DeepState

DeepState is a framework that provides C and C++ developers with a common interface to various symbolic execution and fuzzing engines. Users can write one test harness using a Google Test-like API, then execute it using multiple backends without having to learn the complexities of the underlying engines. It supports writing unit tests and API sequence tests, as well as automatic test generation. Read more about the goals and design of DeepState in our paper.

The 2018 IEEE Cybersecurity Development Conference included a full tutorial on effective use of DeepState.

There are two blog posts on using DeepState: Part 1 and Part 2.

Overview of Features

• Tests look like Google Test, but can use symbolic execution/fuzzing to generate data (parameterized unit testing)
  • Easier to learn than binary analysis tools/fuzzers, but provides similar functionality
• Already supports Manticore, Angr, libFuzzer, file-based fuzzing with e.g., AFL or Eclipser; more back-ends likely in future
  • Switch test generation tool without re-writing test harness
    • Work around show-stopper bugs
    • Find out which tool works best for your code under test
    • Different tools find different bugs/vulnerabilities
    • Fair way to benchmark/bakeoff tools
• Provides test replay for regression plus effective automatic test case reduction to aid debugging
• Supports API-sequence generation with extensions to Google Test interface
  • Concise readable way (OneOf) to say "run one of these blocks of code"
  • Same construct supports fixed value set non-determinism
  • E.g., writing a POSIX file system tester is pleasant, not painful as in pure Google Test idioms
• Provides high-level strategies for improving symbolic execution/fuzzing effectiveness
  • Pumping (novel to DeepState) to pick concrete values when symbolic execution is too expensive
  • Automatic decomposition of integer compares to guide coverage-driven fuzzers

To put it another way, DeepState sits at the intersection of property-based testing, traditional unit testing, fuzzing, and symbolic execution. It lets you perform property-based unit testing using fuzzing or symbolic execution as a back end to generate data, and saves the results so that what DeepState finds can easily be used in deterministic settings such as regression testing or CI.

Supported Platforms

DeepState currently targets Linux, with macOS support in progress (the fuzzers work fine, but symbolic execution is not well-supported yet, without a painful cross-compilation process).

Dependencies

Build:

• CMake
• GCC and G++ with multilib support
• Python 3.6 (or newer)
• Setuptools

Runtime:

• Python 3.6 (or newer)
• Z3 (for the Manticore backend)

Building on Ubuntu 16.04 (Xenial)

First make sure you install Python 3.6 or greater. Then use this command line to install additional requirements and compile DeepState:

sudo apt update && sudo apt-get install build-essential gcc-multilib g++-multilib cmake python3-setuptools libffi-dev z3
git clone https://github.com/trailofbits/deepstate deepstate
mkdir deepstate/build && cd deepstate/build
cmake ../
make

Installing

Assuming the DeepState build resides in $DEEPSTATE, run the following commands to install the DeepState python package:

virtualenv venv
. venv/bin/activate
python $DEEPSTATE/build/setup.py install

The virtualenv-enabled $PATH should now include two executables: deepstate and deepstate-angr. These are executors, which are used to run DeepState test binaries with specific backends (automatically installed as Python dependencies). The deepstate or deepstate-manticore executor uses the Manticore backend while deepstate-angr uses angr. They share a common interface where you may specify a number of workers and an output directory for saving backend-generated test cases.

If you try using Manticore, and it doesn't work, but you definitely have the latest Manticore installed, check the .travis.yml file. If that grabs a Manticore other than the master version, you can try using the version of Manticore we use in our CI tests. Sometimes Manticore makes a breaking change, and we are behind for a short time.

