Fuzzing non-x86_64 targets using libFuzzer on Azure makes use of two Open Source capabilities:
TL/DR: check out our libfuzzer-aarch64-crosscompile example
There are a few notable limitations when using qemu-user
based libFuzzer targets in OneFuzz.
- Live reproduction of crashes does not work.
- The
libfuzzer-coverage
task does not work. - Only Linux is supported at this time.
- Only
aarch64
CPU emulation has been tested. PRs are welcome to support other architectures. - Custom setup scripts are not supported, though you can provide your own sysroot using
--sysroot
.
As such, a libfuzzer qemu_user
template is available, which only uses the libfuzzer_fuzz
and libfuzzer_crash_report
. As these issues are resolve, the template will be updated to include the additional tasks.
Let's build a simple aarch64
target using GCC as a cross-compiler (See our example).
- Make sure you have QEMU and the appropriate cross compiler installed:
sudo apt update
sudo apt install -y qemu-user g++-aarch64-linux-gnu
- Check out the libFuzzer libraries from
compiler-rt
. Note, GCC requires the use ofpc-guard
for instrumentation, which was removed by thecompiler-rt
project. As such, we need an older version of the library:
git clone https://github.com/llvm-mirror/compiler-rt
(cd compiler-rt; git checkout daa6759576548a2f3825faddaa6811cabbfb45eb)
- Build the libFuzzer libraries without ASAN:
mkdir -p fuzz-libs
(cd fuzz-libs; aarch64-linux-gnu-g++ -c ../compiler-rt/lib/fuzzer/*.cpp)
- Build our target:
aarch64-linux-gnu-g++ -pthread -lasan -o fuzz.exe fuzz-libs/*.o fuzz.c -fsanitize=address -fsanitize-coverage=trace-pc
- Verify our target built correctly:
ASAN_OPTIONS=:detect_leaks=0 qemu-aarch64 -L /usr/aarch64-linux-gnu ./fuzz.exe -help=1
NOTE:
LSAN
does not work inqemu-user
, so we need to disable that.
Now we're ready to deploy this target to OneFuzz. Note, if we have custom libraries or want to run on a different version of linux, we'll need to provide our own sysroot.
- Now we can fuzz!
Execute our
fuzz.exe
withqemu-aarch64
ourinputs
directory:
ASAN_OPTIONS=:detect_leaks=0 qemu-aarch64 -L /usr/aarch64-linux-gnu ./fuzz.exe ./inputs
In a few seconds, you'll see output that looks something like this:
INFO: Seed: 113138795
INFO: 1 files found in ./inputs/ INFO: -max_len is not provided; libFuzzer will not generate inputs larger than 4096 bytes
INFO: seed corpus: files: 1 min: 3b max: 3b total: 3b rss: 314Mb
#2 INITED cov: 5 ft: 5 corp: 1/3b lim: 4 exec/s: 0 rss: 317Mb
#3 NEW cov: 5 ft: 9 corp: 2/6b lim: 4 exec/s: 0 rss: 317Mb L: 3/3 MS: 1 ChangeBit-
#4 NEW cov: 5 ft: 13 corp: 3/9b lim: 4 exec/s: 0 rss: 317Mb L: 3/3 MS: 1 ChangeBit-
#8 NEW cov: 10 ft: 21 corp: 4/13b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 4 ShuffleBytes-ChangeBit-ChangeBinInt-CrossOver-
#9 NEW cov: 10 ft: 26 corp: 5/17b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 1 CopyPart-
#10 NEW cov: 10 ft: 31 corp: 6/21b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 1 ChangeBit-
#11 NEW cov: 10 ft: 36 corp: 7/25b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 1 ShuffleBytes-
#12 NEW cov: 10 ft: 37 corp: 8/28b lim: 4 exec/s: 0 rss: 317Mb L: 3/4 MS: 1 EraseBytes-
#16 NEW cov: 10 ft: 45 corp: 9/32b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 4 CopyPart-CopyPart-CrossOver-ShuffleBytes-
#18 NEW cov: 11 ft: 46 corp: 10/36b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 2 CopyPart-ChangeBit-
#19 NEW cov: 11 ft: 47 corp: 11/40b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 1 ShuffleBytes-
#20 REDUCE cov: 11 ft: 47 corp: 11/39b lim: 4 exec/s: 0 rss: 317Mb L: 2/4 MS: 1 EraseBytes-
#30 NEW cov: 11 ft: 52 corp: 12/43b lim: 4 exec/s: 0 rss: 317Mb L: 4/4 MS: 5 ChangeBit-CrossOver-CrossOver-CrossOver-CrossOver-
#32 NEW cov: 11 ft: 56 corp: 13/45b lim: 4 exec/s: 0 rss: 317Mb L: 2/4 MS: 2 CrossOver-ChangeBit-
#38 REDUCE cov: 11 ft: 56 corp: 13/44b lim: 4 exec/s: 0 rss: 317Mb L: 1/4 MS: 1 EraseBytes-
These commands launches the a qemu-user based libFuzzer job in OneFuzz. Note, we've added the arguments --wait_for_running --wait_for_files inputs
such that we can monitor our job until we've seen at least one new input found via fuzzing.
