macosx: call input_changed in extensions
This is obnoxiously complicated. If anyone cares about playing_changed or meta_changed, something similar will probably have to be done. This is a pretty bizarre two-step system to inform the extension manager that the input has changed, but it's necessary to avoid a series of possible deadlocks and other issues. Here are other possible approaches that don't work: - Just call into the extension manager in -PlaylistItemChanged on the main thread. This can pretty easily cause a deadlock if we call -PlaylistItemChanged twice in quick succession. The first call will poke the condvar the extension is waiting on, causing the extension thread to wake up and run extension code; many parts of it -- including the dialog code -- must be run on the main thread. The extension thread goes back to sleep while blocking on the main thread to become available, while holding the extension lock. Meanwhile the main thread goes into the second call of -PlaylistItemChanged, attempts to lock the extension, and that's a deadlock. - Restructure the dialog manager to never block on the main thread while holding the extension lock. This should work, but as it turns out doesn't because the main thread will attempt to lock the same lock twice. What happens is that -performSelectorOnMainThread works by injecting an event into the main event loop of the main thread. For some unknown reason, as part of its processing, when creating an NSAttributedString with HTML, it runs the main event loop, which means we can end up executing one -performSelectoOnMainThread as part of another. Since the dialog manager uses attributed strings with HTML (since dialogs are HTML), we deadlock here too. This seems strictly like a flaw in NSAttributedString and/or in -performSelectorOnMainThread and is documented elsewhere: http://mrrsoftware.com/blog/tag/nsattributedstring/ https://www.bluestatic.org/blog/2010/05/31/nsattributedstring-spins-a-nested-run-loop/ - Change around this bit of code to not force it to run on the main thread. This would probably work, but, as a newcomer to VLC, I don't quite know the implications of doing this, particularly since a lot of code here seems to serailize on the main thread as a way of thread safety; it would likely require some somewhat intricate restructuring and adding of locks. - Let the extension manager deal with listening for events the same way that we do here. That would work, but would require duplicating a nontrivial amount of code from here to deal with tracking the current input. - So, instead, we just serialized all calls to -PlaylistItemChanged (so we make sure to process them in order, with no one trampling p_input_changed), do most of the work on the main thread as before, and then actually inform the extension manager out here where we don't block the main thread. It seems likely that there are other pre-existing deadlock possibilities here -- the main thread can't lock an extension! -- but it at least tends to work in my testing. Signed-off-by:Felix Paul Kühne <fkuehne@videolan.org>
Showing
Please register or sign in to comment