CVE-2021-4102: Chrome incorrect node elision in Turbofan leads to unexpected WriteBarrier elision
Brendon Tiszka, Chrome
The Basics
Disclosure or Patch Date: December 13, 2021
Product: Google Chrome
Affected Versions: pre 96.0.4664.110
First Patched Version: 96.0.4664.110
Issue/Bug Report: crbug.com/1278387
Patch CL: https://chromium-review.googlesource.com/c/v8/v8/+/3329790
Bug-Introducing CL: https://codereview.chromium.org/1908093002
Reporter(s): Anonymous
The Code
Proof-of-concepts:
let mark_sweep = () => {
new ArrayBuffer(2**34);
}
let scavenge = () => {
let ref = new ArrayBuffer(16777216);
ref = new ArrayBuffer(1);
ref = new ArrayBuffer(1);
ref = null;
}
function noop() {}
{
// Ensure `o` is a non-const signed field
var o = {'a':5};
o.a = 5;
// Move `o` to OldSpace
mark_sweep();
var b;
function opt() {
var a1;
for (var i = 0; i < 2; i++) {
b = i + -0;
a1 = Math.abs(b);
for (var j = 0; j < 2; j++) {}
for (i = 0; i < 2; i++) {}
o.a = Math.abs(a1);
noop();
}
}
%PrepareFunctionForOptimization(opt);
opt();
opt();
%OptimizeFunctionOnNextCall(opt);
opt();
// Trigger Scavenger, use-after-free
scavenge();
mark_sweep();
var ret = o.a;
ret[0];
}
Exploit sample: No
Did you have access to the exploit sample when doing the analysis? No
The Vulnerability
Bug class: WriteBarrier elision (use-after-free)
Vulnerability details:
Prerequisites:
- V8 has two garbage collection cycles: Scavenger (minor) and Mark-Compact (major). For the purposes of this RCA, the scavenger only collects garbage in the young-generation
NewSpace
and mark-compact collects garbage in all generationsOldSpace
andNewSpace
. - Objects in
OldSpace
can hold references to objects inNewSpace
. WriteBarriers are used to maintain a list of old-to-new references which are iterated during Scavenge (minor-gc). - WriteBarriers are only required for pointers to other
HeapObjects
. V8 will ignore WriteBarrier slots that point toSmis
. - Turbofan's
SimplifiedLoweringPhase
determines if a WriteBarrier is needed when visiting aStoreField
node by checking the object's field representation. If the field representation is not a tagged or a tagged pointer, then the WriteBarrier will be elided.
An optimization in EarlyOptimizationPhase
that elides ChangeTaggedToInt32
-> ChangeInt31ToTaggedSigned
caused a HeapNumber
to be stored to an object in a field with a TaggedSigned
representation. A check before EarlyOptimizationPhase
in SimplifiedLoweringPhase
makes the correct assumption that write barriers can be elided when storing anything to a field in an object with the TaggedSigned
representation. This leads to a use-after-free if the HeapNumber
being mistakenly stored to the object lives in NewSpace
, and the object being stored to lives in OldSpace
.
case IrOpcode::kChangeInt31ToTaggedSigned:
case IrOpcode::kChangeInt32ToTagged: {
...
if (m.IsChangeTaggedToInt32() || ...) {
return Replace(m.InputAt(0));
}
}
Parts 1 through 4 of the vulnerability details go deep into how to reach this vulnerability.
(Part 1): Eliding Write Barriers in Turbofan
The bug itself lives in SimplifiedOperatorReducer
, which runs during the EarlyOptimizationPhase
. EarlyOptimizationPhase
is notably after SimplifiedLoweringPhase
(where the WriteBarrier
calculation for StoreField
happens):
case IrOpcode::kChangeInt31ToTaggedSigned: /*** 1 ***/
case IrOpcode::kChangeInt32ToTagged: {
Int32Matcher m(node->InputAt(0));
if (m.HasResolvedValue()) return ReplaceNumber(m.ResolvedValue());
if (m.IsChangeTaggedToInt32() || m.IsChangeTaggedSignedToInt32()) { /*** 2 ***/
return Replace(m.InputAt(0)); /*** 3 ***/
}
break;
}
The bug is simple. When the SimplifiedOperatorReducer
visits a ChangeInt31ToTaggedSigned
[1] node, it will elide [3] ChangeInt32ToTagged
-> ChangeInt31ToTaggedSigned
[2]. Consider this theoretical sequence of nodes after the SimplifiedLoweringPhase
which can be emitted when storing to a TaggedSigned
field:
The input to ChangeTaggedToInt32
can be either a Smi
or a TaggedPointer
. If the input to ChangeTaggedToInt32
is a TaggedPointer
, and ChangeTaggedToInt32
-> ChangeInt31ToTaggedSigned
is elided, then it will result in an object with a TaggedPointer
type being stored to an object field whose representation is TaggedSigned
during StoreField
.
