Acquiring a transport is needed in two different situations: either
we are initiating the audio stream locally, or the remote side initiated
it and thus we are just reacting. In the second case, we would expect
the stream is already available, and otherwise the operation should
fail. This means the media API needs to be extended in order to make
this difference.
This issue is specially relevant in the case of SCO, because the current
approach is racy. With HFP, for example (say BlueZ has the HS role), the
following race condition could be met:
1. Phone has an incoming call and thus starts in-band ringing.
2. SCO connection is accepted and stablished by BlueZ.
3. Gateway interface state is changed to Playing.
4. Exactly afterwards, the user routes the audio to the phone, to have
a private conversation. So the SCO link is closed.
5. In parallel, PulseAudio sees the transition to Playing, and acquires
the transport.
6. BlueZ receives an Acquire() request, but SCO is down. So it tries to
reconnect the SCO link.
The last step is an undesired behavior (the audio is routed back to the
car). BlueZ should be smart enough to know that the SCO connection
shouldn't be reestablished, but this is only possible if the endpoint
provides additional information in the media API.
Note that the API change introduced by this patch is backward
compatible, meaning that older versions of BlueZ will just ignore the
flag. So clients are encouraged to use it without necessarily adding a
dependency to newer versions of BlueZ.
Extend the Media API to expose the transport state in D-Bus, as a
property of the transport. This way the clients do not have to find
out which is the corresponding profile-specific interface for the
transport.
Additionally, this state along with the automatic release of transports
will allow clients to avoid the "optional release" or "accept remote
release" race condition. For example, with HSP/HFP profiles, the problem
is the following:
1. User suspends SCO in the remote end.
2. BlueZ signals the Playing->Connected state change in D-Bus.
3. Exactly afterwards, the user resumes SCO in the remote end.
4. In parallel, PulseAudio sees the aforementioned transition to
Connected, and thus releases the transport.
5. BlueZ receives a Release() request while SCO is up. So the audio
stream will be suspended.
The last step is an undesired behavior since the user explicitly wanted
to route the audio stream through Bluetooth.
The issue is difficult to reproduce but it can easily be solved by
exposing the transport state in D-Bus.
If we use the same hash table to set the new metadata, we have 2
undesired behaviors:
1) New track may contain fields from previous track if it didn't set all
the fields
2) If we fail on parsing the signal, we will still change some of the
fields
This should make Acquire blocking friendly since the client no longer has
to call GetProperties to discover how much it can write/read when using
the acquired file descriptor.
This makes simpler to application which are handling many endpoints
without a context data. It also may be useful in future in case we
allow multiple transports per endpoint.
