Improving Linux-Kernel Tests for LockDoc with Feedback-driven Fuzzing
This work addresses the need for better test coverage in kernel subsystems to improve bug detection for developers, but it is incremental as it builds on existing fuzzing tools.
The paper tackled the problem of low code coverage in Linux kernel tests for LockDoc, a tool that extracts locking rules from execution traces, by repurposing a coverage-guided fuzzer to generate new benchmark programs, resulting in a 26.1% increase in VFS basic-block coverage.
LockDoc is an approach to extract locking rules for kernel data structures from a dynamic execution trace recorded while the system is under a benchmark load. These locking rules can e.g. be used to locate synchronization bugs. For high rule precision and thorough bug finding, the approach heavily depends on the choice of benchmarks: They must trigger the execution of as much code as possible in the kernel subsystem relevant for the targeted data structures. However, existing test suites such as those provided by the Linux Test Project (LTP) only achieve -- in the case of LTP -- about 35 percent basic-block coverage for the VFS subsystem, which is the relevant subsystem when extracting locking rules for filesystem-related data structures. In this article, we discuss how to complement the LTP suites to improve the code coverage for our LockDoc scenario. We repurpose syzkaller -- a coverage-guided fuzzer with the goal to validate the robustness of kernel APIs -- to 1) not aim for kernel crashes, and to 2) maximize code coverage for a specific kernel subsystem. Thereby, we generate new benchmark programs that can be run in addition to the LTP, and increase VFS basic-block coverage by 26.1 percent.