[alsa-devel] The new mixer logic in PA
Heya!
During the last weeks I have been reworking the ALSA mixer and device enumeration handling in PA to be more flexible and support more features of usual mixers/cards then it previously did. (Previously we just picked one simple mixer element and used it for volume and muting; 'Master' for output; 'Capture' for input; and the ALSA device name list we used was hardcoded). I just landed all of this in PA, and would like to pass this at least roughly by you folks, hoping for comments. So here's a rough overview of what's going on:
The current ALSA mixer APIs have been criticized in the past (including by me) for their complexity and stacking of abstraction layers. Some folks (including me) suggested reworking the ALSA mixer APIs to simplify the APIs. After doing my research I am not so sure anymore that this really would be of much benefit -- at least for software like PA.
The reason for that is simply that PA has very specific needs on the mixer interfacing: we are only interested in a subset of the potential functionality: a) we want to do hw volume control when supported for a specific snd_pcm_t with good granularity and range, b) we want mute control, and c) we want to select the output/input port for a snd_pcm_t (i.e. line-in vs. mic/speaker vs. headset). And that's about it.
We generally don't want to support any fancier setups even if the hardware would provide it: we don't want simulatenous output to multiple distinct outports, we want no hw mixing, we don't want to support legacy input sources such as CD/Phone/PCSpk..., no "3D" features, not even input feedback. OTOH PA can do stuff other clients cannot do, e.g. extending in softwre the capabilities of the hw if possible and sensible. e.g. we can extend the volume control range/granularity in software beyond the discrete steps the hw provides.
It has been suggested to make queryable how mixer elements relate to each other via some mixer API, possibly using the HDA codec graph data. While that would be nice to have I believe that this information would be much more than PA could ever need, and also certainly not help simplicity of an API. I am hence not asking for that, what I do ask for however is that a couple of assumptions I make stay valid. (see below)
Given that what PA needs and doesn't need is very different from the needs of almost all other applications needed further abstraction can happen in PA, and doesn't need to be done in ALSA. Which is basically the reason why I decided to do all this mixer abstraction work inside of PA, instead of trying to get that into ALSA.
To be able to control more than just one element, PA needs to have some minimal clue how elements are related to each other. ALSA (as mentioned) currently doesn't provide that info, so PA must guess. Hardcoding assumptions about how mixer elements relate to each other of course sucks, which is why I chose to define the list of elements PA should control plus minimal information about how they work together in config files that can be updated more easily and that during runtime. I tried to encode only minimal and the most obvious assumptions.
These config files simply list a series of (simple) mixer controls and the way how to initialize them/expose them in PA. Switches can either be bound to always off or on, or be used to follow PA's mute status, or be exposed as distinct options in the UI (which however almost no switch should do). Volumes can either be bound to the minimal values, or to 0dB or be integrated into the volume slider PA exposes. Enumeration values can be exposed in the UI, too, but only if that makes sense.
The most interesting part of that is probably the volume handling: for that we go from top to botton in the list of elements in those config files and try to set the volume on the first element marked for it, then divide the actual volume set from our target volume and find the next volume element and apply the remaining volume there, and so on. What's left we then do in attenuate software.
All this mixer handling assumes that PA is the only one in charge of volume control. Also, I assume that while I now control a substantial part of the available mixer controls that those I don't control will be initialized properly by "alsactl init".
Here's an example how that looks like, for implementing a Mic path (which is the most complex path, due to the input selection stuff)
http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat... http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat... http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat...
To simplifiy these rules a bit I'd like to see some cleanups in the names used for controls:
- Some drivers have a "Input Source Select" enumeration, others a "Input Source" and even others "Capture Source". These could be named the same, can't they? Or is there any difference in their behaviour from an apps PoV? (There's also a "Mic Select" enum, which however probably makes sense to keep seperate given that it the name suggests it is only useful if "Mic" is activated in a capture group)
- The options of said enums are really chaotically named. That could really be cleaned up, and would probably be a trivial fix.
- Since the various "boost" switches effectively control the volume I'd prefer if they'd be exposed as volume controls with dB info attached. Some Boosts encode the dB info in the name, which sucks.
- Some cards have an element "TV Tuner", others "Video". I think that could be merged.
I'll prepare a patch for at least some of these suggestions. Of course renaming controls on a big scale will break alsactl restore. Is that a problem?
