This fixes the case when synchronization is established but actually not
valid anymore. In such a case, the code would _first_ write to the ring
buffer (at the wrong position due to the invalid sync), and _then_ detect
the bogus synchronization. Reorder the code blocks to _first_ check the
current sync, then resynchronize if neeeded (or perform initial sync if
no sync is established yet), and _then_ write to the ring buffer.
Until now, the timestamp check was comparing the timestamp delta against
the value of the "quantum" variable. However, the timestamps use clock
samples as units, while the "quantum" variable uses nanoseconds. The
outcome is that this check virtually never returned true. Use the
spa_io_clock duration instead of that quantum nanosecond duration to make
the check actually work.
Also, do not just rely on vast timestamp deltas to detect discontinuities;
instead, check first for the presence of the SPA_IO_CLOCK_FLAG_DISCONT
flag to detect said discontinuities.
Instead of implicitly acting on the current thread if the provided thread
handle is unknown, return an error. This behaviour is not depended on inside
pipewire, and if needed, special casing e.g. `thread == NULL` could be added,
but any random `thread` value shouldn't affect the current thread.
Make a new body.h file with some functions to deal with pod and their
body. Make the iter.h functions use mostly this.
Rework the parser so that it only uses body.h functions. With the separation
of pod+body, we can read and verify the pod once and then use the
verified copy to handle the rest of the body safely.
We do this because iter.h only works in pods in memory that doesn't change
because it is vulnerable to modifications of the data after verifying it.
The new parser is not vulnerable to this and will not cause invalid
memory access when used on shared memory. There is however no need for
atomic operations to read the headers, whever is read is either valid
and useable of invalid and rejected.
See #4822
The parser does not check that POD arrays have the correct size for
their type, so the calling code must do that.
This also enumerates some of the code that cannot handle the size of the
values of an array not being the exact expected size for its type.
There is a lot of it.
Direct timestamp mode was incorrectly using over/underrun detection logic
and fill level tracking logic that is actually meant for the other mode
(referred to from now on as "constant latency mode"). Over/underruns are
tracked implicitly in the direct timestamp mode, and the absolute fill
level is not relevant in that mode, since the latency is not needed to
be constant then.
Also improve log lines and the RTP module documentation to define these
buffer modes clearly and explain their differences and use cases.
Opus and MIDI code get TODOs added, since their direct timestamp mode
implementations still may be incorrect. Fixing those will be done in
a separate commit.
When a stream has some delay, a time t1 + delay has to be read in time
t1 to play it when expected.
Decrease target_buffer by delay to start playback sooner, so sound
is played at correct time when delay is applied.
Signed-off-by: Martin Geier <martin.geier@streamunlimited.com>
We also need to close the SynObj fd we got, just like we close any
DmaBuf or MemFd.
Make sure we get a compiler error when we add more items to the
data type enumeration later.
Fixes#4807
It uses the onnxruntime library to parse the onnx file and construct a
neural network. It uses the label field to setup the plugin and how to
map the various tensors of the model to input, output, control and
notify ports.
Add an example config for how to use the silero VAD ONNX model with the
noise gate.
The return value is always 0, and the `impl` parameter
is not used, so ues the return value to return the boolean
result instead of an out parameter, and get rid of the
unused argument.
Instead of using a new macro with the `PW_` prefix, simply define
`SCHED_RESET_ON_FORK` to be `0` when it is not defined; as the
prefixed variant can be a bit confusing.
Reset buffers when deactivating to avoid having old data in the
ringbuffers, which also adds latency when activated again.
Clear sink_ready and capture_ready when resetting buffers to avoid
calling process() before there is new data to process.
capture and sink streams may start before playback stream so process()
may fail to dequeue a playback buffer. In that case advance the read
pointers to avoid building up latency in the ringbuffers.
