Sound Dynamic Deadlock Prediction in Linear Time

Deadlocks are one of the most notorious concurrency bugs, and significant research has focused on detecting them efficiently. Dynamic predictive analyses work by observing concurrent executions, reason about alternative interleavings that can witness bugs. Such techniques offer scalability sound bug reports, have emerged as an effective approach for detection, such data races. Effective dynamic deadlock prediction, however, proven a challenging task, no predictor currently meets requirements soundness, high-precision, efficiency. In this paper, we first formally establish tradeoff is unavoidable, showing (a) complete prediction intractable, in general, (b) even seemingly simpler task determining presence potential deadlocks, which often serve unsound witnesses actual predictable intractable. The main contribution new class called sync(hronization)-preserving deadlocks. Informally, these deadlocks be predicted reordering observed execution while preserving relative order conflicting critical sections. We present two algorithms based notion. Our algorithm SPDOffline detects all sync-preserving with running time linear per abstract pattern, novel notion also introduced work. second SPDOnline predicts involve threads strictly online fashion, runs overall time, better suited runtime monitoring setting. implemented both our evaluated their ability to perform offline deadlock-prediction large dataset standard benchmarks. results indicate highly effective, (i) it characterize vast majority (ii) detected using online, sound, efficient algorithm.