cipherdyne.org

Michael Rash, Security Researcher



afl-cov - Code coverage for AFL fuzzing test cases

Quick Links

Introduction

afl-cov uses test case files produced by the AFL fuzzer to produce gcov code coverage results of the targeted binary. Code coverage is interpreted from one case to the next by afl-cov in order to determine which new functions and lines are hit by AFL with each new test case. Further, afl-cov allows for specific lines or functions to be searched for within coverage results, and when a match is found the corresponding test case file is displayed. This allows the user to discover which AFL test case is the first to exercise a particular function. In addition, afl-cov produces a "zero coverage" report of functions and lines that were never executed during an AFL fuzzing run.

Although of no use to AFL itself, the main application of afl-cov is to wrap some automation around gcov and thereby provide data on how to maximize code coverage with AFL fuzzing runs. Manual interpretation of cumulative gcov results from AFL test cases is usually still required, but the "fiddly" steps of iterating over all test cases and generating code coverage reports (along with the "zero coverage" report) is automated by afl-cov.

Producing code coverage data for AFL test cases is an important step to try and maximize code coverage, and thereby help to maximize the effectiveness of AFL. For example, some binaries have code that is reachable only after a complicated (or even cryptographic) test is passed, and AFL may not be able to exercise this code without taking special measures. These measures commonly include patching the project code to bypass such tests. (For example, there is a patch to solve this problem for a CRC test in libpng included in the AFL sources at experimental/libpng_no_checksum/libpng-nocrc.patch.) When a project implements a patch to assist AFL in reaching code that would otherwise be inaccessible, a natural question to ask is whether the patch is effective. Code coverage results can help to verify this.

Prerequisites

afl-cov requires the following software:

  • afl-fuzz
  • python
  • gcov, lcov, genhtml
Note that afl-cov can parse files created by afl-fuzz from a different system, so technically afl-fuzz does not need to be installed on the same system as afl-cov. This supports scenarios where fuzzing output is collected, say, within a git repository on one system, and coverage results are produced on a different system. However, most workflows typically focus on producing afl-cov results quickly for current fuzzing runs on the same system.

Workflow

At a high level, the general workflow for afl-cov is:

  1. Create a spare copy of the project sources compiled with gcov profiling support.
  2. Run afl-cov while afl-fuzz is building test cases.
  3. Review the cumulative code coverage results in the final web report.
Now, in more detail:

- Copy the project sources to two different directories /path/to/afl-fuzz-output/ and /path/to/project-gcov/. The first will contain the project binaries compiled for AFL fuzzing, and the second will contain the project binaries compiled for gcov profiling support. For the /path/to/project-gcov/ directory, compile the project with gcov profiling support (gcc -fprofile-arcs -ftest-coverage).

- Start up afl-cov in --live mode before also starting the afl-fuzz fuzzing cycle. The command line arguments to afl-cov must specify the path to the output directory used by afl-fuzz, and the command to execute along with associated arguments. This command and arguments should closely resemble the manner in which afl-fuzz executes the targeted binary during the fuzzing cycle. Note that if there is already an existing directory of AFL fuzzing results, then just omit the --live argument to process these results. Here is an example:
$ cd /path/to/project-gcov/
$ afl-cov -d /path/to/afl-fuzz-output/ --live --coverage-cmd \
"cat AFL_FILE | LD_LIBRARY_PATH=./lib/.libs ./bin/.libs/somebin -a -b -c" \
--code-dir .
Note the AFL_FILE string above refers to the test case file that AFL will build in the queue/ directory under /path/to/project-fuzz. Just leave this string as-is - afl-cov will automatically substitute it with each AFL queue/id:NNNNNN* in succession as is builds the code coverage reports.

