When shutting down the chain, we can get a deadlock when removing
a pad, if that chain was being busy streaming but blocked (eg, while
waiting for a queue to have free space).
https://bugzilla.gnome.org/show_bug.cgi?id=746480
In case upstream does not provide videorate with framerate information,
it will detect the current framerate from the buffer it received,
but if downstream forces the use of variable framerate (most probably
through the use of a caps filter with framerate = 0 / 1), videorate will
respect that.
And add some unit tests
https://bugzilla.gnome.org/show_bug.cgi?id=734424
In the case the framerate is variable (represented by framerate=0/1),
we currently end up loop pushing the first buffer and then recompute
diff1 and diff2 without updating the videorate->next_ts at all
leading to infinitely looping pushing that first buffer.
In the case of variable framerate, we should just compute the next_ts
as previous_pts + previous_duration.
https://bugzilla.gnome.org/show_bug.cgi?id=734424
The patch calculates a second channel mixing matrix from the current one. The
matrix contains the original values * (2^10) as integers. This matrix is used
when integer-formatted channels are mixed.
On a ARM Cortex-A8, single core, 800MHz this improves performance in a
testcase from 29s to 9s for downmixing 6 channels to stereo.
https://bugzilla.gnome.org/show_bug.cgi?id=747005
If a new pad is added after playbin has been put to READY/NULL it
should ignore new pads as it is shutting down.
This can happen when the pipeline fails to preroll (is still in READY)
and the user gives up on waiting or an error that doesn't reach
the demuxer occurs (on some event handling) and it will continue to
work and exposing pads while playbin has been put to NULL.
Without this check an input-selector is created and set to PAUSED
state, preventing playbin from properly shutting down in case it
has data blocked inside it.
audio_convert_convert unpacks to default format (signed) before calling
quantize, and the unsigned variants were equivalent to signed anyway,
so we just get rid of them.
Since range size is always 2^n, we can simply use modulo (implemented
with a bitmask).
The previous implementation used 64-bit integer division, which is
done in software on ARMv7. Although the divisor was constant, the
division could not be transformed into "multiplication by magic number"
since the dividend was 64-bit.
The now-unused and not-so-fast gst_fast_random_(u)int32_range functions
were removed.
Also, implementing bug fixes:
1) ADD_DITHER_TPDF_HF_I no longer discards bias.
2) We change TPDF's noise range to be the same as RPDF's. Previously,
RPDF's noise ranged:
{ bias - dither, bias + dither }
while TPDF's noise ranged:
{ bias/2 - dither/2, bias/2 + dither/2 - 1 } +
{ bias/2 - dither/2, bias/2 + dither/2 - 1 } =
{ bias - dither, bias + dither - 2 }
Now, both range:
{ bias - dither, bias + dither - 1 }
https://bugzilla.gnome.org/show_bug.cgi?id=746661
This fixes a race where the use-buffering property on a multiqueue was
set before the queue depth was changed from it's high preroll limits to
lower playback limits. This resulted in buffering messages being emitted
by the multiqueue in the short window between use-buffering being
set and the queue depth being reset.
https://bugzilla.gnome.org/show_bug.cgi?id=744308
The variables could have changed when the lock was released
to push a gap event. Streamsynchronizer needs to check them
again before going to sleep.
Bonus: fix a comment typo
multisocketsink now understands the new GstNetControlMessageMeta to allow
sending control messages (ancillary data) with data when writing to Unix
domain sockets.
Thanks to glib's `GSocketControlMessage` abstraction the code introduced
in this commit is entirely portable and doesn't introduce and additional
dependencies or conditionally compiled code, even if it is unlikely to be
of much use on non-UNIX systems.
multisocketsink now understands the new GstNetControlMessageMeta to allow
sending control messages (ancillary data) with data when writing to Unix
domain sockets.
A later commit will introduce a new socketsrc element which will similarly
understand `GstNetControlMessageMeta`. This, when used with a
`GSocketControlMessage` of type `GUnixFDMessage` will allow GStreamer to
send and receive file-descriptions in ancillary data, the first step to
using memfds to implement zero-copy video IPC.
Thanks to glib's `GSocketControlMessage` abstraction the code introduced
in this commit is entirely portable and doesn't introduce and additional
dependencies or conditionally compiled code, even if it is unlikely to be
of much use on non-UNIX systems.
This provides notification that the socket in use was closed by the peer
and gives an opportunity to replace it with a new one which is not
closed, allowing reading from many sockets in order.
I use this in pulsevideo to implement reconnection logic to handle the
pulsevideo service dieing, such that is can be restarted without
disrupting downstream.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=739546
* Don't bother polling, just do a blocking read, the `GCancellable` will
take care of unlocking. This should also be faster on MS Windows where
the GIO documentation for `g_socket_get_available_bytes` states: "Note
that on Windows, this function is rather inefficient in the UDP case".