There's one assumption I am currently encoding in the files I am not really sure about. Which is that I assume that the volume level of the surround outputs is influenced by the following volume elements if that they exist:
PCM -> Master -> Front, Rear, Surround, Side, Center, LFE
while if switched to headphones I assume:
PCM -> Master -> Headphone
May I assume that? Or is sometimes the routing PCM -> Headphones, and 'Master' does not have influence?
There's something else I was wondering, in contrast to the usual code that makes use of the ALSA mixer's dB data I use them to do actual calculations while previously they were just used for "enriching" the information shown on the screen. I wonder if this is going to backfire? How reliable is the dB data actually? [1]
For more information on how all this really works you might find this mail I posted to the PA ML interesting:
https://tango.0pointer.de/pipermail/pulseaudio-discuss/2009-June/004229.html
It goes more into details in things which are relevant for PA folks, probably not so much for the ALSA folks.
I'd be very interested in opinions on all of this!
Thanks,
Lennart
[1] To verify and measure the dB data I actually wrote a little tool that can measure the data, which seemed to yield quite correct information on my HDA chip:
http://git.0pointer.de/?p=dbmeasure.git;a=tree http://git.0pointer.de/?p=dbmeasure.git;a=blob;f=README
At Sat, 20 Jun 2009 02:21:15 +0200, Lennart Poettering wrote:
Heya!
During the last weeks I have been reworking the ALSA mixer and device enumeration handling in PA to be more flexible and support more features of usual mixers/cards then it previously did. (Previously we just picked one simple mixer element and used it for volume and muting; 'Master' for output; 'Capture' for input; and the ALSA device name list we used was hardcoded). I just landed all of this in PA, and would like to pass this at least roughly by you folks, hoping for comments. So here's a rough overview of what's going on:
The current ALSA mixer APIs have been criticized in the past (including by me) for their complexity and stacking of abstraction layers. Some folks (including me) suggested reworking the ALSA mixer APIs to simplify the APIs. After doing my research I am not so sure anymore that this really would be of much benefit -- at least for software like PA.
The reason for that is simply that PA has very specific needs on the mixer interfacing: we are only interested in a subset of the potential functionality: a) we want to do hw volume control when supported for a specific snd_pcm_t with good granularity and range, b) we want mute control, and c) we want to select the output/input port for a snd_pcm_t (i.e. line-in vs. mic/speaker vs. headset). And that's about it.
We generally don't want to support any fancier setups even if the hardware would provide it: we don't want simulatenous output to multiple distinct outports, we want no hw mixing, we don't want to support legacy input sources such as CD/Phone/PCSpk..., no "3D" features, not even input feedback. OTOH PA can do stuff other clients cannot do, e.g. extending in softwre the capabilities of the hw if possible and sensible. e.g. we can extend the volume control range/granularity in software beyond the discrete steps the hw provides.
This is an expected thing, just the same way Windows went in the past :)
It has been suggested to make queryable how mixer elements relate to each other via some mixer API, possibly using the HDA codec graph data. While that would be nice to have I believe that this information would be much more than PA could ever need, and also certainly not help simplicity of an API. I am hence not asking for that, what I do ask for however is that a couple of assumptions I make stay valid. (see below)
Sounds reasonable. Exposing all topology and parse it is no easy job, speaking from my experience at writing the usb-audio and HD-audio drivers, both of which require things alike.
Given that what PA needs and doesn't need is very different from the needs of almost all other applications needed further abstraction can happen in PA, and doesn't need to be done in ALSA. Which is basically the reason why I decided to do all this mixer abstraction work inside of PA, instead of trying to get that into ALSA.
To be able to control more than just one element, PA needs to have some minimal clue how elements are related to each other. ALSA (as mentioned) currently doesn't provide that info, so PA must guess. Hardcoding assumptions about how mixer elements relate to each other of course sucks, which is why I chose to define the list of elements PA should control plus minimal information about how they work together in config files that can be updated more easily and that during runtime. I tried to encode only minimal and the most obvious assumptions.
These config files simply list a series of (simple) mixer controls and the way how to initialize them/expose them in PA. Switches can either be bound to always off or on, or be used to follow PA's mute status, or be exposed as distinct options in the UI (which however almost no switch should do). Volumes can either be bound to the minimal values, or to 0dB or be integrated into the volume slider PA exposes. Enumeration values can be exposed in the UI, too, but only if that makes sense.
