|
When importing modules - in the QML loader thread - with plugins we keep
globally track of the Qt plugins that we have loaded that contain QML
modules, to ensure that we don't call the engine-independent
registerTypes() function on the plugin multiple times. After
registerTypes() we may also call initializeEngine() on the plugin for
the engine-specific initialization, which - as a QQmlEngine is provided
as parameter - must happen in the gui thread. For that we issue a
thread-blocking call that waits until the gui thread has woken up and
processed the event/call.
During that time the global plugin lock is held by that QML loader
thread.
If meanwhile the gui thread instantiates a second QQmlEngine and
attempts to issue a synchronous type compilation (using
QQmlComponent::CompilationMode::PreferSynchronous), then gui thread is
blocking and waiting for its own QML loader thread to complete the type
compilation, which may involve processing an import that requires
loading a plugin. Now this second QML loader thread is blocked by trying
to acquire the global plugin registry lock
(qmlEnginePluginsWithRegisteredTypes()->mutex) in qqmlimports.cpp.
Now the first QML loader thread is blocked because the gui thread is not
processing the call events for the first engine. The gui thread is
blocked waiting for the second QML loader thread, which in turn is stuck
trying to acquire the lock held by the first QML loader thread.
The provided test case triggers this scenario, although through a
slightly different way. It's not possible to wait in the gui thread for
the plugin lock to be held in a loader thread via the registerTypes
callback, as that also acquires the QQmlMetaType lock that will
interfere with the test-case. However the same plugin lock issue appears
when the first QML engine is located in a different thread altogether.
In that case the dispatch to the engine thread /works/, but it won't be
the gui thread but instead the secondary helper thread of the test case
that will sit in our initializeEngine() callback.
This bug was spotted in production customer code with backtraces
pointing into the three locations described above: One QML loader thread
blocking on a call to the gui thread, the gui thread blocking on a
second QML loader thread and that one blocking on acquisition of the
plugin lock held by the first.
Fortunately it is not necessary to hold on to the global plugin lock
when doing the engine specific initialization. That allows the second
QML loader thread to complete its work and finally resume the GUI
thread's event loop.
Change-Id: If757b3fc9b473f42b266427e55d7a1572b937515
Reviewed-by: Ulf Hermann <ulf.hermann@qt.io>
|