Abstract
Recent works on Control-Flow Integrity (CFI) have mainly focused on Context-Sensitive CFI policies to provide higher security guarantees. They utilize a debugging hardware feature in modern Intel CPUs, Processor Trace (PT), to efficiently collect runtime contextual information. These PT-based CFI mechanisms offload the processing of the collected PT trace and CFI enforcement onto idle cores. However, a processor does not always have idle cores due to the commonly-used multi-threaded applications such as web browsers. In fact, dedicating one or more cores for CFI enforcement reduces the number of available cores for running user programs. Our evaluation with a state-of-the-art CFI mechanism (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu $$\end{document}CFI) shows that the performance overhead of a CFI mechanism can substantially increase (up to 652% on a single-core processor) when there is no idle core for CFI enforcement. To improve the performance of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu $$\end{document}CFI, we propose to leverage a hardware monitor that unlike PT does not incur trace processing overhead. We show that the hardware monitor can be used to efficiently collect program traces (<1% overhead) in their original forms and apply \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu $$\end{document}CFI. We prototype the hardware-monitor based \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu $$\end{document}CFI on a single-core RISC-V processor. Our analysis show that hardware-monitor based \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu $$\end{document}CFI incurs, on average, 43% (up to 277%) performance overhead.