We need to keep the sources around until all sources are dispatched. If
not: source A and B are active after poll, A is dispatched first and
removes B, when B is then dispatched, the memory is gone. We don't
free the source but simply mark the fd invalid and move it do a
destroy list. After all sources are dispatched we destroy.
This is safe because removing a source is either done from the poll
context (with invoke) or when holding the right locks.
A client stream is a more specialized way to send 1 stream to pipewire.
On the client side and receiver side it can do conversion and the
buffer size of the client can be choosen arbitrarily.
Implement audioconvert as a complex element of fmtconver,
channelmix and resample.
Make copying resample just to test.
Plug the converter into pw_stream.
Separate buffer sizes in the info. Make it possible to not set
data pointers for when we use memfd for the memory.
Improve format conversion.
Work on in-place conversion in test-convert2
reset number of fds before we begin reading.
When we destroy the proxy, just clear the id. We then remove the id
completely when the server sends a remove_id request. This avoid
reusing the proxy id before the server has finished with it.
Add some debug
Simplify the stream API. make just 2 methods to queue and dequeue
buffers. Make just one callback when new buffers can be dequeued.
Add support for driver nodes such as the video-src.
Pass a pw_buffer structure to add/remove_buffer and make it possible
to attach metadata to it. This makes it a lot easier to implement
the gstreamer pipewire pool.
Call the stream process function from the main loop and use a lockfree
ringbuffer to pass buffers between the threads. Make it possible to
also call process from the RT thread.
unmap the buffer data when needed.
