Stephen Hines | 0e2c34f | 2015-03-23 12:09:02 -0700 | [diff] [blame^] | 1 | =========================================== |
| 2 | Control Flow Integrity Design Documentation |
| 3 | =========================================== |
| 4 | |
| 5 | This page documents the design of the :doc:`ControlFlowIntegrity` schemes |
| 6 | supported by Clang. |
| 7 | |
| 8 | Forward-Edge CFI for Virtual Calls |
| 9 | ================================== |
| 10 | |
| 11 | This scheme works by allocating, for each static type used to make a virtual |
| 12 | call, a region of read-only storage in the object file holding a bit vector |
| 13 | that maps onto to the region of storage used for those virtual tables. Each |
| 14 | set bit in the bit vector corresponds to the `address point`_ for a virtual |
| 15 | table compatible with the static type for which the bit vector is being built. |
| 16 | |
| 17 | For example, consider the following three C++ classes: |
| 18 | |
| 19 | .. code-block:: c++ |
| 20 | |
| 21 | struct A { |
| 22 | virtual void f1(); |
| 23 | virtual void f2(); |
| 24 | virtual void f3(); |
| 25 | }; |
| 26 | |
| 27 | struct B : A { |
| 28 | virtual void f1(); |
| 29 | virtual void f2(); |
| 30 | virtual void f3(); |
| 31 | }; |
| 32 | |
| 33 | struct C : A { |
| 34 | virtual void f1(); |
| 35 | virtual void f2(); |
| 36 | virtual void f3(); |
| 37 | }; |
| 38 | |
| 39 | The scheme will cause the virtual tables for A, B and C to be laid out |
| 40 | consecutively: |
| 41 | |
| 42 | .. csv-table:: Virtual Table Layout for A, B, C |
| 43 | :header: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 |
| 44 | |
| 45 | A::offset-to-top, &A::rtti, &A::f1, &A::f2, &A::f3, B::offset-to-top, &B::rtti, &B::f1, &B::f2, &B::f3, C::offset-to-top, &C::rtti, &C::f1, &C::f2, &C::f3 |
| 46 | |
| 47 | The bit vector for static types A, B and C will look like this: |
| 48 | |
| 49 | .. csv-table:: Bit Vectors for A, B, C |
| 50 | :header: Class, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 |
| 51 | |
| 52 | A, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0 |
| 53 | B, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 |
| 54 | C, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 |
| 55 | |
| 56 | To emit a virtual call, the compiler will assemble code that checks that |
| 57 | the object's virtual table pointer is in-bounds and aligned and that the |
| 58 | relevant bit is set in the bit vector. |
| 59 | |
| 60 | For example on x86 a typical virtual call may look like this: |
| 61 | |
| 62 | .. code-block:: none |
| 63 | |
| 64 | 159a: 48 8b 03 mov (%rbx),%rax |
| 65 | 159d: 48 8d 15 6c 33 00 00 lea 0x336c(%rip),%rdx |
| 66 | 15a4: 48 89 c1 mov %rax,%rcx |
| 67 | 15a7: 48 29 d1 sub %rdx,%rcx |
| 68 | 15aa: 48 c1 c1 3d rol $0x3d,%rcx |
| 69 | 15ae: 48 83 f9 51 cmp $0x51,%rcx |
| 70 | 15b2: 77 3b ja 15ef <main+0xcf> |
| 71 | 15b4: 48 89 ca mov %rcx,%rdx |
| 72 | 15b7: 48 c1 ea 05 shr $0x5,%rdx |
| 73 | 15bb: 48 8d 35 b8 07 00 00 lea 0x7b8(%rip),%rsi |
| 74 | 15c2: 8b 14 96 mov (%rsi,%rdx,4),%edx |
| 75 | 15c5: 0f a3 ca bt %ecx,%edx |
| 76 | 15c8: 73 25 jae 15ef <main+0xcf> |
| 77 | 15ca: 48 89 df mov %rbx,%rdi |
| 78 | 15cd: ff 10 callq *(%rax) |
| 79 | [...] |
| 80 | 15ef: 0f 0b ud2 |
| 81 | |
| 82 | The compiler relies on co-operation from the linker in order to assemble |
| 83 | the bit vectors for the whole program. It currently does this using LLVM's |
| 84 | `bit sets`_ mechanism together with link-time optimization. |
| 85 | |
| 86 | .. _address point: https://mentorembedded.github.io/cxx-abi/abi.html#vtable-general |
| 87 | .. _bit sets: http://llvm.org/docs/BitSets.html |
| 88 | |
| 89 | Optimizations |
| 90 | ------------- |
| 91 | |
| 92 | The scheme as described above is the fully general variant of the scheme. |