Also, in the above command, this handles the case where the AFL fuzzing cycle is fuzzing the targeted binary via stdin. This explains the cat AFL_FILE | ... ./bin/.lib/somebin ... invocation. For the other style of fuzzing with AFL where a file is read from the filesystem, here is an example:
$ cd /path/to/project-gcov/
$ afl-cov -d /path/to/afl-fuzz-output/ --live --coverage-cmd \
"LD_LIBRARY_PATH=./lib/.libs ./bin/.libs/somebin -f AFL_FILE -a -b -c" \
--code-dir .
With afl-cov running, open a separate terminal/shell, and launch afl-fuzz:
$ cd /path/to/project-fuzzing/
$ LD_LIBRARY_PATH=./lib/.libs afl-fuzz -T somebin -t 1000 -i ./test-cases/ \
-o /path/to/afl-fuzz-output/ ./bin/.libs/somebin -a -b -c
The familiar AFL status screen will be displayed, and afl-cov will start generating code coverage data. AFL fuzzing status screen Here is a sample of what the afl-cov output looks like:
$ afl-cov -d /path/to/afl-fuzz-output/ --live --coverage-cmd \
"LD_LIBRARY_PATH=./lib/.libs ./bin/.libs/somebin -f AFL_FILE -a -b -c" --code-dir .
[+] Imported 184 files from: /path/to/afl-fuzz-output/queue
[+] AFL file: id:000000,orig:somestr.start (1 / 184)
    lines......: 18.6% (1122 of 6032 lines)
    functions..: 30.7% (100 of 326 functions)
    branches...: 14.0% (570 of 4065 branches)
[+] AFL file: id:000001,orig:somestr256.start (2 / 184)
    lines......: 18.7% (1127 of 6032 lines)
    functions..: 30.7% (100 of 326 functions)
    branches...: 14.1% (572 of 4065 branches)
[+] Coverage diff id:000000,orig:somestr.start id:000001,orig:somestr256.start
    Src file: /path/to/project-gcov/lib/proj_decode.c
      New 'line' coverage: 140
      New 'line' coverage: 141
      New 'line' coverage: 142
    Src file: /path/to/project-gcov/lib/proj_util.c
      New 'line' coverage: 217
      New 'line' coverage: 218
[+] AFL file: id:000002,orig:somestr384.start (3 / 184)
    lines......: 18.8% (1132 of 6032 lines)
    functions..: 30.7% (100 of 326 functions)
    branches...: 14.1% (574 of 4065 branches)
[+] Coverage diff id:000001,orig:somestr256.start id:000002,orig:somestr384.start
    Src file: /path/to/project-gcov/lib/proj_decode.c
      New 'line' coverage: 145
      New 'line' coverage: 146
      New 'line' coverage: 147
    Src file: /path/to/project-gcov/lib/proj_util.c
      New 'line' coverage: 220
      New 'line' coverage: 221
[+] AFL file: id:000003,orig:somestr.start (4 / 184)
    lines......: 18.9% (1141 of 6032 lines)
    functions..: 31.0% (101 of 326 functions)
    branches...: 14.3% (581 of 4065 branches)
[+] Coverage diff id:000002,orig:somestr384.start id:000003,orig:somestr.start
    Src file: /path/to/project-gcov/lib/proj_message.c
      New 'function' coverage: validate_cmd_msg()
      New 'line' coverage: 244
      New 'line' coverage: 247
      New 'line' coverage: 248
      New 'line' coverage: 250
      New 'line' coverage: 255
      New 'line' coverage: 262
      New 'line' coverage: 263
      New 'line' coverage: 266
.
.
.
[+] Generating lcov web report in: /path/to/afl-fuzz-output/cov/web/lcov-web-final
[+] Coverage diff id:000182,src:000000,op:havoc,rep:64 id:000184,src:000000,op:havoc,rep:4
[+] Final zero coverage report in: /path/to/afl-fuzz-output/cov/zero-cov/zero-cov-final
[+] Processed 184 / 184 files
In the last few lines above, the locations of the final web coverage and zero coverage reports are shown. The zero coverage reports contains function names that were never executed across the entire afl-fuzz run.

The code coverage results in /path/to/afl-fuzz-output/cov/web/lcov-web-final represent cumulative code coverage across all AFL test cases. This data can then be reviewed to ensure that all expected functions are indeed exercised by AFL - just point a web browser at /path/to/afl-fuzz-output/cov/web/lcov-web-final/index.html.

