We currently send data to the RTSP connection from multiple threads:
whenever a command is to be handled and whenever RTCP is generated. This
can cause data corruption or worse if both happen at the same time.
As such, protect gst_rtsp_connection_send() and gst_rtsp_connection_receive()
calls with a mutex. While this means that we hold a mutex during the IO
operation, this is not actually a problem as the IO operation can be
interrupted (gst_rtsp_connection_flush()) at any time and is blocking by
itself anyway.
The last entry will most likely get new samples added to it in "robust"
muxing mode, changing the samples_per_chunk and thus making it wrong to
keep the last two entries merged. It will run into an assertion later
when adding a new sample to the chunk.
Thanks to gdiener@cardinalpeak.com for the analysis of the bug and
proposal for a solution.
There might be other chunks after the data chunk, so clipping the chunk
size with the data size can lead to a negative number and all following
calculations go wrong and cause crashes or worse.
This was introduced in 3ac119bbe2.
https://bugzilla.gnome.org/show_bug.cgi?id=783760
They can cause us to deadlock, while we're waiting for a new frame and
upstream is waiting for the allocation query to be answered before
sending a frame
https://bugzilla.gnome.org/show_bug.cgi?id=783753
There is no difference between pushing out a buffer directly
with gst_rtp_base_depayload_push() and returning it from the
process function. The base class will just call _depayload_push()
on the returned buffer as well.
So instead of marshalling buffers through three layers and back,
just push them from one place in handle_nal() and always return
NULL from the process vfunc. This simplifies the code a little.
Also rename _push_fragmentation_unit() to _finish_fragmentation_unit()
for clarity. Push sounds like it means being pushed out, whereas
it might just be pushed into an adapter.
This change has the side-effect that multiple NALs in a single STAP
(such as SPS/PPS) may no longer be pushed out as a single buffer if
we output NALs in byte-stream format (i.e. not aggregate AUs), but
that shouldn't really make any difference to anyone.
Use the ::process_rtp_packet() vfunc to avoid mapping the
RTP buffer twice.
gst_rtp_buffer_get_payload_buffer() returns a new sub-buffer
which will always be writable, so no need to make it writable.
Every g_quark_from_static_string() is a hash table lookup serialised
on the global quark lock in GLib. Let's just look up the two quarks
we need once and cache them locally for future use. While we're at it,
add new utility functions for the two most commonly used tags
(audio + video). Make first argument a gpointer so we don't have to
cast and make the code ugly. These are used for logging purposes
only anyway.
Since the move from CVS the property name of the documentation example
has been filename instead of location. Users trying the gst-launch
command as is will get:
no property name "filename" in element
Fixing it.
If a non-reference stream is behind the reference stream by an amount of
time smaller than the alignment threshold (in nsec), it counts as being
after it.
https://bugzilla.gnome.org/show_bug.cgi?id=782563
Timecode trak is only supported for mov right now, not for mp4. That
code would otherwise create an invalid trak if the muxed video contained
timecode metadata.
https://bugzilla.gnome.org/show_bug.cgi?id=782684
We only accept new caps if they are basically the same. We don't want to
reset anything as if the caps are new, otherwise various state could get
out of sync with the current run.
We have some padding added after the initial moov, so a bigger updated
moov can be handled to some degree and is expected. Previously we just
ignored the padding and errored out in cases when the padding would've
just been enough.
This sets up a moov with the correct sample positions beforehand and
only works with constant framerate, I-frame only streams.
Currently only support for ProRes and raw audio is implemented but
adding new codecs is just a matter of defining appropriate maximum frame
sizes.
https://bugzilla.gnome.org/show_bug.cgi?id=781447
When muxing raw audio, we have no way of storing timestamps but are just
storing a continuous stream of audio samples. If the difference between
the expected and the real timestamp becomes to big, we should error out
instead of silently creating files with wrong A/V sync.
https://bugzilla.gnome.org/show_bug.cgi?id=780679
Re-arrange order of index entry struct members to avoid padding
bytes in the middle of the struct, thus potentially reducing the
overall size of the struct and reducing memory used by the index.
