Master of Science

thesisOperating system (OS) kernel extensions, particularly device drivers, are one of the primary sources of vulnerabilities in commodity OS kernels. Vulnerabilities in driver code are often exploited by attackers, leading to attacks like privilege escalation, denial-of-service, and arbitrary code execution. Today, kernel extensions are fully trusted and operate within the core kernel without any form of isolation. But history suggests that this trust is often misplaced, emphasizing a need for some isolation in the kernel. We develop a new framework for isolating device drivers in the Linux kernel. Our work builds on three fundamental principles: (1) strong isolation of the driver code; (2) reuse of existing driver while making no or minimal changes to the source; and (3) achieving same or better performance compared to the nonisolated driver. In comparison to existing driver isolation schemes like driver virtual machines and user-level device driver implementations, our work strives to avoid modifying existing code and implements an I/O path without incurring substantial performance overhead. We demonstrate our approach by isolating a unmodified driver for a null block device in the Linux kernel, achieving near-native throughput for block sizes ranging from 512B to 256KB and outperforming the nonisolated driver for block sizes of 1MB and higher