Background:
Whenever a caps event is received by appsink, the caps are stored in the
same internal queue as buffers. Only when enough buffers have been
popped from the queue to reach the caps, `priv->sample` gets its caps
updated to match, so that they are correct for the following buffers.
Note that as far as upstream elements are concerned, the caps of appsink
are updated immediately when the CAPS event is sent. Samples pulled from
appsink retain the old caps until a later buffer -- one that was sent by
upstream elements after the new caps -- is pulled.
The race condition:
When a flush is received, appsink clears the entire internal queue. The
caps of `priv->sample` are not updated as part of this process, and
instead remain as those of the sample that was last pulled by the user.
This leaves open a race condition where:
1. Upstream sends a new caps event, and possibly some buffers for the
new caps.
2. Upstream sends a flush (possibly from a different thread).
3. Upstream sends a new buffer for the new caps. Since as far as
upstream is concerned, appsink caps are the new caps already, no new
CAPS event is sent.
4. The appsink user pulls a sample, having not pulled before enough
samples to reach the buffers sent in step 1.
Bug: the pulled sample has the old caps instead of the new caps.
Fixing the race condition:
To avoid this problem, when a buffer is received after a flush,
`priv->sample`'s caps should be updated with the current caps before the
buffer is added to the internal queue.
Interestingly, before this patch, appsink already had code for this, in
gst_app_sink_render_common():
/* queue holding caps event might have been FLUSHed,
* but caps state still present in pad caps */
if (G_UNLIKELY (!priv->last_caps &&
gst_pad_has_current_caps (GST_BASE_SINK_PAD (psink)))) {
priv->last_caps = gst_pad_get_current_caps (GST_BASE_SINK_PAD (psink));
gst_sample_set_caps (priv->sample, priv->last_caps);
GST_DEBUG_OBJECT (appsink, "activating pad caps %" GST_PTR_FORMAT,
priv->last_caps);
}
This code assumes `priv->last_caps` is reset when a flush is received,
which makes sense, but unfortunately, there was no code in the flush
code path resetting it.
This patch adds such code, therefore fixing the race condition. A unit
test demonstrating the bug and testing its behavior with the fix has
also been added.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2442>
If for some reason the encoder produces frames with a pts higher than
the input one, we were dropping all the video encoder frames and ended
up crashing when trying to access the pts of a NULL pointer returned by
gst_video_encoder_get_oldest_frame().
I hit this scenario by feeding a decreasing timestamp to vp8enc which
seem to confuse the encoder.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2436>
gst_value_serialize() does more than what's needed to printf-ing
especially when given GValue is already string. Just print string
value as-is without gst_value_serialize() to avoid unreadable
string print, especially for multi-bytes character encoding cases.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2416>
The timestamp in the tfdt refers to the first trun box and if there are
multiple trun boxes then the distance between the first timestamps will
grow.
At some point this distance reaches a threshold and triggers the
resetting of the first sample's timestamp of this trun box to be reset
to the tfdt.
This threshold is implemented for files where there is a jump in the
timeline between fragments and where this can be detected via a jump
between the end timestamp of the previous fragment and the tfdt of the
next. This behaviour is preserved.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2414>
Until March 2022, the FFmpeg MXF muxer would write the various index table
segments with the same instance ID, which should only be used if it is a
duplicate/repeated table.
In order to cope with those, we first compare the other index table segment
properties (body/index SID, start position) before comparing the instance
ID. This will ensure that we don't consider them as duplicate, but can still
detect "real" duplicates (which would have the same other properties).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2408>
Hantro H1 and Rockchip VEPU2 drivers will pad the width/height to a
multiple of 16. In order to obtain the right JPEG size, the image needs
to be cropped using the S_SELECTION API. This support is added as best
effort since older drivers may emulate this by looking at the capture
queue width/height.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2392>
The gst_v4l2_object_set_crop() is used for removing buffer
alignment padding. Give it a name that better reflects
that usage. This helps to distinguish from cropping of the
input image (e.g. cropping at the image sensor on a captre
device), which can be unrelated to the memory buffer padding,
especially if scaling is involved.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2392>
mp4mux can't negotiate caps with upstream/downstream and always outputs
specific caps based on the input streams. This will always happen before
it produces the first buffers.
