When we're doing a state change from PLAYING to NULL, first we invoke
gst_rtspsrc_loop_send_cmd_and_wait (..., CMD_CLOSE, ...) during
PAUSED_TO_READY which will schedule a TEARDOWN to happen async on the
task thread.
The task thread will call gst_rtspsrc_close(), which will send the
TEARDOWN and once it's complete, it will call gst_rtspsrc_cleanup()
without taking any locks, which frees src->streams.
At the same time however, the state change in the app thread will
progress further and in READY_TO_NULL it will call gst_rtspsrc_stop()
which calls gst_rtspsrc_close() a second time, which accesses
src->streams (without a lock again), which leads to simultaneous
access of src->streams, and a segfault.
So the state change and the cleanup are racing, but they almost always
complete sequentially. Either the cleanup sets src->streams to NULL or
_stop() completes first. Very rarely, _stop() can start while
src->streams is being freed in a for loop. That causes the segfault.
This is unlocked access is unfixable with more locking, it just leads
to deadlocks. This pattern has been observed in rtspsrc a lot: state
changes and cleanup in the element are unfixably racy, and that
foundational issue is being addressed separately via a rewrite.
The bandage fix here is to prevent gst_rtspsrc_stop() from accessing
src->streams after it has already been freed by setting src->state to
INVALID.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6346>
And also re-timestamp them with the current buffer's PTS.
Not doing so keeps the timestamps of event packets as
GST_CLOCK_TIME_NONE or the timestamp of the previous buffer, both of
which are bogus.
Making sure that (especially) the first packet has a valid timestamp
allows putting e.g. the NTP timestamp RTP header extension on it.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6294>
Parse the speed and scale in the server's response
*before* the range, so that the range start/stop
are swapped (or not swapped) correctly based
on the server's actual chosen values. Otherwise,
the old rate from the segment is used - what the
last seek asked for, but not necessarily what
the server chooses.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6287>
After a flushing seek, rtspsrc doesn't reset the last_ret value for
streams, so might immediately shut down again when it resumes pushing
buffers to pads due to a cached `GST_FLOW_FLUSHING` result
Prevent a stored flushing value from immediately stopping
playback again by resetting pad flows before (re)starting
playback.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6216>
When level value is greater than 127, it was being clamped but this clamped
value was not the one being actually used. For level values greater than 127
this resulted in an incorrect value being used. As an example, a level value
of 187, after and'ed with 0x7F, it would result in 0x3B being reported as the
level value.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5894>
When doing a segment seek, the base offset in the new segment
would be increased by segment.position which is basically the
timestamp of the last packet. This does not include the duration
of the last packet though, so might be slightly shorter than the
actual duration of the clip or the requested segment.
Increase the base offset by the segment duration instead when
accumulating segments, which is more correct as it doesn't cut
off the last frame and makes the effective loop segment duration
consistent with the actual duration returned from a duration
query.
In case a segment stop was specified it's also possible that
some data was sent beyond the stop that's necessary for decoding
so the base offset increment should be based on that then and
not on the timestamp of the last buffer pushed out.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5787>
Because we treat raw audio chunks/samples as keyframes, they were interfering
with seek time adjustment.
Became apparent when the accompanying video stream was I-frame only,
for example ProRes.
Since raw audio streams can be seeked freely, it's fine to just ignore them here,
giving priority to the real keyframes in the video stream.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5674>
With one regular image file path provided (without %05d),
the element was stuck in a dead loop counting the frames:
gst_image_sequence_src_count_frames
This allows to display any image file out of the element
for a given number of buffers.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5487>
We were already converting the pad last timestamp to running time but
not the segment position.
This segment position is used by gst_aggregator_simple_get_next_time()
to compute the waiting time when aggregating.
Those waiting times were wrong in my live pipeline using the system
clock, resulting in the aggregator to never wait at all.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5465>
The timestamp offset can be negative, and it can be a bigger negative
number than the latency introduced by the rtpjitterbuffer so the overall
timeout offset can be negative.
Using the negative offset for calculating how many packets can still
arrive in time when encountering a lost packet in an equidistant stream
would then overflow and instead of considering fewer packets lost a lot
more packets are considered lost.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5318>
Fixes https://gitlab.freedesktop.org/gstreamer/gstreamer/-/issues/2771
This EOS branch exists so that if a seek with a stop is made, qtdemux
stops accepting bytes from the sink after the entire requested playback
range is demuxed, as otherwise we could keep download content that is
not being used.
This patch fixes two flaws that were present in that EOS check:
1) A comparison was made between track time and movie time without conversion.
This made the check trigger early in files with edit lists. This patch fixes
this by converting the track PTS to movie PTS (stream time) for the check.
2) To avoid sending a EOS prematurely when the segment stop is within a GOP and
B-frames are present, the check for EOS should only be done for keyframes. I
gather this was already the intention with the existing code, but because it
used `stream->on_keyframe` instead of the local variable `keyframe` the old
code was checking if the *previous* frame was a keyframe.
