Software-based Control-Flow Error Detection with Hardware Performance Counters in ARM Processors

The recent trend in processor manufacturing technologies has significantly increased the susceptibility of safety-critical systems against soft errors in harsh environments. Such errors result in control-flow errors (CFEs) that can disturb systems' execution and cause severe financial, human, or environmental disasters. Therefore, there is a severe need for efficient techniques to detect CFEs and keep the systems fault-tolerant. Although numerous control-flow error detection techniques have been proposed, they impose considerable overheads, making them inappropriate for today's safety-critical systems with restricted resources. Several techniques attempt to insert fewer control-flow checking instructions to reduce overheads. However, they limit fault coverage. This paper proposes a software-based technique for ARM processors to detect CFEs. The technique leverages the Hardware Performance Counters (HPCs), which exist in most modern processors, to count micro-architecture events and generate HPC-based signatures. Based on these signatures that capture the correct control flow of the program, the proposed technique can detect CFEs once the correct control flow is violated. We evaluate the detection capability of the proposed technique by performing many fault injection experiments applied on different benchmark programs. Moreover, we compare the proposed technique with common signature-based CFE detection techniques with respect to fault coverage and imposed overheads. The results demonstrate that the proposed technique on average can achieve ~99% fault coverage which is 23.57% higher than that offered by the employed signature-based techniques. Moreover, the memory overhead imposed by the proposed technique is 4.85% lower, and the performance overhead is ~19% lower than that of the studied signature-based techniques.