View on GitHub



Artifact evaluation for Dogfood


Dogfood is a prototype tool to generate workloads and test file systems. In this artifact, we introduce its basic usage, how to use it to reproduce the experiment results in the paper, and how to reuse it for further file system testing.

Obtain the artifact

We have upload a docker image containing all necessary files and subjects. HERE

Dogfood should be used under a Linux operating system. Ubuntu-16.04 and Ubuntu-18.04 are recommendations as we have tested Dogfood under them. Note that a virtual machine such as virtual-box or vmware may not work because Dogfood leverages the KVM, which may be unavailable in a virtual environment (see

Please follow the instructions in to complete the installation.

Reproduce results in the paper

Since there are three testing scenarios are presented in the paper, we introduce steps to reproduce the results, respectively.

The sequential scenario

Reproduce the detected bugs

The tool tmux is preferred to open multiple terminals:

tmux -2 # Start tmux
Ctrl+b c # Start another terminal window
Ctrl+b w # Switch between the windows; learn more from the tmux documentations

As shown in the, start QEMU first:

$qemu: ./ctrl start

Create a virtual machine snapshot:

$test: ./ctrl init

Run bug case scripts:

$test: ./bugcase/<subdirectory>/

The status of each bug is listed in the file workspace/bugcase/

Compare with Syzkaller for long-time running

Run dogfood is easy:

$qemu: ./ctrl start # Start QEMU
$test: ./ctrl init
$test: ./food --fs=<file system> --test

After running for 12 hours, use the workspace/ to extract the number of crashes.

The <file system> should be one of (ext4, btrfs, f2fs, reiserfs, jfs, gfs2, ocfs2).

Run Syzkaller need some extra steps:

  1. Install Syzkaller, follow the instructions
  2. Create a disk formatted by a file system (we provide scripts to make it easy, run):
cd workspace/objs/debian-image
./ <file system>

Which means:

  1. Run Syzkaller

We have provided a Syzkaller configuration and a script.

If Syzkaller is installed according to the above-mentioned instructions successfully, just run:


After running for 12 hours, check the results.

The crash consistency scenario

Crashmonkey works on specific kernels. To ease the burden of setting, we provide a VM image, download here, which is based on Ubuntu-16.04 but with the specific kernel and containing crashmonkey.

tar -xzvf ubuntu-16.04-clone.qcow2.tar.gz
sudo apt-get install virt-manager
sudo virt-manager

Import the VM image (in virt-manager, click):

File -> New Virtual Machine -> Importing existing disk image -> Forward -> Browse -> Browse Local -> Choose the path of qcow2 file, Choose OS type: Linux, Version: Ubuntu 16.04 -> Forward -> Forward -> Finish

Start up the VM and login via ssh:

ssh icse20@
# password: icse

If the IP address of the VM is not sure, use the following commands to check it:

sudo virsh net-list

That should display like this:

Name     State   Autostart Persistent
default  active  yes       yes

Then, use the following command to check the IP address.

sudo virsh net-dhcp-leases default # `default` is the name from the last result

Now, we work in the VM under directory /home/icse20.

All workloads generate by ACE is located at directory /home/icse20/crashmonkey/code/tests/generated_workloads.

All workloads generated by Dogfood is located at directory /home/icse20/subjects/cmp-cases and /home/icse20/subjects/large-cases. The former is used to compared with ACE and the latter consists of a great number of workloads (presented in the paper).

Run workloads by ACE:

cd crashmonkey

sudo python -f /dev/vda -d /dev/cow_ram0 -t <fs> -e 102400 -u build/tests/generated_workloads/ > outfile

<fs> should be one of (f2fs, ext4, btrfs).

Run workloads by Dogfood:

cp subjects/cmp-cases/yy-*.cpp dogfood/ # comparison cases
# or
cp subjects/large-cases/zz-*.cpp dogfood/ # large cases

cd crashmonkey

sudo python -f /dev/vda -d /dev/cow_ram0 -t <fs> -e 102400 -u build/tests/dogfood/ > outfile

Run the command to count #detected issues


Note that before each run, clean redundant files by


You can also use script dogfood/py-code/ to generate more subjects (generated in direcotry output/) and copy these subjects to the directory dogfood in the VM.

The concurrency scenario

We have generate 500 workloads for concurrency testing, which is located at workspace/con-subjects after being extracted.

tar -xzvf con-subjects.tar.gz

Configure FS in and execute it:


Reuse Dogfood: automatic testing

Reuse dogfood to test other file systems. We can create a fresh workspace to test multiple file systems in parallel. The script workspace/bundle is used.

TEST_DIR=<A testing directory>
FS=<File system to test>
./bundle $TEST_DIR

Now, we can use dogfood as mentioned before. Start the QEMU and testing.


$qemu: ./ctrl start # start QEMU

$test: cd $TEST_DIR
$test: ./ctrl init
$test: ./food --fs=$FS --test

If you do not want to start qemu manually, using

$test: ./food --fs=$FS --test --qemu

This will start QEMU as a daemon thread automatically, but the logs/outputs are also not displayed in the terminal. We must check the log vm.log to analyze the results.

Notes for multiple testing in parallel

When testing multiple file systems in parallel, you should use bundle to create testing directories for each file system and edit file $TEST_DIR/mngr-tools/

You MUST set TELNET_PORT, TCP_PORT, and SSH_PORT differently for each file system testing.

Analyze the results

$test: ./ # Count #crashes
$test: ./ | sort | uniq # Count unique bugs

Reuse Dogfood: how to reproduce a new bug

When finding a bug logged in the log file workspace/vm.log, you can make a test case to reproduce it. The script workspace/bug-extract is used to extract bug cases automatically. When finding a bug in the log, you can make a test case to reproduce it. The following section introduces manual steps to reproduce a bug:

1. Locate the bug, like

[ 34.786048] kernel BUG at fs/f2fs/data.c:317!

2. Check the log from the bug location backward to find the test case name, like

TEST CASE[ ../workspace/C-output/2019_07_31-20:06:15-cGKYX/2.c with f2fs-0.img ]

This is the C program file to manifest the bug and its disk image.

3. Collect all mount options in MOUNT OPT [*] between the location of test case name and bug location, like

MOUNT OPT [ background_gc=off,nouser_xattr,noacl,active_logs=2,inline_data,inline_dentry,extent_cache,noextent_cache,noinline_data,data_flush,mode=lfs,io_bits=3,usrquota,grpquota,quota,noquota,fsync_mode=nobarrier ]

4. Copy the C program file and disk image found in step 2 to the directory template

5. Change the script template/

6. Copy the mount options line by line (the order should be the same as the log file) in the file template/mountfood, like

"..." # Other options
"..." # Other options

7. Run it as the demo in section Test QEMU

$test: ./template/