.. | .. |
---|
9 | 9 | #include <asm/nospec-branch.h> |
---|
10 | 10 | #include <asm/unwind_hints.h> |
---|
11 | 11 | #include <asm/frame.h> |
---|
| 12 | +#include <asm/nops.h> |
---|
12 | 13 | |
---|
13 | | - .section .text.__x86.indirect_thunk |
---|
| 14 | + .section .text..__x86.indirect_thunk |
---|
14 | 15 | |
---|
15 | 16 | .macro RETPOLINE reg |
---|
16 | 17 | ANNOTATE_INTRA_FUNCTION_CALL |
---|
.. | .. |
---|
73 | 74 | */ |
---|
74 | 75 | #ifdef CONFIG_RETHUNK |
---|
75 | 76 | |
---|
76 | | - .section .text.__x86.return_thunk |
---|
| 77 | +/* |
---|
| 78 | + * srso_alias_untrain_ret() and srso_alias_safe_ret() are placed at |
---|
| 79 | + * special addresses: |
---|
| 80 | + * |
---|
| 81 | + * - srso_alias_untrain_ret() is 2M aligned |
---|
| 82 | + * - srso_alias_safe_ret() is also in the same 2M page but bits 2, 8, 14 |
---|
| 83 | + * and 20 in its virtual address are set (while those bits in the |
---|
| 84 | + * srso_alias_untrain_ret() function are cleared). |
---|
| 85 | + * |
---|
| 86 | + * This guarantees that those two addresses will alias in the branch |
---|
| 87 | + * target buffer of Zen3/4 generations, leading to any potential |
---|
| 88 | + * poisoned entries at that BTB slot to get evicted. |
---|
| 89 | + * |
---|
| 90 | + * As a result, srso_alias_safe_ret() becomes a safe return. |
---|
| 91 | + */ |
---|
| 92 | +#ifdef CONFIG_CPU_SRSO |
---|
| 93 | + .section .text..__x86.rethunk_untrain |
---|
| 94 | + |
---|
| 95 | +SYM_START(srso_alias_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE) |
---|
| 96 | + UNWIND_HINT_FUNC |
---|
| 97 | + ASM_NOP2 |
---|
| 98 | + lfence |
---|
| 99 | + jmp srso_alias_return_thunk |
---|
| 100 | +SYM_FUNC_END(srso_alias_untrain_ret) |
---|
| 101 | +__EXPORT_THUNK(srso_alias_untrain_ret) |
---|
| 102 | + |
---|
| 103 | + .section .text..__x86.rethunk_safe |
---|
| 104 | +#else |
---|
| 105 | +/* dummy definition for alternatives */ |
---|
| 106 | +SYM_START(srso_alias_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE) |
---|
| 107 | + ANNOTATE_UNRET_SAFE |
---|
| 108 | + ret |
---|
| 109 | + int3 |
---|
| 110 | +SYM_FUNC_END(srso_alias_untrain_ret) |
---|
| 111 | +#endif |
---|
| 112 | + |
---|
| 113 | +SYM_START(srso_alias_safe_ret, SYM_L_GLOBAL, SYM_A_NONE) |
---|
| 114 | + lea 8(%_ASM_SP), %_ASM_SP |
---|
| 115 | + UNWIND_HINT_FUNC |
---|
| 116 | + ANNOTATE_UNRET_SAFE |
---|
| 117 | + ret |
---|
| 118 | + int3 |
---|
| 119 | +SYM_FUNC_END(srso_alias_safe_ret) |
---|
| 120 | + |
---|
| 121 | + .section .text..__x86.return_thunk |
---|
| 122 | + |
---|
| 123 | +SYM_CODE_START(srso_alias_return_thunk) |
---|
| 124 | + UNWIND_HINT_FUNC |
---|
| 125 | + ANNOTATE_NOENDBR |
---|
| 126 | + call srso_alias_safe_ret |
---|
| 127 | + ud2 |
---|
| 128 | +SYM_CODE_END(srso_alias_return_thunk) |
---|
| 129 | + |
---|
| 130 | +/* |
---|
| 131 | + * Some generic notes on the untraining sequences: |
---|
| 132 | + * |
---|
| 133 | + * They are interchangeable when it comes to flushing potentially wrong |
---|
| 134 | + * RET predictions from the BTB. |
---|
| 135 | + * |
---|
| 136 | + * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the |
---|
| 137 | + * Retbleed sequence because the return sequence done there |
---|
| 138 | + * (srso_safe_ret()) is longer and the return sequence must fully nest |
