[Gnucash-changes] Eliminate a double free of memory.

[Chris Shoemaker]

On Thu, Jun 02, 2005 at 06:55:18PM -0400, David Hampton wrote:
> On Thu, 2005-06-02 at 17:33 -0400, Derek Atkins wrote:
> > Quoting David Hampton <hampton-gnucash at rainbolthampton.net>:
> > 
> > > 
> > > I think this will work because gtk_widget_destroy correctly handles
> > > recursive invocation.  If the widget is already in the process of being
> > > destroyed then the call to gtk_widget_destroy from
> > > gnc_customer_window_close_handler is a noop, otherwise it will actually
> > > destroy the dialog.
> > 
> > What if the gtk_widget_destroy() finishes and returns before the CM callback
> > happens?  For example, if the CM events are suspended, the CM system will cache
> > the event and then return, causing gtk_widget_destroy() to finish.  Then later
> > when the events are resumed it will call the CM callback which will call
> > gtk_widget_destroy() on the already-freed widget.  Wouldn't that be a problem?
> 
> It might be possible to add a call to g_signal_connect_after(cw->dialog,
> "destroy", gtk_widget_destroyed, &cw->dialog) when the dialog is created
> to resolve this case.   If the window destruction completes before the
> CM close function is called, then this new callback will have zeroed the
> cw->dialog pointer.  
> 
> Of course, since the gnc_customer_window_close_handler function frees
> cw, it would have to remove this callback so the original invocation
> order we were discussing won't write to freed memory.  I don't know if
> this will work though as I don't know if glib handles deleting callback
> functions from within a callback.  It appears that glib builds an
> ordered list of all handlers before calling the first handler, so the
> above solution will not work for all cases.
> 
> I suppose this could be flipped and the gnc_customer_window_destroy_cb
> callback could be connected "after" the gtk_widget_destroyed handler.
> In this case if the dialog has been closed via the window manager close
> button cw->dialog will always be NULL regardless of whether the CM
> callback runs immediately or at some later time.  If the dialog has been
> closed some other way, then cw->dialog will be non-NULL and this
> function can call gtk_widget_destroy.  I think this works. 

Maybe there's an even easier way: sink and ref the dialog when it gets
created, and then unref it only in the CM close handler, right before
the gtk_destroy_widget() call.  Then you know the dialog won't be
freed until you're done with it.

-chris