1======================
2Control Flow Integrity
3======================
4
5.. toctree::
6   :hidden:
7
8   ControlFlowIntegrityDesign
9
10.. contents::
11   :local:
12
13Introduction
14============
15
16Clang includes an implementation of a number of control flow integrity (CFI)
17schemes, which are designed to abort the program upon detecting certain forms
18of undefined behavior that can potentially allow attackers to subvert the
19program's control flow. These schemes have been optimized for performance,
20allowing developers to enable them in release builds.
21
22To enable Clang's available CFI schemes, use the flag ``-fsanitize=cfi``.
23You can also enable a subset of available :ref:`schemes <cfi-schemes>`.
24As currently implemented, all schemes rely on link-time optimization (LTO);
25so it is required to specify ``-flto``, and the linker used must support LTO,
26for example via the `gold plugin`_.
27
28To allow the checks to be implemented efficiently, the program must
29be structured such that certain object files are compiled with CFI
30enabled, and are statically linked into the program. This may preclude
31the use of shared libraries in some cases.
32
33The compiler will only produce CFI checks for a class if it can infer hidden
34LTO visibility for that class. LTO visibility is a property of a class that
35is inferred from flags and attributes. For more details, see the documentation
36for :doc:`LTO visibility <LTOVisibility>`.
37
38The ``-fsanitize=cfi-{vcall,nvcall,derived-cast,unrelated-cast}`` flags
39require that a ``-fvisibility=`` flag also be specified. This is because the
40default visibility setting is ``-fvisibility=default``, which would disable
41CFI checks for classes without visibility attributes. Most users will want
42to specify ``-fvisibility=hidden``, which enables CFI checks for such classes.
43
44Experimental support for :ref:`cross-DSO control flow integrity
45<cfi-cross-dso>` exists that does not require classes to have hidden LTO
46visibility. This cross-DSO support has unstable ABI at this time.
47
48.. _gold plugin: https://llvm.org/docs/GoldPlugin.html
49
50.. _cfi-schemes:
51
52Available schemes
53=================
54
55Available schemes are:
56
57  -  ``-fsanitize=cfi-cast-strict``: Enables :ref:`strict cast checks
58     <cfi-strictness>`.
59  -  ``-fsanitize=cfi-derived-cast``: Base-to-derived cast to the wrong
60     dynamic type.
61  -  ``-fsanitize=cfi-unrelated-cast``: Cast from ``void*`` or another
62     unrelated type to the wrong dynamic type.
63  -  ``-fsanitize=cfi-nvcall``: Non-virtual call via an object whose vptr is of
64     the wrong dynamic type.
65  -  ``-fsanitize=cfi-vcall``: Virtual call via an object whose vptr is of the
66     wrong dynamic type.
67  -  ``-fsanitize=cfi-icall``: Indirect call of a function with wrong dynamic
68     type.
69  -  ``-fsanitize=cfi-mfcall``: Indirect call via a member function pointer with
70     wrong dynamic type.
71
72You can use ``-fsanitize=cfi`` to enable all the schemes and use
73``-fno-sanitize`` flag to narrow down the set of schemes as desired.
74For example, you can build your program with
75``-fsanitize=cfi -fno-sanitize=cfi-nvcall,cfi-icall``
76to use all schemes except for non-virtual member function call and indirect call
77checking.
78
79Remember that you have to provide ``-flto`` or ``-flto=thin`` if at
80least one CFI scheme is enabled.
81
82Trapping and Diagnostics
83========================
84
85By default, CFI will abort the program immediately upon detecting a control
86flow integrity violation. You can use the :ref:`-fno-sanitize-trap=
87<controlling-code-generation>` flag to cause CFI to print a diagnostic
88similar to the one below before the program aborts.
89
90.. code-block:: console
91
92    bad-cast.cpp:109:7: runtime error: control flow integrity check for type 'B' failed during base-to-derived cast (vtable address 0x000000425a50)
93    0x000000425a50: note: vtable is of type 'A'
94     00 00 00 00  f0 f1 41 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  20 5a 42 00
95                  ^
96
97If diagnostics are enabled, you can also configure CFI to continue program
98execution instead of aborting by using the :ref:`-fsanitize-recover=
99<controlling-code-generation>` flag.
