Despite AMD’s growing market share with Zen CPUs, Rowhammer attacks were absent due to challenges in reverse engineering DRAM addressing, synchronizing with refresh commands, and achieving sufficient row activation throughput.
Researchers addressed these through ZENHAMMER, the first Rowhammer attack on recent AMD CPUs.
ZENHAMMER reverse engineers non-linear addressing uses crafted access patterns for synchronization, and schedules instructions carefully to increase throughput while bypassing mitigations.
Evaluations demonstrated ZENHAMMER finding bit flips on 7 out of 10 DDR4 devices on Zen 2/3 CPUs, enabling Rowhammer exploitation on current AMD platforms.
Besides this, it also triggered the first Rowhammer bit flips on a DDR5 device.
ZENHAMMER – First Rowhammer Attack
There have been cases of recent Rowhammer attacks that were used to bypass in-DRAM mitigations on Intel CPUs by exploiting particular architectural details, though such attacks have not been recorded against modern AMD Zen microarchitecture CPUs.
However, several crucial aspects including physical-to-DRAM address mapping, DRAM command observability, and memory instructions behavior on AMD platforms through extensive experiments were discovered.
Researchers used this information to design ZENHAMMER, it’s the first-ever successful Rowhammer attack against AMD Zen CPUs.
The goal of the researchers was to trigger bit flips on AMD Zen platforms using DDR4 memory, allowing comparison with well-studied Intel systems.
A crucial requirement for effective Rowhammer is knowledge of the DRAM address mapping from physical addresses to DRAM locations, enabling precise attacker row selection.
Since AMD and Intel memory controllers use different mappings, determining the AMD mapping posed the researchers’ first key challenge in constructing a Rowhammer attack on these platforms.
While Intel systems have all DRAM-adding bits within the lower 21 bits, AMD Zen systems utilize up to 34 bits, making exploitation challenging without knowing these bits.
Experts describe a technique combining the bank conflict side channel with reverse-engineered DRAM mappings to detect consecutive same-bank rows crucial for Rowhammer.
By coloring 2MB transparent huge pages (THPs) based on bank conflicts and using known address functions on the lower 21 bits, experts can identify same-bank rows within each THP color.
On a Zen 3 system, THP coloring takes around 39 seconds per attack, while detecting same-bank rows is a one-time 18ms cost per memory configuration.
The evaluation results reveal how well ZENHAMMER’s optimizations for causing bit flips on AMD Zen 2 and Zen 3 processors work as compared to the earlier methods.
By refining hammering instruction sequences and fence scheduling policies, ZENHAMMER dramatically raised the number of devices showing bit flips and the patterns that triggered them, particularly in the case of Zen 3 where no bit flips were reported before.
In comparison with Intel Coffee Lake on some devices, ZENHAMMER was less effective though its optimizations have shown themselves more powerful for some DIMMs even exceeding Coffee Lake’s best-performance bit flip counts.
These findings indicate that successful Rowhammer attacks require platform-specific optimizations beyond just increasing activation rates.
Stay updated on Cybersecurity news, Whitepapers, and Infographics. Follow us on LinkedIn & Twitter.
