Actually negative DTS (i.e. < 0) was handled correctly before when
queueing buffers, but this didn't handle the case correctly where the
very same stream has DTS/PTS shifted to be positive (i.e. start DTS is
before earliest PTS).
To solve this, pass through all DTS as signed values and at the very end
for every fragment calculate the offset between the two in addition to
the PTS/DTS offset of each buffer, and adjust the composition time
offsets accordingly.
This way the final PTS of each buffer after demuxing is equal to the
actual running time of the very same buffer before muxing. Previously
all PTS would've been shifted forwards by a couple of frames, which then
broke A/V synchronization.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2232>
While this is supposed to be a decode time, this doesn't really match
the description in the specification. It starts at 0 and increases by
the duration of each fragment, i.e. is the sum of all previous sample
durations.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2223>
Updates the API with necessary plumbing to allow to track consumers
being added/removed, as well as starting a producer session targeted at
a consumer peer waiting for one.
The example page has also been updated to demonstrate this
functionality.
Can be tested by launching a webrtcsrc without a peer-id specified.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2206>
It's not really scalable in a repository with lots of different pieces
of code. What is an intentional typo (`fiel`) in one directory might be
an actual typo (`field`) elsewhere, and unfortunately this can't be
configured in a useful way on a per-directory basis.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2208>
Switching to a lockable session re-introduced a similar deadlock to the
one fixed by f82a731b3a: we would call
encoder-setup with the session lock, the default handler would try to
lock the settings, thus not respecting the locking order (settings,
state, session).
We fix this by setting up the whole encoding chain outside of any locks,
then calling a reduxed version of Session.connect_input_stream() (sans
encoder-setup / payloader-setup signal calls).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2215>
When a session was removed while its input streams were still getting
connected, a race condition was possible where:
* connect_input_stream was getting called from webrtcbin's pc thread
* the state of the pipeline was getting set to NULL in another thread
* connect_input_stream tried to sync state to NULL, and deadlocked
because webrtcbin tried to stop its pc thread synchronously
This commit fixes this by making sure we hold the session lock when
setting the state of the pipeline to NULL, thus ensuring
`connect_input_stream` isn't getting called at the same time.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2215>
An error message is to be considered fatal by its receiver, but sending
it to webrtcsink when it sends a message for an unknown session leaves
it vulnerable to a race condition where the server has received an
endSession message from a consumer, but hasn't notified webrtcsink yet.
This commit simply removes the sending of an error message for unknown
sessions, it could instead send a new UnknownSession message, or revamp
the error message to include a code, but that isn't critically needed as
webrtcsink will anyway receive the endSession message eventually.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2215>
This commit removes windows-targets 0.48.5 which does not build for
Windows 7 by updating the AWS Rust SDK; post March 11, 2025,
they've switched to using rustls for TLS, and bindgen is compulsory
for non-standard platforms.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/2199>
