Align the allocated memory to 16 bytes. When doing XSHM we are already aligned
to a page boundary but without, we use plain g_malloc, which could allocate
aligned on 8 bytes only.
See https://bugzilla.gnome.org/show_bug.cgi?id=680796
Don't ever block when acquiring a buffer from the bufferpool in the fallback
mode. If we block, we might deadlock when going to PAUSED because we never
unlock when going to paused.
The acquire can block when there are no more buffers in the pool, this is a
sign that the pool is too small. Since we are the only ones using the pool in
the fallback case and because we scale the buffer, someone else must be using
our pool as well and is doing something bad.
Pick delta pad earlier during header parsing, and pick it based
on whether it's a video stream or not rather than some rather
byzantine signalling from theoraenc etc. which would set the delta
flag on header packets which oggmux would then pick up and determine
that this is a "delta-able" stream.
Since the new videodecoder-based theoraenc didn't do that any more,
we would only see the first delta flag on the second video packet,
which is after we've already muxed a few audio packets flagged as
key units, which trips up the unit test.
Fixes pipelines/oggmux unit test.
https://bugzilla.gnome.org/show_bug.cgi?id=679958
There won't be a tag messages on the bus, because tags
are now sent downstream for sinks to post on the bus,
and there's no sink involved here that would do that.
Secondly, the audio decoder base class only sends the
tags out once it has received some non-header data as
input, which is not something we're providing here.
Make sure to send a CAPS event downstream when we get our
first input caps. This fixes not-negotiated errors and
adder use with downstream elements other than fakesink.
Even gst-launch-1.0 audiotestsrc ! adder ! pulsesink works now.
Also, flag the other sink pads as FIXED_CAPS when we receive
the first CAPS event on one of the sink pads (in addition to
setting those caps on the the sink pads), so that a caps query
will just return the fixed caps from now on.
There's still a race between other upstreams checking if
caps are accepted and sending a first buffer with possibly
different caps than the first caps we receive on some other
pad, but such is life.
Also need to take into account optional fields better/properly.
https://bugzilla.gnome.org/show_bug.cgi?id=679545
Fix invalid memory access caused by broken pointer arithmetic.
If we have a uint16_t *tmpbuf and add n * dest->stride to it, we
skip twice as much as we intended to because dest->stride is in
bytes and not in pixels. This made us write beyond the end of
our allocated temp buffer, and made the unit test crash.
