CVE-2019-5788

Code txt

There's a comment in FileSystemOperationRunner::BeginOperation

OperationID FileSystemOperationRunner::BeginOperation(
    std::unique_ptr<FileSystemOperation> operation) {
  OperationID id = next_operation_id_++;

  // TODO(https://crbug.com/864351): Diagnostic to determine whether OperationID
  // wrap-around is occurring in the wild.
  DCHECK(operations_.find(id) == operations_.end());

  // ! If id already in operations_, this will free operation
  operations_.emplace(id, std::move(operation));
  return id;
}

The id is an int, and it can wrap, and if it does this will cause a use-after-free in the browser process, since the normal usage of BeginOperation is the following:

OperationID FileSystemOperationRunner::Truncate(const FileSystemURL& url,
                                                int64_t length,
                                                StatusCallback callback) {
  base::File::Error error = base::File::FILE_OK;
  std::unique_ptr<FileSystemOperation> operation = base::WrapUnique(
      file_system_context_->CreateFileSystemOperation(url, &error));
  // ! take a raw pointer to the contents of the unique_ptr
  FileSystemOperation* operation_raw = operation.get();
  // ! call BeginOperation passing the move'd unique_ptr, freeing operation
  OperationID id = BeginOperation(std::move(operation));
  base::AutoReset<bool> beginning(&is_beginning_operation_, true);
  if (!operation_raw) {
    DidFinish(id, std::move(callback), error);
    return id;
  }
  PrepareForWrite(id, url);
  // ! use the raw free'd pointer here.
  operation_raw->Truncate(url, length,
                          base::BindOnce(&FileSystemOperationRunner::DidFinish,
                                         weak_ptr_, id, std::move(callback)));
  return id;
}

I think that to trigger this, you'd need either a malformed blob in the blob registry, or access to the FileWriter api, so at present this would require a compromised renderer.

I've attached two PoCs that should trigger this issue; it looks like the runtime for either approach from javascript should take ~2 days on my machine. (I'd suggest patching the OperationId typedef to short to reproduce, unless you are extremely patient).

$ python ./copy_mojo_js_bindings.py /path/to/chrome/.../out/Asan/gen
$ python -m SimpleHTTPServer&
$ /ssd/chrome_trunk/src/out/Asan/chrome --enable-blink-features=MojoJS --user-data-dir=/tmp/aa 'http://localhost:8000/id_overflow_no_filewriter.html'


Proof of Concept:
https://gitlab.com/exploit-database/exploitdb-bin-sploits/-/raw/main/bin-sploits/46571.zip