xref: /art/tools/dexfuzz/README

DexFuzz
=======

DexFuzz is primarily a tool for fuzzing DEX files. Fuzzing is the introduction of
subtle changes ("mutations") to a file to produce a new test case. These test cases
can be used to test the various modes of execution available to ART (Interpreter,
Optimizing compiler) to check for bugs in these modes of execution.
This is done by differential testing - each test file is executed with each mode of
execution, and any differences between the resulting outputs may be an indication of
a bug in one of the modes.

For a wider overview of DexFuzz, see:

http://community.arm.com/groups/android-community/blog/2014/11/26/the-art-of-fuzz-testing

In typical operation, you provide DexFuzz with a set of DEX files that are the "seeds"
for mutation - e.g. some tests taken from the ART test suite - and point it at an
ADB-connected Android device, and it will fuzz these seed files, and execute the
resulting new tests on the Android device.

How to run DexFuzz
==================

DexFuzz can run its test programs on either an ADB-connected device, or a host-build of
ART locally.

Execution on an ADB-connected device
------------------------------------

1. Build dexfuzz with mmma tools/dexfuzz from within art/.
2. Make sure you have an Android device connected via ADB, that is capable of
   having DEX files pushed to it and executed with the dalvikvm command.
3. Make sure you're in the Android build environment!
   (That is, . build/envsetup.sh && lunch)
4. Create a new directory, and place some DEX files in here. These are the seed files
   that are mutated to form new tests.
5. Create a directory on your device that mutated test files can be pushed to and
   executed in, using dalvikvm. For example, /data/art-test/
6. If you currently have multiple devices connected via ADB, find out the name of
   your device using "adb devices -l".
7. Run this command:

dexfuzz --inputs=<seeds dir> --execute --repeat=<attempts> \
    --dump-output <combination of ISA(s) and and backend(s)>

You MUST specify one of the following ISAs:
  --arm
  --arm64
  --x86
  --x86_64

And also at least two of the following backends:
  --interpreter
  --optimizing

Note that if you wanted to test both ARM and ARM64 on an ARM64 device, you can use
--allarm. Also in this case only one backend is needed, if i.e., you wanted to test
ARM Optimizing Backend vs. ARM64 Optimizing Backend.

Some legal examples:
  --arm --optimizing --interpreter
  --x86 --optimizing --interpreter
  --allarm --optimizing

Add in --device=<device name, e.g. device:generic> if you want to specify a device.
Add in --execute-dir=<dir on device> if you want to specify an execution directory.
  (The default is /data/art-test/)

Host Execution
--------------

DexFuzz now supports execution on your host machine.
Follow steps 1, 3, 4, and 7 as above, but also observe the following:
 - instead of specifying an ISA, use --host
 - ANDROID_DATA must be set, pointing to a location where dex2oat will place
   OAT files after compilation.
 - Files will always be executed in the same directory where you are executing DexFuzz.

Fuzzer Operation
----------------

As the fuzzer works, you'll see output like:

|-----------------------------------------------------------------|
|Iterations|VerifyFail|MutateFail|Timed Out |Successful|Divergence|
|-----------------------------------------------------------------|
| 48       | 37       | 4        | 0        | 6        | 1        |

Iterations - number of attempts we've made to mutate DEX files.
VerifyFail - the number of mutated files that ended up failing to verify, either
             on the host, or the target.
MutateFail - because mutation is a random process, and has attempt thresholds to
             avoid attempting to mutate a file indefinitely, it is possible that
             an attempt to mutate a file doesn't actually mutate it. This counts
             those occurrences.
Timed Out  - mutated files that timed out for one or more backends.
             Current timeouts are:
               Optimizing - 5 seconds
               Interpreter - 30 seconds
              (use --short-timeouts to set all backends to 2 seconds.)
Successful - mutated files that executed and all backends agreed on the resulting
             output. NB: if all backends crashed with the same output, this would
             be considered a success - proper detection of crashes is still to come.
Divergence - mutated files that executed and some backend disagreed about the
             resulting output. Divergent programs are run multiple times with a
             single backend, to check if they diverge from themselves, and these are
             not included in the count. If multiple architectures are being used
             (ARM/ARM64), and the divergences align with different architectures,
             these are also not included in the count.

8. Check report.log for the full report, including input file and RNG seed for each
   test program. This allows you to recreate a bad program with, e.g.:

dexfuzz --input=<input file> --seed=<seed value>

Check dexfuzz --help for the full list of options.

NOTE: DEX files with unicode strings are not fully supported yet, and DEX files with
JNI elements are not supported at all currently.

Mutation Likelihoods
====================

Each bytecode mutation has a chance out of 100% of firing. Following is the listing
of each mutation's probability. If you wish to easily adjust these values, copy
these values into a file called likelihoods.txt, and run dexfuzz with
--likelihoods=likelihoods.txt.

ArithOpChanger 75
BranchShifter 30
CmpBiasChanger 30
ConstantValueChanger 70
ConversionRepeater 50
FieldFlagChanger 40
InstructionDeleter 40
InstructionDuplicator 80
InstructionSwapper 80
InvokeChanger 30
NewArrayLengthChanger 50
NewInstanceChanger 10
NewMethodCaller 10
NonsenseStringPrinter 10
OppositeBranchChanger 40
PoolIndexChanger 30
RandomBranchChanger 30
RandomInstructionGenerator 30
RegisterClobber 10
SwitchBranchShifter 30
TryBlockShifter 40
ValuePrinter 40
VRegChanger 60