人工智能 redundancy elimination reducer in v8 and 34c3 ctf v9

sh1san · May 25, 2020 · 29 hits

0x00 Overview

RedundancyElimination reducer is used to remove nodes that are not necessary, such as redundant kCheckXXX node. For example, if there is already a kCheckXXX before another kCheckXXX, the later kCheckXXX may be removed if we are sure this check will never be violated. However, if this assumption is wrong, vulnerability may arise. V9 in 34C3 CTF is an example for this.

As I covered in previous articles, GraphReducer reduces the graph using depth-first-search, starting from End node. However, when we tries to reduce a redundant node, we may need to have the information about its input effect nodes (e.i. nodes before), which we cannot have if we search the graph from End node.

Therefore, the approach this reducer applies is to do nothing until Start node is reached, and use the revisit trick to perform search again starting from Start, and this time we will have the information about input effect nodes.

Here is the Reduce function of RedundancyElimination:

Reduction RedundancyElimination::Reduce(Node* node) {
  if (node_checks_.Get(node)) return NoChange();
  switch (node->opcode()) {
    case IrOpcode::kCheckBigInt:
    case IrOpcode::kCheckBounds:
    case IrOpcode::kCheckEqualsInternalizedString:
    /* ... other check nodes ... */
      return ReduceCheckNode(node);
    /* ... some other nodes handling ... */
    case IrOpcode::kStart:
      return ReduceStart(node);
  }
  return NoChange();
}

The ReduceCheckNode is the one that I am going to focus on. Let’s put the conclusion first, field node_checks_ is used to record all effect input predecessor nodes of every node. node_checks_.Get is used to map a node to all of its effect input predecessors, using nodeid as index to a vector. EffectPathChecks is essentially a linked-list of Node* pointers.

Reduction RedundancyElimination::ReduceCheckNode(Node* node) {
  Node* const effect = NodeProperties::GetEffectInput(node);
  EffectPathChecks const* checks = node_checks_.Get(effect);
  // If we do not know anything about the predecessor, do not propagate just yet
  // because we will have to recompute anyway once we compute the predecessor.
  if (checks == nullptr) return NoChange();
  // See if we have another check that dominates us.
  if (Node* check = checks->LookupCheck(node)) {
    ReplaceWithValue(node, check);
    return Replace(check);
  }

  // Learn from this check.
  return UpdateChecks(node, checks->AddCheck(zone(), node));
}

Initially, everything is empty, so node_checks_.Get will just return nullptr, so ReduceCheckNode will just return NoChange(). This always holds until Start node is reached.

Reduction RedundancyElimination::ReduceStart(Node* node) {
  return UpdateChecks(node, EffectPathChecks::Empty(zone()));
}

UpdateChecks essentially calls node_checks_.Set(node, checks) to update the checks of given node, but this is only done when new checks is different from original one. Also, if there is any change, return Changed(node) as Reduction.

Reduction RedundancyElimination::UpdateChecks(Node* node,
                                              EffectPathChecks const* checks) {
  EffectPathChecks const* original = node_checks_.Get(node);
  // Only signal that the {node} has Changed, if the information about {checks}
  // has changed wrt. the {original}.
  if (checks != original) {
    if (original == nullptr || !checks->Equals(original)) {
      node_checks_.Set(node, checks);
      return Changed(node);
    }
  }
  return NoChange();
}

Therefore, after reducing Start node, the result of reduction will be a in-place replacement. Remember, if the result is in-place replacement, all use nodes will be revisited again.

// GraphReducer::ReduceTop
// Check if we have a new replacement.
if (replacement != node) {
  Replace(node, replacement, max_id);
} else {
  // Revisit all uses of the node.
  for (Node* const user : node->uses()) {
    // Don't revisit this node if it refers to itself.
    if (user != node) Revisit(user);
  }
}

Thus the use nodes will be visited again, but this time effect input node_checks_.Get(effect) does not return nullptr but UpdateChecks(node, checks->AddCheck(zone(), node)).

AddCheck simply add the node to the linked-list, and return the newly created EffectPathChecks, defined as below.

RedundancyElimination::EffectPathChecks const*
RedundancyElimination::EffectPathChecks::AddCheck(Zone* zone,
                                                  Node* node) const {
  Check* head = new (zone->New(sizeof(Check))) Check(node, head_);
  return new (zone->New(sizeof(EffectPathChecks)))
      EffectPathChecks(head, size_ + 1);
}

This means EffectPathChecks of a node stores itself plus EffectPathChecks of its effect node, and base case is Start node whose EffectPathChecks is empty. In other words, it stores all previous nodes on the effect input chain, including itself but except Start node. Note that this is only the case if there are kCheckXXX nodes only, for other kinds of nodes, things might be different. // this might be inaccurate?

Then, since UpdateChecks which adds a new node always returns Changed(node), all use nodes of this node (one of the use nodes of Start node) will also be pushed into revisit queue. I would say this is a bit similar to BFS starting from Start node, but in reality things might not work in this way because there are also other reducers and because DFS is also still ongoing.

0x02 Eliminate the Redundancy

There is a piece of code in ReduceCheckNode that I have not discussed yet.

