Guest-transparent instruction authentication for self-patching kernels

Abstract

Attackers can exploit vulnerable programs that are running with elevated permissions to insert kernel rootkits into a system. Security mechanisms have been created to prevent kernel rootkit implantation by relocating the vulnerable physical system to a guest virtual machine and enforcing a W ⊕ KX memory access control policy from the host virtual machine monitor. Such systems must also be able to identify and authorize the introduction of known-good kernel code. Previous works use cryptographic hashes to verify the integrity of kernel code at load-time. The hash creation and verification procedure depends on immutable kernel code. However, some modern kernels contain self-patching kernel code; they may overwrite executable instructions in memory after load-time. Such dynamic patching may occur for a variety of reason including: CPU optimizations, multiprocessor compatibility adjustments, and advanced debugging. The previous hash verification procedure cannot handle such modifications. We describe the design and implementation of a procedure that verifies the integrity of each modified instruction as it is introduced into the guest kernel. Our experiments with a self-patching Linux guest kernel show that our system can correctly detect and verify all valid instruction modifications and reject all invalid ones. In most cases our patch-level verification procedure incurs only nominal performance impact.