A Practical Attack Framework for Precise Enclave Execution Control

SGX-Step is an open-source framework to facilitate side-channel attack research on Intel x86 processors in general and Intel SGX platforms in particular. SGX-Step consists of an adversarial Linux kernel driver and a small user-space operating system library that allows to configure untrusted page table entries and/or x86 APIC timer interrupts completely from user space. SGX-Step has been leveraged in our own research, as well as by independent researchers, to enable several new and improved enclaved execution attacks that gather side-channel observations at a maximal temporal resolution (i.e., by interrupting the victim enclave after every single instruction).

License. SGX-Step is free software, licensed under GPLv3. The SGX-Step logo is derived from Eadweard Muybridge's iconic public domain "Sallie Gardner at a Gallop" photo series, which, like our enclave single-stepping goal, breaks down the galloping horse dynamics into a series of individual photo frames to reveal overall horse gait properties.

Publications. SGX-Step has been employed by several independent research groups and has enabled a new line of high-resolution SGX attacks. A full up-to-date list of known projects using SGX-Step is included at the bottom of this README. A copy of the original paper is available here.

@inproceedings{vanbulck2017sgxstep,
    title     = {{SGX-Step}: A Practical Attack Framework for Precise Enclave Execution Control},
    author    = {Van Bulck, Jo and Piessens, Frank and Strackx, Raoul},
    booktitle = {2nd Workshop on System Software for Trusted Execution {(SysTEX)}},
    publisher = {{ACM}},
    pages     = {4:1--4:6},
    month     = Oct,
    year      = 2017,
}

Demo. The app/memcmp directory contains a small demo application that illustrates the power of SGX-Step attacks by single-stepping a sample enclave that contains subtle, non-constant-time memcmp password comparison logic. As opposed to traditional, notoriously noisy timing attacks, SGX-Step can deterministically brute-force the password character-per-character in linear time:

Overview

Crucial to the design of SGX-Step, as opposed to previous enclave preemption proposals, is the creation of user-space virtual memory mappings for physical memory locations holding page table entries, as well as for the local APIC memory-mapped I/O configuration registers and the x86 Interrupt Descriptor Table (IDT). This allows an untrusted, attacker-controlled host process to easily (i) track or modify enclave page table entries, (ii) configure the APIC timer one-shot/periodic interrupt source, (iii) trigger inter-processor interrupts, and (iv) register custom interrupt handlers completely within user space.

The above figure summarizes the sequence of hardware and software steps when interrupting and resuming an SGX enclave through our framework.

1. The local APIC timer interrupt arrives within an enclaved instruction.
2. The processor executes the AEX procedure that securely stores execution context in the enclave’s SSA frame, initializes CPU registers, and vectors to the (user space) interrupt handler registered in the IDT.
3. At this point, any attack-specific, spy code can easily be plugged in.
4. The library returns to the user space AEP trampoline. We modified the untrusted runtime of the official SGX SDK to allow easy registration of a custom AEP stub. Furthermore, to enable precise evaluation of our approach on attacker-controlled benchmark debug enclaves, SGX-Step can optionally be instrumented to retrieve the stored instruction pointer from the interrupted enclave's SSA frame (using Linux's /proc/self/mem interface and the EDBGRD instruction).
5. Thereafter, we configure the local APIC timer for the next interrupt by writing into the initial-count memory-mapped I/O register, just before executing (6) ERESUME.

Source code overview

This repository is organized as follows:

.
├── app        -- Collection of sample client applications using SGX-Step to
│                 attack different victim enclave scenarios.
├── doc        -- Papers and reference material.
├── kernel     -- Minimal dynamically loadable Linux kernel driver to export
│                 physical memory to user space and bootstrap `libsgxstep`.
├── libsgxstep -- Small user-space operating system library that implements the
│                 actual SGX-Step functionality, including x86 page-table and
│                 APIC timer manipulations.
└── sdk        -- Bindings to use SGX-Step with different SGX SDKs and libOSs.

Framework features and applications

SGX-Step is a universal execution control framework that enables the precise interleaving of victim enclave instructions with arbitrary attacker code. Some of the main use cases of the SGX-Step framework are summarized in the figure below (see also the bottom of this README for an up-to-date list of publications using SGX-Step).

Building and running

0. System requirements

SGX-Step requires an SGX-capable Intel processor, and an off-the-shelf Linux kernel. Our original evaluation was performed on i7-6500U/6700 CPUs, running Ubuntu 18.04 with a stock Linux 4.15.0 kernel. More recent Linux kernels and distributions are also supported. We summarize Linux kernel parameters below.

