Async song loading can fail without user feedback. This change adds return codes
to these async load functions. It will now produce an error dialog in simple
scenarios (test case is user selecting a file that is not readable). Other cases,
such as directories and playlists, aren't yet covered.
QTimeLine duration must be greater than 0. If set to 0, a default of 1000ms will
be used. To avoid this, enforce a minimum of 1ms for pause and cross fade values
if those fades are enabled.
There are a number of cases where gst_pipeline_get_bus,
gst_element_get_static_pad, and g_object_get are called without releasing
references. In addition to memory usage, some of these elements hold file
descriptors. In normal operation, two file descriptors are leaked for each
played track. The default fd ulimit for many linux distros is 1024. This
is likely the cause of the crash reported in issue 6309.
This change fixes the obvious and consistent leaks, but it's probably not a
complete solution. There are many error and corner conditions that need to be
examined.
In the case that an error occurs in ReplaceDecodeBin before the bin is added to
the pipeline, unreference the object to allow cleanup. This change also separates
CreateDecodeBinFromUrl from ReplaceDecodeBin, following the pattern of
CreateDecodeBinFromString.
If ReplaceDecodeBin fails from TransitionToNext, uridecodebin_ will not be
replaced with a new element. Since TransitionToNext does not check the return
value, it unknowingly deletes uridecodebin_.
As reported in issue 6302, playing a stream that causes gstreamer to error at
start can cause a crash. The problem occurs when the MoodbarPipeline receives a
pad-added signal after it has handled an error callback. In the error callback,
the builder_ is freed. In the pad-added handler (NewPadCallback), this object
is accessed.
This change adds a running_ flag that is set when the pipeline is started and
cleared on an error, end of stream, or object destruction. We check this flag at
the beginning of NewPadCallback. For sanity sake, we also check the builder_
pointer before dereferencing. Note that checking the state of the pipeline
wasn't an option since the pipeline is in the process of changing states during
the pad-added callback and gst_element_get_state wants to block during a state
change.
This solution is not complete as there are still some syncronization issues.
With this specific situation, the error and new pad callbacks appear to always
occur on the same thread, but that's probably not true for all error conditions.
The object is also destroyed by a different thread, so it may be possible that a
callback can occur at the wrong time during or after the deletion of the object.
See https://github.com/clementine-player/Clementine/issues/6302
There is a small chance that a device lister is able to discover and add a
previously known device before it is added by the database loader thread.
In this case, copy the data that is user-settable to the existing DeviceInfo
object and destroy the object created from the database query.
This adds and utilizes a new FindEquivalentDevice method that compares the
device unique IDs. This could probably be made more robust as the unique
IDs for some listers may change. However, this is a problem with the database
storage implementation in general.
When DeviceManager initializes, it creates a thread to load device information
from the database. Part of this process includes use of QPixMap for icons which
produced a warning message:
22:32:53.763 WARN unknown QPixmap: It is not safe to use pixmaps outside the GUI thread
In addition, the device is added to the view using beginInsertRows and
endInsertRows. This could contend with a device added by a lister signaling
PhysicalDeviceAdded.
To solve these problems, this change moves the icon loading and insertion to the
main thread. LoadAllDevices reads the data from the database and creates the
DeviceInfo object, then sends a signal to the main thread. In the signal
handler, the icon is loaded and the device is added to the master list and view.
When unmounting a device, the ConnectedDevice object is destroyed. The
FileSystemDevice destructor waits on its worker thread. If a scan is in
progress, this will block until completion.
There is an existing Stop method in the LibraryWatcher class that is intended to
stop long running operations. To fix, or at least significantly shorten this
hang, we'll call this before waiting for the thread to exit. Also add a
stop_requested check in the cover art scan.
In addition, add a call to Stop in the Library destructor, which has a similar
usage.
