LLVM Multicompiler

This repo is based off the official LLVM git mirror: http://llvm.org/git/llvm.git. We have added passes which randomize the implementation details of the code to combat code-reuse attacks.

Test Environment

Ubuntu 14.04 LTS 64 bit

gcc 4.8.4

Python 2.7.6

GNU M4 1.4.17

GNU Autoconf 2.69

GNU Automake 1.14.1

libtool 2.4.2

zlib 1.2.3.4

binutils 2.24 (build instructions below)

Installation

Below are all instructions explained above.

Installing prerequisites

apt-get install -y libssl-dev libxml2-dev libpcre3-dev

Checking out LLVM, Clang, compiler-rt, poolalloc, and SVF:

git clone [email protected]:securesystemslab/multicompiler.git llvm

git clone [email protected]:securesystemslab/multicompiler-clang.git llvm/tools/clang

git clone [email protected]:securesystemslab/multicompiler-compiler-rt.git llvm/projects/compiler-rt

git clone [email protected]:securesystemslab/poolalloc.git llvm/projects/poolalloc

Link-Time Optimization (LTO)

Function randomization requires link-time optimization, since LTO ensures that functions are shuffled not within each source file but across the whole program, achieving higher entropy.

Installing prerequisites

apt-get install -y flex bison texinfo

Building binutils with Gold for LLVM LTO support and global shuffling

For Binutils 2.26, clone:

git clone [email protected]:/securesystemslab/binutils.git

Binutils 2.24 is also available in the cfar-2_24 branch

If you want to use Readactor execute-only features, apply the Readactor Gold patch now:

patch -p1 < PATH_TO_READACTOR_SOURCES/linker/binutils-gold-xonly.patch

Configure binutils with these flags:

--enable-gold --enable-plugins --prefix=prefix --disable-werror

(prefix is not required, you can run gold/ld-new directly from the build directory by symlinking or copying it to the LLVM build/install binary directory.)

The randomizing gold patch requires openssl < v1.1. To explicitly build against openssl-1.0, add the following environment variables before both the configure and make commands: CPPFLAGS="-I/usr/include/openssl-1.0" LDFLAGS="-L/usr/lib/openssl-1.0".

then

make

make install

Building and Compiling LLVM

Patch printf.h from glibc:

There is a bug in glibc printf.h: https://sourceware.org/bugzilla/show_bug.cgi?id=18907. Apply the following patch to /usr/include/printf.h to fix the bug.

--- /usr/include/printf.h.orig	2016-12-13 21:34:35.897301441 +0000
+++ /usr/include/printf.h	2016-12-13 21:35:40.374031243 +0000
@@ -111,13 +111,13 @@
    it returns a positive value representing the bit set in the USER
    field in 'struct printf_info'.  */

-extern int register_printf_modifier (const wchar_t *__str) __wur __THROW;
+extern int register_printf_modifier (const wchar_t *__str) __THROW __wur;


 /* Register variable argument handler for user type.  The return value
    is to be used in ARGINFO functions to signal the use of the
    type.  */
-extern int register_printf_type (printf_va_arg_function __fct) __wur __THROW;
+extern int register_printf_type (printf_va_arg_function __fct) __THROW __wur;


 /* Parse FMT, and fill in N elements of ARGTYPES with the

Recommended: Building LLVM and Clang using cmake:

Linux

1. cmake .. -DLLVM_TARGETS_TO_BUILD="X86" -DCMAKE_INSTALL_PREFIX=... -DCMAKE_BUILD_TYPE=Release -DLLVM_BINUTILS_INCDIR=...

To build cross-checking support, add the following to the previous command -DMULTICOMPILER_RAVEN_SRCDIR=<raven sources> -DMULTICOMPILER_RAVEN_OUTDIR=<raven build> where <raven sources> is the directory containing the raven MVEE source code (containing rbuff/rbuff.h) and <raven build> is the directory containing the built raven MVEE (containing rbuff/librbuff.so).

The multicompiler RNG requires openssl < v1.1. If you have both openssl-1.0 and openssl-1.1 installed, you need to force cmake to recognize the 1.0 version with the following additional arguments: -DOPENSSL_INCLUDE_DIR=/usr/include/openssl-1.0 -DOPENSSL_CRYPTO_LIBRARY=/usr/lib/libcrypto.so.1.0.0. If you wish to enable LLVM_OPTIMIZED_TABLEGEN, you'll also need to patch NATIVE/CMakeCache.txt in the build directory manually after the build fails because there is no good way to set cmake variables for the second cmake invocation the LLVM build system uses to build optimized tablegen for a debug build. Manually overwrite OPENSSL_CRYPTO_LIBRARY and OPENSSL_INCLUDE_DIR with the values from CMakeCache.txt. This has been tested on current Arch Linux, exact paths may vary for other distributions.

