Documentation for running and understanding the implementation of mutation-guided fuzzing provided here.
For more questions, feel free to email Bella Laybourn, Vasu Vikram, Rafaello Sanna, or Rohan Padhye.
This repository works together with the sort-benchmarks
repository . The current dependency structure is as follows:
sort-benchmarks --> jqf-fuzz (for API only)
mu2 --> jqf-fuzz
mu2 --> sort-benchmarks (for testing only)
git clone https://github.com/cmu-pasta/mu2 && cd mu2
mvn install -DskipTests
cd ..
git clone https://github.com/cmu-pasta/sort-benchmarks && cd sort-benchmarks
mvn install
cd ..
Install mu2 with:
cd mu2
mvn install
cd ..
If you don't install the sort-benchmarks
, you can install mu2
via mvn install -DskipTests
. This is not recommended.
When making changes to mu2 or JQF, you only need to install that particular project's SNAPSHOT for the changes to take effect. You do not have to do the long bootstrap build as above during regular development.
To test changes to mu2
, run the integration tests as follows from the mu2
repository:
mvn verify
Coverage reports should be available in target/site/jacoco-integration-test-coverage-report/index.html
.
To use the differential fuzzing goal mu2:diff
, instrument your tests with @Diff
and @Comparison
annotations instead of @Test
and use the DiffGoal
and MutateDiffGoal
as described below.
In the differential testing framework, each test is a pair of functions, one annotated with @Diff
and the other with @Comparison
.
The function annotated with @Diff
should not be static
and should return an object (neither void nor a primitive, though boxed primitives work).
The result of this diff function will be compared with other results using the comparison function (annotated with @Comparison
).
The comparison function should be static
, take two arguments of the same type as the return type of any associated diff functions, and return a Boolean
(Boolean.TRUE
if the two inputs are equivalent, Boolean.FALSE
otherwise).
Multiple functions with these annotations can be included in a single test class, so, to uniquely match the diff function with a comparison function,
pass the name of the comparison function as an argument to @Diff
as cmp
(see below for example). If no argument is passed to cmp
, Objects.equals()
will be used as a default.
Each diff function is associated with up to one comparison function, but arbitrarily many diff functions may use a comparison function.
Example of a differential test for a function populateList(List<Integer>)
when the comparison only cares about the size of the populated list:
@Diff(cmp = "compare")
public List<Integer> testPopulate(List<Integer> input) {
return populateList(input);
}
@Comparison
public static Boolean compare(List<Integer> list1, List<Integer> list2) {
return list1.size() == list2.size();
}
For running differential fuzzing. Currently defaults to mutation-guided fuzzing.
Example:
mvn mu2:diff -Dclass=package.class -Dmethod=method -Dincludes=prefix
For repro on mutation guidance with diff-based tests.
Example:
mvn mu2:mutate -Dclass=package.class -Dmethod=method -Dincludes=prefix -Dinput=path/to/corpus
For repro on other guidances with diff-based tests.
Example:
mvn mu2:repro -Dclass=package.class -Dmethod=method -Dincludes=prefix -Dinput=path/to/corpus
You can use the -Dincludes
flags to select which code will be mutated.
mvn mu2:diff -Dclass=package.class -Dmethod=method -Dincludes=prefix1,prefix2
Additionally, setting the -DtargetIncludes
flag to a comma-separated list that includes every mutated class
as well as all classes that depend on a mutable class anywhere in their dependency trees improves efficiency by
loading classes not in the list with a separate intermediate ClassLoader.
Example using a test of Scala's chess library:
mvn mu2:diff -Dclass=edu.berkeley.cs.jqf.examples.chess.FENTest -Dmethod=testWithGenerator -Dincludes=chess.format.Forsyth,chess.Situation -DtargetIncludes=edu.berkeley.cs.jqf.examples.chess,chess
A MutationInstance
is an instance of a mutation: e.g. the class in which the mutation resides, the kind of mutation performed (the Mutator
- see below), and the location in which the mutation takes places (given as the n'th opportunity for the Mutator
to be applied to the given class). Each instance also has an ID, which can be used as its index in the MutationInstance.mutationInstances
array.
