Skip to contentSkip to conversation

Trouble understanding ALSA's DMA buffers

DISCUSSION Timur Tabi 4 participants 7 messages Jun 11, 2007 – Jun 12, 2007

lore
Share & tools
lore

Timur Tabi

author
#0 lore raw

I'm writing an ALSA SOC driver for an I2S-based device, and I'm having a really hard time understanding how ALSA uses the DMA buffers. And yes, I've read the documentation and studied some sample source code.

I used to write audio drivers for a living, but that was many years ago, and it wasn't for Linux. Perhaps the concepts in my head are outdated, but I just don't see enough explanation as to how DMA buffers are supposed to work.

Back then, audio drivers used "ping pong" DMA buffers. A single DMA buffer is allocated, and the audio hardware is programmed to read from that buffer in a loop. That is, it would automatically restart reading from the beginning of the buffer without any reprogramming. The hardware would also be programmed to issue an interrupt when it got to the end of the buffer, and when it got to the half-way point.

To start playback, you first filled the whole buffer with audio data, and then told the hardware to start playing. After the hardware got to the half-way point, it would issue an interrupt. You would then tell the OS you need more data, and you'd get it. You then copy that data into the first half of the buffer *while* the hardware was playing the second half. Later, the hardware would interrupt you when it got to the end of the buffer. You'd then copy more data to the 2nd half while the hardware is playing the first half.

And so - hardware plays one half while you copy data to the other half. Hence, "ping pong".

So how do I implement this in ALSA? The "Writing an ALSA Driver" document doesn't even contain the words "ping" or "pong".

Signature · 2 lines
Timur Tabi
Linux Kernel Developer @ Freescale
#1 lore raw

Hi,

Now ALSA (audio ) buffer is divided into periods, i.e. a chain of small packets.

periods size is configurable. Data transfer to the codec starts only after reaching start_threshold point (start_threshold is in periods), this time DMA trigger is called.

This trigger onwards application will get notification from the kernel saying that period buffer is empty you can write into it.Till the end of music file.

Nobin Mathew

Quoted 35 lines · Timur Tabi

I'm writing an ALSA SOC driver for an I2S-based device, and I'm having a really hard time understanding how ALSA uses the DMA buffers. And yes, I've read the documentation and studied some sample source code.

I used to write audio drivers for a living, but that was many years ago, and it wasn't for Linux. Perhaps the concepts in my head are outdated, but I just don't see enough explanation as to how DMA buffers are supposed to work.

Back then, audio drivers used "ping pong" DMA buffers. A single DMA buffer is allocated, and the audio hardware is programmed to read from that buffer in a loop. That is, it would automatically restart reading from the beginning of the buffer without any reprogramming. The hardware would also be programmed to issue an interrupt when it got to the end of the buffer, and when it got to the half-way point.

To start playback, you first filled the whole buffer with audio data, and then told the hardware to start playing. After the hardware got to the half-way point, it would issue an interrupt. You would then tell the OS you need more data, and you'd get it. You then copy that data into the first half of the buffer *while* the hardware was playing the second half. Later, the hardware would interrupt you when it got to the end of the buffer. You'd then copy more data to the 2nd half while the hardware is playing the first half.

And so - hardware plays one half while you copy data to the other half. Hence, "ping pong".

So how do I implement this in ALSA? The "Writing an ALSA Driver" document doesn't even contain the words "ping" or "pong".

--
Timur Tabi
Linux Kernel Developer @ Freescale
_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

#2 lore raw

At Tue, 12 Jun 2007 01:59:45 +0530,

Quoted 5 lines · Nobin Mathew

Hi,

Now ALSA (audio ) buffer is divided into periods, i.e. a chain of small packets.

Quoted 7 lines

periods size is configurable. Data transfer to the codec starts only after reaching start_threshold point (start_threshold is in periods), this time DMA trigger is called.

This trigger onwards application will get notification from the kernel saying that period buffer is empty you can write into it.Till the end of music file.

