In this talk, I will present a new Spectre/Meltdown mitigation that I have prototyped for the Hedron microhypervisor. This prototype has also been used to quantify the runtime overhead of the proposed mitigation.
Processor-level vulnerabilities, such as Meltdown and Spectre v1/v2, allow attackers in userspace to leak information from the kernel address space. This is particularly devastating for kernel designs where the kernel address space is identical for all processes and thus allows the attacker to break the system's confidentiality boundaries.
Previous mitigation attempts, such as kernel page-table isolation (formerly KAISER) for Meltdown and various branch predictor/speculation barriers for Spectre v1/v2, introduce costly instructions into performance critical parts of the operating system kernel. Especially mitigations related to the branch predictor are only possible if the CPU vendor has exposed special functionality.
During the last six months I investigated an alternative mitigation strategy on the kernel design level that shows good mitigation properties, but adds negligible runtime overhead. This alternative mitigation involves moving process-related information in the kernel into a process-local part of the kernel address space. A userspace attacker that can infer the content of its associated kernel page table can thus only read information about its own process. Switching between these kernel address spaces is done as part of the normal address space switch when a thread in a different process is scheduled and thus comes with no additional cost.