100
101Forward-Edge CFI for Virtual Calls
102==================================
103
104This scheme checks that virtual calls take place using a vptr of the correct
105dynamic type; that is, the dynamic type of the called object must be a
106derived class of the static type of the object used to make the call.
107This CFI scheme can be enabled on its own using ``-fsanitize=cfi-vcall``.
108
109For this scheme to work, all translation units containing the definition
110of a virtual member function (whether inline or not), other than members
111of :ref:`blacklisted <cfi-blacklist>` types or types with public :doc:`LTO
112visibility <LTOVisibility>`, must be compiled with ``-flto`` or ``-flto=thin``
113enabled and be statically linked into the program.
114
115Performance
116-----------
117
118A performance overhead of less than 1% has been measured by running the
119Dromaeo benchmark suite against an instrumented version of the Chromium
120web browser. Another good performance benchmark for this mechanism is the
121virtual-call-heavy SPEC 2006 xalancbmk.
122
123Note that this scheme has not yet been optimized for binary size; an increase
124of up to 15% has been observed for Chromium.
125
126Bad Cast Checking
127=================
128
129This scheme checks that pointer casts are made to an object of the correct
130dynamic type; that is, the dynamic type of the object must be a derived class
131of the pointee type of the cast. The checks are currently only introduced
132where the class being casted to is a polymorphic class.
133
134Bad casts are not in themselves control flow integrity violations, but they
135can also create security vulnerabilities, and the implementation uses many
136of the same mechanisms.
137
138There are two types of bad cast that may be forbidden: bad casts
139from a base class to a derived class (which can be checked with
140``-fsanitize=cfi-derived-cast``), and bad casts from a pointer of
141type ``void*`` or another unrelated type (which can be checked with
142``-fsanitize=cfi-unrelated-cast``).
143
144The difference between these two types of casts is that the first is defined
145by the C++ standard to produce an undefined value, while the second is not
146in itself undefined behavior (it is well defined to cast the pointer back
147to its original type) unless the object is uninitialized and the cast is a
148``static_cast`` (see C++14 [basic.life]p5).
149
150If a program as a matter of policy forbids the second type of cast, that
151restriction can normally be enforced. However it may in some cases be necessary
152for a function to perform a forbidden cast to conform with an external API
153(e.g. the ``allocate`` member function of a standard library allocator). Such
154functions may be :ref:`blacklisted <cfi-blacklist>`.
155
156For this scheme to work, all translation units containing the definition
157of a virtual member function (whether inline or not), other than members
158of :ref:`blacklisted <cfi-blacklist>` types or types with public :doc:`LTO
159visibility <LTOVisibility>`, must be compiled with ``-flto`` or ``-flto=thin``
160enabled and be statically linked into the program.
161
162Non-Virtual Member Function Call Checking
163=========================================
164
165This scheme checks that non-virtual calls take place using an object of
166the correct dynamic type; that is, the dynamic type of the called object
167must be a derived class of the static type of the object used to make the
168call. The checks are currently only introduced where the object is of a
169polymorphic class type.  This CFI scheme can be enabled on its own using
170``-fsanitize=cfi-nvcall``.
171
172For this scheme to work, all translation units containing the definition
173of a virtual member function (whether inline or not), other than members
174of :ref:`blacklisted <cfi-blacklist>` types or types with public :doc:`LTO
175visibility <LTOVisibility>`, must be compiled with ``-flto`` or ``-flto=thin``
176enabled and be statically linked into the program.
177
178.. _cfi-strictness:
179
180Strictness
181----------
182
183If a class has a single non-virtual base and does not introduce or override
184virtual member functions or fields other than an implicitly defined virtual
185destructor, it will have the same layout and virtual function semantics as
186its base. By default, casts to such classes are checked as if they were made
187to the least derived such class.
188
189Casting an instance of a base class to such a derived class is technically
190undefined behavior, but it is a relatively common hack for introducing
191member functions on class instances with specific properties that works under
192most compilers and should not have security implications, so we allow it by
193default. It can be disabled with ``-fsanitize=cfi-cast-strict``.
194
195Indirect Function Call Checking
196===============================
197
198This scheme checks that function calls take place using a function of the
199correct dynamic type; that is, the dynamic type of the function must match
200the static type used at the call. This CFI scheme can be enabled on its own
201using ``-fsanitize=cfi-icall``.