Yes. And the "ping-poing" is the case that you have two periods in a ring buffer.

Most hardwares support more periods practically. That's why "periods" (corresponds to "fragments" in OSS) was introduced, as more generic abstraction.

Takashi

Quoted 46 lines

Nobin Mathew

Quoted 27 lines · Timur Tabi

I'm writing an ALSA SOC driver for an I2S-based device, and I'm having a really hard time understanding how ALSA uses the DMA buffers. And yes, I've read the documentation and studied some sample source code.

I used to write audio drivers for a living, but that was many years ago, and it wasn't for Linux. Perhaps the concepts in my head are outdated, but I just don't see enough explanation as to how DMA buffers are supposed to work.

Back then, audio drivers used "ping pong" DMA buffers. A single DMA buffer is allocated, and the audio hardware is programmed to read from that buffer in a loop. That is, it would automatically restart reading from the beginning of the buffer without any reprogramming. The hardware would also be programmed to issue an interrupt when it got to the end of the buffer, and when it got to the half-way point.

To start playback, you first filled the whole buffer with audio data, and then told the hardware to start playing. After the hardware got to the half-way point, it would issue an interrupt. You would then tell the OS you need more data, and you'd get it. You then copy that data into the first half of the buffer *while* the hardware was playing the second half. Later, the hardware would interrupt you when it got to the end of the buffer. You'd then copy more data to the 2nd half while the hardware is playing the first half.

And so - hardware plays one half while you copy data to the other half. Hence, "ping pong".

So how do I implement this in ALSA? The "Writing an ALSA Driver" document doesn't even contain the words "ping" or "pong".

Quoted 8 lines

--
Timur Tabi
Linux Kernel Developer @ Freescale
_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

#3 lore raw
Quoted 89 lines · Takashi Iwai

At Tue, 12 Jun 2007 01:59:45 +0530,

Quoted 4 lines · Nobin Mathew

Hi,

Now ALSA (audio ) buffer is divided into periods, i.e. a chain of small

packets.

Quoted 8 lines

periods size is configurable. Data transfer to the codec starts only after reaching start_threshold point (start_threshold is in periods), this time DMA trigger is called.

This trigger onwards application will get notification from the kernel saying that period buffer is empty you can write into it.Till the end of music file.

Yes. And the "ping-poing" is the case that you have two periods in a ring buffer.

Most hardwares support more periods practically. That's why "periods" (corresponds to "fragments" in OSS) was introduced, as more generic abstraction.

Takashi

Quoted 7 lines

Nobin Mathew

Quoted 1 line · Timur Tabi

I'm writing an ALSA SOC driver for an I2S-based device, and I'm having

a really hard time

Quoted 1 line
Quoted 1 line

understanding how ALSA uses the DMA buffers. And yes, I've read the

documentation and

Quoted 3 lines
Quoted 3 lines

studied some sample source code.

I used to write audio drivers for a living, but that was many years

ago, and it wasn't for

Quoted 1 line
Quoted 1 line

Linux. Perhaps the concepts in my head are outdated, but I just don't

see enough

Quoted 3 lines
Quoted 3 lines

explanation as to how DMA buffers are supposed to work.

Back then, audio drivers used "ping pong" DMA buffers. A single DMA

buffer is allocated,

Quoted 1 line
Quoted 1 line

and the audio hardware is programmed to read from that buffer in a

loop. That is, it

Quoted 1 line
Quoted 1 line

would automatically restart reading from the beginning of the buffer

without any

Quoted 1 line
Quoted 1 line

reprogramming. The hardware would also be programmed to issue an

interrupt when it got to

Quoted 3 lines
Quoted 3 lines

the end of the buffer, and when it got to the half-way point.