On Linux x86_64 the size goes down from 32 bytes to 24 bytes for
each index entry.
If no clock was provided directly by rtspsrc. This behaviour was removed
by f8013487c9 and results in rtspsrc not
providing the system clock via the rtpjitterbuffer.
As a result, if another element like an audio sink, provides a clock,
the pipeline would select that (when going to PAUSED/PLAYING again later).
Audio clocks usually don't progress in PAUSED, and thus our live source
won't be able to use the clock to produce data, making the sink never
preroll and everything is stuck.
... unless the muxer uses the same audio pad template name as
splitmuxsink. We can't request a pad called "audio_0" on a muxer that
wants pads to be "sink_%d".
In push mode we process as much as possible in the adapter. When we receive
a DISCONT buffer which we can't match to an actual sample (based on the existing
sample table) and there is still data remaining in the incoming adapter,there is
one of two cases happening:
1) We are doing reverse playback, in which case we should flush out all pending
data
2) We have leftover data from the previous incoming buffer... which we can't do
anything about.
For the second case, make sure we flush out the remaining data so that we can start
parsing again from scratch.
https://bugzilla.gnome.org/show_bug.cgi?id=781319
They should have ideally the same timescale of the video track, which we
can't guarantee here as in theory timecode configuration and video
framerate could be different. However we should set a correct timescale
based on the framerate given in the timecode configuration, and not just
use the framerate numerator.
Make sure offset and neededbytes are properly resetted when all
streams are EOS in push-mode.
Avoids cases when some data might still be pushed by upstream (because
it didn't yet see the resulting GST_FLOW_EOS yet) and qtdemux gets
completely lost.
https://bugzilla.gnome.org/show_bug.cgi?id=781266
buf is the current pad->last_buf value. If ever it gets copied/unreffed,
we need to make sure to write back the new pointer to the last_buf
variable.
Fixes using wrong pointer values in the case of decrasing DTS value
Before pushing a sample, check if there was a change in the current
stsd entry. This patch also assumes that the first stsd entry is
used as default for the first sample. It might cause an uneeded
caps renegotiation when this isn't the case.
stsd can have multiple format entries, parse them all.
This is required to play DVB DASH profile that uses multiple entries
to identify the different available bitrates/options on dash streams
The stream format-specific data is not stored into QtDemuxStreamStsdEntry
Instead of using the stsd as a base pointer, use the actual stsd
entry as the stsd can have multiple entries. This is rarely used
for file playback but is a possible profile with in DVB DASH specs.
This still doesn't support stsd with multiple entries but makes it
easier to do so.
AudioSpecifigConfig is used in a variety of AAC streams but was
being parsed differently. Instead, make everyone use the same parsing.
* Remove unused 'bits' field (it was always set to 0 if present)
* Add proper GAConfig parsing (to know the number of samples per frame
if present).
Fixes wrong rate/channels configuration in streams coming from qtdemux
https://bugzilla.gnome.org/show_bug.cgi?id=780966
According to ISO/IEC:14496-2:2009 , in the case of HE-AACv2 (audioObjecType
29) parametric stereo is used (a single mono track is used and then
transformations are applied to it to provide a stereo output).
We therefore report two channels in the case where there is one reported
in the audioChannelConfiguration.
Fixes the various issues where a demuxer would report two channels, but
then the parser would say there's only one channel, and then the decoder
would output two channels.
last_buf is the one we're going to write next, not buf. As such we
should check timestamps against that one if there is one to select the
earliest pad.
Also remember the currently selected pad in the very beginning when
storing the first last_buf.
This both solves some edge cases where not the correct next pad was
selected corresponding to the target interleave.
This is an update of d78d589627
We still exit as early as possible in case of non-ok/non-unlinked combined
flow, but we first make sure that we update the internal position variables.
This ensures that if upstreams "ignores" the flow return (and carries on pushing),
we don't end up processing data with completely bogus variables/positions.
If self->channel_positions == NULL (which seems unlikely),
self->default_channels_ordering_map will be used unintialised.
