In 2014, a group of researchers from Carnegie Mellon University and Intel published a new kind of disturbance attack on DRAM: rowHammer . At the difference of SRAM (static), DRAM (dynamic) need regular refreshing to keep their memory. DRAM are organized by rows. Indeed, when reading or writing to an address, the circuit access the full row rather than only one specific cell. Cells are susceptible to inter-cell crosstalk (like any electronic elements). The researchers discovered the fast, repetitive reading of two rows they could generate a high rate of disturbances that produce errors in the memory. The actual code to produce errors is simple and short. It is a simple loop that reads two addresses, flushes the registers and the instruction cache. A typical 1 million iterations takes less than one second. The code does not need to be root. They tested 129 different DDR3 DRAM commercial modules. They induced errors in 110 modules.
Thus, they demonstrate that with simple software, it was possible to wreck DRAM memory.
This month, Google researchers went one step further. They used the rowHammer technique to create actual fault injection. On a standard x86-64 bit machine, they demonstrated two exploits .
- Native Client (NACl) is a sandboxing system that allows only a limited subset of instructions. They were able to have ‘blacklisted’ instructions to execute in the NACl environment.
- They succeeded to escalate the privilege to Kernel privilege on a standard Linux.
Of course, these exploits have some limitations. The escalation was done only on a Linux machine without some sandboxing mechanisms. Nevertheless, they highlight that rowHammer may become a powerful fault injection tool. The interesting part of rowHammer is that it is purely software.
Currently, they have only experimented rowHammer on standard DRAM commercial modules. This may be an interesting way to bypass some trusted execution environment that isolate the DRAM space.
DRAM for servers should be more resistant to rowHammer as Error Correction is embedded in the chip. Nevertheless, error correction can only correct a limited amount of simultaneous errors. It may be possible perhaps to also overflow the correction. If rowHammer would be possible on DRAM for servers, then it may be a potential interesting attack vector in the public cloud. The attacker may either bypass the sandbox or impair the memory of another user of the same server.
We may see in coming months more studies and exploits around rowHammer. Will it have the same impact than side channel attacks? To be surveyed…
The two papers are worthwhile to read. Read them in the chronological order.
 Y. Kim, R. Daly, J. Kim, C. Fallin, J. H. Lee, D. Lee, C. Wilkerson, K. Lai, and O. Mutlu, “Flipping bits in memory without accessing them: An experimental study of DRAM disturbance errors,” in Proceeding of the 41st annual international symposium on Computer architecture, 2014, pp. 361–372.
 C. Evans, “Project Zero: Exploiting the DRAM rowhammer bug to gain kernel privileges,” Project Zero, 09-Mar-2015.