• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 Google Inc. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #ifndef TENSORFLOW_CORE_KERNELS_FUZZING_FUZZ_SESSION_H_
17 #define TENSORFLOW_CORE_KERNELS_FUZZING_FUZZ_SESSION_H_
18 
19 #include "tensorflow/cc/framework/scope.h"
20 #include "tensorflow/core/graph/graph.h"
21 #include "tensorflow/core/public/session.h"
22 
23 // Standard invoking function macro to dispatch to a fuzzer class.
24 #ifndef PLATFORM_WINDOWS
25 #define STANDARD_TF_FUZZ_FUNCTION(FuzzerClass)                              \
26   extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) { \
27     static FuzzerClass* fuzzer = new FuzzerClass();                         \
28     return fuzzer->Fuzz(data, size);                                        \
29   }
30 #else
31 // We don't compile this for Windows, MSVC doesn't like it as pywrap in Windows
32 // links all the code into one big object file and there are conflicting
33 // function names.
34 #define STANDARD_TF_FUZZ_FUNCTION(FuzzerClass)
35 #endif
36 
37 // Standard builder for hooking one placeholder to one op.
38 #define SINGLE_INPUT_OP_BUILDER(dtype, opName)                          \
39   void BuildGraph(const Scope& scope) override {                        \
40     auto op_node =                                                      \
41         tensorflow::ops::Placeholder(scope.WithOpName("input"), dtype); \
42     (void)tensorflow::ops::opName(scope.WithOpName("output"), op_node); \
43   }
44 
45 namespace tensorflow {
46 namespace fuzzing {
47 
48 // Create a TensorFlow session using a specific GraphDef created
49 // by BuildGraph(), and make it available for fuzzing.
50 // Users must override BuildGraph and FuzzImpl to specify
51 // (1) which operations are being fuzzed; and
52 // (2) How to translate the uint8_t* buffer from the fuzzer
53 //     to a Tensor or Tensors that are semantically appropriate
54 //     for the op under test.
55 // For the simple cases of testing a single op that takes a single
56 // input Tensor, use the SINGLE_INPUT_OP_BUILDER(dtype, opName) macro in place
57 // of defining BuildGraphDef.
58 //
59 // Typical use:
60 // class FooFuzzer : public FuzzSession {
61 //   SINGLE_INPUT_OP_BUILDER(DT_INT8, Identity);
62 //   void FuzzImpl(const uint8_t* data, size_t size) {
63 //      ... convert data and size to a Tensor, pass it to:
64 //      RunInputs({{"input", input_tensor}});
65 //
66 class FuzzSession {
67  public:
FuzzSession()68   FuzzSession() : initialized_(false) {}
~FuzzSession()69   virtual ~FuzzSession() {}
70 
71   // Constructs a Graph using the supplied Scope.
72   // By convention, the graph should have inputs named "input1", ...
73   // "inputN", and one output node, named "output".
74   // Users of FuzzSession should override this method to create their graph.
75   virtual void BuildGraph(const Scope& scope) = 0;
76 
77   // Implements the logic that converts an opaque byte buffer
78   // from the fuzzer to Tensor inputs to the graph.  Users must override.
79   virtual void FuzzImpl(const uint8_t* data, size_t size) = 0;
80 
81   // Initializes the FuzzSession.  Not safe for multithreading.
82   // Separate init function because the call to virtual BuildGraphDef
83   // can't be put into the constructor.
InitIfNeeded()84   Status InitIfNeeded() {
85     if (initialized_) {
86       return Status::OK();
87     }
88     initialized_ = true;
89 
90     Scope root = Scope::DisabledShapeInferenceScope().ExitOnError();
91     SessionOptions options;
92     session_ = std::unique_ptr<Session>(NewSession(options));
93 
94     BuildGraph(root);
95 
96     GraphDef graph_def;
97     TF_CHECK_OK(root.ToGraphDef(&graph_def));
98 
99     Status status = session_->Create(graph_def);
100     if (!status.ok()) {
101       // This is FATAL, because this code is designed to fuzz an op
102       // within a session.  Failure to create the session means we
103       // can't send any data to the op.
104       LOG(FATAL) << "Could not create session: " << status.error_message();
105     }
106     return status;
107   }
108 
109   // Runs the TF session by pulling on the "output" node, attaching
110   // the supplied input_tensor to the input node(s), and discarding
111   // any returned output.
112   // Note: We are ignoring Status from Run here since fuzzers don't need to
113   // check it (as that will slow them down and printing/logging is useless).
RunInputs(const std::vector<std::pair<string,Tensor>> & inputs)114   void RunInputs(const std::vector<std::pair<string, Tensor> >& inputs) {
115     RunInputsWithStatus(inputs).IgnoreError();
116   }
117 
118   // Same as RunInputs but don't ignore status
RunInputsWithStatus(const std::vector<std::pair<string,Tensor>> & inputs)119   Status RunInputsWithStatus(
120       const std::vector<std::pair<string, Tensor> >& inputs) {
121     return session_->Run(inputs, {}, {"output"}, nullptr);
122   }
123 
124   // Dispatches to FuzzImpl;  small amount of sugar to keep the code
125   // of the per-op fuzzers tiny.
Fuzz(const uint8_t * data,size_t size)126   int Fuzz(const uint8_t* data, size_t size) {
127     Status status = InitIfNeeded();
128     TF_CHECK_OK(status) << "Fuzzer graph initialization failed: "
129                         << status.error_message();
130     // No return value from fuzzing:  Success is defined as "did not
131     // crash".  The actual application results are irrelevant.
132     FuzzImpl(data, size);
133     return 0;
134   }
135 
136  private:
137   bool initialized_;
138   std::unique_ptr<Session> session_;
139 };
140 
141 // A specialized fuzz implementation for ops that take
142 // a single string.  Caller must still define the op
143 // to plumb by overriding BuildGraph or using
144 // a plumbing macro.
145 class FuzzStringInputOp : public FuzzSession {
FuzzImpl(const uint8_t * data,size_t size)146   void FuzzImpl(const uint8_t* data, size_t size) final {
147     Tensor input_tensor(tensorflow::DT_STRING, TensorShape({}));
148     input_tensor.scalar<tstring>()() =
149         string(reinterpret_cast<const char*>(data), size);
150     RunInputs({{"input", input_tensor}});
151   }
152 };
153 
154 }  // end namespace fuzzing
155 }  // end namespace tensorflow
156 
157 #endif  // TENSORFLOW_CORE_KERNELS_FUZZING_FUZZ_SESSION_H_
158