We avoid that by keeping track of the channel_mask, which is set when
the ordering map is initialised.
https://bugzilla.gnome.org/show_bug.cgi?id=780331
When there are more than 64 channels, we don't want to exceed the
bounds of the ordering_map buffer, and in these cases we don't want to
rempa at all. Here we avoid doing that.
https://bugzilla.gnome.org/show_bug.cgi?id=780331
TFDTs with time 0 are being ignored since commit 1fc3d42f. They're
mistaken with the case of not having TFDT, but those two cases
must be distinguished in some way.
This patch passes an extra boolean flag when the TFDT is present.
This is now the condition being evaluated, instead of checking for
0 time.
https://bugzilla.gnome.org/show_bug.cgi?id=780410
If we have multiple tracks with timecodes, or it's not the first track
that has timecodes, or not the first buffer, we already started a chunk
for media data. We now need to "close" that chunk because we wrote data
for the timecode track and a new chunk has to be started for the
original track the next time it has data.
Similar to what was done in adaptivedemux, ignore seek
events we've already handled - such as when they are received
on every srcpad of files with lots of streams.
Otherwise mdatleft will have a value calculated from the initial
mdatsize minus the parts of the stream that we saw, which is not
including all the parts of the stream that might've been skipped.
This breaks gst-validate on the build server (though not locally),
and a unit test, and I can't run unit tests right now for some
unrelated reason.
This reverts commit 0747b56f8e.
This debug statement is meant to print the time since the last (early)
RTCP transmission, not the last regular RTCP transmission (which also
happens to be set a few lines above to current_time, so the debug output
is just confusing)
Take into account the atoms at the end of the 'trak' atom when
recovering it. So that its size (already computed and added in the trak
size) isn't making offsets wrong.
https://bugzilla.gnome.org/show_bug.cgi?id=771478
Fix the check for whether the start time of the segment has
been reached when playing in reverse. Otherwise, playback
stops after reaching the start of any file part, instead of
continuing until all parts within the segment have played
We parse the next moof in advance of having pushed
all samples from the previous one in some cases, and
we'll still need the crypto info from the previous
fragment so keep around any unused crypto info entries
when adding new ones
qtdemux.c: In function ‘qtdemux_parse_samples’:
qtdemux.c:8450:39: error: ‘*’ in boolean context, suggest ‘&&’ instead [-Werror=int-in-bool-context]
if (stream->samples_per_frame * stream->bytes_per_frame) {
~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~
gstmpegaudioparse.c: In function ‘gst_mpeg_audio_parse_reset’:
gstmpegaudioparse.c:209:3: error: ‘memset’ used with length equal to number of elements without multiplication by element size [-Werror=memset-elt-size]
memset (mp3parse->xing_seek_table_inverse, 0, 256);
^~~~~~
gstmpegaudioparse.c: In function ‘gst_mpeg_audio_parse_handle_first_frame’:
gstmpegaudioparse.c:951:7: error: ‘memset’ used with length equal to number of elements without multiplication by element size [-Werror=memset-elt-size]
memset (mp3parse->xing_seek_table_inverse, 0, 256);
^~~~~~
This prevents storing an infinite amount of e.g. comment headers if they
come without a new initialization header in front of them. There can
only be one header of each type.
If we also replace all headers when receiving any possibly following
comments header, we would throw away the config header before being able
to make use of it.
A sparse stream's ending timestamp can be considerably smaller
than the ending timestamps of the other streams, which can lead
to skipping considerable time from the next part.
https://bugzilla.gnome.org/show_bug.cgi?id=761086
In function rtp_jitter_buffer_calculate_pts: If gap in incoming RTP
timestamps is more than (3 * jbuf->clock_rate) we call
rtp_jitter_buffer_reset_skew which resets pts to 0. So components down
the pipeline (playes, mixers) just skip frames/samples until pts becomes
equal to pts before gap.
In version 1.10.2 and before this checking was bypassed for packets with
"estimated dts", and gaps were handled correctly.
https://bugzilla.gnome.org/show_bug.cgi?id=778341
The payloader needs to reset and update the vorbis config data which is
pushed on the network if it receives new headers, or at least, it may
have to do so.
