Accurately Classifying Data Races with Portend

Even though most data races are harmless, the harmful ones are at the heart of some of the worst concurrency bugs. Eliminating all data races from programs is impractical (e.g., system performance could suffer severely), yet spotting just the harmful ones is like finding a needle in a haystack: state-of-the-art data race detectors and classifiers suffer from high false positive rates of 37%–84%. We present Portend, a technique and system for automatically triaging suspect data races based on their potential consequences: Could they lead to crashes or hangs? Alter system state? Could their effects be externalized? Or are they harmless? Our proposed technique achieves very high accuracy by efficiently analyzing multiple paths and multiple thread schedules in combination, and by performing symbolic comparison between program states. We ran Portend on several dozen data races from real-world applications, and it correctly classified all of them, with no human effort. It also produced easy-to-understand evidence of the consequences of harmful races, thus proving their harmfulness and making debugging easier. We envision using Portend for testing and debugging, as well as for automatically triaging bug reports.