// See if we have another check that dominates us.
if (Node* check = checks->LookupCheck(node)) {
  ReplaceWithValue(node, check); 
  // remove node, use check new value node
  return Replace(check);
  // identical to `return Reduction(check)`
  // nothing will be done in GraphReducer::Replace later on
  // because node has been removed,
  // and it has no use nodes
}

The intuition of this is to check whether node can be covered by previous input effect nodes, and if so, remove this node, because we know it is redundant.

LookupCheck iterates over the list stored in checks, which stores effect predecessor nodes. If one of the predecessor nodes subsumes the given node, return it.

Node* RedundancyElimination::EffectPathChecks::LookupCheck(Node* node) const {
  for (Check const* check = head_; check != nullptr; check = check->next) {
    if (CheckSubsumes(check->node, node) && TypeSubsumes(node, check->node)) {
      DCHECK(!check->node->IsDead());
      return check->node;
    }
  }
  return nullptr;
}

// Does check {a} subsume check {b}?
bool CheckSubsumes(Node const* a, Node const* b) {
  if (a->op() != b->op()) {
    if (a->opcode() == IrOpcode::kCheckInternalizedString &&
        b->opcode() == IrOpcode::kCheckString) {
      // CheckInternalizedString(node) implies CheckString(node)
    } else if (a->opcode() == IrOpcode::kCheckSmi &&
               b->opcode() == IrOpcode::kCheckNumber) {
      // CheckSmi(node) implies CheckNumber(node)
    }  /* ... some other else-if checks ... */
      else if (a->opcode() != b->opcode()) {
      return false;
    } else {
      switch (a->opcode()) {
        case IrOpcode::kCheckBounds:
        case IrOpcode::kCheckSmi:
        case IrOpcode::kCheckString:
        case IrOpcode::kCheckNumber:
        case IrOpcode::kCheckBigInt:
          break;
        /* ... other cases ... */
      }
    }
  } // check if opcode of a could subsume opcode of b
  for (int i = a->op()->ValueInputCount(); --i >= 0;) {
    if (a->InputAt(i) != b->InputAt(i)) return false;
  } // check that inputs are identical
  return true;
}

0x03 V9

V9 is a challenge from 34C3 CTF, the patch I am going to use here is v9_7.2.patch.

Root Cause

In the patch, case IrOpcode::kCheckMaps: is added to Reduce function, so that for this opcode, ReduceCheckNode will be called, and a piece of codes is added to CheckSubsumes function.

case IrOpcode::kCheckMaps: {
    // CheckMaps are compatible if the first checks a subset of the second.
    ZoneHandleSet<Map> const& a_maps = CheckMapsParametersOf(a->op()).maps();
    ZoneHandleSet<Map> const& b_maps = CheckMapsParametersOf(b->op()).maps();
    if (!b_maps.contains(a_maps)) {
        return false;
    } // TODO: investigate CheckMapsParametersOf and ZoneHandleSet<Map>
    break;
}

What this does is essentially check if kCheckMaps b includes kCheckMaps a, if so we don’t need the b because it thinks b must hold as long as a holds. However, the problem is map of object can actually change during the execution, so even if it could satisfy a before, it does not have to satisfy b later on. Thus the assumption is wrong, and using this we can have type confusion.

let’s see how this function is reduced by RedundancyElimination.

function f(o, g)
{
    const x = o.x;// = 0xdead;
    g();
    o.d = 0x2000;
}
for (var i = 0; i < 0x1000; i++)
{
    f({x:1,y:2,z:3,l:4,a:5,b:6,c:7,d:8,e:9}, ()=>1);
    f({x:1,y:2,z:3,l:4,a:5,b:6,c:7,d:8,e:9}, ()=>2);
    f({x:1,y:2,z:3,l:4,a:5,b:6,c:7,d:8,e:9}, ()=>3);
}

Since the first phase that use RedundancyElimination is LoadElimination, we can look at LoopPeeling to see the graph before RedundancyElimination.

// pic

As we can see, there are 2 CheckMaps nodes. Node 34 is effect input of LoadField[+24], and node 37 is effect input of StoreField[+80].

After LoadElimination phase, the second CheckMaps disappears, so StoreField[+80] is executed without checking, which is problematic.

//pic

This is also clear if we look at assembly generated.

call r10 ; call g()
REX.W movq rax,0x200000000000
REX.W movq rbx,[rbp+0x18] ; load object pointer
REX.W movq [rbx+0x4f],rax ; write to +0x50 field
; clearly there is no check

Exploitation

In function g(), if we can shrink the size of obj, we can have a out-of-bound write. The idea is to put an array there so that we can rewrite its length field, which gives an array with OOB access. Now the exploitation becomes regular: use a signature array to leak web assembly function address and useArrayBuffer to achieve arbitrary R&W.

The difference is, after shrinking size of obj, positions that is originally used to store fields will be filled with FILLER_TYPE objects, and it is useless to rewrite these things. Therefore, we need to invoke garbage collection so that these offsets will store interesting data.

Full exploit with comments is here.

No Reply at the moment.
You need to Sign in before reply, if you don't have an account, please Sign up first.