Other Examples

The workflow above is probably the main strategy for using afl-cov. However, additional use cases are supported such as:

1) Suppose there are a set of wrapper scripts around afl-fuzz to run fuzzing cycles against various aspects of a project. By building a set of corresponding afl-cov wrappers, and then using the --disable-coverage-init option on all but the first of these wrappers, it is possible to generate code coverage results across the entire set of afl-fuzz fuzzing runs. (By default, afl-cov resets gcov counters to zero at start time, but the --disable-coverage-init stops this behavior.) The end result is a global picture of code coverage across all invocations of afl-fuzz.

2) Specific functions can be searched for in the code coverage results, and afl-cov will return the first afl-fuzz test case where a given function is executed. This allows afl-cov to be used as a validation tool by other scripts and testing infrastructure. For example, a test case could be written around whether an important function is executed by afl-fuzz to validate a patching strategy mentioned in the introduction.

Here is an example where the first test case that executes the function validate_cmd_msg() is returned (this is after all afl-cov results have been produced in the main workflow above):

./afl-cov -d /path/to/afl-fuzz-output --func-search "validate_cmd_msg"
[+] Function 'validate_cmd_mag()' executed by: id:000002,orig:somestr384.start id:000003,orig:somestr.start

Directory Structure

afl-cov creates a few directories for coverage results within the specified afl-fuzz directory (-d). These directories are displayed below, and all are contained within the /path/to/afl-fuzz-output/cov/ directory:
  • cov/diff/ - contains new code coverage results when queue/id:NNNNNN* file causes afl-fuzz to execute new code.
  • cov/lcov/ - contains raw code coverage data produced by the lcov front-end to gcov.
  • cov/web/ - contains code coverage results in web format produced by genhtml.
  • cov/zero-cov/ - contains data for each queue/id:NNNNNN* on functions (and optionally lines) that are never executed by afl-fuzz.

Usage Information

Basic --help output appears below:
usage: afl-cov [-h] [-e COVERAGE_CMD] [-d AFL_FUZZING_DIR] [-c CODE_DIR] [-O]
           [--disable-cmd-redirection] [--disable-lcov-web]
           [--disable-coverage-init] [--coverage-include-lines] [--live]
           [--sleep SLEEP] [--lcov-web-all] [--func-search FUNC_SEARCH]
           [--line-search LINE_SEARCH] [--src-file SRC_FILE]
           [--afl-queue-id-limit AFL_QUEUE_ID_LIMIT] [-v] [-V] [-q]

optional arguments:
  -h, --help            show this help message and exit
  -e COVERAGE_CMD, --coverage-cmd COVERAGE_CMD
                        set command to exec (including args, and assumes code
                        coverage support)
  -d AFL_FUZZING_DIR, --afl-fuzzing-dir AFL_FUZZING_DIR
                        top level AFL fuzzing directory
  -c CODE_DIR, --code-dir CODE_DIR
                        directory where the code lives (compiled with code
                        coverage support)
  -O, --overwrite       overwrite existing coverage results
  --disable-cmd-redirection
                        disable redirection of command results to /dev/null
  --disable-lcov-web    disable generation of all lcov web code coverage
                        reports
  --disable-coverage-init
                        disable initialization of code coverage counters at
                        afl-cov startup
  --coverage-include-lines
                        include lines in zero-coverage status files
  --live                process a live AFL directory, and afl-cov will exit
                        when it appears afl-fuzz has been stopped
  --sleep SLEEP         In --live mode, # of seconds to sleep between checking
                        for new queue files
  --lcov-web-all        generate lcov web reports for all id:NNNNNN* files
                        instead of just the last one
  --func-search FUNC_SEARCH
                        search for coverage of a specific function
  --line-search LINE_SEARCH
                        search for coverage of a specific line number
                        (requires --src-file)
  --src-file SRC_FILE   restrict function or line search to a specific source
                        file
  --afl-queue-id-limit AFL_QUEUE_ID_LIMIT
                        limit the number of id:NNNNNN* files processed in the
                        AFL queue/ directory
  -v, --verbose         verbose mode
  -V, --version         print version and exit
  -q, --quiet           quiet mode

License

afl-cov is released as open source software under the terms of the GNU General Public License (GPL v2). The latest release can be found at github.

Contact

All feature requests and bug fixes are managed through github issues tracking. However, you can email me (michael.rash_AT_gmail.com), or reach me through Twitter @michaelrash.