Native binary for testing Android phones for the Rowhammer bug
C++ Makefile
Latest commit 4c5ef24 Nov 3, 2016 @m000 m000 committed on GitHub Merge pull request #6 from Aaahh/patch-1
Correct path

README.md

Drammer

This software is the open-source component of our paper "Drammer: Deterministic Rowhammer Attacks on Mobile Devices", published in ACM Computer and Communications Security (CCS) 2016. It allows you to test whether an Android device is vulnerable to the Rowhammer bug. It does not allow you to root your device.

This code base contains our native, C/C++-based mobile Rowhammer test implementation.

Disclaimer

If, for some weird reason, you think running this code broke your device, you get to keep both pieces.

Android GUI app

If you don't want to build the test yourself, we also provide an Android app as a GUI for the native component that may or may not be currently available on the Google Play Store depending on the store's policy.

The app supports relaxed and aggressive hammering, which corresponds to the number of seconds to run 'defrag' (-d command line option described below): you can choose a timeout between 0 (no defrag) and 60 seconds, although higher timeouts likely cause the app to become unresponsive.

The app optionally collects basic statistics on the type of device and test results so that we can gain insights into the number and type of vulnerable devices in the wild, so please consider sharing them for science.

Native installation

To build the native binary, you need an Android NDK toolchain. I used android-ndk-r11c:

wget https://dl.google.com/android/repository/android-ndk-r11c-linux-x86_64.zip
unzip android-ndk-r11c-linux-x86_64.zip
cd android-ndk-r11c
./build/tools/make-standalone-toolchain.sh --ndk-dir=`pwd` \
  --arch=arm --platform=android-24 \
  --install-dir=./sysroot-arm/ \
  --verbose

You can then build the program setting STANDALONE_TOOLCHAIN variable to point to the toolchain:

STANDALONE_TOOLCHAIN=path/to/android-ndk-r11c/sysroot-arm/bin make

This gives you a stripped ARMv7 binary that you can run on both ARMv7 (32-bit) and ARMv8 (64-bit) devices. The Makefile provides an install feature that uses the Android Debug Bridge (adb) to push the binary to your device's /data/local/tmp/ directory. You can install adb by doing a sudo apt-get install android-tools-adb (on Ubuntu) or by installing the Android SDK via android.com. Then do a:

make install
make test

to install and start the Rowhammer test binary. Once installed, you may also invoke it from the shell directly:

adb shell
cd /data/local/tmp
./rh-test

Command line options

The native binary provides a number of command line options:

  • -a
    Do templating with all patterns. Without this option, only the patterns 010 and 101 are used, meaning that we hammer each row twice: once with it's aggressor rows containing all zeros while the victim row holds only ones, and once with the aggressor rows holding ones while the victim consists of zeros only. Enabling this option hammers each row with the following configurations: 000, 001, 010, 011, 100, 101, 110, 111, 00r, 0r0, 0rr, r00, r0r, rr0, rrr (where r is random and changed every 100 iterations).

  • -c
    Number of memory accesses per hammer round, defaults to 1000000. It is said that 2500000 yields the most flips.

  • -d
    Number of seconds to run 'defrag' (disabled by default). This tricks the system into freeing more ION memory that can be used for templating. Since Android tries to keep as many background processes in memory as possible, the amount of memory available for ION allocations may be very small (all of the memory is either in use, or cached in the operating system). By allocating many ION chunks, this option forces Android's low memory killer to kill background processes, giving us more (contiguous) memory to hammer in the templating phase.
    Use this option with caution: setting it too high likely hangs your device and trigger a reboot. My advice is to first try without -d (or with -d0), see how much memory you get, if not enough, hit CTRL^C, and restart with -d3. If this still does not give you enough memory, I usually repeat the sequence of breaking with CTRL^C and restarting with -d3 again in favor of using a higher timeout value. To answer the question of "how much is enough": on a Nexus 5, that comes with 2GB of memory, you should be able to get 400 to 600 MB of ION memory.

  • -f
    Write results not only to stdout but also to this file.

  • -h
    Dump the help screen.

  • -i
    Run an ION heap-type detector function.

  • -q
    Pin the program to this CPU. Some big.LITTLE architectures require you to pin the program to a big core, to make sure memory accesses are as fast as possible.

  • -r
    The rowsize in bytes. If this value is not provided, the program tries to find it using a timing side-channel (described in the paper) which may not always work. The most common value seems to be 65536 (64KB).

  • -s Hammer more conservatively. By default, we hammer each page, but this option moves less bytes (currently set to 64 bytes).

  • -t
    Stop hammering after this many seconds. The default behavior is to hammer all memory that we were able to allocate.

Description of source files

The native code base is written in C and abuses some C++ functionality. There are some comments in the source files that, combined with run-time output dumped on stdout, should give you an indication of what is happening. The main output of a run consists of numbers that indicate the average DRAM access time (in nanoseconds).

What follows is a short description of all source files.

  • Makefile
    Build system.

  • helper.h
    Inline helper functions defined in a header file.

  • ion.cc and ion.h
    Implements all ION related functionality: allocate, share, and free. By using a custom ION data data structure defined in ion.h, we also provide some functions on top of these core ION ionctls: bulk (bulk allocations), mmap, clean, and clean_all. It is required to call ION_init() before performing any ION related operations, as this function takes care of opening the /dev/ion file and reads /proc/cpuinfo to determine which ION heap to use. Note that the latter functionality is likely incomplete.

  • massage.cc and massage.h
    Implements exhaust (used for exhausting ION chunks: allocate until nothing is left) and defrag functions.

  • rh-test.cc
    Implements main() and is in charge of parsing the command line options and starting a template session.

  • rowsize.cc and rowsize.h
    Implements the auto detect function for finding the rowsize (described in more detail in the paper, Sections 5.1 and 8.1, and Figure 3)

  • templating.cc and templating.h
    Implements the actual Rowhammer test and builds template_t data structures (defined in templating.h, which might include some redundant fields). The is_exploitable() function checks whether a given template is in fact exploitable with Drammer. The main function is TMPL_run which loops over all hammerable ION chunks.