See https://datatracker.ietf.org/doc/html/draft-ietf-rmcat-gcc-02
This commit implements the bandwidth estimation as a GStreamer element
that is then used in webrtcbin through the new `request-bandwidth-estimator`
signal.
This keeps our Homegrown congestion controller but removes the possibility
to switch CC algorithm at runtime.
In some applications people might need to be able to pass more metadata
than the only "display-name" this commit allow that by adding a field in
all signalling structure where it is free form json and removing the
"display-name" as the free form `meta` field allows application to pass
this kind of information without any problem.
On the "webrtcsink" side this commit adds a `meta` property as a
gst::Structure which will be passed as a json string to the signalling
server so we enforce data to be structured.
As specified in Google Congestion Control we should run the packet loss
estimation algorithm "every time feedback from the receiver is
received".
And, also as defined by GCC, we now have 2 different estimated bitrates,
one for the delay-based controller value and one for the loss-based one,
and we use the minimum value between those 2 as our current estimation.
[GCC]: https://datatracker.ietf.org/doc/html/draft-ietf-rmcat-gcc-02
1. Working scenario:
T1 -> Caps event (all caps have been received)
T1 -> Start discovering
T2 -> Change state to Playing
T2 -> The signaller is not started as:
- Sink current_state() == Paused as it will be set to
playing after the change_state vmethod returns
- Discovery is not done anyway
T1 -> Discovery is done
=> The signaller is started, and **everything works well**.
2. Failing scenario:
T1 -> Caps event (all caps have been received)
T1 -> Start discovering
T1 -> Discovery is done
T1 -> The signaller is not started as:
- Current state == Paused (it will be set to playing
after the change_state vmethod returns)
- Discovery is not done anyway
T2 -> Change state to Playing
T2 -> The signaller is not started as:
- Sink current_state == Paused as it will be set to
playing after the we return from the change_state
vmethod
In that case the signaller never starts.
The change was merged inadvertently alongside the display-name
API extension, we will probably eventually expose API on webrtcsink
pads to control that, but we won't just set it to pad.name() anyway.
Fixes#33
We no longer connect to on-negotiation-needed, this in order to call the
consumer-added signal without holding the state lock:
Going to Ready triggers synchronous emission of the on-negotiation-needed
signal, during which time the application may add a data channel, causing
renegotiation, which we do not support at this time.
This is completely safe, as we know that at that point all conditions are
gathered: webrtcbin is in the Ready state, and all its transceivers have
codec_preferences.
