Class Mocks Not Thread-Safe

[alanterranova]

Hi Erik,

I’m debugging thread safety in OCMock 3 and have an example test case that demonstrates some faulty behavior. In short, it looks like invoking a class method from a background thread while a class mock is being restored causes crashes due to unrecognized selectors.

Strangely, this isn’t a problem with OCMock 2. I suspect 1ad7292 may have something to do with it.

I realize this is a contrived example, but it’s flaky enough to prevent me from upgrading my project from v2 to v3.

Thanks,

Alan

[kballard]

I just hit this same issue, with OCMock 3.3 (which the release notes claims makes mocking thread-safe).

I believe the root issue here is OCMock's class mocking isa-swizzles the metaclass prior to installing methods. This means that background threads may start interacting with the metaclass while it's in the process of installing methods, which is pretty bad.

[kballard]

Since I'm not sure if moving the metaclass swizzle to the end is safe (I don't know if any of the intervening code relies on it already being installed), I believe a safe fix is to reorder two lines in -setupForwarderForClassMethodSelector::

diff --git a/Pods/OCMock/Source/OCMock/OCClassMockObject.m b/Pods/OCMock/Source/OCMock/OCClassMockObject.m
index f8a8180..39b5f22 100644
--- a/Pods/OCMock/Source/OCMock/OCClassMockObject.m
+++ b/Pods/OCMock/Source/OCMock/OCClassMockObject.m
@@ -148,8 +148,8 @@ - (void)setupForwarderForClassMethodSelector:(SEL)selector
 
     Class metaClass = object_getClass(mockedClass);
     IMP forwarderIMP = [originalMetaClass instanceMethodForwarderForSelector:selector];
-    class_replaceMethod(metaClass, selector, forwarderIMP, types);
     class_addMethod(metaClass, aliasSelector, originalIMP, types);
+    class_replaceMethod(metaClass, selector, forwarderIMP, types);
 }

[kballard]

@erikdoe I assume you're getting notifications for issue updates anyway, but just in case, I'm mentioning you as this is a nasty issue that has been causing spurious test failures for us for a long time and I'm surprised that this ticket never got a response.

[kballard]

@erikdoe Can you please take a look at this? It's a real issue with a trivial fix.

FWIW, it turns out that uninstalling class mocks isn't thread safe either, which doesn't seem to be a solvable short of simply never destroying the dynamically-created class pairs, but in my experience the uninstall issue crops up a lot less frequently than the install issue did.

[erikdoe]

Sorry. I should have responded earlier. My concern is that the failing test provided in the issue tests a scenario that, as far as I can tell, cannot be supported by OCMock. It has always been clear that OCMock is only thread-safe to a certain degree. Repeatedly calling a method while setting up and removing a mock would, as far as I am aware, require most of the operations on the class to be atomic, and I don't see a way how to achieve this.

The fix provided by @kballard looks benign, but it doesn't fix the issue reported. In fact, there is no test that I can think of that would fail without the change but pass with it. That said, the change is small and it feels like it brings the two operations into a "more correct" sequence. So I have accepted it now.