[whatwg] MessagePort close event and discarding a Document

[Ian Hickson]

On Tue, 11 Nov 2008, Alexey Proskuryakov wrote:
> 
> Currently, HTML5 specifies that when a Document is discarded, "close" 
> event should be asynchronously dispatched to MessagePorts that are 
> entangled with ports belonging (in some specific sense) to this 
> document.
> 
> There is a race with garbage collection inherent to this requirement. 
> Below I describe the issue in detail, and propose a solution.
> 
> Suppose we have two browsing contexts, frameA and frameB, and two 
> MessagePorts, portA and portB, owned by these contexts respectively. The 
> ports are entangled with each other, and their queues are started.
> 
> Further, portA is not reachable from live code, but has a "close" event 
> listener (set via either onclose or addEventListener). Port portB is 
> also unreachable, so garbage collection can destroy both ports.
> 
> Now, we close frameB. This results in its document being discarded, so a 
> "close" event is dispatched on portA (portA becomes GC protected when 
> the task to dispatch it is posted). However, no event is dispatched if 
> portA and portB are collected before the document is destroyed. The 
> problem is that observable behavior depends on GC order, which it 
> shouldn't.
> 
> In practice, it is normally the same GC pass that destroys the Document 
> and both unreachable ports, so a straightforward implementation is 
> likely to crash due to using semi-deleted objects.
> 
> I propose to remove the requirement to dispatch "close" event on the 
> surviving port (#3 in paragraph 7.5.3.1). This is the only fix I can 
> see, as making GC dispatch "close" to fix it from other side would 
> suffer from basically the same logical problem.

I've done as you describe and in fact simplified a lot of the text around 
here at the same time to make this simpler now that we have the concept of 
"frozen" script groups.

-- 
Ian Hickson               U+1047E                )\._.,--....,'``.    fL
http://ln.hixie.ch/       U+263A                /,   _.. \   _\  ;`._ ,.
Things that are impossible just take longer.   `._.-(,_..'--(,_..'`-.;.'