For some reason, nsWebShellWindow's destructor clears out the nsIWidget member inherited from nsXULWindow, accidentally bypassing nsXULWindow's destruction code. Previously someone has always managed to invoke Destroy manually, however as of bug 777063 it's definitely possible to tickle the code so that nsWebShellWindow gets destroyed before Destroy gets invoked. This prevents nsXULWindow from clearing the pointers in the tree owners which are now stale and promptly crash if any of the tree owner's methods get called. As it happens, Destroy has an identical copy of the code to clear that member.
Assignee: nobody → neil
Status: NEW → ASSIGNED
Attachment #646761 - Flags: review?(roc)
Attachment #646769 - Flags: review?(roc) → review+
Sorry, I backed this out on inbound because it looks like it introduced a perma-orange in make check on Windows debug builds: https://hg.mozilla.org/integration/mozilla-inbound/rev/72966f49cf09 https://tbpl.mozilla.org/php/getParsedLog.php?id=13975366&tree=Mozilla-Inbound mozilla::ShutdownXPCOM[xul +0x1253D82] ScopedXPCOM::~ScopedXPCOM[TestAppShellSteadyState +0x37E0] main[TestAppShellSteadyState +0x5970] __tmainCRTStartup[TestAppShellSteadyState +0x97EF] mainCRTStartup[TestAppShellSteadyState +0x961F] BaseThreadInitThunk[kernel32 +0x133CA] RtlInitializeExceptionChain[ntdll +0x39ED2] RtlInitializeExceptionChain[ntdll +0x39EA5] ###!!! Deadlock may happen NOW! ###!!! ASSERTION: Potential deadlock detected: Cyclical dependency starts at Mutex : nsWindowMediator.mListLock (currently acquired) Cycle completed at Mutex : nsWindowMediator.mListLock (currently acquired)
Annoying. The test shuts down XPCOM while the window is still open, causing the window mediator to unregister all of its windows. Because it's the owner of last resort, this causes the destructor to run, which (with this patch) causes the nsXULWindow to try to unregister itself again.
Sounds like we should make sure we don't re-enter nsXULWindow::Destroy.
(In reply to Justin Lebar from comment #7) > Sounds like we should make sure we don't re-enter nsXULWindow::Destroy. We don't, this is the only call to nsXULWindow::Destroy in the test.
Okay, let's try this. https://tbpl.mozilla.org/?tree=Try&rev=120d646c3cf3
As Neil pointed out on IRC, we don't actually need this additional block scope, so long as the array is declared before the auto-lock. I think it clarifies things, but I'm not attached to it.
Attachment #648005 - Flags: review?(roc)
Attachment #648005 - Flags: review?(roc) → review+
Comment on attachment 648005 [details] [diff] [review] Part 1a, v1: Release the references after we release the lock. https://hg.mozilla.org/integration/mozilla-inbound/rev/3111cd3bc5f5
Status: ASSIGNED → RESOLVED
Closed: 7 years ago
Resolution: --- → FIXED
Target Milestone: --- → mozilla17
You need to log in before you can comment on or make changes to this bug.