Skip to content

README.md

Latest commit

 

History

History
194 lines (163 loc) · 7.91 KB

README.md

File metadata and controls

194 lines (163 loc) · 7.91 KB

Oreo Artifact

This repository is the artifact of Oreo: Protecting ASLR Against Microarchitectural Attacks.

Please check our paper for details:

@inproceedings{song-oreo-ndss-2025,
  author = {Song, Shixin and Zhang, Joseph and Yan, Mengjia},
  title = {Oreo: Protecting ASLR Against Microarchitectural Attacks},
  booktitle = {The Network and Distributed System Security (NDSS) Symposium 2025},
  series = {NDSS 2025},
  year = {2025}
}

Supported Platforms

The artifact should work on most modern Linux distributions and CPUs. It has been currently tested on the following platforms:

CPUs

  • Intel(R) Xeon(R) Gold 5220R
  • Intel(R) Core(TM) i9-14900K
  • AMD Ryzen Threadripper 2970WX

Linux distributions

  • Ubuntu 22.04 (Linux kernel 5.15)
  • Manjaro Linux 24.0 (Linux kernel 6.1 and 6.6)

Dependencies

The following software dependencies should be correctly installed and configured before evaluating the artifact:

  • Docker v24+ (with docker-compose v1.29+)
  • Vagrant v2.4
  • VirtualBox v7.0
  • SPEC 2017 Benchmarks (cpu2017-1.1.9.iso is not included in the artifact because of copyright issues)

Artifact Structure

The artifact consists of three main parts, which are managed by git submodules:

  • gem5 is a modified version of gem5 v24.0 and our major changes are in the following subdirectories:
    • configs/example/gem5_library/gem5-configs: specialized gem5 config files for the artifact
    • scripts: helper scripts to build gem5 and run the experiments
    • src: implementation of Oreo's microarchitecture changes
  • linux is a modified version of Linux Kernel v6.6 and our major changes are in the following subdirectories:
    • compile_scripts: helper scripts to build the Linux kernel
    • other subdirectories: implementation of Oreo's kernel changes
  • experiments contains a Vagrantfile that automatically setups an Ubuntu 24.04 virtual machine with Vagrant and VirtualBox. The disk image of the VM is then converted to raw format with qemu-img for gem5 simulation. The experiment scripts are included in the VM.

Before evaluating the artifact, run the following commands to clone the git submodules from our public repositories:

git submodule update --init --recursive

NOTE: cpu2017-1.1.9.iso is the official image for SPEC 2017 benchmarks. It is not released with the artifact since it is not open-source. You should obtain it by yourself and place it in the root directory of the artifact.

The file structure should be as follows:

artifact
    | experiments
    | gem5
    | linux
    | cpu2017-1.1.9.iso
    | ...

Start docker

In the artifact root directory, set up and enter the docker by running:

cd linux
docker-compose up -d
docker-compose exec x86_fs /bin/bash

If you are using a newer version of Docker and docker-compose, you may need to replace docker-compose with docker compose.

Build Linux

In the docker's shell, build Linux by running:

cd /root/linux
python3 compile_scripts/compile.py --num-cores=80

Please replace 80 with an appropriate number of CPU cores used for the build based on CPU and memory resources.

Build gem5

In the docker's shell, build gem5 by running:

cd /root/gem5
python3 scripts/compile.py --num-cores=80

Please replace 80 with an appropriate number of CPU cores used for the build based on CPU and memory resources.

Build disk image

We use gem5 full system mode simulation, so we need to put all experiment files into a disk image. We use Vagrant and VirtualBox to generate the disk image. Note that in this part, we run all commands on the host machine (not in the docker).