TARGET_PROJECT=AARCH64
TARGET_NAME=Example
TARGET_BUILD=1
FUZZ_POOL=linux
onefuzz template libfuzzer qemu_user ${TARGET_PROJECT} ${TARGET_NAME} ${TARGET_BUILD} ${FUZZ_POOL} --wait_for_running --wait_for_files inputs
When we run this, we'll see output similar to:
WARNING:onefuzz:qemu_user jobs are a preview feature and may change in the future
INFO:onefuzz:creating job (runtime: 24 hours)
INFO:onefuzz:created job: fa5b7870-a51b-4f79-924f-2ef11a9830a0
INFO:onefuzz:using container: oft-setup-5346d5f33bc35c3d94cbc70f7815b85e
INFO:onefuzz:using container: oft-inputs-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:using container: oft-crashes-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:using container: oft-reports-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:using container: oft-unique-reports-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:using container: oft-no-repro-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:uploading target exe `fuzz.exe`
INFO:onefuzz:uploading /tmp/tmp_9_f9kc3/setup.sh
INFO:onefuzz:uploading /tmp/tmp_9_f9kc3/fuzz.exe-wrapper.sh
INFO:onefuzz:creating libfuzzer_fuzz task
INFO:onefuzz:creating libfuzzer_crash_report task
INFO:onefuzz:done creating tasks
- waiting on: libfuzzer_crash_report:init, libfuzzer_fuzz:init
- waiting on: libfuzzer_crash_report:waiting, libfuzzer_fuzz:scheduled
| waiting on: libfuzzer_crash_report:waiting, libfuzzer_fuzz:setting_up
/ waiting on: libfuzzer_crash_report:waiting
| waiting on: libfuzzer_crash_report:scheduled
\ waiting on: libfuzzer_crash_report:setting_up
INFO:onefuzz:tasks started
\ waiting for new files: oft-inputs-9c31136dc16a5aab8edf7666a614a285
INFO:onefuzz:new files found
{
"config": {
"build": "1",
"duration": 24,
"name": "Example",
"project": "AARCH64"
},
"end_time": "2021-02-27T16:41:24+00:00",
"job_id": "fa5b7870-a51b-4f79-924f-2ef11a9830a0",
"state": "enabled",
"task_info": [
{
"state": "stopped",
"task_id": "15cc52b9-b15b-4cf7-9fa7-669db67c8e0b",
"type": "libfuzzer_fuzz"
},
{
"state": "running",
"task_id": "b3466396-7047-46f5-a58d-a21ada881e97",
"type": "libfuzzer_crash_report"
}
],
"user_info": {
"application_id": "e3b350d1-7863-4bd5-a4c0-83e6436c9c09",
"object_id": "232a2ac6-f8fc-4eb3-b427-0c91bbab7eea",
"upn": "example@contoso.com"
}
}
- afl-unicorn - AFL-based application fuzzing using Unicorn
- TriforceAFL - AFL-based full-system fuzzing using a fork of QEMU