As we do that for serialized events as well, and the subclass will
most likely need to access pad->segment to make its decisions,
doing that from the sinkpad's streaming threads was racy.
The following case can happen when two thread try to activate and
deactivate a pad at the same time:
T1: starts to deactivate, calls pre_activate(), sets in_activation
to TRUE and carries on
T2: starts to activate, calls pre_activate(), in_activation is TRUE
so it waits on the GCond
T1: calls post_activate(), tries to acquire the streaming lock ..
but can't because T2 is currently holding it
With this patch, the deadlock will no longer happen but does not
solve the problem that:
T2: will resume activation of the pad, set the pad mode to the target
one (PUSH or PULL) and eventually the streaming lock gets released.
T1: is able to finish calling post_activate() ... but ... the pad
wasn't deactivated (T2 was the last one to "activate" the pad.
https://bugzilla.gnome.org/show_bug.cgi?id=792341
Set up all ten pipelines and preroll them first, and only set
them to playing to run wild after they're all set up. If we set
them to PLAYING directly and let those threads run wild, then
it might take ages (many seconds) for the other pipelines to
even get up and running, especially on machines with only one
or two cores, and operating systems that suck at scheduling.
Now the fakesink test takes 19 secs instead of 71 secs on a
single-cpu windows machine.
This is a better fit given that the function docs say this
should (only) be used for interval measurements, but also
this seems to give much better granularity on Windows
systems, where before this change there would often be
10-20 lines of debug log with the same timestamp up front.
Scale the number of threads used in the stress tests according to
the number of cores/cpus. We want some contention, but we also
don't want too much contention, as some operating systems are
better at handling 100 threads running wild on a single core
than others.
Fix refcounting issue when plugin was loaded already.
gst_plugin_load() is supposed to return a ref, so it
must always return a ref.
This also fixes the gstplugin unit test on windows where
fork is not available and where test_load_coreelements()
would unref a plugin ref it didn't get and then mess up
the internal registry plugin list state for the next test,
in case where the test registry does not exist yet.
Add header with structure sizes for 64-bit windows as well.
They're almost the same as on Linux, but it looks like things
like padding unions get aligned slightly differently so there
are a handful of differences:
sizeof(GstGhostPad) is 528, expected 536
sizeof(GstPad) is 512, expected 520
sizeof(GstPadProbeInfo) is 64, expected 72
sizeof(GstProxyPad) is 520, expected 528
The test checks that categories not covered by the pattern in the
GST_DEBUG string have debug level GST_LEVEL_DEFAULT set, but previous
tests mess with the default threshold, which made this test fail on
Windows or when run with CK_FORK=no. Fix this by resetting everything
at the beginning, and then also do a sanity check afterwards.
Start task on new source pads added at runtime after they
have been added to the element, not during activation.
This ensures the pads can post their CREATE stream-status
messages and the application can set thread priorities.
https://bugzilla.gnome.org/show_bug.cgi?id=756867
Sub-class may want to decide to go passthrough/in-place by inspecting
the support meta APIs. This patch duplicates the check for this mode,
so we still don't do uneeded allocation query while we allow sub-classes
to switch the behaviour during it's own decide_allocation call.
Notice that such sub-class need to reset the class to non-passthrough in
set_caps() in order for decide_allocation to be called again. This is
needed otherwise we'd be doing an allocation query in element in which
it make no sense (notably capsfilter).
https://bugzilla.gnome.org/show_bug.cgi?id=791453
