Static Detection of Atomicity Violations in Object-Oriented Programs.

Abstract

Violations of atomicity are possible sources of errors in parallel programs. A violation occurs if the effect of a method execution depends on the execution of concurrent threads that operate on the same or overlapping parts of a shared data structure. All accesses to shared data are assumed to be ordered through synchronization, hence common techniques for data race and deadlock detection are not able to find such errors. We have developed a static analysis that infers atomicity constraints and identifies potential violations. The analysis is based on an abstract model of threads and data. A symbolic execution tracks object locking and access and provides information that is finally used to determine potential violations of atomicity. We provide a detailed evaluation of our algorithm for a number of Java benchmarks. Although the algorithm does not guarantee to find all violations of atomicity, our experience shows that our method is efficient and effective in determining several known synchronization problems in a set of application programs and the Java library. The problem of overreporting that is commonly encountered due to conservatism in static analyses is moderate.