While the bug is simple, producing the correct sequence of nodes after SimplifiedLoweringPhase
and coercing SimplifiedOperatorReducer
into visiting ChangeInt31ToTaggedSigned
before ChangeTaggedToInt32
is difficult.
(Part 2): Arranging Nodes Pre-Simplified-Lowering
Let's take a look at a simplified poc that can be used to create the arrangement of nodes mentioned above after SimplifiedLoweringPhase
:
var o = {'a': 5}; o.a = 6;
var b;
function opt() {
for (var i = 0; i < 2; i++) {
b = i + -0;
o.a = Math.abs(b);
}
}
Step 1
var o = {'a': 5}; o.a = 6;
We start by creating a JSObject
with a field that has a mutable TaggedSigned
representation.
Step 2
var b;
for (var i = 0; i < 2; i++) {
b = i + -0;
[...]
}
JSNativeContextSpecialization
will emit a CheckHeapObject node as input to StoreField because we are storing a HeapNumber
to the global property b
.
Step 3
for (var i = 0; i < 2; i++) {
[...]
o.a = Math.abs(b);
}
Passing b
to Math.abs
causes LoadEliminationPhase
to pass the CheckHeapObject
node as input to both the NumberAbs
and the global StoreField
above. Math.abs
will be reduced to NumberAbs
and then JSSetNamedProperty
will be reduced to a CheckSmi node as input to StoreField because the field that's being stored to has the field representation TaggedSmi
.
Putting all of this together this results in this graph after LoadEliminationPhase
:
(Part 3): Arranging Nodes during SimplifiedLowering
These nodes are lowered as follows in the SimplifiedLoweringPhase
:
NumberAdd
is lowered toInt32Add
because its type isRange(0, 2)
.CheckHeapObject
is lowered toChangeUint32ToFloat64
->ChangeFloat64ToTaggedPointer
because the type ofNumberAdd
isRange(0, 2)
.NumberAbs
is lowered toChangeTaggedToInt32
because the type ofCheckHeapObject
isRange(0, 2)
.CheckSmi
is lowered toChangeInt32ToTaggedSigned
because the type of theNumberAbs
node isRange(0, 2)
.
This results in the above graph which appears to satisfy all of the conditions that we need to trigger this vulnerability and store a TaggedPointer
to a field with a smi
representation without a write barrier because ChangeUint64ToFloat64
-> ChangeFloat64ToTaggedPointer
-> ChangeTaggedToInt32
-> ChangeInt31ToTaggedSigned
-> StoreField
should be elided directly to ChangeUint64ToFloat64
-> ChangeFloat64ToTaggedPointer
-> StoreField
.
(Part 4): Traversing the Graph
Turbofan's GraphReducer algorithm is a modified version of depth-first-search. It starts from the kEnd
node and recurses up node inputs from left to right until it reaches a node that has either seen before or until all of its inputs have been visited. The order in which nodes are visited is important triggering this bug. For example, the simple proof-of-conept from Part2/Part3 does not reach the vulnerability because ChangeTaggedToInt32
is visited before the ChangeInt31ToTaggedSigned
.
