In case many packets fit on a page, we may not see a granpos for
a while, and granpos interpolation can wrap the 'frames since last
keyframe' part of the granpos, generating a granpos which is smaller
than what it should be.
This is fixed by detecting keyframe packets (at least for Theora),
and updating the last keyframe granpos from this.
This may still be generating potentially wrong granpos for streams
which have a Theora like granpos (keyframes, a max keyframe distance
and a count of frames since last keyframe), and which allow implicit
granules on packets. For these streams, a custom keyframe detection
routine should be plugged into their GstOggStream mapper.
https://bugzilla.gnome.org/show_bug.cgi?id=669164
When I first implemented push mode seeking, I removed the chain
freeing there as it could be used later. The current code does not
seem to do that though, so I'm restoring the previous freeing,
which plugs the leak while apparently not reintroducing use of
freed data with chained and normal files, both with gst-launch
playbin2 and Totem.
This reverts commit 5df30c1b90.
I must have dreamt the Valgrind logs, reverting this reintroduces
no leak, and gets rid of the test failures it introduced :S
A first hang was happening when trying to locate a page backwards,
where we'd sync forever on the same page.
With that fixed, a second hang would happen after preparing an EOS
event, but with no chain created yet to send it to, the pipeline
would stay idle forever.
An element error is now emitted for this case.
Pads are initialized twice: when requesting pads and when
initializing collectpads. Avoid double initialization by
checking if collectpads are still going to be initialized when
creating request pads.
This prevents trying to seek and failing, then ending up unable
to stream because we can't get back at the headers.
A more robust way would be to find a good place to reinject the
headers when a seek fails, but I can't seem to get this to work.
Add private replacements for deprecated functions such as
g_mutex_new(), g_mutex_free(), g_cond_new() etc., mostly
to avoid the deprecation warnings. We'll change these
over to the new API once we depend on glib >= 2.32.
Replace g_thread_create() with g_thread_try_new().
If we already saw the keyframes that we need to find,
we do not need to bisect to find them.
This will always be the case for streams with audio only,
where each frame acts as a keyframe, but will occasionally
also happen for streams with video.
https://bugzilla.gnome.org/show_bug.cgi?id=662475
Opus streams outside of Ogg may not have headers, and oggstream
may be used by oggmux to mux an Opus stream which does not come
from Ogg - thus without headers.
Determining headerness by packet count would strip the first two
packets from such an Opus stream, leading to a very small amount
of audio being clipped at the beginning of the stream.
In push mode, we determine duration by doing a seek to the end of the
stream. However, a skeleton stream with an index will cause the duration
to be known already, and we end up never setting the push_time_duration
variable which we use to know duration has been determined.
https://bugzilla.gnome.org/show_bug.cgi?id=662049
The codec setup headers are a lot more likely to have correct information,
especially as it's easy to remux a skeleton in a file where streams don't
have the same parameters (I've even seen a file with two skeletons).
Still, this is useful in the case we have a codec we can't decode, so we
can at least (theoretically) convert granpos to time, so we discard this
information if the codec setup has already provided it.
This fixes playback on (at lesat) the original archive.org encoding of
"The Night of the Living Dead" (now replaced by another encoding).
https://bugzilla.gnome.org/show_bug.cgi?id=612443
This could happen when testing with navseek, and pressing
right and left at roughly the same time. The current chain
is temporarily moved away, and this caused the flush events
not to be sent to the source pads, which would cause the
data queues downstream to reject incoming data after the
seek, and shut down, wedging the pipeline.
Now, I can't really decide whether this is a nasty steaming
hack or a good fix, but it certainly does fix the issue, and
does not seem to break anything else so far.
https://bugzilla.gnome.org/show_bug.cgi?id=621897
This patch implements seeking in push mode (eg, over the net)
in Ogg, using the double bisection method.
As a side effect, it also fixes duration determination of network
streams, by seeking to the end to check the actual duration.
Known issues:
- Getting an EOS while seeking stops the streaming task, I can't
find a way to prevent this (eg, by issuing a seek in the event
handler).
- Seeking twice in a VERY short succession with playbin2 fails
for streams with subtitles, we end up pushing in a dataqueue
which is flushing. Rare in normal use AFAICT.
- Seeking is slow on slow links - byte ranges guesses could be
made better, decreasing the number of required requests
- If no granule position is found in the last 64 KB of a stream,
duration will be left unknown (should be pretty rare)
https://bugzilla.gnome.org/show_bug.cgi?id=621897
The first packet of a sparse stream may arrive after an initial
delay in the stream. If ogg_stream_packetout reports a discontinuity
in a sparse stream, do not propagate it to other streams in the
chain unnecessarily.
https://bugzilla.gnome.org/show_bug.cgi?id=621897
Remove the _ in front of the endianness prefix.
Remove the _3 postfix for the 24 bits formats.
Add a _32 postfix after the formats that occupy extra space beyond their
natural size.
The result is that the GST_AUDIO_NE() macro can simply append the endianness
after all formats and that we only specify a different sample width when it is
different from the natural size of the sample. This makes things more consistent
and follows the pulseaudio conventions instead of the alsa ones.
After all, we do hope to find actual data for these streams.
However, warn if we could not set up a chain when we find a
non BOS page, as that means we don't have a valid Ogg stream.
https://bugzilla.gnome.org/show_bug.cgi?id=657151
While the casual reader might end up bewildered by just why this
change might increase clarity, it just happens than, in the libogg
and associated sources, op is the canonical name for an ogg_packet
whlie og is the canonical name for an ogg_page, and reading this
code confuses me.
https://bugzilla.gnome.org/show_bug.cgi?id=657151
Headers are inherently durationless.
Instead, set duration to 0 to avoid increasing tracked granpos,
and do not warn about it, since it is totally expected.
https://bugzilla.gnome.org/show_bug.cgi?id=657151
Version written is 3.0
Base times are left empty for now.
Content-Type should be the MIME type of the stream. It is set to
the GStreamer media type for now, which is probably the same for
the streams oggmux supports.
https://bugzilla.gnome.org/show_bug.cgi?id=563251
If ints are 64 bits, 32 bits should get promoted in varargs anyway,
and we don't care about 16 bit ints.
This makes the code a lot more readable, and still gets us nice
hexadecimal 32 bit serialnos.
https://bugzilla.gnome.org/show_bug.cgi?id=656775
This decreases the number of buffers held on each pad by one,
eliminating next_buffer. Simplifies the logic by relying solely
on CollectPads to let us know when a pad is in EOS. As a side
benefit, the collect pads related code is structured more like
other CollectPad users.
The previous code would occasionally mark the wrong pad as EOS,
causing the code to get in a state where all the streams were
finished, but EOS hadn't been sent to the source pad.
Ogg mandates the first header packet must determine a stream's type.
However, some streams (such as VP8) do not include such a header
when muxed in other containers, and thus do not include this header
as a buffer, but only in caps. We thus use headers from caps when
available to determine a new stream's type.
https://bugzilla.gnome.org/show_bug.cgi?id=647856
