| Description |
Many of the operating system kernels we use today are monolithic. They consist of numerous file systems, device drivers, and other subsystems interacting with no isolation and full trust. As a result, a vulnerability or bug in one part of a kernel can compromise an entire machine. Our work is motivated by the following observations: (1) introducing some form of isolation into the kernel can help confine the effects of faulty code, and (2) modern hardware platforms are better suited for a decomposed kernel than platforms of the past. Platforms today consist of numerous cores, large nonuniform memories, and processor interconnects that resemble a miniature distributed system. We argue that kernels and hypervisors must eventually evolve beyond their current symmetric mulitprocessing (SMP) design toward a corresponding distributed design. But the path to this goal is not easy. Building such a kernel from scratch that has the same capabilities as an equivalent monolithic kernel could take years of effort. In this work, we explored the feasibility of incrementally isolating subsystems in the Linux kernel as a path toward a distributed kernel. We developed a design and techniques for moving kernel modules into strongly isolated domains in a way that is transparent to existing code, and we report on the feasibility of our approach. |