To start playback, you first filled the whole buffer with audio data,

and then told the

Quoted 1 line
Quoted 1 line

hardware to start playing. After the hardware got to the half-way

point, it would issue

Quoted 1 line
Quoted 1 line

an interrupt. You would then tell the OS you need more data, and

you'd get it. You then

Quoted 1 line
Quoted 1 line

copy that data into the first half of the buffer *while* the hardware

was playing the

Quoted 1 line
Quoted 1 line

second half. Later, the hardware would interrupt you when it got to

the end of the

Quoted 1 line
Quoted 1 line

buffer. You'd then copy more data to the 2nd half while the hardware

is playing the first

Quoted 3 lines
Quoted 3 lines

half.

And so - hardware plays one half while you copy data to the other

half. Hence, "ping pong".

Quoted 2 lines
Quoted 2 lines

So how do I implement this in ALSA? The "Writing an ALSA Driver"

document doesn't even

Quoted 10 lines
Quoted 10 lines

contain the words "ping" or "pong".

--
Timur Tabi
Linux Kernel Developer @ Freescale
_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Quoted 5 lines

_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Quoted 5 lines

_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Hi

yet another newbie question about periods here:

1. AFAIK, the period size is closely dependent on the h/w, but after reading some docs I collected
that, they can be given values depending on how much we care about the latency. Does it mean
that, I can vary it without paying any attention to what h/w manual says just because I want low or
high latency? I hope this question is clear.

2. As periods correspond to fragment size from OSS world, what the other periods related fields
correspond to i.e. what do the following fields mean?

period_bytes_min,
period_bytes_max,
periods_min,
periods_max,

I know what they mean after looking at them but I want to know the relationship between various
fields.

For e.g.

I have,

buffer_bytes_max = 8192 * 8
i.e. = AUDIO_FRAGSIZE_DEFAULT * AUDIO_NBFRAGS_DEFAULT

here AUDIO_FRAGSIZE_DEFAULT is size of period right? Then to get the max buffer size we should multiply it
by number of periods, is this correct? Also, these are default values of the period and no of periods, then do
I need to see the h/w manual to decide the periods_min/periods_max and period_bytes_min/period_bytes_max
fields?

#4 lore raw

I found the following doc, it talks about periods in depth with a figure.

http://delivery.acm.org/10.1145/1020000/1017977/6735.html?key1=1017977&key2=0592661811&coll=GUIDE&dl=GUIDE&CFID=15151515&CFTOKEN=6184618

-pharaoh.

Quoted 137 lines · Pharaoh .
Quoted 89 lines · Takashi Iwai

At Tue, 12 Jun 2007 01:59:45 +0530,

Quoted 4 lines · Nobin Mathew

Hi,

Now ALSA (audio ) buffer is divided into periods, i.e. a chain of small

packets.

Quoted 8 lines

periods size is configurable. Data transfer to the codec starts only after reaching start_threshold point (start_threshold is in periods), this time DMA trigger is called.

This trigger onwards application will get notification from the kernel saying that period buffer is empty you can write into it.Till the end of music file.

Yes. And the "ping-poing" is the case that you have two periods in a ring buffer.

Most hardwares support more periods practically. That's why "periods" (corresponds to "fragments" in OSS) was introduced, as more generic abstraction.

Takashi

Quoted 7 lines

Nobin Mathew

Quoted 1 line · Timur Tabi

I'm writing an ALSA SOC driver for an I2S-based device, and I'm having

a really hard time

Quoted 1 line
Quoted 1 line

understanding how ALSA uses the DMA buffers. And yes, I've read the

documentation and

Quoted 3 lines
Quoted 3 lines

studied some sample source code.

I used to write audio drivers for a living, but that was many years

ago, and it wasn't for

Quoted 1 line
Quoted 1 line

Linux. Perhaps the concepts in my head are outdated, but I just don't

see enough

Quoted 3 lines
Quoted 3 lines

explanation as to how DMA buffers are supposed to work.

Back then, audio drivers used "ping pong" DMA buffers. A single DMA

buffer is allocated,

Quoted 1 line
Quoted 1 line

and the audio hardware is programmed to read from that buffer in a

loop. That is, it

Quoted 1 line
Quoted 1 line

would automatically restart reading from the beginning of the buffer

without any

Quoted 1 line
Quoted 1 line

reprogramming. The hardware would also be programmed to issue an

interrupt when it got to

Quoted 3 lines
Quoted 3 lines

the end of the buffer, and when it got to the half-way point.

