This is the source code accompanying the paper "Telling Your Secrets Without Page Faults: Stealthy Page Table-Based Attacks on Enclaved Execution" which appears in the 26th USENIX security symposium. A copy of the paper is available at https://people.cs.kuleuven.be/~jo.vanbulck/usenix17.pdf.
Van Bulck, J., Weichbrodt, N., Kapitza, R., Piessens, F., and Strackx, R. Telling your secrets without page faults: Stealthy page table-based attacks on enclaved execution. In 26th USENIX Security Symposium (2017), USENIX Association.
Protected module architectures, such as Intel SGX, enable strong trusted computing guarantees for hardware-enforced enclaves on top a potentially malicious operating system. However, such enclaved execution environments are known to be vulnerable to a powerful class of controlled-channel attacks. Recent research convincingly demonstrated that adversarial system software can extract sensitive data from enclaved applications by carefully revoking access rights on enclave pages, and recording the associated page faults. As a response, a number of state-of-the-art defense techniques has been proposed that suppress page faults during enclave execution.
This paper shows, however, that page table-based threats go beyond page faults. We demonstrate that an untrusted operating system can observe enclave page accesses without resorting to page faults, by exploiting other side-effects of the address translation process. We contribute two novel attack vectors that infer enclaved memory accesses from page table attributes, as well as from the caching behavior of unprotected page table memory. We demonstrate the effectiveness of our attacks by recovering EdDSA session keys with little to no noise from the popular Libgcrypt cryptographic software suite.
We based our attack framework on on commit #df4af24 from the upstream
Graphene-SGX project. Note that we use the environment variables $GRAPHENE
and $GRAPHENE_SGX
to refer to respectively the root directory containing the Graphene git directory and $GRAPHENE/Pal/src/host/Linux-SGX/
.
The following lists the major modifications:
-
$GRAPHENE_SGX/sgx_attacker.c
: untrusted user space runtime modifications to create and synchronize spy/victim threads. -
$GRAPHENE_SGX/sgx-driver/gsgx_attacker_*
: untrusted gsgx kernel driver modifications implementing attacker thread to spy on victim Page Table Entries (PTEs). -
$GRAPHENE/LibOS/shim/test/apps/hello
: simple microbenchmark application to quantify Inter Processor Interrupt (IPI) latency in terms of the number of instructions executed by the enclave after accessing a page, and before being interrupted by the kernel. -
$GRAPHENE/LibOS/shim/test/apps/libgcrypt/
: minimal client application and Graphene manifest to run unmodified Libgcrypt v1.6.3/v1.7.5 libraries in an SGX enclave.
Build PAL (partly trusted/untrusted), Graphene SGX driver (untrusted), and libOS (trusted) (https://github.com/oscarlab/graphene/wiki/SGX-Quick-Start).
- Prepare a signing key:
$ cd $GRAPHENE_SGX/signer
$ openssl genrsa -3 -out enclave-key.pem 3072
- Build PAL/libOS (with debug output enabled):
$ cd $GRAPHENE/Pal/src/
$ make SGX=1 DEBUG=1
$ cd $GRAPHENE/LibOS/
$ make SGX=1 DEBUG=1
-
Make sure you have a working linux-sgx-driver. The microbenchmark code requires the patches in the
$GRAPHENE_SGX/sgx-driver/isgx-patches/
directory, so as to be able to read the memory of a debug enclave from the gsgx driver. We applied the patches to isgx v1.7 (commit #51b2884). Note that EDBGRD is only used to quantify the latency of Inter Processor Interrupts by retrieving the stored instruction pointer of an interrupted microbenchmark debug enclave. -
Build and load gsgx driver (including our attacker spy code):
$ cd $GRAPHENE_SGX/sgx-driver/
$ make load
Graphene somehow wants to map enclaves in low virtual memory (from 0x0).
This has to be explicitly allowed (https://wiki.debian.org/mmap_min_addr);
make load
should automatically take care of this.
