If I startup and shutdown mail news w/o selecting anything, all our datasources are freed. Similarly, if I expand a server w/o selecting it, all our datasources are freed. But if I select a server, we leak the datasources. This is because we leak an nsXULTreeElement, as near as I can tell. I'm still investigating, but here's the memory leak report: Type # created # leaked ---- --------- -------- RDF_InMemoryDataSource 17 5 RDF_RDFContainerUtilsImpl 1 1 RDF_Assertion 823 643 RDF_LiteralImpl 667 192 nsBaseWidget 39 1 I'm going to see if rerooting the tree is what's causing this, or just the act of selecting. DeviceContextImpl 8 1 nsWebShell 6 1 RDF_DateImpl 117 103 RDF_RDFXMLDataSourceImpl 7 2 AtomImpl 13908 180 StyleContextImpl 298 15 RDF_LocalStoreImpl 1 1 RDF_nsXULContentUtils 1 1 RDF_RDFElementImpl 690 38 RDF_nsXULAttribute 1675 107 nsMsgAccountManagerDataSource 1 1 nsMsgFolderDataSource 1 1 RDF_nsXULKeyListenerImpl 1 1 RDF_CompositeDataSourceImpl 9 1 RDF_nsRDFDOMNodeList 901 2 RDF_nsXULTreeElement 5 1 nsMsgFolder 135 135

I disabled the tree rerooting code, and we stil leak. This is good news, I think, since it means it's most likely the tree selection code. But I've been wrong before...

Notice that we leak a web shell in the second case.

I think the web shell leaks because the data source leaks - our datasources hold onto a msg status feedback object, which hold onto a web shell. Perhaps this causes some circularity in some situations. I can look at making one of these a weak ptr, probably the web shell held onto by the msg status feedback object.

so much for that theory - making the msg status feedback not hold onto the web shell or dom window didn't fix anything.

I've traced this down some more - we leak one nsEventStateManager, which is holding onto some content nodes when you do a selection. These all tend to be an RDFElementImpl, which holds onto the usual littany of suspects, composite datasources, our datasource, etc. I believe we leak the nsEventStateManager because we leak a web shell - this seems highly circular. Also, the nsEventStateManager code is highly suspicious. It doesn't use nsCOMPtr's, and often assigns to interface pointers without releasing the original value.

add joki to cc list. tom: any ideas?

OK, I was a fool to think I could fix this. I'm going to try to ameliorate the damage done by adding a Close method to our datasources and calling it from the RDF shutdown notification. The close method will just null out our nsCOMPtrs, which tends to fix all the leaks caused by the ds leaks, which is something, anyway.

I'm not seeing this leak anymore - it must have gotten fixed by someone!

ok, I'll mark verified since David B. reported the bug and is not seeing this any longer.