| 93 | Most of the time we are able to apply one or more of the following |
| 94 | optimizations to improve binary size or performance. |
| 95 | |
| 96 | In fact, if you try the above example with the current version of the |
| 97 | compiler, you will probably find that it will not use the described virtual |
| 98 | table layout or machine instructions. Some of the optimizations we are about |
| 99 | to introduce cause the compiler to use a different layout or a different |
| 100 | sequence of machine instructions. |
| 101 | |
| 102 | Stripping Leading/Trailing Zeros in Bit Vectors |
| 103 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 104 | |
| 105 | If a bit vector contains leading or trailing zeros, we can strip them from |
| 106 | the vector. The compiler will emit code to check if the pointer is in range |
| 107 | of the region covered by ones, and perform the bit vector check using a |
| 108 | truncated version of the bit vector. For example, the bit vectors for our |
| 109 | example class hierarchy will be emitted like this: |
| 110 | |
| 111 | .. csv-table:: Bit Vectors for A, B, C |
| 112 | :header: Class, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 |
| 113 | |
| 114 | A, , , 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, , |
| 115 | B, , , , , , , , 1, , , , , , , |
| 116 | C, , , , , , , , , , , , , 1, , |
| 117 | |
| 118 | Short Inline Bit Vectors |
| 119 | ~~~~~~~~~~~~~~~~~~~~~~~~ |
| 120 | |
| 121 | If the vector is sufficiently short, we can represent it as an inline constant |
| 122 | on x86. This saves us a few instructions when reading the correct element |
| 123 | of the bit vector. |
| 124 | |
| 125 | If the bit vector fits in 32 bits, the code looks like this: |
| 126 | |
| 127 | .. code-block:: none |
| 128 | |
| 129 | dc2: 48 8b 03 mov (%rbx),%rax |
| 130 | dc5: 48 8d 15 14 1e 00 00 lea 0x1e14(%rip),%rdx |
| 131 | dcc: 48 89 c1 mov %rax,%rcx |
| 132 | dcf: 48 29 d1 sub %rdx,%rcx |
| 133 | dd2: 48 c1 c1 3d rol $0x3d,%rcx |
| 134 | dd6: 48 83 f9 03 cmp $0x3,%rcx |
| 135 | dda: 77 2f ja e0b <main+0x9b> |
| 136 | ddc: ba 09 00 00 00 mov $0x9,%edx |
| 137 | de1: 0f a3 ca bt %ecx,%edx |
| 138 | de4: 73 25 jae e0b <main+0x9b> |
| 139 | de6: 48 89 df mov %rbx,%rdi |
| 140 | de9: ff 10 callq *(%rax) |
| 141 | [...] |
| 142 | e0b: 0f 0b ud2 |
| 143 | |
| 144 | Or if the bit vector fits in 64 bits: |
| 145 | |
| 146 | .. code-block:: none |
| 147 | |
| 148 | 11a6: 48 8b 03 mov (%rbx),%rax |
| 149 | 11a9: 48 8d 15 d0 28 00 00 lea 0x28d0(%rip),%rdx |
| 150 | 11b0: 48 89 c1 mov %rax,%rcx |
| 151 | 11b3: 48 29 d1 sub %rdx,%rcx |
| 152 | 11b6: 48 c1 c1 3d rol $0x3d,%rcx |
| 153 | 11ba: 48 83 f9 2a cmp $0x2a,%rcx |
| 154 | 11be: 77 35 ja 11f5 <main+0xb5> |
| 155 | 11c0: 48 ba 09 00 00 00 00 movabs $0x40000000009,%rdx |
| 156 | 11c7: 04 00 00 |
| 157 | 11ca: 48 0f a3 ca bt %rcx,%rdx |
| 158 | 11ce: 73 25 jae 11f5 <main+0xb5> |
| 159 | 11d0: 48 89 df mov %rbx,%rdi |
| 160 | 11d3: ff 10 callq *(%rax) |
| 161 | [...] |
| 162 | 11f5: 0f 0b ud2 |
| 163 | |
| 164 | If the bit vector consists of a single bit, there is only one possible |
| 165 | virtual table, and the check can consist of a single equality comparison: |
| 166 | |
| 167 | .. code-block:: none |
| 168 | |
| 169 | 9a2: 48 8b 03 mov (%rbx),%rax |
| 170 | 9a5: 48 8d 0d a4 13 00 00 lea 0x13a4(%rip),%rcx |
| 171 | 9ac: 48 39 c8 cmp %rcx,%rax |
| 172 | 9af: 75 25 jne 9d6 <main+0x86> |
| 173 | 9b1: 48 89 df mov %rbx,%rdi |
| 174 | 9b4: ff 10 callq *(%rax) |
| 175 | [...] |
| 176 | 9d6: 0f 0b ud2 |
| 177 | |
| 178 | Virtual Table Layout |
| 179 | ~~~~~~~~~~~~~~~~~~~~ |
| 180 | |
| 181 | The compiler lays out classes of disjoint hierarchies in separate regions |
| 182 | of the object file. At worst, bit vectors in disjoint hierarchies only |