The most interesting part of that is probably the volume handling: for that we go from top to botton in the list of elements in those config files and try to set the volume on the first element marked for it, then divide the actual volume set from our target volume and find the next volume element and apply the remaining volume there, and so on. What's left we then do in attenuate software.
All this mixer handling assumes that PA is the only one in charge of volume control. Also, I assume that while I now control a substantial part of the available mixer controls that those I don't control will be initialized properly by "alsactl init".
Here's an example how that looks like, for implementing a Mic path (which is the most complex path, due to the input selection stuff)
http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat... http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat... http://git.0pointer.de/?p=pulseaudio.git;a=blob;f=src/modules/alsa/mixer/pat...
To simplifiy these rules a bit I'd like to see some cleanups in the names used for controls:
- Some drivers have a "Input Source Select" enumeration, others a "Input Source" and even others "Capture Source". These could be named the same, can't they? Or is there any difference in their behaviour from an apps PoV? (There's also a "Mic Select" enum, which however probably makes sense to keep seperate given that it the name suggests it is only useful if "Mic" is activated in a capture group)
As mentioned before, these are basically identical. A different name was given just because of the strange behavior of the mixer abstraction implementation, which always wants to expand "Capture Source" elements. I prefer to keep it as enum, so "Input Source" is there.
So, renaming "Capture Source" to "Input Source" would be safe. This won't break existing setup. But, not from "Input Source" to "Capture Source". Then you'll get errors when multiple elements exist.
- The options of said enums are really chaotically named. That could really be cleaned up, and would probably be a trivial fix.
Yes.
- Since the various "boost" switches effectively control the volume I'd prefer if they'd be exposed as volume controls with dB info attached. Some Boosts encode the dB info in the name, which sucks.
Agreed. The switch had been already there before the dB information was introduced. Now it can be covered better by dB value.
- Some cards have an element "TV Tuner", others "Video". I think that could be merged.
Likely.
I'll prepare a patch for at least some of these suggestions. Of course renaming controls on a big scale will break alsactl restore. Is that a problem?
It's not a problem if you'll handle all reported bugs ;) In practice, I don't think it'd be a big issue. The system (not the driver, though) will (should) reset to sane values when uninitialized.
There's one assumption I am currently encoding in the files I am not really sure about. Which is that I assume that the volume level of the surround outputs is influenced by the following volume elements if that they exist:
PCM -> Master -> Front, Rear, Surround, Side, Center, LFE
while if switched to headphones I assume:
PCM -> Master -> Headphone
May I assume that? Or is sometimes the routing PCM -> Headphones, and 'Master' does not have influence?
In the perfect world, Front, Surround ,etc are logical mixer volumes. That is, they correspond to the output volumes of these channels, independent from the routing, thus Front influences on Headphone as well. But, it's case-by-case. Sometimes "Front" only refers to the front line-out. We'll need to check in the deep details of each driver implementation.
There's something else I was wondering, in contrast to the usual code that makes use of the ALSA mixer's dB data I use them to do actual calculations while previously they were just used for "enriching" the information shown on the screen. I wonder if this is going to backfire? How reliable is the dB data actually? [1]
This is *very* hard to tell. Many drivers have been written / provided without the hardware. The dB information was taken just from the hardware information. Thus, it's the question whether you can trust the hardware. (And I don't :)
Of course, we can give any corrections to specific hardware chip / model if needed in the driver. But, as said, it can't be done without the hardware, so I cannot answer whether it can be trusted.
thanks,
Takashi
For more information on how all this really works you might find this mail I posted to the PA ML interesting:
https://tango.0pointer.de/pipermail/pulseaudio-discuss/2009-June/004229.html
It goes more into details in things which are relevant for PA folks, probably not so much for the ALSA folks.
I'd be very interested in opinions on all of this!
Thanks,
Lennart
[1] To verify and measure the dB data I actually wrote a little tool that can measure the data, which seemed to yield quite correct information on my HDA chip:
http://git.0pointer.de/?p=dbmeasure.git;a=tree http://git.0pointer.de/?p=dbmeasure.git;a=blob;f=README-- Lennart Poettering Red Hat, Inc. lennart [at] poettering [dot] net http://0pointer.net/lennart/ GnuPG 0x1A015CC4
participants (2)
-
Lennart Poettering -
Takashi Iwai