Pass the desired boot parameters to the kernel as follows:

  # if you don't have vim, use nano instead
$ sudo vim /etc/default/grub
  # Add the following line: GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nox2apic iomem=relaxed no_timer_check clearcpuid=308,295,514 pti=off isolcpus=1 nmi_watchdog=0 rcupdate.rcu_cpu_stall_suppress=1 msr.allow_writes=on vdso=0"

$ sudo update-grub && reboot

To check that the currently running kernel is configured correctly, execute:

$ sudo ./check_sys.sh
.. Checking recommended SGX-Step parameters          [OK]
.. Checking unknown kernel parameters                [OK]
.. Checking CPU features                             [OK]
.. Checking kernel page-table isolation              [OK]

Finally, to improve overall execution time stability, you may opt to additionally disable C-States and SpeedStep technology in the BIOS configuration.

1. Build and load /dev/sgx-step

SGX-Step comes with a loadable kernel module that exports an IOCTL interface to the libsgxstep user-space library. The driver is mainly responsible for (i) hooking the APIC timer interrupt handler, (ii) collecting untrusted page table mappings, and optionally (iii) fetching the interrupted instruction pointer for benchmark enclaves.

To build and load the /dev/sgx-step driver, execute:

$ cd kernel/
$ ./install_SGX_driver.sh              # tested on Ubuntu 20.04/22.04
$ make clean load

Note (/dev/sgx_enclave). SGX-Step supports both the legacy Intel /dev/isgx out-of-tree driver that should work on all platforms, as well as well as the upstream /dev/sgx_enclave driver for platforms with recent Linux kernels >5.11 plus hardware support for flexible-launch control. The install_SGX_driver.sh script should automatically detect whether an in-tree /dev/sgx_enclave driver is available, and, if not, build and load the out-of-tree /dev/isgx driver via the git submodule that points to an unmodified v2.14 linux-sgx-driver.

Note (/dev/mem). We rely on Linux's virtual /dev/mem device to construct user-level virtual memory mappings for APIC physical memory-mapped I/O registers and page table entries of interest. Recent Linux distributions typically enable the CONFIG_STRICT_DEVMEM option which prevents such use, however. Our /dev/sgx-step driver therefore includes an approach to bypass devmem_is_allowed checks, without having to recompile the kernel.

2. Patch and install SGX SDK

To enable easy registration of a custom Asynchronous Exit Pointer (AEP) stub, we modified the untrusted runtime of the official Intel SGX SDK. Proceed as follows to checkout linux-sgx v2.23 and apply our patches.

$ cd sdk/intel-sdk/
$ ./install_SGX_SDK.sh                 # tested on Ubuntu 20.04/22.04
$ source /opt/intel/sgxsdk/environment # add to ~/.bashrc to preserve across terminal sessions
$ sudo service aesmd status            # stop/start aesmd service if needed

The above install scripts are tested on Ubuntu 22.04 LTS. For other GNU/Linux distributions, please follow the instructions in the linux-sgx project to build and install the Intel SGX SDK and PSW packages. You will also need to build and load an (unmodified) linux-sgx-driver SGX kernel module in order to use SGX-Step.

Note (local installation). The patched SGX SDK and PSW packages can be installed locally, without affecting a compatible system-wide 'linux-sgx' installation. For this, the example Makefiles support an SGX_SDK environment variable that points to the local SDK installation directory. When detecting a non-default SDK path (i.e., not /opt/intel/sgxsdk), the "run" Makefile targets furthermore dynamically link against the patched libsgx_urts.so untrusted runtime built in the local linux-sgx directory (using the LD_LIBRARY_PATH environment variable).

Note (32-bit support). Instructions (outdated and not maintained!) for building 32-bit versions of the SGX SDK and SGX-Step can be found in README-m32.md.

3. Build and run test applications

User-space applications can link to the libsgxstep library to make use of SGX-Step's single-stepping and page table manipulation features. Have a look at the example applications in the "app" directory.

First, check the APIC and interrupt-descriptor table setup:

$ cd app/idt
$ make run    # fires interrupts in an infinite loop to stress-test stability; exit with CTRL-C

For example, to build and run an elementary example application to test page table manipulation features and SDK patches:

$ cd app/aep-redirect
$ make run

To test timer single-stepping functionality, try for example building and running the strlen attack from the paper for a benchmark enclave that processes the secret string 100 repeated times:

$ cd app/bench
$ NUM=100 STRLEN=1 make parse   # alternatively vary NUM and use BENCH=1 or ZIGZAG=1
$ # (above command defaults to the Dell Inspiron 13 7359 evaluation laptop machine;
$ # use DESKTOP=1 to build for a Dell Optiplex 7040 machine)
$ # use SGX_SDK=/home/jo/sgxsdk/ for a local SDK installation
$ # use M32=1 To produce a 32-bit executable

The above command builds libsgxstep, the benchmark victim enclave, and the untrusted attacker host process, where the attack scenario and instance size are configured via the corresponding environment variables. The same command also runs the resulting binary non-interactively (to ensure deterministic timer intervals), and finally calls an attack-specific post-processing Python script to parse the resulting enclave instruction pointer benchmark results.

Note (performance). Single-stepping enclaved execution incurs a substantial slowdown. We measured execution times of up to 15 minutes for the experiments described in the paper. SGX-Step's page table manipulation features allow to initiate single-stepping for selected functions only, for instance by revoking access rights on specific code or data pages of interest.

