There were two issues with the previous decodebin2 group switching algorithm:
Issue 1: It operated with no memory of what has been drained or not, leading to
multiple checks for chains/groups that were already drained.
Issue 2: When receiving an EOS, it only detected that a higher-level chain
was drained if it contained the pad receiving the EOS.
The following modifications have been applied:
- a new drained property has been added to GstDecodeChain
- both drained properties of chain/group are set as soon as they are detected
- the algorithm now tests agains these values
See https://bugzilla.gnome.org/show_bug.cgi?id=685938
Should fix "cannot register existing type `GstPlaybinSelectorPad'" warnings
and subsequent errors when creating multiple players at the same time.
Conflicts:
gst/playback/gststreamselector.c
This allows the following use-cases to expose the group and pads
before an ALLOCATION query comes through:
* Single stream use-cases
* Multi stream use-cases where all streams sent the CAPS event before
the first ALLOCATION query
Some cases will still make the initial ALLOCATION query fail though,
which isn't optimal, but not fatal (it will recover when pads are
exposed, a RECONFIGURE event is sent upstream and elements can
re-send an ALLOCATION query which will reach downstream elements).
https://bugzilla.gnome.org/show_bug.cgi?id=680262
A caps event is also used to establish that a stream has prerolled.
Without this, we end up allowing negotiation queries to fail, ending
in decoders (and other elements) to not be configured right from the
start with the most optimal settings.
...and in playbin2 additionally prefer sinks over parsers.
This makes sure that we a) always directly plug a sink if it supports
the (compressed) format and b) always plug parsers in front of decoders.
In order to allow for proper functionality when a decoder only supports
one instance at a time (dsp), we must block the demuxer pads when they
get created if they are not part of the active group, preventing buffers
from being sent to the decoder (and initializing it through setcaps),
then after we switch to a new group, we unblock the demuxer pads for
the active groups. In the callback for the unblock, we prune the old
groups, making sure the previous decoder instance is destroyed before
we push a buffer to the new instance.
Since caps are no longer 'shared' between two pads (but forwarded from
source pad to sink pad) we end up with the first chain pad not having
specified caps (i.e. typefind:src).
This solves the issues by getting the pad's peer caps.
It is not optimal since it will (for most demuxers) return the pad
template caps, which might contain non-fixed caps (ex : with
qtdemux "video/quicktime; video/mj2; audio/x-m4a; application/x-3gp")
https://bugzilla.gnome.org/show_bug.cgi?id=667337
... to also properly indicate chain's endpad if no elements are in the
chain (due to the endpad being a raw demuxer pad, or one setup without
decoders since uridecodebin or higher up decided not to need those).
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().
After preroll the multiqueue limits are still set to the preroll
limits if use-buffering is set to TRUE. In that case we only want
time limits on the multiqueue if upstream is seekable.
Such streams were detected as seekable, as the query on the typefind
element was testing the m3u8 file listing the actual streams, and
not going through the demuxer(s).
We now check for seekability for each multiqueue following a demuxer,
so the query will flow through the elements which might prevent seeking.
https://bugzilla.gnome.org/show_bug.cgi?id=647769
This allows us to easily get ahold of all pads on a stream-topology message, including
pre-decoder ones, while "pad" only gives us access to the raw pads (as used by discoverer).
This is made possible by filtering errors. This is required to let
harware accelerated element query the video context. The video context
is used to determine if the HW is capable, and thus if the element is
supported or not.
Fixes bug #662330.
With unfixed caps we can't reliably decide if the final caps
are going to be "raw" (e.g. supported by a sink) or not.
We will get here again later when the caps are fixed.
If subdrained isn't initialized to FALSE then a chain might think
that its group is drained when in fact it's not and this can cause
a switch too early or even cause a deadlock.
g_value_get_object() does not give us our own ref.
Fixes "Trying to dispose object "flacparse", but it still has a parent "registry0".
You need to let the parent manage the object instead of unreffing the object directly."
and similar warnings.
https://bugzilla.gnome.org/show_bug.cgi?id=658416
This is done by adding a capsfilter after every parser/converter that contains
all possible caps supported by downstream elements. A capsfilter is necessary
here because the decoder is only selected after the parser selected a format
and the parser can't know what downstream would support otherwise.
When we have a multi-stream (i.e. audio and video) input and the demuxer
adds/removes pads for a new stream (common in a mpeg-ts stream when the
program stream mapping is updated), the algorithm for EOS handling was
previously wrong (it would only drop the EOS of the *last* pad but would
let the EOS on the other pads go through).
The logic has only been changed a tiny bit for EOS handling resulting in:
* If there is no next group, let the EOS go through
* If there is a next group, but not all pads are drained in the active
group, drop the EOS event
* If there is a next group and all pads are drained, then the ghostpads
will be removed and the EOS event will be dropped automatically.
