Tutorial Two
Testing a Simple Regular Expression Library
This is an example of using KLEE to test a simple regular expression matching function. You can find the basic example in the source tree under examples/regexp
.
Regexp.c
contains a simple regular expression matching function, and the bare bones testing harness (in main()
) needed to explore this code with KLEE. You can see a version of the source code here.
This example shows how to build and run the example using KLEE, as well as how to interpret the output, and some additional KLEE features that can be used when writing a test driver by hand.
We’ll start by showing how to build and run the example, and then explain how the test harness works in more detail.
Building the example
The first step is to compile the source code using a compiler which can generate object files in LLVM bitcode format.
From within the examples/regexp
directory execute:
which should create a Regexp.bc file in LLVM bitcode format. The -I
argument
is used so that the compiler can find
klee/klee.h
,
which contains definitions for the intrinsic functions used to interact with
the KLEE virtual machine. -c
is used because we only want to compile the code
to an object file (not to a native executable), and finally -g
causes additional
debug information to be stored in the object file, which KLEE will use to
determine source line number information. -O0 -Xclang -disable-O0-optnone
is
used to compile without any optimisation, but without preventing KLEE from
performing its own optimisations, which compiling with -O0
would.
If you have the LLVM tools installed in your path, you can verify that this step worked by running llvm-nm on the generated file:
$ llvm-nm Regexp.bc
U klee_make_symbolic
---------------- T main
---------------- T match
---------------- t matchhere
---------------- t matchstar
Normally, before running this program, we would need to link it to create a native executable. However, KLEE runs directly on LLVM bitcode files. Since this program only has a single file there is no need for linking. For “real” programs with multiple inputs, the llvm-link tools can be used in place of the regular link step to merge multiple LLVM bitcode files into a single module which can be executed by KLEE.
Executing the code with KLEE
The next step is to execute the code with KLEE (number of instructions varies between LLVM versions and optimisation levels):
$ klee --only-output-states-covering-new Regexp.bc
KLEE: output directory = "klee-out-0"
KLEE: ERROR: Regexp.c:23: memory error: out of bound pointer
KLEE: NOTE: now ignoring this error at this location
KLEE: ERROR: Regexp.c:25: memory error: out of bound pointer
KLEE: NOTE: now ignoring this error at this location
KLEE: done: total instructions = 4848112
KLEE: done: completed paths = 6675
KLEE: done: partially completed paths = 763
KLEE: done: generated tests = 16
On startup, KLEE prints the directory used to store output (in this case klee-out-0
). By default KLEE will use the first free klee-out-N
directory and also create a klee-last
symlink which will point to the most recent created directory. You can specify a directory to use for outputs using the -output-dir=path
command line argument.
While KLEE is running, it will print status messages for “important” events, for example when it finds an error in the program. In this case, KLEE detected two invalid memory accesses on lines 23 and 25 of our test program. We’ll look closer at this in a moment.
Finally, when KLEE finishes execution it prints out a few statistics about the run. Here we see that KLEE executed a total of ~4.8 million instructions, explored 7,438 paths, and generated 16 test cases. KLEE only generates 16 test cases because we limited the test generation to states that actually covered new code with --only-output-states-covering-new
. If we would omit this flag, KLEE would create 6,677 test cases! Still, KLEE does not create a test case for every path. Whenever it finds a bug, it creates a test case for the first state that reaches the bug. All other paths that reach the bug at the same location are terminated silently and reported as partially completed paths. If you don’t mind the duplication of error cases, use --emit-all-errors
to generate test cases for all 7,438 paths.
Note that many realistic programs have an infinite (or extremely large) number of paths through them, and it is common that KLEE will not terminate. By default KLEE will run until the user presses Control-C (i.e. klee gets a SIGINT
), but there are additional options to limit KLEE’s runtime and memory usage:
-max-time=<time span>
: Halt execution after the given amount of time, e.g.10min
or1h5s
.-max-forks=N
: Stop forking afterN
symbolic branches, and run the remaining paths to termination.-max-memory=N
: Try to limit memory consumption toN
megabytes.
KLEE error reports
When KLEE detects an error in the program being executed it will generate a test case which exhibits the error, and write some additional information about the error into a file testN.TYPE.err
, where N
is the test case number, and TYPE
identifies the kind of error. Some types of errors KLEE detects include:
- ptr: Stores or loads of invalid memory locations.
- free: Double or invalid
free()
. - abort: The program called
abort()
. - assert: An assertion failed.
- div: A division or modulus by zero was detected.
- user: There is a problem with the input (invalid klee intrinsic calls) or the way KLEE is being used.
- exec: There was a problem which prevented KLEE from executing the program; for example an unknown instruction, a call to an invalid function pointer, or inline assembly.
- model: KLEE was unable to keep full precision and is only exploring parts of the program state. For example, symbolic sizes to malloc are not currently supported, in such cases KLEE will concretize the argument.
KLEE will print a message to the console when it detects an error, in the test run above we can see that KLEE detected two memory errors. For all program errors, KLEE will write a simple backtrace into the .err
file. This is what one of the errors above looks like:
Each line of the backtrace lists the frame number, the instruction line (this is the line number in the assembly.ll
file found along with the run output), the function and arguments (including values for the concrete parameters), and the source information.
Particular error reports may also include additional information. For memory errors, KLEE will show the invalid address, and what objects are on the heap both before and after that address. In this case, we can see that the address happens to be exactly one byte past the end of the previous object.
Changing the test harness
The reason KLEE is finding memory errors in this program isn’t because the regular expression functions have a bug, rather it indicates a problem in our test driver. The problem is that we are making the input regular expression buffer completely symbolic, but the match function expects it to be a null terminated string. Let’s look at two ways we can fix this.
The simplest way to fix this problem is to store ‘\0’ at the end of the buffer, after making it symbolic. This makes our driver look like this:
Making a buffer symbolic just initializes the contents to refer to symbolic variables, we are still free to modify the memory as we wish. If you recompile and run klee on this test program, the memory errors should now be gone.
Another way to accomplish the same effect is to use the klee_assume
intrinsic function. klee_assume
takes a single argument (an unsigned integer) which generally should be some kind of conditional expression, and “assumes” that expression to be true on the current path (if that can never happen, i.e. the expression is provably false, KLEE will report an error).
We can use klee_assume
to cause KLEE to only explore states where the string is null terminated by writing the driver like this:
In this particular example both solutions work fine but in general klee_assume
is more flexible:
- By explicitly declaring the constraint, this will force test cases to have the
'\0'
in them. In the first example where we write the terminating null explicitly, it doesn’t matter what the last byte of the symbolic input is and KLEE is free to generate any value. In some cases where you want to inspect the test cases by hand, it is more convenient for the test case to show all the values that matter. klee_assume
can be used to encode more complicated constraints. For example, we could useklee_assume(re[0] != '^')
to cause KLEE to only explore states where the first byte is not'^'
.
NOTE: There is one important caveat when using klee_assume
with multiple conditions. Remember that boolean conditionals like '&&'
and '||'
may be compiled into code which branches before computing the result of the expression. In such situations KLEE will branch the process before it reaches the call to klee_assume
, which may result in exploring unnecessary additional states. For this reason it is good to use as simple expressions as possible to klee_assume
(for example splitting a single call into multiple ones), and to use the '&'
and '|'
operators instead of the short-circuiting ones.