_start() function prepares the device to start. Once the play buffer is full, i.e. sio_par.bufsz samples are queued with sio_write(), playback starts automatically. If record-only mode is selected, then sio_start() starts recording immediately. In full-duplex mode, playback and recording will start synchronously as soon as the play buffer is full.

The sio_stop() function puts the audio subsystem in the same state as before sio_start() was called. It stops recording, drains the play buffer and then stops playback. If samples to play are queued but playback hasn't started yet then playback is forced immediately; playback will actually stop once the buffer is drained. In no case are samples in the play buffer discarded.

The sio_flush() function stops playback and recording immediately, possibly discarding play buffer contents, and puts the audio subsystem in the same state as before sio_start() was called.

Playing and recording

When record mode is selected, the sio_read() function must be called to retrieve recorded data; it must be called often enough to ensure that internal buffers will not overrun. It will store at most nbytes bytes at the addr location and return the number of bytes stored. Unless the nbio_flag flag is set, it will block until data becomes available and will return zero only on error.

Similarly, when play mode is selected, the sio_write() function must be called to provide data to play. Unless the nbio_flag is set, sio_write() will block until the requested amount of data is written.

Non-blocking mode operation

If the nbio_flag is set on sio_open(), then the sio_read() and sio_write() functions will never block; if no data is available, they will return zero immediately.

The poll(2) system call can be used to check if data can be read from or written to the device. The sio_pollfd() function fills the array pfd of pollfd structures, used by poll(2), with events; the latter is a bit-mask of POLLIN and POLLOUT constants; refer to poll(2) for more details. The sio_revents() function returns the bit-mask set by poll(2) in the pfd array of pollfd structures. If POLLIN is set, recorded samples are available in the device buffer and can be read with sio_read(). If POLLOUT is set, space is available in the device buffer and new samples to play can be submitted with sio_write(). POLLHUP may be set if an error occurs, even if it is not selected with sio_pollfd().

The size of the pfd array, which the caller must pre-allocate, is provided by the sio_nfds() function.

Synchronizing non-audio events to the audio stream in real-time

In order to perform actions at precise positions of the audio stream, such as displaying video in sync with the audio stream, the application must be notified in real-time of the exact position in the stream the hardware is processing.

The sio_onmove() function can be used to register the cb() callback function called at regular time intervals. The delta argument contains the number of frames the hardware played and/or recorded since the last call of cb(). It is called by sio_read(), sio_write(), and sio_revents(). When the first sample is played and/or recorded, right after the device starts, the callback is invoked with a zero delta argument. The value of the arg pointer is passed to the callback and can contain anything.

If desired, the application can maintain the current position by starting from zero (when sio_start() is called) and adding to the current position delta every time cb() is called.

Measuring the latency and buffers usage

The playback latency is the delay it will take for the frame just written to become audible, expressed in number of frames. The exact playback latency can be obtained by subtracting the current position from the number of frames written. Once playback is actually started (first sample audible), the latency will never exceed the bufsz parameter (see the sections above). There's a phase during which sio_write() only queues data; once there's enough data, actual playback starts. During this phase talking about latency is meaningless.

In any cases, at most bufsz frames are buffered. This value takes into account all buffers. The number of frames stored is equal to the number of frames written minus the current position.

The recording latency is obtained similarly, by subtracting the number of frames read from the current position.

Note that sio_write() might block even if there is buffer space left; using the buffer usage to guess if sio_write() would block is false and leads to unreliable programs – consider using poll(2) for this.

Handling buffer overruns and underruns

When the application cannot accept recorded data fast enough, the record buffer (of size appbufsz) might overrun; in this case recorded data is lost. Similarly if the application cannot provide data to play fast enough, the play buffer underruns and silence is played instead. Depending on the xrun parameter of the sio_par structure, the audio subsystem will behave as follows:

SIO_IGNORE

The device pauses during overruns and underruns, thus the current position (obtained through sio_onmove()) stops being incremented. Once the overrun and/or underrun condition is gone, the device resumes; play and record are always kept in sync. With this mode, the application cannot notice underruns and/or overruns and shouldn't care about them.

This mode is the default. It's suitable for applications, like audio players and telephony, where time is not important and overruns or underruns are not short.

SIO_SYNC

If the play buffer underruns, then silence is played, but in order to reach the right position in time, the same amount of written samples will be discarded once the application is unblocked. Similarly, if the record buffer overruns, then samples are discarded, but the same amount of silence will be returned later. The current position (obtained through sio_onmove()) is still incremented. When the play buffer underruns the play latency might become negative; when the record buffer overruns, the record latency might become larger than bufsz.

This mode is suitable for applications, like music production, where time is important and where underruns or overruns are short and rare.

SIO_ERROR

With this mode, on the first play buffer underrun or record buffer overrun, playback and/or recording is terminated and no other function than sio_close() will succeed.

This mode is mostly useful for testing.

Controlling the volume

The sio_setvol() function can be used to set playback attenuation. The vol parameter takes a value between 0 (maximum attenuation) and SIO_MAXVOL (no attenuation). It specifies the weight the audio subsystem will give to this stream. It is not meant to control hardware parameters like speaker gain; the mixerctl(8) interface should be used for that purpose instead.

An application can use the sio_onvol() function to register a callback function that will be called each time the volume is changed, including when sio_setvol() is used. The callback is always invoked when sio_onvol() is called in order to provide the initial volume. An application can safely assume that once sio_onvol() has returned a non-zero value, the callback has been invoked and thus the current volume is available. If there's no volume setting available, sio_onvol() returns 0 and the callback is never invoked and calls to sio_setvol() are ignored.

The sio_onvol() function can be called with a NULL argument to check whether a volume knob is available.

Error handling

Errors related to the audio subsystem (like hardware errors, dropped connections) and programming errors (e.g. call to sio_read() on a play-only stream) are considered fatal. Once an error occurs, all functions taking a sio_hdl argument, except sio_close() and sio_eof(), stop working (i.e. always return 0). The sio_eof() function can be used at any stage.