202
203For this scheme to work, each indirect function call in the program, other
204than calls in :ref:`blacklisted <cfi-blacklist>` functions, must call a
205function which was either compiled with ``-fsanitize=cfi-icall`` enabled,
206or whose address was taken by a function in a translation unit compiled with
207``-fsanitize=cfi-icall``.
208
209If a function in a translation unit compiled with ``-fsanitize=cfi-icall``
210takes the address of a function not compiled with ``-fsanitize=cfi-icall``,
211that address may differ from the address taken by a function in a translation
212unit not compiled with ``-fsanitize=cfi-icall``. This is technically a
213violation of the C and C++ standards, but it should not affect most programs.
214
215Each translation unit compiled with ``-fsanitize=cfi-icall`` must be
216statically linked into the program or shared library, and calls across
217shared library boundaries are handled as if the callee was not compiled with
218``-fsanitize=cfi-icall``.
219
220This scheme is currently supported on a limited set of targets: x86,
221x86_64, arm, arch64 and wasm.
222
223``-fsanitize-cfi-icall-generalize-pointers``
224--------------------------------------------
225
226Mismatched pointer types are a common cause of cfi-icall check failures.
227Translation units compiled with the ``-fsanitize-cfi-icall-generalize-pointers``
228flag relax pointer type checking for call sites in that translation unit,
229applied across all functions compiled with ``-fsanitize=cfi-icall``.
230
231Specifically, pointers in return and argument types are treated as equivalent as
232long as the qualifiers for the type they point to match. For example, ``char*``,
233``char**``, and ``int*`` are considered equivalent types. However, ``char*`` and
234``const char*`` are considered separate types.
235
236``-fsanitize-cfi-icall-generalize-pointers`` is not compatible with
237``-fsanitize-cfi-cross-dso``.
238
239.. _cfi-canonical-jump-tables:
240
241``-fsanitize-cfi-canonical-jump-tables``
242----------------------------------------
243
244The default behavior of Clang's indirect function call checker will replace
245the address of each CFI-checked function in the output file's symbol table
246with the address of a jump table entry which will pass CFI checks. We refer
247to this as making the jump table `canonical`. This property allows code that
248was not compiled with ``-fsanitize=cfi-icall`` to take a CFI-valid address
249of a function, but it comes with a couple of caveats that are especially
250relevant for users of cross-DSO CFI:
251
252- There is a performance and code size overhead associated with each
253  exported function, because each such function must have an associated
254  jump table entry, which must be emitted even in the common case where the
255  function is never address-taken anywhere in the program, and must be used
256  even for direct calls between DSOs, in addition to the PLT overhead.
257
258- There is no good way to take a CFI-valid address of a function written in
259  assembly or a language not supported by Clang. The reason is that the code
260  generator would need to insert a jump table in order to form a CFI-valid
261  address for assembly functions, but there is no way in general for the
262  code generator to determine the language of the function. This may be
263  possible with LTO in the intra-DSO case, but in the cross-DSO case the only
264  information available is the function declaration. One possible solution
265  is to add a C wrapper for each assembly function, but these wrappers can
266  present a significant maintenance burden for heavy users of assembly in
267  addition to adding runtime overhead.
268
269For these reasons, we provide the option of making the jump table non-canonical
270with the flag ``-fno-sanitize-cfi-canonical-jump-tables``. When the jump
271table is made non-canonical, symbol table entries point directly to the
272function body. Any instances of a function's address being taken in C will
273be replaced with a jump table address.
274
275This scheme does have its own caveats, however. It does end up breaking
276function address equality more aggressively than the default behavior,
277especially in cross-DSO mode which normally preserves function address
278equality entirely.
279
280Furthermore, it is occasionally necessary for code not compiled with
281``-fsanitize=cfi-icall`` to take a function address that is valid
282for CFI. For example, this is necessary when a function's address
283is taken by assembly code and then called by CFI-checking C code. The
284``__attribute__((cfi_canonical_jump_table))`` attribute may be used to make
285the jump table entry of a specific function canonical so that the external
286code will end up taking a address for the function that will pass CFI checks.
287
288``-fsanitize=cfi-icall`` and ``-fsanitize=function``
289----------------------------------------------------
290
291This tool is similar to ``-fsanitize=function`` in that both tools check
292the types of function calls. However, the two tools occupy different points
293on the design space; ``-fsanitize=function`` is a developer tool designed
294to find bugs in local development builds, whereas ``-fsanitize=cfi-icall``
295is a security hardening mechanism designed to be deployed in release builds.