By having the default aggregator negotiation enabled the same caps
would be pushed twice in the beginning, and again every time a
reconfigure event is received.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2381>
mxfmux can't negotiate caps with upstream/downstream and always outputs
specific caps based on the input streams. This will always happen before
it produces the first buffers.
By having the default aggregator negotiation enabled the same caps
would be pushed twice in the beginning, and again every time a
reconfigure event is received.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2381>
mpegtsmux can't negotiate caps with upstream/downstream and always outputs
specific caps based on the input streams. This will always happen before
it produces the first buffers.
By having the default aggregator negotiation enabled the same caps
would be pushed twice in the beginning, and again every time a
reconfigure event is received.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2381>
flvmux can't negotiate caps with upstream/downstream and always outputs
specific caps based on the input streams. This will always happen before
it produces the first buffers.
By having the default aggregator negotiation enabled the same caps
would be pushed twice in the beginning, and again every time a
reconfigure event is received.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2381>
Otherwise setting the srcpad caps based on the sinkpad caps event will
already push a segment event downstream before the upstream segment is
known.
If the upstream segments are just forwarded when the upstream segment
event arrives this would result in two segment events being sent
downstream, of which the first one will usually be simply wrong.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2381>
Baseclass calls get_preferred_output_delay() in a chain of
sequence header parsing and then new_sequence() is called
with required DPB size (includes render-delay) information.
Thus latency query should happen before the sequence header
parsing for subclass to report required render-delay accordingly
via get_preferred_output_delay() method.
(e.g., zero delay in case of live pipeline)
This commit is to fix wrong liveness signalling in case of
upstream packetized format.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2380>
GstD3D11ScreenCapture object is pipeline-independent global object
and the object can be shared by multiple src elements,
in order to overcome a limitation of DXGI Desktop Duplication API.
Note that the API allows only single capture session in a process for
a monitor.
Therefore GstD3D11ScreenCapture object must be able to handle a case
where a src element holds different GstD3D11Device object. Which can
happen when GstD3D11Device context is not shared by pipelines.
What's changed:
* Allocates capture texture with D3D11_RESOURCE_MISC_SHARED for the
texture to be able to copied into other device's texture
* Holds additional shader objects per src element and use it when drawing
mouse
Fixes: https://gitlab.freedesktop.org/gstreamer/gstreamer/-/issues/1197
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2378>
Our decoder implementation does not use downstream d3d11 pool for
decoding because of special requirement of D3D11/DXVA. So preallocation
using the downstream buffer pool will waste GPU memory in most cases.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2377>
If there weren't any moved/dirty regions in the captured frame, the
viewport of the ID3D11DeviceContext would be left at whatever previous
value it had, which could lead to the cursor being drawn in a wrong
position and/or in an incorrect size.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2371>
If a file includes a new version of a plugin that exits in the
registry, the output of gst-inspect is incorrect. The output has the
correct version but incorrect filename, and element description.
This seems to have also fixed some documentation issues.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2357>
Some streams have 2 PMT sections in a single TS packet. The first one is "valid"
but doesn't contain/define any streams. That causes an unrecoverable issue when
we try to activate the 2nd (valid) PMT.
Instead of doing that, pre-emptively refuse to process PMT without any streams
present within. We still do post that section on the bus to inform applications.
Fixes#1181
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2327>
In push mode (streaming), if the received chunk buffer size from _chain is bigger
than output buffer size, the flags of the divided-buffers are propagated to the
DISCONT flag from first received chunk buffer. This unexpected buffers contained DISCONT
flags are abnormally transformed when changing the sampling rate by audioresample element.
So unset unnecessary DISCONT flag before pad_push().
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2322>
The decision to store the input buffer depends on whether extensions
are to be added to the output buffer, I assume as an optimization.
This creates an issue for subclasses that call negotiate(), where
header_exts is actually populated, from their handle_buffer()
implementation: at chain time, no header extension has been negotiated
yet, which means that we don't add extensions to the first batch of
buffers that comes out.
Keep track of whether negotiate has been called (this is different
from the negotiated field) and always store the input buffer until
then. This fixes the issue while largely preserving the optimization.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/2321>
