Monitoramento do fluxo de controle de processadores embarcados baseado em profiling de software

摘要

In the recent years, the society observes with enthusiasm the rapid proliferation of a vast diversity of embedded systems targeted to safe-critical applications like health-care systems, telecommunication, automotive and aerospace. As a consequence, besides the need for low-cost components, mainly memory, it is also mandatory the use of more robustness hardware and software parts that integrate these systems. Among the possible types of faults, those that change the control-flow of the processors that carry out embedded applications are focused on this work. Very often, these types of faults induce in catastrophic system failure. By catastrophic failure, we mean those faults that in addition to drive the system to an unexpected behavior, it is also needed to reinitialize the system to recover from the faulty state. Thus, the use of techniques capable of detecting these types of faults prevents them from spreading through the system and ultimately, generating incorrect outputs. Unfortunately, the use of fault detection techniques increase memory overhead and degrades system performance. These collateral effects may be critical, preventing real-time embedded systems from attaining their goals. As possible solutions to the mentioned problems, three hypotheses were investigated, and one of them was implemented. Therefore, this work proposes an approach based on software profiling that analyses the control-flow graph of applications, to optimize the number of checkpoints to be inserted along with the application code. In order to validate the proposed approach, we performed fault injection of three types of faults: jump, nop and bit-flips. This fault injection process was accelerated by means of hardware prototyping of the system. In this case, we used a FPGA (Field-Programmable Gate Array) mounted on a Xilinx commercial board. Detailed analysis of the obtained results indicates that the proposed approach reduces the number of checkpoints to be inserted along with the application code, thus, minimizing memory overhead and system performance degradation, while maintaining approximately unchanged the fault detection coverage, when compared to another existing approaches in the literature.