Without this, the stream configuration could change without the
payloader sending the new configuration to the other side.
This reverts commit 107902ec51.
This commit intended to ensure that keyframe seeks land at the
start timestamp of a keyframe, rather than in the middle of one,
but they cause trouble on files with sparse streams, or with
JPEG 'cover art' tracks that have only one or a few JPEG samples
with very long durations.
That's still desirable for doing seamless cutting of videos,
but needs a rethink for implementation.
https://bugzilla.gnome.org/show_bug.cgi?id=778690
Add a new boolean surround-delay property that makes
audioecho just apply a delay to certain channels to create
a surround effect, rather than an echo on all
channels. This is useful when upmixing from stereo - for example.
Add a surround-mask property to control which channels
are considered surround sound channels when adding a
delay with surround-delay = true
Original patch from Jochen Henneberg <jh@henneberg-systemdesign.com>
This goes around the inefficient control message based filtering and
does all the filtering kernel-side. Unfortunately this is Linux-only and
there is no IPv6 variant of it (yet).
Some radio streams uses StreamTitle='' to reset the title after a
track stopped playing, e.g. while the host talks between tracks or
during news segments.
This change forces an empty tag object to be distributed if
StreamTitle or StreamUrl is received with empty value, thus allowing
downstream elements to get notified about this.
https://bugzilla.gnome.org/show_bug.cgi?id=778437
Upstream elements like videoflip can transform caps, such as changing width and height.
When an imagefreeze downstream receives an ACCEPT_CAPS query it will NOW return
all caps that it can accept.
https://bugzilla.gnome.org/show_bug.cgi?id=778389
Used signed calculations when measuring the max_ts of an input
fragment, so as to calculate the correct duration and offset
when buffers have timestamps preceding their segment
The n_frames field (frames per second) should follow the nominal frame
rate for drop-frame timecodes.
Also, the trak's timescale (and duration, accordingly) should follow the
STSD entry's timescale and frame duration (fps_n and fps_d accordingly),
not the other way around.
https://bugzilla.gnome.org/show_bug.cgi?id=777832
In case wavparse receives a manually injected FLUSH_STOP event
while operating in pull mode we get criticals because we'd try
to clear a NULL adapter.
https://bugzilla.gnome.org/show_bug.cgi?id=777123
Insert VPS/SPS/PPS before the first NAL unit containing an I-frame in an
access unit only. If an access unit consists of several such NAL units
(tiles) VPS/SPS/PPS should only be inserted before the first of them so
that parameters are only updated between frames.
Do not insert VPS/SPS/PPS before P-frames when config-interval is -1.
https://bugzilla.gnome.org/show_bug.cgi?id=775817
qtdemux_handle_xmp_taglist() requires a writable taglist,
but qtdemux->tag_list can become non-writable, specifically
after sending global tags (qtdemux.c:958), which adds a
second reference. Ensure the list is made writable before
calling (make_writable will copy the list if necessary).
https://bugzilla.gnome.org/show_bug.cgi?id=766177
These are usually much bigger than icon size and required by
iTunes to be certain fairly large sizes. In qtmux it is also
the IMAGE tags which we write out as 'covr' atoms.
When reset, don't restart request pad numberings, as
request pads can survive across state changes. Only
restart at 0 if all request pads are handed back first.
https://bugzilla.gnome.org/show_bug.cgi?id=777174
'stream-format' and 'alignment' are defined in pad template caps so
there is no need to check them again here. Also remove bitrate parsing from
caps as bitrate in caps doesn't make sense but from tags, which is
actually the case.
https://bugzilla.gnome.org/show_bug.cgi?id=777181
Needed for QuickTime 7 to properly play files.
Also write the clap atom for MOV files always, not only when ProRes is
used as a video codec. It's mandatory for MOV.
https://bugzilla.gnome.org/show_bug.cgi?id=777100
The seqh buffer allocated in qtdemux_parse_svq3_stsd_data() needs to
be freed by the caller after use.
https://bugzilla.gnome.org/show_bug.cgi?id=777157
Signed-off-by: Andre McCurdy <armccurdy@gmail.com>
If a fragmented stream doesn't have a tfdt, don't
reset the output timestamps at each fragment boundary
by erroneously using the default value of 0. Introduced
by commit 69fc48
https://bugzilla.gnome.org/show_bug.cgi?id=754230
Majorly change the way that splitmuxsink collects
incoming data and sends it to the output, so that it
makes all decisions about when / where to split files
on the input side.
