[pulseaudio-discuss] RFC: Public API for managing nodes

[David Henningsson]

On 07/16/2013 03:20 PM, Tanu Kaskinen wrote:
> On Tue, 2013-07-16 at 14:01 +0200, David Henningsson wrote:
>> On 07/13/2013 05:48 PM, Tanu Kaskinen wrote:
>>> Hi all,
>>>
>>> I've written up a proposal for a public API for controlling routing with
>>> nodes:
>>> http://www.freedesktop.org/wiki/Software/PulseAudio/RFC/RoutingAPI/
>>>
>>> Comments would be very welcome.
>>>
>>> Also, if anyone wants to write a proof-of-concept GUI for exercising the
>>> API, or extend pavucontrol, that would be awesome. I probably won't do
>>> it.
>>
>> I've tried to read it through and initially I'm trying to grasp the
>> concepts here.
>>
>> A few points:
>>
>>    * The docs say "since 5.0", are you really planning that, or given our
>> new schedule, should it rather be since 6.0?
>
> Very likely 6.0, but let's see what we have when the freeze starts.
>
>>    * I also wonder about the basic rules for edges/connections here. Can
>> we line them up:
>>      - Edges are directed (stream flows from node A to node B)
>
> Obviously audio flows from input to output. If this is what you meant,
> then yes, they are directed.
>
> There exists also some sort of "initiation ownership" of connections,
> that is, a connection usually (but not necessarily always) originates
> from a stream node (either input or output), and the device node is a
> more passive participant in the relationship. This ownership is explicit
> with the default connections (a node has "own" and "other" default
> connections), but for explicit connections there is no such ownership
> (or it's implied by the node type: if a stream is left without
> connections, it needs to be "rescued", but there's no need to rescue a
> sink when it's left without connections).
>
>>      - A node can have more than one incoming edge (mixes input to single
>> output)
>
> Yep.
>
>>      - Can a node have more than one outgoing edge, and how is this then
>> handled for playback (clock deviation etc)? Automatic insert of combine
>> modules?
>
> Yes, automatic use of the combine module is planned (the combining
> functionality would move to the core, because what's the point of having
> code in a module if the core directly depends on that code).

Okay, so combine stuff would then not have its own node or set of nodes, 
but essentially be as integrated as mixing is today?

>>      - And recording has the same thing but different, i e, spreading
>> data is not a problem but mixing requires clock synchronisation
>
> Yes. I don't know what the solution will be (or whether we even need any
> solution in the near future), but I guess this could be implemented
> relatively easily with looping back the sources to a null sink and
> recording from the null sink's monitor.

Yeah, that sounds possible although I'm not that happy about the added 
latency of the loopback stuff. I don't know if there's a much better 
solution though.

>>      - Can there be more than one edge between the same nodes? E g, one
>> default connection and one explicit connection? Or how does this work?
>
> Yes, there can be both an explicit and a default connection between two
> nodes, or it can be also thought as being one connection that is both
> explicit and default. I don't know what is a better way to think about
> it, but perhaps it doesn't matter anyway.

I would prefer the latter way. I would then see "default" and "explicit" 
as a property of the edge.

>>      - Can edges have properties, e g, a volume?
>
> I don't know. Currently I'm not aware of a requirement to have
> properties on the connections. My current thinking regarding volume
> specifically is that we shouldn't attach volume to nodes: let's keep the
> nodes for routing only.

Ok, that makes sense.

>>    * As for possible node operations, a move operation would be the same
>> as a "batched remove + add" operation. Maybe therefore the batch
>> solution would be better, i e, the client API inputs an array of node
>> operations?
>
> I don't like the idea that if a client wants to move a stream, it has to
> break the operation down to "remove + add", and then the server tries to
> guess what the client really meant. Having a "move" operations keeps the
> client intention obvious.

Fair point, although the move can be on both sides, whereas one type of 
move would be "I want node A to take data from C instead of B" and the 
other type would be "I want node A to output data to node C instead of B".
Or perhaps you want to swap, so that if you're currently on "A -> B and 
C -> D" and you want "A -> D and B -> C" and you want to do all of this 
atomically. The number of operations you want to do might grow out of 
hand unless you have a "batch mode".

>
>>    * The operations on changing the default connections does not make
>> sense to me. If the definition for default connections are those made
>> automatically by the routing system, if you change them then you broke
>> the definition...?
>
> What operations do you mean? Moving or removing a default connection is
> not supported as such, but if the client tries that anyway, we can
> implicitly convert the connection to an explicit one and disable default
> connections, or we can require the client to do these operations
> explicitly, but I think the latter would be too inconvenient for the
> client.

Ok, I think I didn't read the proposal well enough. Having done that, I 
understand that you're suggesting a global switch "default connections 
on/off" only. Or is it a per-node switch?

I have another idea that might be worth considering: how about that the 
"explicit" layer can both enable and disable connections? So that there 
could be a default connection between A and B, but there is also some 
sort of explicit override that disables it. This would be more flexible 
than a more global on/off switch.

> Adding a default connection is of course a totally impossible operation.
>
>>    * Now, streams can be moved using either the new node API or the
>> existing stream/sink API. Will either cause consequences for the other,
>> e g changing node connections will cause an update event for the sink,
>> and changing stream/sink connection will cause update events for the
>> relevant nodes?
>
> Yes, clients can do routing with either API, and changes done with one
> API will cause events in the other API.
>
>>    * And an bird's eye thought - I was thinking that these nodes were
>> something for policy/routing modules primarily. Is a client API the next
>> step here?
>
> pactl support is very useful for testing and debugging while developing,
> so I think it makes sense to have that before writing policy modules
> that operate on nodes.(There's also the aspect that Janos wanted to
> work on the policy module side. He's currently on vacation, though, but
> this means that I will not work on any policy module before Janos has
> some code ready.)

Ok, I guess you could use pacmd for debugging too though.


-- 
David Henningsson, Canonical Ltd.
https://launchpad.net/~diwic