This repository contains the artifact for the ASPLOS’25 paper:

Sishuai Gong, Rui Wang, Deniz Altınbüken, Pedro Fonseca, Petros Maniatis, “Snowplow: Effective Kernel Fuzzing with a Learned White-box Test Mutator”.

The following instructions have been tested on an Ubuntu 20.04 virtual machine.

Getting started

Compile the fuzzer

The syzkaller/ directory contains Snowplow’s fuzzer implementation, which is based on Syzkaller. To compile the fuzzer, follow these steps:

$ cd syzkaller/
$ go version
# go version go1.21.3 linux/amd64
$ make

Once compiled, the fuzzer binaries will be available in syzkaller/bin/.

Compile the target kernel

Follow the instructions in the official Syzkaller repo to compile the target kernel.

For convenience, we provide an example configuration as a reference. Once the kernel is compiled, set the environment variable BZIMAGE_PATH to the path of the kernel bzImage.

Perform static analysis on the compiled kernel

1. Follow the instructions under ./scripts/kernel-analysis/ to perform static analysis on the compiled kernel binary.
2. Copy the output files and the file ./prerequisite/kernel/asm-token-dict to a designated folder.
3. Set the environment variable KERNEL_ANALYSIS_PATH to point to that folder.

Prepare the VM image

Follow the instructions to prepare a VM image. Then, configure the following environment variables:

• VM_IMAGE_PATH: Path to the VM image
• VM_KEY_PATH: Path to the VM SSH key

Set up the inference environment

Please refer to the conda configuration file to set up the inference environment.

Start the model inference service

On the inference machine, follow these steps to start the inference service:

cd ./prerequisite/inference
torchserve --stop
torchserve --start --ncs --model-store ./checkpoint/ --models PMModel.mar --enable-model-api --disable-token-auth --ts-config config.properties

Set the environment variable INFERENCE_SERVER_IP to the internal IP of the inference machine.

Run the fuzzer

On the fuzzer machine, follow these steps to start testing the kernel.

Create the fuzzer configuration file

The fuzzer requires a JSON configuration file, which should follow the structure below:

{
    "target": "linux/amd64",
    "http": "127.0.0.1:1234",
    "workdir": "./workdir",
    "init_seed": 1,
    "kernel_obj": "",
    "kernel_analysis": "$KERNEL_ANALYSIS_PATH",
    "use_ml": true,
    "mlserver_addr": "$INFERENCE_SERVER_IP:7070",
    "mlsmash_exec_cnt": 2,
    "reproduce": false,
    "image": "$VM_IMAGE_PATH",
    "sshkey": "$VM_KEY_PATH",
    "syzkaller": "./syzkaller/",
    "procs": 4,
    "type": "qemu",
    "vm": {
        "count": 42,
        "kernel": "$BZIMAGE_PATH",
        "cpu": 2,
        "mem": 2048
    }
}

Start the fuzzer

Execute the following command to start the fuzzer:

./syzkaller/bin/syz-manager -config=fuzz.cfg