Bug 21609 – Make MessagePorts protect their peers across heaps

• Status: RESOLVED
• Resolution: FIXED
• Component: WebCore JavaScript

[Alexey Proskuryakov]

Each MessagePort needs to GC protect its entangled one, because even if the latter does not have a reachable wrapper, a message dispatched via the former will have a reference to it when dispatched.

This already works in the case when both wrappers are in the same JS heap, but as MessagePorts will be used for cross-thread communication, a different mechanism is needed.

[Alexey Proskuryakov]

Created attachment 24358 [details]
proposed patch

Unfortunately, this introduces knowledge about bindings into the implementation MessagePort object. Also, I do not quite understand how this will work in presence of other bindings (currently, an ObjC binding is a stub, but it exists nonetheless).

[Darin Adler]

Comment on attachment 24358 [details]
proposed patch

+ objects in the heap, makin git possible to implement cross-heap dependencies.

Misplaces space here "makin git".

+ void markCrossHeapDependentObjectsForDocument(JSGlobalData& globalData, Document* doc)

I'm slightly annoyed that we sometimes use references and other times pointers for JSGlobalData. Which style do you think is preferred? I also prefer document as the name for a document argument rather than doc.

It's a shame that m_jsWrapperIsKnownInaccessible is such a long name, but not correct grammar. The "known inaccessible" part isn't good grammar. I don't have a better name to suggest, but I think it's worth pondering a little more. Maybe just "KnownToBeInaccessible"?

+ bool jsWrapperIsKnownInaccessible() { return m_jsWrapperIsKnownInaccessible; }

Should be const?

+    JSGlobalObject* globalObject = m_globalData->head;
+    do {
+        globalObject->markCrossHeapDependentObjects();
+        globalObject = globalObject->next();
+    } while (globalObject != m_globalData->head);

The loop is just complicated enough that I think it should either be in its own function, or paragraphed separately from the other marking work.

r=me

[Alexey Proskuryakov]

Committed revision 37631.

(In reply to comment #2)

> I'm slightly annoyed that we sometimes use references and other times pointers
> for JSGlobalData. Which style do you think is preferred?

I prefer references for things that can never be NULL. But that's hardly our prevailing style.

> It's a shame that m_jsWrapperIsKnownInaccessible is such a long name, but not
> correct grammar.

I went with m_jsWrapperIsInaccessible.