Accurate Data Race Prediction in the Linux Kernel through Sparse Fourier Learning
summary
A system called HBFourier uses sparse Fourier learning to predict feasibility of ordered memory-access pairs from a coverage-guided trace corpus, estimating sparse Fourier coefficients per Boolean indicator and combining that with conservative alias/lockset analysis to avoid expensive replays. Evaluated on a recent Linux development kernel, it finds 44 more races and achieves 15.7% higher accuracy than the next-best tool, reproduces races with sanitizers and schedules, and yields five new kernel race bugs confirmed by developers.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:46Focus on kernel data races
- 3:33Multi-thread race prediction challenges
- 6:04Learning-based race detection
- 9:41Improvements over prior work
- 13:40Evaluation and comparison
- 16:08Summary and future work
- 16:30Q&A: Feature selection
- 17:11Q&A: Confirming races
related talks
22mData race detection
48mProfiling data structures
35mLinux Tracing updates
50mBreaking Kernel Address Space Layout Randomization (KASLR) With Intel TSX
24mRacerD - Compositional Static Race Detection
40mTowards a Tool for Access-Affinity Based Structure Reordering in the Linux Kernel
01h 02mProfiling with Heaptrack and Hotspot on Embedded Linux
28mExtreme OS Kernel Testing