For streams at low bitrates we need to set a limit in time because the limit
in bytes might not reached too late, sometimes more than 30 seconds.
This limit can only be set if upstream is seekable (see #584104)
Closes#647769
Parsers are the only element class that are not changing the data and
could lead to an infinite loop. Other element classes like demuxers,
e.g. id3demux, can be used multiple times in a row and sometimes are.
Previously we only checked against the raw caps but we should also
check against the return value of autoplug-continue. Additionally fix
a thread-safety issue with accessing the raw caps.
...instead of copying the array. Returning NULL will result
in the original factories array to be used and prevents a useless
array copy in most use cases.
Add notes about the behaviour if multiple signal handlers are connected.
For most autoplug-* signals only the first signal handler will ever
be invoked.
Also add to the autoplug-sort docs that the signal handler can return NULL
to specify that the order should change and other handlers get the chance
to sort the array.
Where it was previously located, we would get async-done for the first
unknown-type, even if other valid streams would appear afterwards.
decode_bin_expose() will take care of posting async-done when the group
is exposed.
But we still want to post it in case the typefinding returned an unknown
type, in which case we will post it after posting an error.
These two changes ensure we do as much as possible before posting async-done.
If an error happens, the PAUSED state will never be reached. If an
application re-uses decodebin2 (like totem) where one would normally
set to READY between each file, the cleanup that normally happens in
the PAUSED=>READY codepath will never be called, resulting in the
following file to re-use the previous demuxer/decoder/...
https://bugzilla.gnome.org/show_bug.cgi?id=622807
Use the pad caps when they are available to continue the autoplugging. If the
pad caps are set, they are fixed and then we can directly continue autoplugging.
Use an accumulator for the autoplug-sort signal so that we can stop the emission
when a signal handler produced a valid result. This avoids the object handler
to overwrite the results from user signals.
Fixes#621161
If a file contains raw streams (not requiring a decoder) that we do
not want (expose-all-streams == FALSE), we would previously consider
those of unknown-type (missing a decoder) ... whereas in fact it was just
because they don't need decoders.
This only applies if expose-all-streams is FALSE.
API : expose-all-streams
If disabled:
* only the streams that CAN be decoded and match the final caps will have a
decoder plugged in and be exposed.
* the streams that COULD HAVE BEEN decoded but do not match the finals caps
will not have a decoder plugged in and will not be exposed.
If no decoder is available to decode a certain stream, then the missing element
message will still be emitted regardless of the value of the property.
https://bugzilla.gnome.org/show_bug.cgi?id=617868
Otherwise the ghostpad will still be linked to the peer and there
will still be a reference kept, leading to nothing being unlinked
and destroyed until decodebin2 is finalized.
This fixes reuse of decodebin2 if a raw stream is connected to
its sinkpad.
In some cases (all buffers dropped by a parser) a decodebin2
chain might receive an EOS before it gets enough data to
expose a decoded pad. In the case that no streams can expose
a pad we should error out instead of hang.
Fixes#542758
When a decodebin2 receives no-more-pads of a group it
can set that group's multiqueue buffering thresholds to
'playing' buffering method, avoiding that it buffers
too long and cause problems when using with queue2.
See the associated bug for details.
Fixes#600787
Using the object lock here can and will lead to deadlocks because
of deep-notifies of property changes: the deep-notify handler will
get the parent of objects, which will take the object lock again.
Fixes bug #600479.
There's not much point in using GST_DEBUG_FUNCPTR with GObject
virtual functions such as get_property, set_propery, finalize and
dispose, since they'll never be used by anyone anyway. Saves a
few bytes and possibly a sixteenth of a polar bear.
instead of printing an error that no corresponding group could
be found. no-more-pads from non-demuxer elements doesn't give
any additional information because there can only be a single srcpad.
Fixes bug #598288.
This allows partial group changes, i.e. demuxer2 in the example below
goes EOS but has a next group and audio2 stays the same.
/-- >demuxer2---->video
demuxer--- \--->audio1
\--->audio2
This now keeps track of everything that is going on, creates
a tree of chains and groups to allow "demuxer after demuxer" scenarios
and allows chained Oggs with multiple streams (needs oggdemux or playbin2 fixes).
Also document everything in detail and give a general overview of what
decodebin2 is doing at the top of the sources.
Fixes bug #596183, #563828 and #591677.
Pad blocks should never be done on external pads as outside elements
might want to use their own pad blocks on them and this will lead to
conflicts and deadlocks.
When we are probing for streams, we want to set the queue size in such a way
that we can scan a maximum amount of data without consuming too much memory.
Therefore, remove the time limit on the queue and only stop scanning after 2MB
of data.
See #584104.
The 2s limit is way too small for a lot of files (which have an interleave
in time of between 3 and 5s). Instead, leave it to the initial 5s value
and reduce the other limits (allowing us to stay memory-efficient).