* Implement `GstPushSrc.fill` rather than `GstPushSrc.create`. This means
that we will be using the downstream allocator which may be more
efficient. It also means that socketsrc is likely to respect its
"blocksize" property (assuming that there is enough data available).
See https://bugzilla.gnome.org/show_bug.cgi?id=739546
`socketsrc` can be considered a source counterpart to `multisocketsink`.
It can be considered a generalization of `tcpclientsrc` and
`tcpserversrc`: it contains all the logic required to communicate over
the socket but none of the logic for creating the sockets/establishing
the connection in the first place, allowing the user to accomplish this
externally in whatever manner they wish making it applicable to other
types of sockets besides TCP.
This commit essentially copies the implementation directly from
tcpserversrc. Later patches will tidy the implementation up and
re-implement `tcpclientsrc` and `tcpserversrc` in terms of `socketsrc`.
See https://bugzilla.gnome.org/show_bug.cgi?id=739546
If a buffer is made up of non-contiguous `GstMemory`s `gst_buffer_map`
has to copy all the data into a new `GstMemory` which is contiguous. By
mapping all the `GstMemory`s individually and then using scatter-gather
IO we avoid this situation.
This is a preparatory step for adding support to multisocketsink for
sending file descriptors, where a GstBuffer may be made up of several
`GstMemory`s, some of which are backed by a memfd or file, but I think this
patch is valid and useful on its own.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=746150
When we modify a GList (via g_list_delete_link), always reassign the
new head to the original GList. Otherwise we end up with
filtered_errors being corrupt (the head might have been the element
removed)
This function is static, and only ever called with the expose lock
taken. It thus has no reason to take this lock itself.
This was introduced by one of my locking fixes from 741355.
https://bugzilla.gnome.org/show_bug.cgi?id=741355
Check if dbin->decode_chain is NULL before running drain_and_switch_chains()
because if it is, we shouldn't run that function or it will segfault.
CID #1271074
Otherwise if there are multiple parsers we would most likely break negotiation
of the stream-format/alignment wanted by the decoders as parsers generally
support all possible stream-formats and alignments.
If caps on a newly added pad are NULL, analyze_new_pad will try to
acquire the chain lock to add a probe to the pad so the chain can
be built later. This comes from the streaming thread, in response
to headers or other buffers causing this pad to be added, so the
stream lock is taken.
Meanwhile, another thread might be destroying the chain from a
downward state change. This will cause the chain to be freed with
the chain lock taken, and some elements are set to NULL here, which
can include the parser. This causes pad deactivation, which tries
to take the element's pad's stream lock, deadlocking.
Fix this by keeping track of which elements need setting to NULL,
and only do this after the chain lock is released. Only the chain
manipulation needs to be locked, not the elements' state changes.
https://bugzilla.gnome.org/show_bug.cgi?id=741355
There was a deadlock between a thread changing decodebin/demuxer
state from PAUSED to READY, and another thread pushing data
when starting.
From the stack trace at
https://bug741355.bugzilla-attachments.gnome.org/attachment.cgi?id=292471,
I deduce the following is happening, though I did not reproduce the
problem so I'm not sure this patch fixes it.
The streaming thread (thread 2 in that stack trace) takes the demuxer's
sink pad's stream lock in gst_ogg_demux_perform_seek_pull and will
activate a new chain. This ends up causing the expose lock being taken
in _pad_added_cb in decodebin.
Meanwhile, a state changed is triggered on thread 1, which takes the
expose lock in decodebin in gst_decode_bin_change_state, then frees
the previous chain, which ends up calling gst_pad_stop_task on the
demuxer's task, which in turn takes the demuxer's sink pad's stream
lock, deadlocking as both threads are now waiting for each other.
https://bugzilla.gnome.org/show_bug.cgi?id=741355
Also improve the waiting condition for stream switches, which was assuming
before that the condition variable will only stop waiting once when it is
signaled. But the documentation says that there might be spurious wakeups.
https://bugzilla.gnome.org/show_bug.cgi?id=736655
Change the GAP events that are currently sent from the chain function of
the current pad to all other EOS pads. They should instead be sent from
their own streaming threads.
https://bugzilla.gnome.org/show_bug.cgi?id=736655
Wait in the event function when EOS is received until all pads are EOS
and then forward the EOS event from each pads own event function.
Also send a new GAP event for EOS pads from the event function whenever
going from PLAYING->PAUSED by shortly waking up the GCond. This is needed
to allow sinks to pre-roll again, as they did not receive EOS yet because
we blocked that, but also will never get data again.
https://bugzilla.gnome.org/show_bug.cgi?id=736655
In gst_video_scale_fixate_caps () it can goto done without freeing the memory
of the tmp GstStructure. This makes it go out of scope and leak.
CID #1265766
Ignore chroma subsampling and color matrix transformations like the
old videoscale used to do. This is to make the performance like it was
before.
See https://bugzilla.gnome.org/show_bug.cgi?id=741987