You can check your build using the test binaries that were (by default) built and emitted to deepstate/build/examples. For example, to use angr to symbolically execute the IntegerOverflow test harness with 4 workers, saving generated test cases in a directory called out, you would invoke:

deepstate-angr --num_workers 4 --output_test_dir out $DEEPSTATE/build/examples/IntegerOverflow

The resulting out directory should look something like:

out
└── IntegerOverflow.cpp
   ├── SignedInteger_AdditionOverflow
   │   ├── a512f8ffb2c1bb775a9779ec60b699cb.fail
   │   └── f1d3ff8443297732862df21dc4e57262.pass
   └── SignedInteger_MultiplicationOverflow
       ├── 6a1a90442b4d898cb3fac2800fef5baf.fail
       └── f1d3ff8443297732862df21dc4e57262.pass

To run these tests, you can just use the native executable, e.g.:

$DEEPSTATE/build/examples/IntegerOverflow --input_test_dir out

to run all the generated tests, or

$DEEPSTATE/build/examples/IntegerOverflow --input_test_files_dir out/IntegerOverflow.cpp/SignedInteger_AdditionOverflow --input_which_test SignedInteger_AdditionOverflow

to run the tests in one directory (in this case, you want to specify which test to run, also). You can also run a single test, e.g.:

$DEEPSTATE/build/examples/IntegerOverflow --input_test_file out/IntegerOverflow.cpp/SignedInteger_AdditionOverflow/a512f8ffb2c1bb775a9779ec60b699cb.fail--input_which_test SignedInteger_AdditionOverflow

In the absence of an --input_which_test argument, DeepState defaults to the last-defined test. Run the native executable with the --help argument to see all DeepState options.

If you want to use DeepState in C/C++ code, you will likely want to run sudo make install from the $DEEPSTATE/build directory as well. The examples mentioned below (file system, databases) assume this has already been done.

Usage

DeepState consists of a static library, used to write test harnesses, and command-line executors written in Python. At this time, the best documentation is in the examples and in our paper. A more extensive example, using DeepState and libFuzzer to test a user-mode file system, is available here; in particular the Tests.cpp file and CMakeLists.txt show DeepState usage. Another extensive example is a differential tester that compares Google's leveldb and Facebook's rocksdb.

Example Code

#include <deepstate/DeepState.hpp>

using namespace deepstate;

/* Simple, buggy, run-length encoding that creates "human readable"
  * encodings by adding 'A'-1 to the count, and splitting at 26.
  * e.g., encode("aaabbbbbc") = "aCbEcA" since C=3 and E=5 */

char* encode(const char* input) {
  unsigned int len = strlen(input);
  char* encoded = (char*)malloc((len*2)+1);
  int pos = 0;
  if (len > 0) {
    unsigned char last = input[0];
    int count = 1;
    for (int i = 1; i < len; i++) {
      if (((unsigned char)input[i] == last) && (count < 26))
	count++;
      else {
	encoded[pos++] = last;
	encoded[pos++] = 64 + count;
	last = (unsigned char)input[i];
	count = 1;
      }
    }
    encoded[pos++] = last;
    encoded[pos++] = 65; // Should be 64 + count
  }
  encoded[pos] = '\0';
  return encoded;
}

char* decode(const char* output) {
  unsigned int len = strlen(output);
  char* decoded = (char*)malloc((len/2)*26);
  int pos = 0;
  for (int i = 0; i < len; i += 2) {
    for (int j = 0; j < (output[i+1] - 64); j++) {
      decoded[pos++] = output[i];
    }
  }
  decoded[pos] = '\0';
  return decoded;
}

// Can be (much) higher (e.g., > 1024) if we're using fuzzing, not symbolic execution
#define MAX_STR_LEN 6

TEST(Runlength, BoringUnitTest) {
  ASSERT_EQ(strcmp(encode(""), ""), 0);
  ASSERT_EQ(strcmp(encode("a"), "aA"), 0);
  ASSERT_EQ(strcmp(encode("aaabbbbbc"), "aCbEcA"), 0);
}

TEST(Runlength, EncodeDecode) {
  char* original = DeepState_CStrUpToLen(MAX_STR_LEN, "abcdef0123456789");
  char* encoded = encode(original);
  ASSERT_LE(strlen(encoded), strlen(original)*2) << "Encoding is > length*2!";
  char* roundtrip = decode(encoded);
  ASSERT_EQ(strncmp(roundtrip, original, MAX_STR_LEN), 0) <<
    "ORIGINAL: '" << original << "', ENCODED: '" << encoded <<
    "', ROUNDTRIP: '" << roundtrip << "'";
}

