Deadlocks as Runtime Exceptions

Deadlocks are a common type of concurrency bug. When a deadlock occurs, it is difficult to clearly determine whether there is an actual deadlock or if the application is slow or hanging due to a different reason. It is also difficult to establish the cause of the deadlock. In general, developers deal with deadlocks by using analysis tools, introducing application-specific deadlock detection mechanisms, or simply by using techniques to avoid the occurrence of deadlocks by construction. In this paper we propose a different approach. We believe that if deadlocks manifest at runtime, as exceptions, programmers will be able to identify these deadlocks in an accurate and timely manner. We leverage two insights to make this practical: (i) most deadlocks occurring in real systems involve only two threads acquiring two locks (TTTL deadlocks); and (ii) it’s possible to detect TTTL deadlocks efficiently enough for most practical systems. We conducted a study on bug reports and found that more than 90% of identified deadlocks were indeed TTTL. We extended Java’s ReentrantLock class to detect TTTL deadlocks and measured the performance overhead of this approach with a conservative benchmark. For applications whose execution time is not dominated by locking, the overhead is estimated as below 6%. Empirical usability evaluation in two experiments showed that students finished tasks 16.87% to 30.7% faster on the average using our approach with the lock being the most significant factor behind it, and, in one of the experiments they were able to identify the defects more accurately with a signficant 81.25% increase in the number of correct answers when deadlock exceptions where present.