To start playback, you first filled the whole buffer with audio data,

and then told the

Quoted 1 line
Quoted 1 line

hardware to start playing. After the hardware got to the half-way

point, it would issue

Quoted 1 line
Quoted 1 line

an interrupt. You would then tell the OS you need more data, and

you'd get it. You then

Quoted 1 line
Quoted 1 line

copy that data into the first half of the buffer *while* the hardware

was playing the

Quoted 1 line
Quoted 1 line

second half. Later, the hardware would interrupt you when it got to

the end of the

Quoted 1 line
Quoted 1 line

buffer. You'd then copy more data to the 2nd half while the hardware

is playing the first

Quoted 3 lines
Quoted 3 lines

half.

And so - hardware plays one half while you copy data to the other

half. Hence, "ping pong".

Quoted 2 lines
Quoted 2 lines

So how do I implement this in ALSA? The "Writing an ALSA Driver"

document doesn't even

Quoted 10 lines
Quoted 10 lines

contain the words "ping" or "pong".

--
Timur Tabi
Linux Kernel Developer @ Freescale
_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Quoted 5 lines

_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Quoted 5 lines

_______________________________________________
Alsa-devel mailing list
[email protected]
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Hi

yet another newbie question about periods here:

1. AFAIK, the period size is closely dependent on the h/w, but after reading some docs I collected
that, they can be given values depending on how much we care about the latency. Does it mean
that, I can vary it without paying any attention to what h/w manual says just because I want low or
high latency? I hope this question is clear.

2. As periods correspond to fragment size from OSS world, what the other periods related fields
correspond to i.e. what do the following fields mean?

period_bytes_min,
period_bytes_max,
periods_min,
periods_max,

I know what they mean after looking at them but I want to know the relationship between various
fields.

For e.g.

I have,

buffer_bytes_max = 8192 * 8
i.e. = AUDIO_FRAGSIZE_DEFAULT * AUDIO_NBFRAGS_DEFAULT

here AUDIO_FRAGSIZE_DEFAULT is size of period right? Then to get the max buffer size we should multiply it
by number of periods, is this correct? Also, these are default values of the period and no of periods, then do
I need to see the h/w manual to decide the periods_min/periods_max and period_bytes_min/period_bytes_max
fields?

author
#5 lore raw
In reply to Takashi Iwai
Quoted 2 lines · Takashi Iwai

Yes. And the "ping-poing" is the case that you have two periods in a ring buffer.

Ok, I understand the 'periods' concept now, thanks.

So when/where does ALSA copy data to the DMA buffer? Isn't there supposed to be some kind of callback where ALSA calls the driver and says, "here's some data, please copy it to your DMA buffer?"

Or does ALSA do all the copying itself whenever the driver calls snd_pcm_period_elapsed()?

Signature · 2 lines
Timur Tabi
Linux Kernel Developer @ Freescale
#6 lore raw

At Tue, 12 Jun 2007 11:36:33 -0500,

Quoted 6 lines · Timur Tabi
Quoted 2 lines · Takashi Iwai

Yes. And the "ping-poing" is the case that you have two periods in a ring buffer.

Quoted 7 lines

Ok, I understand the 'periods' concept now, thanks.

So when/where does ALSA copy data to the DMA buffer? Isn't there supposed to be some kind of callback where ALSA calls the driver and says, "here's some data, please copy it to your DMA buffer?"

Or does ALSA do all the copying itself whenever the driver calls snd_pcm_period_elapsed()?

It depends on the implementation, but usually the latter case. The whole data-transfer task is done by ALSA PCM middle layer, and each driver has to take care of the DMA setup -- as long as your hardware can have a buffer on RAM. If you need to copy the data to hardware buffer manually in some way, the things become complicated...

Takashi