In the artifact root directory, pack Linux source files for faster disk image build.

tar -cvf linux.tar linux

Run the following commands to generate the disk image, which takes about 1 hour on an Intel(R) Xeon(R) Gold 5220R CPU.

cd experiments/disk-image
vagrant up
vagrant halt # To shut down the current VM
vagrant global-status --prune # To shut down other VMs

Next, we convert VirtualBox image to the raw image for gem5

ls ~/VirtualBox\ VMs
# Find vm directory denoted as <vm-directory>
cd experiments/disk-image
qemu-img convert -f vmdk -O raw ~/VirtualBox\ VMs/<vm-directory>/ubuntu-jammy-22.04-cloudimg.vmdk experiments.img

The disk image is then generated at experiments/disk-image/experiments.img.

Experiments

Basic Workflow

We run our experiments in the docker in the working directory /root/gem5. Each experiment follows the following workflow:

  1. Boot Linux with gem5 KVM with different setups (baseline and Oreo) and generate a checkpoint.
  2. Resume simulation using gem5 O3 to simulate the baseline and Oreo with a script specifying which program/benchmark to run.

Run Example Programs

Run the following command to simulate two example programs hello and hello_invalid. Their source code can be found at /root/experiments/disk-image/experiments/experiments/src.

python3 scripts/gen_checkpoint.py && python3 scripts/run_example.py

hello is valid program, so gem5 should finish simulation without any exceptions. Check result/restore_ko_111_0c0c00/hello_/board.pc.com_1.device for its output. hello_invalid is a malicious program that accesses an invalid address. In the directory result/restore_ko_111_0c0c00/hello_invalid_, check board.pc.com_1.device for its output; check stderr.log for gem5 exception message on committing invalid memory accesses. The exception message can be located by running:

cat result/restore_ko_111_0c0c00/hello_invalid_/stderr.log | grep "ASLR violation"

Run Performance Evaluation with SPEC2017 IntRate

In the paper we use the following commands to run SPEC2017 benchmarks and parse the results. With the following parameters, we warm up each benchmark with 10 billion user instructions and measure the performance of the following 1 billion instructions.

python3 scripts/run_spec.py --gen-cpt --begin-cpt=1 --num-cpt=1 --num-cores=80 --user-delta=32 --spec-inst-warmup-step=10
python3 scripts/parse_spec.py --parse-raw --begin-cpt=1 --num-cpt=1 --roi-idx=11 --expected-stats=12
python3 scripts/parse_spec.py --begin-cpt=1 --num-cpt=1 --roi-idx=11 --expected-stats=12

The above full experiments takes several days to finish. For artifact evaluation, we offer a scaled-down option to warm up with 1 billion user instructions as follows.

python3 scripts/run_spec.py --gen-cpt --begin-cpt=1 --num-cpt=1 --num-cores=80 --user-delta=32 --spec-inst-warmup-step=1
python3 scripts/parse_spec.py --parse-raw --begin-cpt=1 --num-cpt=1 --roi-idx=2 --expected-stats=3
python3 scripts/parse_spec.py --begin-cpt=1 --num-cpt=1 --roi-idx=2 --expected-stats=3

Check scripts/spec_output/merge_input_mean_user_cpi_1_2.pdf for CPI overhead introduced by Oreo.

Run Performance Evaluation with LEBench

# Run the experiments
python3 scripts/run_perf.py --gen-cpt --num-cpt=8 --begin-cpt=0 --num-cores=80
# Parse the output
python3 scripts/parse_perf.py --suffix-range=0,8 --plot --parse

Check output files at /root/gem5/scripts/plot.

Run Security Evaluation

# Run the prefetch attack, and attack gadgets for leakge paths 1 and 2
python3 scripts/gen_checkpoint.py && python3 scripts/run_sec.py
# Parse output of the prefetch attack
python3 scripts/parse_prefetch.py
# Parse the output trace of attack gadgets for leakage paths 1 and 2
python3 scripts/parse_trace.py

Check output files at /root/gem5/scripts/plot.

Ratio of Masked Memory Accesses

The masked ratio of SPEC benchmarks can be found in file scripts/spec_output/merge_input_oreo_user_cpi_1_2.csv. Run the following commands, and the masked ratio of LEBench can be found in file scripts/lebench_output/test_mask_ratio.csv.

python3 scripts/parse_perf_stats.py --begin-cpt=0 --num-cpt=8