| 183 | need to cover their disjoint hierarchy. But the closer that classes in |
| 184 | sub-hierarchies are laid out to each other, the smaller the bit vectors for |
| 185 | those sub-hierarchies need to be (see "Stripping Leading/Trailing Zeros in Bit |
| 186 | Vectors" above). The `GlobalLayoutBuilder`_ class is responsible for laying |
| 187 | out the globals efficiently to minimize the sizes of the underlying bitsets. |
| 188 | |
| 189 | .. _GlobalLayoutBuilder: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Transforms/IPO/LowerBitSets.h?view=markup |
| 190 | |
| 191 | Alignment |
| 192 | ~~~~~~~~~ |
| 193 | |
| 194 | If all gaps between address points in a particular bit vector are multiples |
| 195 | of powers of 2, the compiler can compress the bit vector by strengthening |
| 196 | the alignment requirements of the virtual table pointer. For example, given |
| 197 | this class hierarchy: |
| 198 | |
| 199 | .. code-block:: c++ |
| 200 | |
| 201 | struct A { |
| 202 | virtual void f1(); |
| 203 | virtual void f2(); |
| 204 | }; |
| 205 | |
| 206 | struct B : A { |
| 207 | virtual void f1(); |
| 208 | virtual void f2(); |
| 209 | virtual void f3(); |
| 210 | virtual void f4(); |
| 211 | virtual void f5(); |
| 212 | virtual void f6(); |
| 213 | }; |
| 214 | |
| 215 | struct C : A { |
| 216 | virtual void f1(); |
| 217 | virtual void f2(); |
| 218 | }; |
| 219 | |
| 220 | The virtual tables will be laid out like this: |
| 221 | |
| 222 | .. csv-table:: Virtual Table Layout for A, B, C |
| 223 | :header: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
| 224 | |
| 225 | A::offset-to-top, &A::rtti, &A::f1, &A::f2, B::offset-to-top, &B::rtti, &B::f1, &B::f2, &B::f3, &B::f4, &B::f5, &B::f6, C::offset-to-top, &C::rtti, &C::f1, &C::f2 |
| 226 | |
| 227 | Notice that each address point for A is separated by 4 words. This lets us |
| 228 | emit a compressed bit vector for A that looks like this: |
| 229 | |
| 230 | .. csv-table:: |
| 231 | :header: 2, 6, 10, 14 |
| 232 | |
| 233 | 1, 1, 0, 1 |
| 234 | |
| 235 | At call sites, the compiler will strengthen the alignment requirements by |
| 236 | using a different rotate count. For example, on a 64-bit machine where the |
| 237 | address points are 4-word aligned (as in A from our example), the ``rol`` |
| 238 | instruction may look like this: |
| 239 | |
| 240 | .. code-block:: none |
| 241 | |
| 242 | dd2: 48 c1 c1 3b rol $0x3b,%rcx |
| 243 | |
| 244 | Padding to Powers of 2 |
| 245 | ~~~~~~~~~~~~~~~~~~~~~~ |
| 246 | |
| 247 | Of course, this alignment scheme works best if the address points are |
| 248 | in fact aligned correctly. To make this more likely to happen, we insert |
| 249 | padding between virtual tables that in many cases aligns address points to |
| 250 | a power of 2. Specifically, our padding aligns virtual tables to the next |
| 251 | highest power of 2 bytes; because address points for specific base classes |
| 252 | normally appear at fixed offsets within the virtual table, this normally |
| 253 | has the effect of aligning the address points as well. |
| 254 | |
| 255 | This scheme introduces tradeoffs between decreased space overhead for |
| 256 | instructions and bit vectors and increased overhead in the form of padding. We |
| 257 | therefore limit the amount of padding so that we align to no more than 128 |
| 258 | bytes. This number was found experimentally to provide a good tradeoff. |
| 259 | |
| 260 | Eliminating Bit Vector Checks for All-Ones Bit Vectors |
| 261 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 262 | |
| 263 | If the bit vector is all ones, the bit vector check is redundant; we simply |
| 264 | need to check that the address is in range and well aligned. This is more |
| 265 | likely to occur if the virtual tables are padded. |