A ClassLoader
which loads a mutated version of a class specified by a MutationInstance
. For all classes except that specified by the MutationInstance
, it loads it normally. For the class specified, it mutates the class as specified by the Mutator
and the mutator index.
This class can be used on its own, or can be used through a MutationClassLoaders
object, which takes the class-loading information as a parameter then provides a map from MutationInstances
to MutationClassLoader
s.
Initial class loading should take place using this ClassLoader. It creates a list of MutationInstances representing all of the mutation opportunities (the "cartograph"). It does this by instrumenting the input class using the Cartographer
: a SafeClassWriter
which both logs which mutants are possible, and transforms the input class into one which measures which mutants are run for later optimization.
A class that contains references to both the CartographyClassLoader
and a mapping of each MutationInstance
to its MutationClassLoader
. A MutationClassLoaders
object is the external entry point, and such an object is passed to a MutationGuidance
during construction.
Each mutator replaces one instruction with another, or with a series of instructions.
They sometimes take the form prefix_oldInstruction_TO_newInstruction(Opcodes.oldInstruction, returnType, [list, of, instruction, calls, ...])
and other times, as appropriate to the type of mutator, are just descriptors.
To make that more clear, three actual examples, broken down:
I_ADD_TO_SUB(Opcodes.IADD, null, new InstructionCall(Opcodes.ISUB))
refers to swapping Integer ADDition for SUBtraction, which requires no relevant return type.
VOID_REMOVE_STATIC(Opcodes.INVOKESTATIC, "V", new InstructionCall(Opcodes.NOP))
refers to removing calls to void functions made with invokestatic
, which requires that the return type of the called function be Void.
S_IRETURN_TO_0(Opcodes.IRETURN, "S", new InstructionCall(Opcodes.POP), new InstructionCall(Opcodes.ICONST_0), new InstructionCall(Opcodes.IRETURN))
refers to making a function with return type Short return the short 0
instead of what it would normally have returned. This requires that the return type of the returning function be Short.
Note how in the second two examples, the opcode alone was not enough to identify the relevant return type, which is why that information was included separately.
When calling functions, the arguments are loaded onto the stack. This means that, when removing a function call, those arguments must be popped off the stack before continuing.
Typically, we would pop off the same number of arguments as the function has, but when the opcode is invokestatic
there's one fewer because there's an implicit this
argument added in other cases that invokestatic
doesn't need.
Just a wrapper for bytecode instructions so their arguments can be included.
A class whose static methods are invoked from mu2-instrumented classes. Used to track infinite loops in mutated code and to perform speculative mutation analysis in the original (cartography) test run.
For the most part, an extension of Zest
.
For each fuzz input, runs first the class loaded by the CartographyClassLoader
, then each of the MutationInstance
s it has generated that hasn't been killed by a previous fuzz input and have been marked as run by the class loaded by the CartographyClassLoader
.
Saves for reason +mutants
along with Zest's +coverage
, etc.
To prevent hanging forever on an infinite loop, both MutationInstances and CartographyClassLoaders outfit their classes with a timeout functionality that essentially kills the program after some maximum number of control jumps. This assumes that if the number of control jumps exceeds that then the program has encountered an infinite loop and will not ascertain any new information by continuing to run.
pip install -r requirements.txt
This is a script for visualizing the various mutant numbers across trials of a given fuzzing session run with -Dengine=mutation
. It takes in an input plot_data
CSV file and
outputs an image plotting found mutants, killed mutants, and optionally seen/infected mutants.
Usage:
python scripts/plot_mutant_data.py <plot_data_file> <output_image_file>
Usage:
./getMutants.sh <testClass> <testMethod> <includesClass> <experiments> <trials> <baseDirectory>
./getMutants.sh edu.berkeley.cs.jqf.examples.commons.PatriciaTrieTest testPrefixMap org.apache.commons.collections4.trie 3 1000 ../../jqf/examples
Venn diagram script alone:
python scripts/venn.py --filters_dir <filters_dir> --num_experiments <num_experiments> --output_img <output_img_name>