296
297``-fsanitize=function`` has a higher space and time overhead due to a more
298complex type check at indirect call sites, as well as a need for run-time
299type information (RTTI), which may make it unsuitable for deployment. Because
300of the need for RTTI, ``-fsanitize=function`` can only be used with C++
301programs, whereas ``-fsanitize=cfi-icall`` can protect both C and C++ programs.
302
303On the other hand, ``-fsanitize=function`` conforms more closely with the C++
304standard and user expectations around interaction with shared libraries;
305the identity of function pointers is maintained, and calls across shared
306library boundaries are no different from calls within a single program or
307shared library.
308
309Member Function Pointer Call Checking
310=====================================
311
312This scheme checks that indirect calls via a member function pointer
313take place using an object of the correct dynamic type. Specifically, we
314check that the dynamic type of the member function referenced by the member
315function pointer matches the "function pointer" part of the member function
316pointer, and that the member function's class type is related to the base
317type of the member function. This CFI scheme can be enabled on its own using
318``-fsanitize=cfi-mfcall``.
319
320The compiler will only emit a full CFI check if the member function pointer's
321base type is complete. This is because the complete definition of the base
322type contains information that is necessary to correctly compile the CFI
323check. To ensure that the compiler always emits a full CFI check, it is
324recommended to also pass the flag ``-fcomplete-member-pointers``, which
325enables a non-conforming language extension that requires member pointer
326base types to be complete if they may be used for a call.
327
328For this scheme to work, all translation units containing the definition
329of a virtual member function (whether inline or not), other than members
330of :ref:`blacklisted <cfi-blacklist>` types or types with public :doc:`LTO
331visibility <LTOVisibility>`, must be compiled with ``-flto`` or ``-flto=thin``
332enabled and be statically linked into the program.
333
334This scheme is currently not compatible with cross-DSO CFI or the
335Microsoft ABI.
336
337.. _cfi-blacklist:
338
339Blacklist
340=========
341
342A :doc:`SanitizerSpecialCaseList` can be used to relax CFI checks for certain
343source files, functions and types using the ``src``, ``fun`` and ``type``
344entity types. Specific CFI modes can be be specified using ``[section]``
345headers.
346
347.. code-block:: bash
348
349    # Suppress all CFI checking for code in a file.
350    src:bad_file.cpp
351    src:bad_header.h
352    # Ignore all functions with names containing MyFooBar.
353    fun:*MyFooBar*
354    # Ignore all types in the standard library.
355    type:std::*
356    # Disable only unrelated cast checks for this function
357    [cfi-unrelated-cast]
358    fun:*UnrelatedCast*
359    # Disable CFI call checks for this function without affecting cast checks
360    [cfi-vcall|cfi-nvcall|cfi-icall]
361    fun:*BadCall*
362
363
364.. _cfi-cross-dso:
365
366Shared library support
367======================
368
369Use **-f[no-]sanitize-cfi-cross-dso** to enable the cross-DSO control
370flow integrity mode, which allows all CFI schemes listed above to
371apply across DSO boundaries. As in the regular CFI, each DSO must be
372built with ``-flto`` or ``-flto=thin``.
373
374Normally, CFI checks will only be performed for classes that have hidden LTO
375visibility. With this flag enabled, the compiler will emit cross-DSO CFI
376checks for all classes, except for those which appear in the CFI blacklist
377or which use a ``no_sanitize`` attribute.
378
379Design
380======
381
382Please refer to the :doc:`design document<ControlFlowIntegrityDesign>`.
383
384Publications
385============
386
387`Control-Flow Integrity: Principles, Implementations, and Applications <https://research.microsoft.com/pubs/64250/ccs05.pdf>`_.
388Martin Abadi, Mihai Budiu, Úlfar Erlingsson, Jay Ligatti.
389
390`Enforcing Forward-Edge Control-Flow Integrity in GCC & LLVM <http://www.pcc.me.uk/~peter/acad/usenix14.pdf>`_.
391Caroline Tice, Tom Roeder, Peter Collingbourne, Stephen Checkoway,
392Úlfar Erlingsson, Luis Lozano, Geoff Pike.
393