Abstract
Software security practitioners are often torn between choosing performance or security. In particular, OS kernels are sensitive to the smallest performance regressions. This makes it difficult to develop innovative kernel hardening mechanisms: they may inevitably incur some run-time performance overhead. Here, we propose building each kernel function with and without hardening, within a single split kernel. In particular, this allows trusted processes to be run under unmodified kernel code, while system calls of untrusted processes are directed to the hardened kernel code. We show such trusted processes run with no overhead when compared to an unmodified kernel. This allows deferring the decision of making use of hardening to the run-time. This means kernel distributors, system administrators and users can selectively enable hardening according to their needs: we give examples of such cases. Although this approach cannot be directly applied to arbitrary kernel hardening mechanisms, we show cases where it can. Finally, our implementation in the Linux kernel requires few changes to the kernel sources and no application source changes. Thus, it is both maintainable and easy to use. Copyright is held by the owner/author(s).