Use separate queues for each stream, so they can be
grown individually and kept as small as possible.
This removes raciness I observed where sometimes
some data would end up put in a different output file
over multiple runs with the same input.
Also fixes hangs with input queues getting full
and causing muxing to stall out.
Add a new signal for formatting the filename, which receives
a GstSample containing the first buffer from the reference
stream that will be muxed into that file.
Useful for creating filenames that are based on the
running time or other attributes of the buffer.
To make it work, opening of files and setting filenames is
now deferred until there is some data to write to it,
which also requires some changes to how async state changes
and gap events are handled.
When performing a key-unit seek, always snap to the start ts
of the keyframe buffer we landed on so that the keyframe is
entirely within the resulting outgoing segment. That seems
the most sensible result, since the user requested snapping
to the keyframe position.
Segments times and seek requests are stored and handled
in raw 'PTS' time, without the cslg_shift - which only applies
to outgoing samples. Omit the cslg_shift portion when
extracting PTS to compare for internal seek snaps.
If the cslg_shift is included, then keyframe+snap-before seeks
generate a segment start/stop time that already includes the
cslg_shift, and it's then added a 2nd time, causing the
first buffer(s) to have timestamps that are out of segment.
Remove an old check from atom_stsc_add_new_entry() that
extends the last entry in the STSC if the samples per chunk
matches, as the new interleave merging logic requires that
the final entry by updateable. There's already code
below which simply merges the final entry into the previous
one when needed, so rely on that instead.
Fixes asserts like:
ERROR:atoms.c:2940:atom_stsc_update_entry: assertion failed:
(atom_array_index (&stsc->entries, len - 1).first_chunk == first_chunk)
Make sure the state of the parser is set to
collecting streams before chaining up to the
parent change_state() method, to close a
small window that can cause playback to
never commence.
Use GQueue instead of a GSList so we don't have to traverse
the whole list to append something every time. And it also
keeps track of the number of items in it for us.
Add a function to add filenames to the list of old files and
use it in more places, so that memory doesn't build up in
other modes either if no max_files limit is specified.
https://bugzilla.gnome.org/show_bug.cgi?id=766991
Technically we weren't leaking the memory, just storing it internally
and never using it until the element is freed. But we'd still use more
and more memory over time, so this is not good over longer periods
of time. Only keep track of files if there's actually a limit set,
so that we will prune the list from time to time.
https://bugzilla.gnome.org/show_bug.cgi?id=766991
Previously, seeking to position y where y is (strictly) within a keyframe
would seek to that keyframe both with SNAP_BEFORE and SNAP_AFTER,
where the latter is now adjusted to really snap to the next keyframe.
Rather amazingly (and equally unnoticed), keyunit seeking resulted in segments
where start != time (which is bogus for simple avi timeline). So, properly
adjust the segment (start) rather than fiddling with segment time (only).
... by using the original seek event's flags rather than the corresponding
segment flags, which do not have such counterpart flags (and
do no longer have them covertly sneaking in nowadays).
With Xiph codecs the stream header buffers are both in the caps and are
usually also at the beginning of each input stream, but it's perfectly
possible that the input stream does not have the stream header buffers
inline in the data. Matroskamux would drop the first N buffers assuming
they're stream headers, but this meant it would drop actual payload data
when the stream didn't contain the stream headers inline. Fix this by
only dropping leading buffers if they're flagged as stream headers. This
fixes issues with streams that are being tapped into after streaming
has started.
https://bugzilla.gnome.org/show_bug.cgi?id=749098
That is, whenever we go through start/stop we have to ensure that on the
next opportunity the buffers are reallocated again. Otherwise the
buffers might be NULL because the element was reused with the same
configuration as before (i.e. set_caps() wouldn't have reinited the
buffers).
https://bugzilla.gnome.org/show_bug.cgi?id=775898
Redirect on PLAY wasn't doing the necessary session cleanup. Fixed by
removing code from gst_rtspsrc_send that changed the state varable upon
encountering a redirect. Better to let the redirect handlers in
gst_rtspsrc_retrieve_sdp and gst_rtspsrc_play do their own
state-dependent cleanup.
https://bugzilla.gnome.org/show_bug.cgi?id=775543
When providing items with a seqnum, there is a (very small) probability
that an element with the same seqnum already exists. Don't forget
to free that item if it wasn't inserted.
