This was causing a memory corruption and crash because the
simple-endpoint-link was not calling finalize of the endpoint-link
and the GWeakRefs pointing to the endpoints were not uninitialized,
causing g_object_unref later to try to write to them (on free'ed memory)
* rework how global objects are stored in the core
* rework how users get notified about global objects
and proxies of remote global objects
The purpose of this change is to have a class that can manage
objects that are registered in the core or signalled through the
registry. This object can declare interest on certain types
of global objects and only keep & signal those objects that it is
interested in. Additionally, it can prepare proxy features and
asynchronously deliver an 'objects-changed' signal, which is
basically telling us that the list of objects has changed.
This is useful to simplify port proxies management in WpAudioStream.
Now the stream object can declare that it is interested in ports
that have "node.id" == X and the object manager will only maintain
a list of those. Additionally, it will emit the 'objects-changed'
signal when the list of ports is complete, so there is no reason to
do complex operations and core syncs in the WpAudioStream class
in order to figure out when the list of ports is ready.
As a side effect, this also reduces resource management. Now we
don't construct a WpProxy for every global that pipewire reports;
we only construct proxies when there is interest in them!
Another interesting side effect is that we can now register an
object manager at any point in time and get immediately notified
about remote globals that already exist. i.e. when you register
an object manager that is interested in nodes, it will be immediately
notified about all the existing nodes in the graph. This is useful
to avoid race conditions between connecting the signal and objects
beting created in pipewire
PipeWire headers were recently cleaned up to reduce the number of
included headers. This leads to a number of functions and types not
being included when needed in wireplumber.
The alsa.pcm devices currently use "device.api" = "alsa:pcm"
but then the node overrides this later to just "alsa", which
is confusing. For now, let's match "alsa:pcm" in the monitor,
even though this does not appear on the node later.
This commit fixes device preferences in wireplumber.conf,
which were broken because the alsa endpoint names were empty
of info about the alsa device
This allows the bluealsa gstreamer helper to link to the speakers
constantly, with both A2DP and HFP streams. If this is not enabled,
only one of those will manage to connect, randomly.
This is definitely something to be superseeded by proper policy
management implemented with configuration files
When the audioconvert starts, it emits 2 ports, but when we set the PortConfig,
it removes them and re-creates them. Previously, the stream class would not
remove the old port proxies from the list and therefore they existed twice.
It is also necessary here to store the proxies earlier, when they are added,
instead of when they are augmented, so that we can ensure they are removed.
Previously we would hit an issue where:
- port proxy is added, augmented
- augment completes but the GTask wants to complete asynchronously:
it stores a ref on the proxy and adds an idle source
- server removes the proxy, we delete it from the core's list
- the GTask now calls the augment callback, which stores the (removed)
proxy on the stream's port_proxies list...
Running audioconvert in merge+split mode is the only way to make this work with
the adapter, since the adapter does not support passing multiple channels on
a single port right now, and if it does at some point, it will be without a
mixing node on the port, which means we will not be able to mix multiple
audioconvert nodes on the same adapter. In the future we need to consider
writing a lighter volume node with multiple channels support to replace
audioconvert.
The new linking algorithm now takes into account the channel positions and makes
sure to link the correct channels together. Also, it avoids passing the port
proxies inside the GVariants, thus making the algorithm a bit more generic
and easier to unit test.
GVariant is instructed to store a 64-bit unsigned integer (mode "t")
but we pass it a pointer to a pointer to read this value out,
which will crash in 32-bit architectures, as the pointer doesn't have
enough bits and g_variant_lookup therefore corrupts the stack.
In practice we always create a remote and connect to pipewire.
Any other scenario is invalid, therefore, it is not justified
to be confused with so many classes for such small functionality.
This simplifies a lot the modules code.
Also, this commit exposes the pw_core and pw_remote objects
out of WpCore. This is in practice useful when dealing with low-level
pw and spa factories, which are used in the monitors. Let's not
add API wrappers for everything... Bindings will never use this
functionality anyway, since it depends on low level pipewire C API.
