Build a Fake Phone, Find Real Bugs: Qualcomm GPU Emulation and Fuzzing with LibAFL QEMU

Mobile phone manufacturers ship competitive hardware supported by increasingly complex software stacks, ranging from firmware and bootloaders to kernel modules, hypervisors, and other TrustZone environments. In an effort to keep their products secure, these companies rely on state-of-the-art testing techniques such as fuzzing. They commonly perform their fuzzing campaigns on-device to find vulnerabilities. Unfortunately, this approach is expensive to scale and does not always provide fine-grained control over the target. To address these issues, we approached the problem through the prism of emulation, by partially reimplementing the hardware as a normal software to run on a computer. That way, we could scale fuzzing instances, and gain full control over the emulated target. The presentation will outline how we made the full emulation of Qualcomm’s Android ecosystem possible by tweaking the complex build system of the Android image and implementing a custom board (including more than 10 custom devices) in QEMU. We will review the steps required and the technical challenges encountered along the way. After providing a quick recap and the latest updates on LibAFL QEMU (presented at 37C3) by one of the LibAFL maintainers, we will delve into the gory details of how we partially emulated the latest version of Adreno—the GPU designed by Qualcomm—and built a fuzzer for its Android kernel driver. In particular, we will show how LibAFL QEMU was integrated into our custom board and the few improvements we made to the kernel to get better coverage with KCOV. Finally, we will demonstrate how our approach enabled us to find a new critical vulnerability in the GPU kernel driver.