The code above (which can be found here) shows an example of a DeepState test harness. Most of the code is just the functions to be tested. Using DeepState to test them requires:

• Including the DeepState C++ header and using the DeepState namespace

• Defining at least one TEST, with names

• Calling some DeepState APIs that produce data

  • In this example, we see the DeepState_CStrUpToLen call tells DeepState to produce a string that has up to MAX_STR_LEN characters, chosen from those present in hex strings.

• Optionally making some assertions about the correctness of the results

  • In Runlen.cpp this is the ASSERT_LE and ASSERT_EQ checks.
  • In the absence of any properties to check, DeepState can still look for memory safety violations, crashes, and other general categories of undesirable behavior, like any fuzzer.

DeepState will also run the "BoringUnitTest," but it (like a traditional hand-written unit test) is simply a test of fixed inputs devised by a programmer. These inputs do not expose the bug in encode. Nor do the default values (all zero bytes) for the DeepState test:

~/deepstate/build/examples$ ./Runlen
TRACE: Running: Runlength_EncodeDecode from /Users/alex/deepstate/examples/Runlen.cpp(55)
TRACE: Passed: Runlength_EncodeDecode
TRACE: Running: Runlength_BoringUnitTest from /Users/alex/deepstate/examples/Runlen.cpp(49)
TRACE: Passed: Runlength_BoringUnitTest

Using DeepState, however, it is easy to find the bug. Just go into the $DEEPSTATE/build/examples directory and try:

deepstate-angr ./Runlen

or

./Runlen --fuzz --exit_on_fail

The fuzzer will output something like:

INFO: Starting fuzzing
WARNING: No seed provided; using 1546631311
WARNING: No test specified, defaulting to last test defined (Runlength_EncodeDecode)
CRITICAL: /Users/alex/deepstate/examples/Runlen.cpp(60): ORIGINAL: '91c499', ENCODED: '9A1AcA4A9A', ROUNDTRIP: '91c49'
ERROR: Failed: Runlength_EncodeDecode

Log Levels

By default, DeepState is not very verbose about testing activity, other than failing tests. The --log_level argument lowers the threshold for output, with 0 = DEBUG, 1 = TRACE (output from the tests, including from printf), 2 = INFO (DeepState messages, the default), 3 = WARNING, 4 = ERROR, 5 = EXTERNAL (output from other programs such as libFuzzer), and 6 = CRITICAL messages. Lowering the log_level can be very useful for understanding what a DeepState harness is actually doing; often, setting --log_level 1 in either fuzzing or symbolic execution will give sufficient information to debug your test harness.

Built-In Fuzzer

Every DeepState executable provides a simple built-in fuzzer that generates tests using completely random data. Using this fuzzer is as simple as calling the native executable with the --fuzz argument. The fuzzer also takes a seed and timeout (default of two minutes) to control the fuzzing. By default fuzzing saves only failing and crashing tests, and these only when given an output directory. If you want to actually save the test cases generated, you need to add a --output_test_dir argument to tell DeepState where to put the generated tests, and if you want the (totally random and unlikely to be high-quality) passing tests, you need to add --fuzz_save_passing.

Note that while symbolic execution only works on Linux, without a fairly complex cross-compilation process, the brute force fuzzer works on macOS or (as far as we know) any Unix-like system.

A Note on MacOS and Forking

Normally, when running a test for replay or fuzzing, DeepState forks in order to cleanly handle crashes of a test. Unfortunately, fork() on macOS is extremely slow. When using the built-in fuzzer or replaying more than a few tests, it is highly recommended to add the --no_fork option on macOS, unless you need the added crash handling (that is, only when things aren't working without that option).