2. make

macOS

To build on macOS, you may need to export OPENSSL_ROOT_DIR so it points to your openssl root folder. Homebrew puts openssl in /usr/local/opt/openssl.

1. Export OPENSSL_ROOT_DIR=/path/to/openssl.

2. cmake .. -DLLVM_TARGETS_TO_BUILD="X86" -DCMAKE_INSTALL_PREFIX=... -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_FLAGS=-I$OPENSSL_ROOT_DIR/include -DCMAKE_CXX_FLAGS=-I$OPENSSL_ROOT_DIR/include

Deprecated: Building LLVM and Clang using configure:

cd llvm

mkdir build

cd build

../configure --prefix=... --enable-optimized --enable-targets=x86,x86_64

make

make install

Options:

The release version is optimized. Switch --enable-optimized with --disable-optimized for a debug (10x slower!!) build.

For LTO:

1. use --with-binutils-include=<binutils 'include' path>

2. to install in your own prefix directory, make sure that both binutils and LLVM have the same prefix, then make install both toolchains to the same prefix.

3. also, copy or link from LLVMgold.so to prefix/lib/bfd-plugins/LLVMgold.so and make sure prefix/bin/ld points to ld.gold, not ld.bfd.

--with-built-clang instructs the test-suite to use the built version of clang rather than llvm-gcc.

For 64-bit targets, use --enable-targets=x86_64 or --enable-targets=x86,x86_64.

Note: On some systems the prefix path must be absolute: /home/myuser/multicompiler/install rather than ../install.

Options

For LTO, all -mllvm -option-here options should be translated to the form -Wl,--plugin-opt,-option-here and -flto added to the compilation (and potentially linking, if they are different) flags.

important: When using build systems that use libtool (e.g. Apache), you must include -flto in the compiler path (CC), rather than CFLAGS/LDFLAGS. Libtool (stupidly and silently) strips unrecognized options such as flto from the CFLAGS/LDFLAGS. Example of correct usage: export CC="/path/to/multicompiler/clang -flto"

General Options

-frandom-seed=# - Set the random seed to #.(The type of # is uint64_t, which means the range should be within 0 - (2^64-1))

For LTO: -Wl,--plugin-opt,-random-seed=#

Stack-layout randomization and reversal

-mllvm -shuffle-stack-frames - Enable stack-layout randomization.

-mllvm -reverse-stack-frames - Reverse layout of each stack frame.

-mllvm -stack-frame-random-seed=SEED - Distinct stack frame randomization seed. Overrides -frandom-seed (or -random-seed above) for this randomization (and stack frame padding).

Insert padding between stack frames

-mllvm -max-stack-pad-size=# - Enable inter-stack-frame padding of which the maximum size is #. (The preferable size is between 0 and 256)

-mllvm -stack-frame-random-seed=SEED - Distinct stack frame randomization seed. Overrides -frandom-seed (or -random-seed above) for this randomization (and stack frame shuffling).

Stack-element padding

Insert padding between elements in the unsafe stack.

-mllvm -stack-element-percentage=# - Percentage of elements in the unsafe stack prepended by paddings.

-mllvm -max-stack-element-pad-size=# - Maximum size of stack element padding (The preferable size is between 0 and 50).

-mllvm -stack-element-padding-random-seed=SEED - Distinct stack element padding randomization seed. Overrides -frandom-seed (or -random-seed above) for this randomization.

Function randomization (LTO req'd)

Using this transformation without LTO is possible but not recommended.

-mllvm -randomize-function-list - Enable function randomization.

Machine register randomization

-mllvm -randomize-machine-registers - Enable machine register randomization.

Safestack

-fsanitize=safe-stack - Enable Safestack. This feature places buffers and other "address-taken" variables on a separate stack to prevent stack smashing.

NOTE: this option should be passed to both compiler and linker as any other sanitizer flags.

Stack-to-heap promotion

-mllvm -stack-to-heap-promotion - Enable stack-to-heap promotion: this feature randomly promotes buffers in stack slots to heap.

-mllvm -stack-to-heap-percentage=# - Percentage of buffers to be promoted to heap.

The promoted slots are malloc'ed in the beginning and then free'd when function returns (performance concerns). In the current implementation, the promoted stack slots and their pointers are not remained in the stack.

TODO: 1) This currently does not promote buffers in safestack if it is used with SafeStack. 2) Promoting DynamicAllocas is currently not supported.

