FIRestarter: Practical Software Crash Recovery with Targeted Library-level Fault Injection

摘要

Despite advances in software testing, many bugs still plague deployed software, leading to crashes and thus service disruption in high-availability production applications. Existing crash recovery solutions are either limited to transient faults or require manual annotations to target predetermined persistent bugs. Moreover, existing solutions are generally inefficient, hindering practical deployment.In this paper, we present FIRestarter (Fault Injection-based Restarter), an efficient and automatic crash recovery solution for commodity user applications. To eliminate the need for manual annotations, FIRestarter injects targeted software faults at the library interface to automatically trigger error handling code for standard library calls already part of the application. In particular, when a crash occurs, we roll back the application state before the last recoverable library call, inject a fault, and restart execution forcing the call to immediately return a predetermined error code. This strategy allows the application to automatically bypass the crashing code upon such a restart and exploits existing error-handling code to recover from even persistent bugs. Moreover, since library calls lie pervasively throughout the code, our design provides a large recovery surface despite the automated approach. Finally, FIRestarter’s recovery windows are small and frequent compared to traditional checkpoint-restart, which enables new optimizations such as the ability to support rollback by means of hybrid hardware/software transactional memory instrumentation and improve performance. We apply FIRestarter to a number of event-driven server applications and show our solution achieves near-instantaneous, state-preserving crash recovery in the face of even persistent crashes. On popular web servers, our evaluation results show a recovery surface of at least 77%, with low performance overhead of at most 17%.