And avoid returning undefined values when dealing with duplicate items
We can't simply assume that the length of the tag value as given
inside the stream is correct but should also check against the amount of
data we have actually available.
https://bugzilla.gnome.org/show_bug.cgi?id=775451
qtdemux.c: In function ‘qtdemux_parse_trak’:
qtdemux.c:10184:38: error: format ‘%lu’ expects argument of type ‘long unsigned int’, but argument 9 has type ‘gint {aka const int}’ [-Werror=format=]
GST_DEBUG_OBJECT (qtdemux, "Found jpeg: len %u, need %lu", len,
^
If an element queries the number of retransmission buffers pushed
*while* the push is still taking place (and before the object lock
is taken just after) it would end up with the wrong statistic
being reported.
Increment it just before the push, avoids races when getting statistics
https://bugzilla.gnome.org/show_bug.cgi?id=768723
39f7e52266 was setting the buffer duration
to 0 if is not valid, under the assumption that this is "the last"
buffer and no others are coming next. This is wrong, last_buf is the
previous buffer and not the very last one.
4e3c13c87c was setting DTS to 0 if there
was none. This will set DTS to 0 for all e.g. audio streams, completely
messing up calculations if streams don't start at 0.
https://bugzilla.gnome.org/show_bug.cgi?id=774840
Solves overreading/writing the given arrays and will error out if the
streams asks to do that.
Also does more error checking that the stream is valid and won't
overrun any allocated arrays. Also mitigate integer overflow errors
calculating allocation sizes.
https://bugzilla.gnome.org/show_bug.cgi?id=774859
After finding a cluster id in the byte reader, we skip ahead the reader
position by one further byte to be able to continue searching from there
inside the same chunk if the cluster candidate was a false positive.
We have to accomodate for that additional byte when resuming the search,
otherwise all following pulls are off-by-one for every resume and we run
into an assertion.
bf43f44fcf was comparing an unsigned
expression to be < 0 which was always false.
gstflxdec.c: In function ‘flx_decode_brun’:
gstflxdec.c:322:33: warning: comparison of unsigned expression < 0 is always false [-Wtype-limits]
if ((glong) row - count < 0) {
^
gstflxdec.c:332:33: warning: comparison of unsigned expression < 0 is always false [-Wtype-limits]
if ((glong) row - count < 0) {
^
https://bugzilla.gnome.org/show_bug.cgi?id=774834
| ../../../git/gst/isomp4/qtdemux.c: In function 'qtdemux_parse_tree':
| ../../../git/gst/isomp4/qtdemux.c:10224:24: error: 'size' may be used uninitialized in this function [-Werror=maybe-uninitialized]
| offset += size;
| ^~
| ../../../git/gst/isomp4/qtdemux.c:10197:25: note: 'size' was declared here
| guint32 size, tag;
| ^~~~
https://bugzilla.gnome.org/show_bug.cgi?id=774747
Always write an edit list for the whole track. In general this is not
necessary except for the case of having a gap or DTS adjustment but
it allows to give the whole track's duration in the usually more
accurate media timescale.
https://bugzilla.gnome.org/show_bug.cgi?id=774403
splitmuxsink requests pad from element using pad template like "video_%u", "audio_%u" and "sink_%d". This is true for most of the muxers.
But splitmuxsink not able to request pad to flvmux as flvmux has "audio" and "video" as pad templates.
fix: splitmuxsink should fallback to "audio" and "video" when template not found.
https://bugzilla.gnome.org/show_bug.cgi?id=774507