NOTE: This transformation should not be applied to signal handlers because it inserts async-signal-unsafe functions: malloc() and free(). To avoid this issue, we whitelist signal handlers defined in ATDSigHandlers.def.

Code-pointer protection (LTO req'd)

-fcode-pointer-protection - Replaces code pointers with pointers to trampolines. Prevents code pointers from leaking the code layout. Cookies are used to authenticate calls and returns through trampolines. Intended to be used in conjunction with X-only memory. Note: this feature works with stack unwinding but likely breaks C++ exception handling.

For LTO: -Wl,--plugin-opt,-pointer-protection -Wl,--plugin-opt,-call-pointer-protection -Wl,--plugin-opt,-cookie-protection

Function pointer trampolines support striding to support disjoint trampoline table indices. By carefully choosing relatively prime offsets for each variant, we can ensure that the same offset from a give trampoline will not be a valid trampoline in all variants.

-mllvm -disjoint-trampoline-spacing=M - Space trampolines apart by M table indices (M*8 bytes on X86).

-mllvm -disjoint-trampoline-multiple=N - Do not emit a trampoline at multiples of index N

With these two options together, we can create disjoint trampoline tables. For example: set -disjoint-trampoline-spacing to 3, 4, 5, and 7 in four variant respectively, and set -disjoint-trampoline-multiple=420 to avoid emitting trampolines to all common multiples of those offsets.

Global padding (LTO req'd)

Using this transformation without LTO is possible but not recommended.

-mllvm -global-padding-percentage=N - Add a randomly size padding global for N% of global variables.

-mllvm -global-padding-max-size=SIZE - Maximum size of global variable padding (size is randomly chosen from (0,SIZE] ).

-mllvm -global-min-count=N - Ensure that there are at least N global variables. If the input (plus any additional padding globals inserted via -global-padding-percentage above) contains more than 1 but fewer than N globals, add enough randomly sized padding globals to ensure N globals. This is particularly useful when shuffling globals to ensure there is sufficient entropy. This option treats normal and common globals separately and ensures there are at least N of each, since these global lists are shuffled independently.

-mllvm -global-randomization-random-seed=SEED - Distinct global randomization seed. Overrides -frandom-seed (or -random-seed above) for this randomization (and global shuffling, below).

Global Shuffling and reversal (LTO req'd)

Using this transformation without LTO is possible but not recommended. This transformation requires a patched version of gold for complete coverage, see below.

-mllvm -shuffle-globals - Randomly permute the ordering of global variables.

-mllvm -reverse-globals - Reverse the ordering of global variables.

-mllvm -global-randomization-random-seed=SEED - Distinct global randomization seed. Overrides -frandom-seed (or -random-seed above) for this randomization (and global padding, above).

Common global symbols, i.e. the compiler was unsure where the global was defined and therefore allocated, cannot be randomized in LLVM/Clang. The linker actually defines and allocates these objects. Therefore, we have to use a patched linker to enforce randomization of these variables. With the patched gold mentioned above, add the following linker flags for common symbol randomization:

-Wl,--sort-common=random - Sort the common variables randomly.

-Wl,--sort-common=random-reverse - Sort the common variables randomly and then reverse their order.

-Wl,--random-seed=SEED - Random seed passed for the linker.

NOP Insertion

Insert NOP instructions before some instructions.

-Xclang -nop-insertion - Enable NOP insertion.

-mllvm -nop-insertion-percentage=# - Percentage of instructions prepended by NOP instructions.

-mllvm -max-nops-per-instruction=# - Maximum number of NOP insertion for a instruction.

-mllvm -NOP-random-seed=# - Distinct NOP insertion seed. Overrides -frandom-seed (or -random-seed above) for this randomization.

MOV-to-LEA

Change “MOV r1, r2” to the equivalent “LEA r1, [r2]".

-mllvm -mov-to-lea-percentage=# - Percentage of “MOV” instructions that are changed to “LEA” instructions.

-mllvm -MOVToLEA-random-seed=# - Distinct “MOV to LEA” seed. Overrides -frandom-seed (or -random-seed above) for this randomization.

VTable randomization (Linux only)

Split vtable into read-only part (rvtable) and randomized execute-only part (xvtable).

-fvtable-rando - Enable VTable randomization including booby trap insertion.

NOTE: this option should be passed to both compiler and linker to link against compiler-rt (libclang-rt.vtable_rando-$arch.so).

This includes the following defaults, which can be changed by adding them after -fvtable-rando:

-Xclang -min-number-vtable-entries=10 - Pad each vtable to have at least 10 entries. Too few entries will have insufficient entropy for effective randomization.