It's interesting to note that these two flaws in the old code mask each other
in most cases: the track PTS will have reached the movie end PTS, but EOS would
only be sent if the previous frame was a keyframe. A simple case where they
wouldn't mask each other, reproducing the bug, is a sequence of 3 frame GOPs
with structure I-B-P.
The following validateflow tests have been added to future-proof the
fix:
* validate.test.mp4.qtdemux_ibpibp_non_frag_pull.default
* validate.test.mp4.qtdemux_ibpibp_non_frag_push.default
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5114>
We were checking if the tag list is writable, but it may actually be
shared through the same event (tee upstream or multiple consumers).
Fix a bug where multiple branches have a videoflip element checking the
taglist. The first one was changing the orientation back to rotate-0
which was resetting the other instances.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5099>
Fix this pipeline where the tag list is not writable:
gst-launch-1.0 videotestsrc ! taginject tags="image-orientation=rotate-90" ! videoflip video-direction=auto \
! autovideosink
GStreamer-CRITICAL **: 12:34:36.310: gst_tag_list_add: assertion 'gst_tag_list_is_writable (list)' failed
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4990>
Refusing an incoming segment in < GST_MATROSKA_READ_STATE_DATA should only be
done if the incoming segment is not in GST_FORMAT_TIME.
In GST_FORMAT_TIME, we are just storing the values and returning, so we can
invert the order of the checks.
Fixes proper segment propagation in matroska/webm DASH use-cases
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4922>
... otherwise streams with constant size samples defined with a single
`sample_size` for all samples in the `stsz` box fall in the category
`chunks_are_samples` in `qtdemux_stbl_init`, overriding the actual
sample count.
`FOURCC_soun` would set this automatically for `compression_id == 0xfffe`,
however `compression_id` is read from the Audio Sample Entry box at an offset
marked as "pre-defined" in some version of the spec and set to 0 both by
GStreamer and FFmpeg for opus streams.
Considering the stream `sampled` flag is set explicitely by other fourcc
variants, doing so for opus seems consistent.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4908>
The "Encapsulation of Opus in ISO Base Media File Format" [1] specifications,
§ 4.3.2 Opus Specific Box, indicates that data must be stored as big-endian.
In `build_opus_extension`, `gst_byte_writer_put*_le ()` variants were used,
causing audio streams conversion to Opus in mp4 to offset samples due to the
PreSkip field incorrect value (29ms early in our test cases).
[1] https://opus-codec.org/docs/opus_in_isobmff.html#4.3.2
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4891>
The muxer used a fixed value of 2 channels because the TR 102 366 spec
says they're to be ignored. However, the demuxer still trusted them,
resulting in bad caps.
Make the muxer fill in the correct channel count anyway (FFmpeg already
does) and make the demuxer ignore the value.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4773>
Since c0bf793c05 ("flvmux: Set PTS based on
running time") the timestamp of the output buffer is already in running
time. So using that for 'srcpad->segment.position' does not work correctly
because gst_aggregator_simple_get_next_time() will convert it again with
gst_segment_to_running_time().
This means that the timestamp returned by
gst_aggregator_simple_get_next_time() may be incorrect. For example, if
flvmux is added to a already runinng pipeline then the timestamp is too
small and gst_aggregator_wait_and_check() returns immediately. As a result,
buffers may be muxed in the wrong order.
To fix this, use the PTS of the incoming buffer instead of the outgoing
buffer. Also add the duration as get_next_time() is supposed to return the
timestamp of the next buffer, not the current one.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4734>
Make splitmuxsrc deal better with stream reordering by
making the largest observed PTS contiguous in the
next fragment. Previously, it selected DTS, but then
aligned that with the segment start of the next fragment,
which holds PTS values - leading to glitches in
streams that don't have PTS = DTS at the start.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4660>
The index is already incremented by 3 every iteration so multiplying it
by 3 additionally on each array access is doing it twice and does not
work.
This caused invalid files to be created if there's more than one CEA608
triplet in a buffer, and out of bounds memory reads.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4645>
This is a fix for a data race leading to:
> GLib-CRITICAL: g_hash_table_foreach:
> assertion 'version == hash_table->version' failed
Identified sequence:
* `rtp_session_on_timeout` acquires the lock on `session` and proceeds with its
processing.
* `rtp_session_process_rtcp` is called (debug log : received RTCP packet) and
attempts to acquire the lock on `session`, which is still held by
`rtp_session_on_timeout`.
* as part of an hash table iterator, `rtp_session_on_timeout` transitively
invokes `source_caps` which releases the lock on `session` so as to call
`session->callbacks.caps`.
* Since `rtp_session_process_rtcp` was waiting for the lock to be released, it
succeeds in acquiring it and proceeds with `rtp_session_process_rr` which
transitively calls `g_hash_table_insert` via `add_source`.
* After `source_caps` re-acquires the lock and gives the control flow back to
`rtp_session_on_timeout`, the hash table iterator is changed, resulting in the
assertion failure.
This commits copies `sess->ssrcs[sess->mask_idx]` and iterates on the copy so
the iterator is not affected by a concurrent change due to the lock being
released in the `source_caps` callback.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4585>
