MOCK: Optimizing Kernel Fuzzing Mutation with Context-aware Dependency

Jiacheng Xu, Xuhong Zhang, Shouling Ji, Yuan Tian, Binbin Zhao, Qinying Wang, Peng Cheng, Jiming Chen

Abstract

Kernels are at the heart of modern operating systems, whereas their development comes with vulnerabilities. Coverage-guided fuzzing has proven to be a promising software testing technique. When applying fuzzing to kernels, the salient aspect of it is that the input is a sequence of system calls (syscalls). As kernels are complex and stateful, specific sequences of syscalls are required to build up necessary states to trigger code deep in the kernels. However, the syscall sequences generated by existing fuzzers fall short in maintaining states to sufficiently cover deep code in the kernels where vulnerabilities favor residing.

In this paper, we present a practical and effective kernel fuzzing framework, called MOCK, which is capable of learning the contextual dependencies in syscall sequences and then generating context-aware syscall sequences. To conform to the statefulness when fuzzing kernel, MOCK adaptively mutates syscall sequences in line with the calling context. MOCK integrates the context-aware dependency with (1) a customized language model-guided dependency learning algorithm, (2) a context-aware syscall sequence mutation algorithm, and (3) an adaptive task scheduling strategy to balance exploration and exploitation. Our evaluation shows that MOCK performs effectively in achieving branch coverage (up to 32% coverage growth), producing high-quality input (50% more interrelated sequences), and discovering bugs (15% more unique crashes) than the state-of-the-art kernel fuzzers. Various setups including initial seeds and a pre-trained model further boost MOCK’s performance. Additionally, MOCK also discovers 15 unique bugs in the most recent Linux kernels, including two CVEs.