-Xclang -min-percent-vtable-boobytraps=25 - Insert boobytrap vtable entries so that at least 25% of the vtable is boobytrapped.

With LTO, -Wl,--plugin-opt,-mark-vtables is required to enable marking of VTables in IR.

Usage Notes:

1. Vtables are randomized at runtime, therefore, the randomization seed and reversal option is controlled via environment variables. Use MVEE_VTABLE_RANDO_SEED to specify the seed (must fit into an unsigned long) and set MVEE_VTABLE_RANDO_REVERSE to any value to reverse the shuffled vtable layout.
2. Vtable randomization appears to work on Ubuntu 14.04 x86-64 but generates errors on Ubuntu 16.04 x86-64. The feature was tested in isolation in googletest-1.8.0; beware of interactions with other randomization and/or compilation options.
3. Determining that vtable randomization was applied correctly:
   • ldd /path/to/your/binary should include libclang_rt.vtable_rando-x86_64.so in its output.
   • nm /path/to/your/binary should include vtablerando_randomize and vtablerando_register_module.
4. RTTI is required for VTable randomization, so do NOT include -fno-rtti in the C++ build flags.

Building shared libraries for VTable randomization:

If a target application for vtable-rando is calling virtual methods in shared libraries, the shared libraries should also be compiled by vtable-rando. The detailed instructions how to build and link against the libraries are here.

PLT randomization (Linux only)

Shuffles the procedure linkage table at load time.

-fplt-rando - Enable PLT randomization.

NOTE: this option should be passed to both compiler and linker to link against compiler-rt (libclang-rt.plt_rando-$arch.so).

Usage Notes:

1. PLTs are randomized at run time, therefore, the randomization seed and reversal option is controlled via environment variables. Use PLT_RANDO_SEED to specify the seed (must fit into an unsigned long) and set OKT_RANDO_REVERSE to any value to reverse the shuffled PLT layout.
2. Determining that PLT randomization was applied correctly:
   • ldd /path/to/your/binary should include libclang_rt.plt_rando-x86_64.so in its output.
   • nm /path/to/your/binary should include pltrando_randomize and pltrando_register_module.

Data randomization (LTO req'd, Linux only)

Randomize data representations by using xor encryption and description with random masks. This is only known to work with Apache 2.0.4 and thttpd.

-data-rando - Enable data randomization pass.

-cs-data-rando - Enable context sensitive data randomization pass.

For LTO: -fdata-rando - Context insensitive data randomization -fcs-data-rando - Context sensitive data randomization

Datarando does not always function correctly when SLP vectorization is enabled. To disable the SLP vectorizer use both of the following options:

During compilation (CFLAGS): -fno-slp-vectorize During LTO linking (LDFLAGS): -Wl,-plugin-opt,disable-vectorization

Data crosschecking

Insert raven crosschecks before branching based on potentially encrypted booleans. This pass is now run by clang during regular compilation, before LTO.

-fsanitize=crosscheck - Enable data crosschecking pass -mllvm -xcheck-data - Enable data cross checks (requires -fsanitize=crosscheck)

The sanitize option is required during both compilation and linking.

To force cross checks at branches instead of after loading cross-checked values, additionally use: -mllvm -data-checks-at-branch (required when combined with struct layout randomization). This requires the previous two options as well.

Periodic cross-checking is enabled by default. To disable, configure cmake with -DMULTICOMPILER_PERIODIC_CROSSCHECKS=On, however, this should not be necessary in normal use.

Linking against the synchronous version of the cross-checking runtime for debugging is enabled by linking with -fsanitize-debug-crosscheck along with the usual -fsanitize=crosscheck. Additional logging of crosschecks for debugging can be enabled with -mllvm -log-xchecks.

Further details of data randomization and crosschecking are here.

Functions can be blacklisted from crosschecking by adding them to a -fsanitize blacklist (see http://releases.llvm.org/3.8.1/tools/docs/SanitizerSpecialCaseList.html)

Control flow crosschecking

Insert raven crosschecks at the beginning of each function to ensure that the same function sequence is called in each variant.

-fsanitize=crosscheck - Enable data crosschecking pass (required for control flow checks) -mllvm -xcheck-cf - Enable function-level control flow crosschecking

Can be freely combined with data crosschecks, just use a single -fsanitize=crosscheck option.

Control-flow crosschecks respect the same sanitizer blacklist as data crosschecks (see above). Blacklist functions with fun:foo in the blacklist to not insert crosschecks into function foo.