-
Before building Graphene untrusted runtime (step 1/3 above):
$GRAPHENE_SGX/sgx_attacker.c
: configure spy/victim pinned CPU numbers$GRAPHENE_SGX/sgx-driver/gsgx_attacker_config.h
: configure spy thread options$GRAPHENE_SGX/sgx-driver/gsgx_attacker_pte_set.c
: configure addresses to monitor
-
Build the trusted PAL, application binary and untrusted loader, based on the configuration in the manifest. Also signs the enclaved binary and supporting libOS to make them ready for shipment.
$ cd $GRAPHENE/LibOS/shim/test/apps/hello
$ make SGX=1 # DEBUG=1 for dmesg-style debug output
- Get the enclave launch token from Intel aesmd service. (Keep on restarting the aesmd service in case it crashes.)
$ make SGX_RUN=1
$ sudo service aesmd status # restart/stop/start as needed
- Launch enclaved application binary.
$ ../pal_loader helloworld
- Retrieve the spy results from the gsgx driver:
$ dmesg | tail
Proceed as follows:
-
$GRAPHENE_SGX/sgx-driver/sgx_attacker_config.h
: configure spy thread before building untrusted Graphene runtime:- Enable
CONFIG_SPY_GCRY
and disableCONFIG_SPY_MICRO
. - Choose to monitor trigger pages with either the A/D or Flush Flush technique. Likewise, PTE sets on IPIs can be constructed with either A/D or Flush Reload. Adjust cache timing threshold values as needed.
- Select victim Libgcrypt version (v1.6.3 or v1.7.5).
- Enable
-
Build unmodified Libgcrypt binaries:
$ cd $GRAPHENE/LibOS/shim/test/apps/libgcrypt/libgcrypt-ecc/
$ make src
- Build simple client application by setting
$GCRY_VERSION
and$MAIN
accordingly in the Makefile.
$ make SGX=1
$ make SGX_RUN=1
- Run the application, extract side-channel measurements, and run the post-processing script to extract the 512-bit EdDSA key:
$ ./run.sh
The helloworld binary ($GRAPHENE/LibOS/shim/test/apps/hello/
)
includes asm code that can be used to quantify Inter
Processor Interrupt (IPI) latency in terms of the number of instructions
executed by the enclave after accessing a page, and before being interrupted
by the kernel.
Proceed as follows:
-
Before building untrusted Graphene runtime:
$GRAPHENE_SGX/sgx_attacker.c
: enableSYSDUMP_CONTROL_SPY
.$GRAPHENE_SGX/sgx-driver/sgx_attacker_config.h
: includes precompiler options to investigate the effect of a.o, disabling the cache on the enclave CPU (CR0.CD) and sending the IPI directly from custom kernel asm code. Also enableCONFIG_SPY_MICRO
and disableCONFIG_SPY_GCRY
here.$GRAPHENE_SGX/sgx-driver/gsgx_attacker_pte_set.c
: Configure the address to monitor&a
and the expected instruction pointer address&asm_microbenchmark_slide
.
-
x86 instruction type and the landing slide length for the microbenchmark experiment can be configured in the
build_asm
Python script. -
Finally, run the enclaved binary in Graphene, retrieve the measurements from the gsgx driver, and parse them as follows:
$ ./parse_microbenchmarks.sh
This creates a file measurements.txt and dumps basic statistics (median, mean, stddev) on stdout using R. Also, a histogram of the distribution is created in plot.pdf (using gnuplot).
Note: the raw microbenchmark data used to generate Table 1 in the paper can
be found in the $GRAPHENE/LibOS/shim/test/apps/hello/paper-data
directory.
The code base is based on the Graphene-SGX project, which is itself licensed under GPLv3 (gramineproject/graphene#1). Libgcrypt is available under the LGPL license (https://gnupg.org/related_software/libgcrypt/).
All our attacker code extensions are equally licensed as free software, under the GPLv3 license.