Skip to content

Assertion fail on MLIR empty edge names

Low
pak-laura published GHSA-jvhc-5hhr-w3v5 Sep 15, 2022

Package

pip tensorflow, tensorflow-cpu, tensorflow-gpu (pip)

Affected versions

< 2.10.0

Patched versions

2.7.4, 2.8.3, 2.9.2, 2.10.0

Description

Impact

When mlir::tfg::ConvertGenericFunctionToFunctionDef is given empty function attributes, it crashes.

// We pre-allocate the array of operands and populate it using the
// `output_name_to_position` and `control_output_to_position` populated
// previously.
SmallVector<Value> ret_vals(func.ret_size() + func.control_ret_size(),
                            Value());
for (const auto& ret_val : func.ret()) {
  auto position = output_name_to_position.find(ret_val.first);
  if (position == output_name_to_position.end())
    return InvalidArgument(
        "Can't import function, returned value references unknown output "
        "argument ",
        ret_val.first);
  ret_vals[position->second] =
      value_manager.GetValueOrCreatePlaceholder(ret_val.second);
}
for (const auto& ret_val : func.control_ret()) {
  auto position = control_output_to_position.find(ret_val.first);
  if (position == control_output_to_position.end())
    return InvalidArgument(
        "Can't import function, returned value references unknown output "
        "argument ",
        ret_val.first);
  Value result = value_manager.GetValueOrCreatePlaceholder(
      (Twine("^") + ret_val.second).str());

ret_val.second cannot be empty. Neither can input.

// Process every node and create a matching MLIR operation
for (const NodeDef& node : nodes) {
  if (node.op().empty()) return InvalidArgument("empty op type");
  OperationState state(unknown_loc, absl::StrCat("tfg.", node.op()));
  // Fetch the inputs, creating placeholder if an input hasn't been visited.
  for (const std::string& input : node.input())
    state.operands.push_back(
        value_manager.GetValueOrCreatePlaceholder(input));

Patches

We have patched the issue in GitHub commit ad069af92392efee1418c48ff561fd3070a03d7b.

The fix will be included in TensorFlow 2.10.0. We will also cherrypick this commit on TensorFlow 2.9.1, TensorFlow 2.8.1, and TensorFlow 2.7.2, as these are also affected and still in supported range.

For more information

Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.

Severity

Low

CVE ID

CVE-2022-36012

Weaknesses

No CWEs