Created attachment 368480 [details]
proposed fix

The following backtrace was reported while running netbeans:

+ Trace 238401

This is most probably because the object was getting freed while the signal was being processed, and thus the hash table was being destroyed. We thus need to keep a reference on the atk object while the signal is processed.

The attached patch does that, by adding a reference in alloc_callback_para, and releasing it in free_callback_para. However, since like explained in https://bugzilla.gnome.org/show_bug.cgi?id=766774 we have to avoid calling methods of applications objects, we can't afford unreferencing from the jaw thread. This patch thus makes it just queue the para struct to be freed, so it's all processed in the application when it makes the next accessibility call. This might actually fix other bugs (deleting a global ref on para->args might call a destructor).