---|
| 139 | + * (end before) the untraining sequence. Therefore, the untraining |
---|
| 140 | + * sequence must fully overlap the return sequence. |
---|
| 141 | + * |
---|
| 142 | + * Regarding alignment - the instructions which need to be untrained, |
---|
| 143 | + * must all start at a cacheline boundary for Zen1/2 generations. That |
---|
| 144 | + * is, instruction sequences starting at srso_safe_ret() and |
---|
| 145 | + * the respective instruction sequences at retbleed_return_thunk() |
---|
| 146 | + * must start at a cacheline boundary. |
---|
| 147 | + */ |
---|
77 | 148 | |
---|
78 | 149 | /* |
---|
79 | 150 | * Safety details here pertain to the AMD Zen{1,2} microarchitecture: |
---|
80 | | - * 1) The RET at __x86_return_thunk must be on a 64 byte boundary, for |
---|
| 151 | + * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for |
---|
81 | 152 | * alignment within the BTB. |
---|
82 | | - * 2) The instruction at zen_untrain_ret must contain, and not |
---|
| 153 | + * 2) The instruction at retbleed_untrain_ret must contain, and not |
---|
83 | 154 | * end with, the 0xc3 byte of the RET. |
---|
84 | 155 | * 3) STIBP must be enabled, or SMT disabled, to prevent the sibling thread |
---|
85 | 156 | * from re-poisioning the BTB prediction. |
---|
86 | 157 | */ |
---|
87 | 158 | .align 64 |
---|
88 | | - .skip 63, 0xcc |
---|
89 | | -SYM_FUNC_START_NOALIGN(zen_untrain_ret); |
---|
| 159 | + .skip 64 - (retbleed_return_thunk - retbleed_untrain_ret), 0xcc |
---|
| 160 | +SYM_FUNC_START_NOALIGN(retbleed_untrain_ret); |
---|
90 | 161 | |
---|
91 | 162 | /* |
---|
92 | | - * As executed from zen_untrain_ret, this is: |
---|
| 163 | + * As executed from retbleed_untrain_ret, this is: |
---|
93 | 164 | * |
---|
94 | 165 | * TEST $0xcc, %bl |
---|
95 | 166 | * LFENCE |
---|
96 | | - * JMP __x86_return_thunk |
---|
| 167 | + * JMP retbleed_return_thunk |
---|
97 | 168 | * |
---|
98 | 169 | * Executing the TEST instruction has a side effect of evicting any BTB |
---|
99 | 170 | * prediction (potentially attacker controlled) attached to the RET, as |
---|
100 | | - * __x86_return_thunk + 1 isn't an instruction boundary at the moment. |
---|
| 171 | + * retbleed_return_thunk + 1 isn't an instruction boundary at the moment. |
---|
101 | 172 | */ |
---|
102 | 173 | .byte 0xf6 |
---|
103 | 174 | |
---|
104 | 175 | /* |
---|
105 | | - * As executed from __x86_return_thunk, this is a plain RET. |
---|
| 176 | + * As executed from retbleed_return_thunk, this is a plain RET. |
---|
106 | 177 | * |
---|
107 | 178 | * As part of the TEST above, RET is the ModRM byte, and INT3 the imm8. |
---|
108 | 179 | * |
---|
.. | .. |
---|
114 | 185 | * With SMT enabled and STIBP active, a sibling thread cannot poison |
---|
115 | 186 | * RET's prediction to a type of its choice, but can evict the |
---|
116 | 187 | * prediction due to competitive sharing. If the prediction is |
---|
117 | | - * evicted, __x86_return_thunk will suffer Straight Line Speculation |
---|
| 188 | + * evicted, retbleed_return_thunk will suffer Straight Line Speculation |
---|
118 | 189 | * which will be contained safely by the INT3. |
---|
119 | 190 | */ |
---|
120 | | -SYM_INNER_LABEL(__x86_return_thunk, SYM_L_GLOBAL) |
---|