Fuzzing with libFuzzer

If you install clang 6.0 or later, and run cmake when you install with the BUILD_LIBFUZZER environment variable defined, you can generate tests using libFuzzer. Because both DeepState and libFuzzer want to be main, this requires building a different executable for libFuzzer. The examples directory shows how this can be done. The libFuzzer executable works like any other libFuzzer executable, and the tests produced can be replayed using the normal DeepState executable. For example, generating some tests of the OneOf example (up to 5,000 runs), then running those tests to examine the results, would look like:

mkdir OneOf_libFuzzer_corpus
./OneOf_LF -runs=5000 OneOf_libFuzzer_corpus
./OneOf --input_test_files_dir OneOf_libFuzzer_corpus

Use the LIBFUZZER_WHICH_TEST environment variable to control which test libFuzzer runs, using a fully qualified name (e.g., Arithmetic_InvertibleMultiplication_CanFail). By default, you get the last test defined (which works fine if there is only one test). Obviously, libFuzzer may work better if you provide a non-empty corpus, but fuzzing will work even without an initial corpus, unlike AFL.

One hint when using libFuzzer is to avoid dynamically allocating memory during a test, if that memory would not be freed on a test failure. This will leak memory and libFuzzer will run out of memory very quickly in each fuzzing session. Using libFuzzer on macOS requires compiling DeepState and your program with a clang that supports libFuzzer (which the Apple built-in probably won't); this can be as simple as doing:

brew install llvm@7
CC=/usr/local/opt/llvm\@7/bin/clang CXX=/usr/local/opt/llvm\@7/bin/clang++ BUILD_LIBFUZZER=TRUE cmake ..
make install

Other ways of getting an appropriate LLVM may also work.

On macOS, libFuzzer's normal output is not visible. Because libFuzzer does not fork to execute tests, there is no issue with fork speed on macOS for this kind of fuzzing.

On any platform, you can see more about what DeepState under libFuzzer is doing by setting the LIBFUZZER_LOUD environment variable, and tell libFuzzer to stop upon finding a failing test using LIBFUZZER_EXIT_ON_FAIL.

Test case reduction

While tests generated by symbolic execution are likely to be highly concise already, fuzzer-generated tests may be much larger than they need to be.

DeepState provides a test case reducer to shrink tests intelligently, aware of the structure of a DeepState test. For example, if your executable is named TestFileSystem and the test you want to reduce is named rmdirfail.test you would use it like this:

deepstate-reduce ./TestFileSystem rmdirfail.test minrmdirfail.test

In many cases, this will result in finding a different failure or crash that allows smaller test cases, so you can also provide a string that controls the criteria for which test outputs are considered valid reductions (by default, the reducer looks for any test that fails or crashes). Only outputs containing the --criteria are considered to be valid reductions:

deepstate-reduce ./TestFileSystem rmdirfail.test minrmdirfail.test --criteria "FATAL: /root/testfs/super.c(252)"

The output will look something like:

ORIGINAL TEST HAS 119 BYTES
ONEOF REMOVAL REDUCED TEST TO 103 BYTES
ONEOF REMOVAL REDUCED TEST TO 87 BYTES
ONEOF REMOVAL REDUCED TEST TO 67 BYTES
ONEOF REMOVAL REDUCED TEST TO 51 BYTES
BYTE RANGE REMOVAL REDUCED TEST TO 50 BYTES
BYTE RANGE REMOVAL REDUCED TEST TO 49 BYTES
BYTE REDUCTION: BYTE 3 FROM 4 TO 0
BYTE REDUCTION: BYTE 43 FROM 4 TO 0
ONEOF REMOVAL REDUCED TEST TO 33 BYTES
ONEOF REMOVAL REDUCED TEST TO 17 BYTES
BYTE REDUCTION: BYTE 7 FROM 2 TO 1
BYTE REDUCTION: BYTE 15 FROM 2 TO 1
NO REDUCTIONS FOUND
PADDING TEST WITH 3 ZEROS

WRITING REDUCED TEST WITH 20 BYTES TO minrmdirfail.test

You can use --which_test <testname> to specify which test to run, as with the --input_which_test options to test replay.

