AutoRFM: Scaling Low-Cost in-DRAM Trackers to Ultra-Low Rowhammer Thresholds
Moinuddin K. Qureshi
Abstract
In-DRAM Rowhammer mitigation has the potential to solve the Rowhammer problem without relying on other parts of the system. In-DRAM mitigation requires space (to identify the aggressor rows) and time (to perform the victim refresh). To reduce the storage overheads of tracking, recent works have developed secure low-cost in-DRAM trackers that can probabilistically identify aggressor rows. To obtain the time required for mitigation, these trackers rely on the Refresh Management (RFM) command introduced in DDR5. As RFM stalls the bank for a latency of 200ns-400ns, frequent use of RFM can cause significant slowdowns. For example, scaling the recent MINT tracker to a threshold of 100 incurs 33% slowdown. The goal of this paper is to enable low-cost trackers to tolerate ultralow thresholds (sub-100) while incurring negligible slowdown. This paper proposes AutoRFM, a transparent RFM mechanism that can provide mitigation time to the DRAM chips without stalling the bank. The key insight in AutoRFM is to leverage the subarray structure (e.g. each bank contains 256 subarrays) and perform mitigation on only one of the subarrays. Operations to all subarrays that are not under mitigation are serviced without any interruption. If activation occurs to the subarray under mitigation, the DRAM chip sends an ALERT signal informing the Memory Controller to retry after a predefined time. As AutoRFM works best if consecutive requests to the same bank do not get mapped to the same subarray, we use Randomized Memory Mapping to break the spatial correlation between memory accesses. Furthermore, we also develop a Fractal Mitigation Algorithm that can tolerate transitive attacks (such as Half-Double) without requiring recursive mitigations to the same subarray. Our design ensures that a declined request does not have to wait more than 200 ns before retrying, thus limiting the slowdown and avoiding any potential for denial of service. Our evaluations, with SPEC, GAP, and stream workloads, show that AutoRFM enables low-cost trackers to tolerate a threshold of as low as 74 while incurring an average slowdown of only 3.1%.