Bug 128840 – ASSERT(isValidAllocation(bytes)) when ObjC API creates custom errors

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Joseph Pecoraro]

* TEST:
JSContext *context = [[[JSContext alloc] init] autorelease];
[[JSValue valueWithInt32:42 inContext:context] toDictionary];

* ASSERT
ASSERTION FAILED: isValidAllocation(bytes)
/Volumes/Data/Code/safari/OpenSource/Source/JavaScriptCore/heap/Heap.h(467) : void *JSC::Heap::allocateWithoutDestructor(size_t)
1   0x10a94ce50 WTFCrash
2   0x10a1ce54b JSC::Heap::allocateWithoutDestructor(unsigned long)
3   0x10a4ea107 void* JSC::allocateCell<JSC::ErrorInstance>(JSC::Heap&, unsigned long)
4   0x10a4e986f void* JSC::allocateCell<JSC::ErrorInstance>(JSC::Heap&)
5   0x10a4e8ffb JSC::ErrorInstance::create(JSC::VM&, JSC::Structure*, WTF::String const&, WTF::Vector<JSC::StackFrame, 0ul, WTF::CrashOnOverflow>)
6   0x10a4e88f7 JSC::createTypeError(JSC::JSGlobalObject*, WTF::String const&)
7   0x10a4e8b35 JSC::createTypeError(JSC::ExecState*, WTF::String const&)
8   0x10a6d5af5 valueToDictionary(OpaqueJSContext*, OpaqueJSValue const*, OpaqueJSValue const**)
9   0x10a6d5921 -[JSValue toDictionary]
10  0x10a1b5f04 main
11  0x7fff8cbbf5f1 start

Specifically, vm->identifierTable != wtfThreadData().currentIdentifierTable.

My guess is that the ObjC API has to grab the APIEntryShim before it calls into JSC (via JSC::createTypeError). Most of the ObjC API uses the C API which already does this implicitly.

[Joseph Pecoraro]

There are a number of create calls throughout the ObjC API. They will all need to be audited:

    shell> ack 'create.*?Error' API/*.mm
    API/JSValue.mm
    784:        *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSArray"));
    800:        *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSDictionary"));

    API/ObjCCallbackFunction.mm
    132:        *exception = toRef(JSC::createTypeError(toJS(contextRef), "Argument does not match Objective-C Class"));
    456:    // (2) We're calling some JSC internals that require us to be on the 'inside' - e.g. createTypeError.
    499:        *exception = toRef(JSC::createTypeError(toJS(contextRef), "Objective-C blocks called as constructors must return an object."));
    565:            *exception = toRef(JSC::createTypeError(toJS(contextRef), "self type check failed for Objective-C instance method"));
    575:            *exception = toRef(JSC::createTypeError(toJS(contextRef), "self type check failed for Objective-C instance method"));

[Geoffrey Garen]

> My guess is that the ObjC API has to grab the APIEntryShim before it calls into JSC (via JSC::createTypeError).

Right.

[Mark Hahnenberg]

Created attachment 224250 [details]
Patch

[Mark Hahnenberg]

(In reply to comment #1)
> There are a number of create calls throughout the ObjC API. They will all need to be audited:
> 
>     shell> ack 'create.*?Error' API/*.mm
>     API/JSValue.mm
>     784:        *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSArray"));
>     800:        *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSDictionary"));
> 
>     API/ObjCCallbackFunction.mm
>     132:        *exception = toRef(JSC::createTypeError(toJS(contextRef), "Argument does not match Objective-C Class"));
>     456:    // (2) We're calling some JSC internals that require us to be on the 'inside' - e.g. createTypeError.
>     499:        *exception = toRef(JSC::createTypeError(toJS(contextRef), "Objective-C blocks called as constructors must return an object."));
>     565:            *exception = toRef(JSC::createTypeError(toJS(contextRef), "self type check failed for Objective-C instance method"));
>     575:            *exception = toRef(JSC::createTypeError(toJS(contextRef), "self type check failed for Objective-C instance method"));

I audited the other call sites and other than these two they are all in places where we already hold the API lock.

[Joseph Pecoraro]

Comment on attachment 224250 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=224250&action=review

r=me

> Source/JavaScriptCore/API/JSValue.mm:785
>          *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSArray"));

Nit: You could ASCIILiteral(...) the string being passed to createTypeError, since it is a string literal turning into a WTF::String.

Is the lock needed by the "tryUnwrapObjcObject" call at the top of these functions?

Otherwise this is identical to the local fix I had.

[Mark Hahnenberg]

(In reply to comment #5)
> (From update of attachment 224250 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=224250&action=review
> 
> r=me
> 
> > Source/JavaScriptCore/API/JSValue.mm:785
> >          *exception = toRef(JSC::createTypeError(toJS(context), "Cannot convert primitive to NSArray"));
> 
> Nit: You could ASCIILiteral(...) the string being passed to createTypeError, since it is a string literal turning into a WTF::String.
> 
Will do.

> Is the lock needed by the "tryUnwrapObjcObject" call at the top of these functions?
> 
I don't think so. We don't create any objects in it, and the object is always reachable from the stack. If I missed something and we do need to take the shim for things that happen inside tryUnwrapObjcObject then the shim should probably go in there instead of out here.

[Mark Hahnenberg]

Committed r164136: <http://trac.webkit.org/changeset/164136>