Test case reduction should work on any OS.

Fuzzing with AFL

DeepState can also be used with a file-based fuzzer (e.g. AFL). There are a few steps to this. First, compile DeepState itself with any needed instrumentation. E.g., to use it with AFL, you will want to set the compilers to afl-gcc and afl-g++ or afl-clang and afl-clang++ when you cmake on your DeepState install:

CC=afl-clang CXX=afl-clang++ cmake ..

Since you may want to use other fuzzers, you might at this point want to do something like:

cp /usr/local/lib/libdeepstate.a /usr/local/lib/libdeepstate_AFL.a

You can then recompile DeepState with a "normal" compiler and use -ldeepstate_AFL when working with AFL.

In either case, compile the DeepState test harness and any code it links to you want instrumented with the same AFL compiler, and link to an AFL-generated version of DeepState. Finally, run the fuzzing via the interface to replay test files. For example, to fuzz the OneOf example, if we were in the deepstate/build/examples directory, you would do something like:

afl-fuzz -d -i corpus -o afl_OneOf -- ./OneOf --input_test_file @@ --abort_on_fail--no_fork

where corpus contains at least one file to start fuzzing from. The file needs to be smaller than the DeepState input size limit, but has few other limitations (for AFL it should also not cause test failure). The abort_on_fail flag makes DeepState crashes and failed tests appear as crashes to the fuzzer. There's no reason to run AFL tests with a fork for better crash reporting, so --no_fork avoids an extra fork.

To replay the tests from AFL:

./OneOf --input_test_files_dir afl_OneOf/crashes
./OneOf --input_test_files_dir afl_OneOf/queue

Finally, if an example has more than one test, you need to specify, with a fully qualified name (e.g., Arithmetic_InvertibleMultiplication_CanFail), which test to run, using the --input_which_test flag to the binary. By default, DeepState will run the last test defined.

Because AFL and other file-based fuzzers only rely on the DeepState native test executable, they should (like DeepState's built-in simple fuzzer) work fine on macOS and other Unix-like OSes. On macOS, you will want to consider doing the work to use persistent mode, or even running inside a VM, due to AFL (unless in persistent mode) relying extensively on forks, which are very slow on macOS.

Fuzzing with Eclipser

Eclipser is a powerful new fuzzer/grey-box concolic tool with some of the advantages of symbolic execution, but with more scalability. DeepState supports Eclipser out of the box. To use it, you just need to

• Install Eclipser as instructed at https://github.com/SoftSec-KAIST/Eclipser (you'll need to be on Linux)
• Set the ECLIPSER_HOME environment variable to where-ever you installed Eclipser (the root, above build)
• Make sure you compile your DeepState native without any sanitizers (QEMU, used by Eclipser, doesn't like them)

After that, you can use Eclipser like this:

deepstate-eclisper <binary> --timeout <how long to test> --output_test_dir <where to put generated tests>

In our experience, Eclipser is quite effective, often better than libFuzzer and sometimes better than AFL, despite having a much slower test throughput than either.

Which Fuzzer Should I Use?

In fact, since DeepState supports libFuzzer, AFL, and Eclipser (and others), a natural question is "which is the best fuzzer?" In general, it depends! We suggest using them all, which DeepState makes easy. libFuzzer is very fast, and sometimes the CMP breakdown it provides is very useful; however, it's often bad at finding longer paths where just covering nodes isn't helpful. AFL is still an excellent general-purpose fuzzer, and often beats "improved" versions over a range of programs. Finally, Eclipser has some tricks that let it get traction in some cases where you might think only symbolic execution (which wouldn't scale) could help.

Contributing

All accepted PRs are awarded bounties by Trail of Bits. Join the #deepstate channel on the Empire Hacking Slack to discuss ongoing development and claim bounties. Check the good first issue label for suggested contributions.

License

DeepState is released under The Apache License 2.0.