| 191 | +SYM_INNER_LABEL(retbleed_return_thunk, SYM_L_GLOBAL) |
---|
121 | 192 | ret |
---|
122 | 193 | int3 |
---|
123 | | -SYM_CODE_END(__x86_return_thunk) |
---|
| 194 | +SYM_CODE_END(retbleed_return_thunk) |
---|
124 | 195 | |
---|
125 | 196 | /* |
---|
126 | 197 | * Ensure the TEST decoding / BTB invalidation is complete. |
---|
.. | .. |
---|
131 | 202 | * Jump back and execute the RET in the middle of the TEST instruction. |
---|
132 | 203 | * INT3 is for SLS protection. |
---|
133 | 204 | */ |
---|
134 | | - jmp __x86_return_thunk |
---|
| 205 | + jmp retbleed_return_thunk |
---|
135 | 206 | int3 |
---|
136 | | -SYM_FUNC_END(zen_untrain_ret) |
---|
137 | | -__EXPORT_THUNK(zen_untrain_ret) |
---|
| 207 | +SYM_FUNC_END(retbleed_untrain_ret) |
---|
| 208 | +__EXPORT_THUNK(retbleed_untrain_ret) |
---|
138 | 209 | |
---|
| 210 | +/* |
---|
| 211 | + * SRSO untraining sequence for Zen1/2, similar to retbleed_untrain_ret() |
---|
| 212 | + * above. On kernel entry, srso_untrain_ret() is executed which is a |
---|
| 213 | + * |
---|
| 214 | + * movabs $0xccccc30824648d48,%rax |
---|
| 215 | + * |
---|
| 216 | + * and when the return thunk executes the inner label srso_safe_ret() |
---|
| 217 | + * later, it is a stack manipulation and a RET which is mispredicted and |
---|
| 218 | + * thus a "safe" one to use. |
---|
| 219 | + */ |
---|
| 220 | + .align 64 |
---|
| 221 | + .skip 64 - (srso_safe_ret - srso_untrain_ret), 0xcc |
---|
| 222 | +SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE) |
---|
| 223 | + .byte 0x48, 0xb8 |
---|
| 224 | + |
---|
| 225 | +/* |
---|
| 226 | + * This forces the function return instruction to speculate into a trap |
---|
| 227 | + * (UD2 in srso_return_thunk() below). This RET will then mispredict |
---|
| 228 | + * and execution will continue at the return site read from the top of |
---|
| 229 | + * the stack. |
---|
| 230 | + */ |
---|
| 231 | +SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL) |
---|
| 232 | + lea 8(%_ASM_SP), %_ASM_SP |
---|
| 233 | + ret |
---|
| 234 | + int3 |
---|
| 235 | + int3 |
---|
| 236 | + /* end of movabs */ |
---|
| 237 | + lfence |
---|
| 238 | + call srso_safe_ret |
---|
| 239 | + ud2 |
---|
| 240 | +SYM_CODE_END(srso_safe_ret) |
---|
| 241 | +SYM_FUNC_END(srso_untrain_ret) |
---|
| 242 | +__EXPORT_THUNK(srso_untrain_ret) |
---|
| 243 | + |
---|
| 244 | +SYM_CODE_START(srso_return_thunk) |
---|
| 245 | + UNWIND_HINT_FUNC |
---|
| 246 | + ANNOTATE_NOENDBR |
---|
| 247 | + call srso_safe_ret |
---|
| 248 | + ud2 |
---|
| 249 | +SYM_CODE_END(srso_return_thunk) |
---|
| 250 | + |
---|
| 251 | +SYM_FUNC_START(entry_untrain_ret) |
---|
| 252 | + ALTERNATIVE_2 "jmp retbleed_untrain_ret", \ |
---|
| 253 | + "jmp srso_untrain_ret", X86_FEATURE_SRSO, \ |
---|
| 254 | + "jmp srso_alias_untrain_ret", X86_FEATURE_SRSO_ALIAS |
---|
| 255 | +SYM_FUNC_END(entry_untrain_ret) |
---|
| 256 | +__EXPORT_THUNK(entry_untrain_ret) |
---|
| 257 | + |
---|
| 258 | +SYM_CODE_START(__x86_return_thunk) |
---|
| 259 | + UNWIND_HINT_FUNC |
---|
| 260 | + ANNOTATE_NOENDBR |
---|
| 261 | + ANNOTATE_UNRET_SAFE |
---|
| 262 | + ret |
---|
| 263 | + int3 |
---|
| 264 | +SYM_CODE_END(__x86_return_thunk) |
---|
139 | 265 | EXPORT_SYMBOL(__x86_return_thunk) |
---|
140 | 266 | |
---|
141 | 267 | #endif /* CONFIG_RETHUNK */ |
---|