case IrOpcode::kChangeTaggedSignedToInt32:
case IrOpcode::kChangeTaggedToInt32: {
NumberMatcher m(node->InputAt(0));
if (m.HasResolvedValue())
return ReplaceInt32(DoubleToInt32(m.ResolvedValue()));
if (m.IsChangeFloat64ToTagged() || m.IsChangeFloat64ToTaggedPointer()) {
return Change(node, machine()->ChangeFloat64ToInt32(), m.InputAt(0)); /*** A ***/
}
if (m.IsChangeInt31ToTaggedSigned() || m.IsChangeInt32ToTagged()) {
return Replace(m.InputAt(0));
}
break;
}
This happens because the graph is traversed with DFS starting from End
, we will always visit ChangeTaggedToInt32
before ChangeInt31ToTaggedSigned
with that graph because ChanteTaggedToInt32
only has one output and ChanteInt32ToTaggedSigned
only has one input. ChangeTaggedToInt32
is guaranteed to be reduced before ChangeInt31ToTaggedSigned
:
When this happens, SimplifedOperatorReducer
elides ChangeUint64ToFloat64
-> ChangeFloat64ToTaggedPointer
-> ChangeTaggedToInt32
-> ChangeInt31ToTaggedSigned
-> StoreField
to ChangeUint64ToFloat64
-> ChangeFloat64ToInt32
-> ChangeInt31ToTaggedSigned
-> StoreField
[A] which is semantically valid and does not introduce a security issue.
To get around this, the final proof of concept intricately and cleverly creates multiple FrameState
, StateValue
, and CheckPoint
nodes that take ChangeTaggedToInt32
as an input, causing the GraphTraversal to push ChangeTaggedToInt32
on the stack before reaching ChangeInt31ToTaggedSigned
. This causes the GraphTraversal for SimplifiedOperatorReducer
to stop at ChangeInt31ToTaggedSigned
because ChangeTaggedToInt32
is already on the stack. Finally, ChanteTaggedToInt32
-> ChanteInt31ToTaggedSigned
is elided.
Patch analysis: Commit 4fae8b1 fixed the bug in SimplifiedOperatorReducer
that could cause ChangeTaggedToInt32
-> ChangeInt31ToTaggedSigned
to be elided.
Thoughts on how this vuln might have been found (fuzzing, code auditing, variant analysis, etc.):
The bug was probably discovered through manual analysis.
(Historical/present/future) context of bug:
There was a bug in Turbofan's handling of WriteBarriers in the past during SimplfiedLoweringPhase
crbug.com/791245. However, CVE-2021-4102 was the first bug of its kind that bypassed that check in a later optimization pass.
The Exploit
(The terms exploit primitive, exploit strategy, exploit technique, and exploit flow are defined here.)
Exploit strategy (or strategies): We did not have an exploit sample while analyzing this vulnerability, but it is reasonable to assume that the use-after-free was used to either leak a sentinel value like TheHole
or UninitializedOddball
or to directly materialize a fake object similar to crbug.com/1307610.
Exploit flow: A typical exploit will first construct an arbitrary read/write primitive, then use that to gain shellcode execution.
Known cases of the same exploit flow: Most other V8 exploits.
Part of an exploit chain? N/A
The Next Steps
Variant analysis
Areas/approach for variant analysis (and why):
- Manual Variant Analysis: Look at other reduction passes after
SimplifiedLoweringPhase
that could be abused to cause unexpected WriteBarrier elision.
Found variants:
- crbug.com/1307610 discovered by Brendon Tiszka is not a direct variant of this vulnerability, however it is another WriteBarrier elision bug and was reported after root-causing this vulnerability.
- crbug.com/1382434 discovered by Sergei Glazunov of Google Project Zero is not a variant of this vulnerability, however it uses WriteBarrier elision to exploit a class of vulnerabilty.
- crbug.com/1423610 discovered by Nan Wang and Zhenghang Xiao of Qihoo360 is another WriteBarrier elision bug in V8's Maglev JIT compiler.
Structural improvements
What are structural improvements such as ways to kill the bug class, prevent the introduction of this vulnerability, mitigate the exploit flow, make this type of vulnerability harder to exploit, etc.?
Ideas to kill the bug class: N/A
Ideas to mitigate the exploit flow: The V8 Sandbox project is designed to break this exploit flow for the vast majority of V8 vulnerabilities, including this one.
Other potential improvements: N/A
0-day detection methods
What are potential detection methods for similar 0-days? Meaning are there any ideas of how this exploit or similar exploits could be detected as a 0-day? N/A
Other References
- A WriteBarrier elision hole leaking technique and a WriteBarrier elision fake object technique - exploit - are publicly documented.