Risk-Aware Leak Detection at Binary Level

Memory leak is a problematic bug that a heap object will never be deallocated after it is last accessed. Memory leaks can be classified into two categories in terms of their risk: high-risk leaks and low-risk ones. High-risk leaks eventually induce failures such as program clashes and performance degradation, whereas low-risk ones may not necessarily cause failures. As promising leak detectors, there have been proposed growth-sensitive or staleness-sensitive leak detectors. However, they have at least one of the major drawbacks: (1) they are not applicable to executable binaries of C/C++ programs, (2) they cannot quickly detect both high-risk and low-risk leaks in distinction, or (3) their run-time overheads are prohibitively high.This paper proposes BIGLeak, a risk-aware, efficient leak detector based on dynamic binary analysis. BIGLeak consists of three components: (1) the BIGLeak algorithm, which enables the accurate detection of high-risk leaks at binary level, (2) the intermittency analysis, which allows for the quick prediction of both high-risk and low-risk leaks in a short period of time, and (3) context-aware execution sampling, which effectively reduces the run-time overheads incurred by the BIGLeak algorithm and the intermittency analysis.Experiments with several synthetic and real programs, including lighttpd and postgres, show that BIGLeak succeeded in accurately detecting both high-risk and low-risk leaks in distinction at binary level in a short period of time. Especially, BIGLeak’s precision and recall of high-risk leak detection were much better than those of an existing staleness detector, SWAT, which works at binary level. Experimental results also indicate that BIGLeak’s run-time overheads were comparable to SWAT, despite BIGLeak’s accuracy outperformed SWAT’s.