Note (timer interval). The exact timer interval value depends on CPU frequency, and hence remains inherently platform-specific (see also app/apic for detailed microbenchmarks assessing the accuracy of various APIC timer modes). Configure a suitable value in /app/bench/main.c. We established precise timer intervals for our evaluation platforms (see table below) by tweaking and observing the NOP microbenchmark enclave instruction pointer trace results, as further outlined below.

Note (stability). In order to avoid the Linux kernel getting stuck or panicking, SGX-Step should automatically restore the interrupt-descriptor table and local APIC timer after exiting the libsgxstep process. You can check if the APIC timer is still firing on all cores as follows:

$ watch -n0.1 "cat /proc/interrupts | grep 'Local timer interrupts'"

Calibrating the single-stepping interval

The table below lists currently supported Intel CPUs, together with their single-stepping APIC timer interval (libsgxstep/config.h). Note that the exact single-stepping interval may depend on the microcode version of the processor when recent transient-execution attack mitigations are in place to flush microarchitectural buffers on enclave entry/exit. Some different microcode versions are provided for reference in the table below.

Note (calibration). Currently, the easiest way to configure a reliable timer interval is to use the app/bench benchmarking tool with a long NOP slide and gradually increase/decrease SGX_STEP_TIMER_INTERVAL. You can probably start around 20 and then execute NUM=100 make parse to get a summary of single-steps, zero-steps, and multi-steps for a NOP slide of 100 instructions (once you have a more or less stable interval you can switch to longer slides). Too many zero-steps indicate that you have to increase the timer interval, whereas multi-steps demand lowering the timer interval.

Note (filtering out zero-steps). Important: do not worry when there are some zero-steps left, as long as you make progress, you can always deterministically filter out zero-steps by looking at the enclave's code PTE accessed bit (which is only set when the instruction actually retires and a single-step occured). Thus, after configuring a conservative timer interval that always precludes multi-steps, SGX-Step can achieve noiseless single-stepping at a perfect, instruction-level granularity.

Note (extending the interrupt "landing window"). As clarified in the root-cause analysis below, the slower the page table walk to resolve the (code) address of the first enclave instruction following ERESUME, the longer the interrupt "landing window" and, hence, the more reliable SGX-Step's single-stepping rate will be. For instance, we found that, on top of clearing the enclave's PMD accessed bit, the landing window can be even further extended by flushing one or more of the unprotected page-table entries from the CPU cache before ERESUME, effectively forcing the CPU to wait for slow memory during the page-table walk. Thus, when you cannot find a reliable timer-interval configuration, make sure to (i) clear the enclave's code PTE/PMD "accessed" bit and (ii) flush (CLFLUSH) one or more enclave page-table entries in the AEP handler.

Note (microcode). Another word of caution relates to recent Foreshadow/ZombieLoad/RIDL/etc microcode mitigations that flush leaky uarch buffers on enclave entry/exit. Be aware that when these mitigations are enabled, the timer interval will have to be increased as enclave entry takes longer (e.g., on my i7-8650U CPU I found the single-step timer interval goes up to 54 with recent ucode, from only 34 with pre-Foreshadow ucode). The additional flushing operations may furthermore somewhat increase the variance of enclave entry time, which implies that you might have to configure the timer more conservatively with more zero-steps (which can be deterministically filtered out as explained above).

SGX-Step root-cause analysis

A detailed root-cause analysis of how exactly SGX-Step succeeds in reliably interrupting the first (possibly very short!) enclave instruction following the notoriously complex ERESUME instruction is described in the AEX-Notify paper. We found that the key to SGX-Step's success lies in its use of the "accessed" (A) bit. Specifically, SGX-Step always clears the A-bit in the victim en clave's page-middle directory (PMD) before arming the APIC to fire a one-shot interrupt. The A-bit is only ever set by the processor when at least one instruction is executed by the enclave and can, hence, be used to deterministically distinguish between zero-steps versus single-steps.

Crucially, as the processor's page-miss handler is optimized for the common fast path and uses a much slower "microcode assist" to handle the less frequent and more complex case where a PMD or PTE needs to be modified, this assist has the effect of prolonging the execution of the first enclave instruction following ERSUME by several hundreds of cycles. This "assist window" thus effectively opens a spacious landing space for the coarse-grained, normally distributed APIC timer interrupt to arrive with high accuracy.

Using SGX-Step in your own projects

The easiest way to get started using the SGX-Step framwork in your own projects, is through git submodules:

$ cd my/git/project
$ git submodule add https://github.com/jovanbulck/sgx-step.git
$ cd sgx-step # Now build `/dev/sgx-step` and `libsgxstep` as described above

Have a look at the Makefiles in the app directory to see how a client application can link to libsgxstep plus any local SGX SDK/PSW packages.

The following is a list of known projects that use SGX-Step. Feel free to open a pull request if your project uses SGX-Step but is not included below.