Fwd: [Cryptography] "Flip Feng Shui: Hammering a Needle in the Software Stack"

[Riad S. Wahby]

Georgi Guninski <guninski at guninski.com> wrote:
> Does Rowhammer work in clouds? According to the popular story it
> affected only laptops.

The answer is "it depends."

Machines with ECC RAM make successful rowhammer attacks considerably
harder, and meanwhile most cloud providers use ECC (e.g., Amazon uses
ECC on all machines according to their FAQ). In fact, the Flip Feng
Shui paper obliquely acknowledges that ECC helps to prevent the attack,
but doesn't quantify beyond "we have observed that Rowhammer can
occasionally induce multiple flips in a single 64-bit word" (\S 6.1.1).

For a better idea of how much harder it makes things, let's have a
look at another paper from USENIX Security this year,
    https://www.usenix.org/conference/usenixsecurity16/technical-sessions/presentation/xiao

There's a bit of decoding to do here: all of the evaluation in
this paper uses machines that *don't* have ECC. Fortunately, we can
extrapolate from figure 13(c). Remember that with ECC, one needs to
flip 3 bits in a word to undetectably change the state of RAM: ECC
will silently fix 1-bit errors and produce a machine check exception
on a 2-bit error. How much harder is it to flip 1 bit than to flip
3? According to Fig. 13(c), it's ~30x harder to flip 2 bits than 1,
and another ~30x harder to flip 3 bits than 2.

As an aside: note that the attack the Xiao paper describes only works
against Xen guests that *don't* use hardware-assisted page tables
(EPT for Intel, NPT for AMD). If you're using hardware-assisted
virtualization (e.g., most Amazon "HVM" instances), this particular
attack won't work; others might, of course.

So if you're paranoid about rowhammer in a cloud setting, one strategy
is to monitor the MCE log and shut down any instance that's getting
a lot of uncorrectable ECC errors, as this may indicate an active
rowhammer attack. But my guess is that if someone is trying to pwn
you with a cross-VM attack, they're going to use something like
cache timing: it's harder to detect and probably easier to pull off,
assuming your cloud box has ECC RAM.

But as always, new discoveries might change the whole game.

-=rsw