When Good Protections Go Bad: Exploiting Anti-DoS Measures to Accelerate Rowhammer Attacks

The rowhammer vulnerability, where repeated accesses to a DRAM row can speed the discharge of neighboring bits, has emerged as a significant security concern in the computing industry. To address the problem, computer and software vendors have: i) doubled DRAM refresh rates, ii) restricted access to virtual-to-physical page mappings, and iii) disabled access to cache-flush operations in sandboxed environments. While recent efforts have shown how to overcome each of these protections individually, machines today are protected from rowhammer attacks if they employ all three of these protections simultaneously. In this paper, we demonstrate the first rowhammer attack that overcomes all three of these protections when used in tandem. Our attack is a virtual-memory based cache-flush free attack that is sufficiently fast to rowhammer with double rate refresh. The most astonishing aspect of our attack is that it is enabled by the recently introduced Cache Allocation Technology, a mechanism designed in part to protect virtual machines from inter-VM denial-of-service attacks. The subtext of this paper asks the question: "Is there any hope for system security, when the protections for one attack enable yet another?" We claim that the solution to this conundrum lies in the approach taken to protecting systems. Adopting a subtractive approach to secure systems, in contrast to additive measures, could go a long way toward building provably secure systems.