1 /* SPDX-License-Identifier: GPL-2.0 */
3 #include <linux/stringify.h>
4 #include <linux/linkage.h>
5 #include <asm/dwarf2.h>
6 #include <asm/cpufeatures.h>
7 #include <asm/alternative.h>
8 #include <asm/export.h>
9 #include <asm/nospec-branch.h>
10 #include <asm/unwind_hints.h>
11 #include <asm/frame.h>
14 .section .text..__x86.indirect_thunk
17 ANNOTATE_INTRA_FUNCTION_CALL
32 .align RETPOLINE_THUNK_SIZE
33 SYM_INNER_LABEL(__x86_indirect_thunk_\reg, SYM_L_GLOBAL)
37 ALTERNATIVE_2 __stringify(RETPOLINE \reg), \
38 __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg; int3), X86_FEATURE_RETPOLINE_LFENCE, \
39 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), ALT_NOT(X86_FEATURE_RETPOLINE)
44 * Despite being an assembler file we can't just use .irp here
45 * because __KSYM_DEPS__ only uses the C preprocessor and would
46 * only see one instance of "__x86_indirect_thunk_\reg" rather
47 * than one per register with the correct names. So we do it
48 * the simple and nasty way...
50 * Worse, you can only have a single EXPORT_SYMBOL per line,
51 * and CPP can't insert newlines, so we have to repeat everything
55 #define __EXPORT_THUNK(sym) _ASM_NOKPROBE(sym); EXPORT_SYMBOL(sym)
56 #define EXPORT_THUNK(reg) __EXPORT_THUNK(__x86_indirect_thunk_ ## reg)
58 .align RETPOLINE_THUNK_SIZE
59 SYM_CODE_START(__x86_indirect_thunk_array)
61 #define GEN(reg) THUNK reg
62 #include <asm/GEN-for-each-reg.h>
65 .align RETPOLINE_THUNK_SIZE
66 SYM_CODE_END(__x86_indirect_thunk_array)
68 #define GEN(reg) EXPORT_THUNK(reg)
69 #include <asm/GEN-for-each-reg.h>
73 * This function name is magical and is used by -mfunction-return=thunk-extern
74 * for the compiler to generate JMPs to it.
79 * srso_alias_untrain_ret() and srso_alias_safe_ret() are placed at
82 * - srso_alias_untrain_ret() is 2M aligned
83 * - srso_alias_safe_ret() is also in the same 2M page but bits 2, 8, 14
84 * and 20 in its virtual address are set (while those bits in the
85 * srso_alias_untrain_ret() function are cleared).
87 * This guarantees that those two addresses will alias in the branch
88 * target buffer of Zen3/4 generations, leading to any potential
89 * poisoned entries at that BTB slot to get evicted.
91 * As a result, srso_alias_safe_ret() becomes a safe return.
93 #ifdef CONFIG_CPU_SRSO
94 .section .text..__x86.rethunk_untrain
96 SYM_START(srso_alias_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
101 jmp srso_alias_return_thunk
102 SYM_FUNC_END(srso_alias_untrain_ret)
103 __EXPORT_THUNK(srso_alias_untrain_ret)
105 .section .text..__x86.rethunk_safe
107 /* dummy definition for alternatives */
108 SYM_START(srso_alias_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
112 SYM_FUNC_END(srso_alias_untrain_ret)
113 __EXPORT_THUNK(srso_alias_untrain_ret)
116 SYM_START(srso_alias_safe_ret, SYM_L_GLOBAL, SYM_A_NONE)
117 lea 8(%_ASM_SP), %_ASM_SP
122 SYM_FUNC_END(srso_alias_safe_ret)
124 .section .text..__x86.return_thunk
126 SYM_CODE_START(srso_alias_return_thunk)
129 call srso_alias_safe_ret
131 SYM_CODE_END(srso_alias_return_thunk)
134 * Some generic notes on the untraining sequences:
136 * They are interchangeable when it comes to flushing potentially wrong
137 * RET predictions from the BTB.
139 * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the
140 * Retbleed sequence because the return sequence done there
141 * (srso_safe_ret()) is longer and the return sequence must fully nest
142 * (end before) the untraining sequence. Therefore, the untraining
143 * sequence must fully overlap the return sequence.
145 * Regarding alignment - the instructions which need to be untrained,
146 * must all start at a cacheline boundary for Zen1/2 generations. That
147 * is, instruction sequences starting at srso_safe_ret() and
148 * the respective instruction sequences at retbleed_return_thunk()
149 * must start at a cacheline boundary.
153 * Safety details here pertain to the AMD Zen{1,2} microarchitecture:
154 * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for
155 * alignment within the BTB.
156 * 2) The instruction at retbleed_untrain_ret must contain, and not
157 * end with, the 0xc3 byte of the RET.
158 * 3) STIBP must be enabled, or SMT disabled, to prevent the sibling thread
159 * from re-poisioning the BTB prediction.
162 .skip 64 - (retbleed_return_thunk - retbleed_untrain_ret), 0xcc
163 SYM_START(retbleed_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
166 * As executed from retbleed_untrain_ret, this is:
170 * JMP retbleed_return_thunk
172 * Executing the TEST instruction has a side effect of evicting any BTB
173 * prediction (potentially attacker controlled) attached to the RET, as
174 * retbleed_return_thunk + 1 isn't an instruction boundary at the moment.
179 * As executed from retbleed_return_thunk, this is a plain RET.
181 * As part of the TEST above, RET is the ModRM byte, and INT3 the imm8.
183 * We subsequently jump backwards and architecturally execute the RET.
184 * This creates a correct BTB prediction (type=ret), but in the
185 * meantime we suffer Straight Line Speculation (because the type was
186 * no branch) which is halted by the INT3.
188 * With SMT enabled and STIBP active, a sibling thread cannot poison
189 * RET's prediction to a type of its choice, but can evict the
190 * prediction due to competitive sharing. If the prediction is
191 * evicted, retbleed_return_thunk will suffer Straight Line Speculation
192 * which will be contained safely by the INT3.
194 SYM_INNER_LABEL(retbleed_return_thunk, SYM_L_GLOBAL)
197 SYM_CODE_END(retbleed_return_thunk)
200 * Ensure the TEST decoding / BTB invalidation is complete.
205 * Jump back and execute the RET in the middle of the TEST instruction.
206 * INT3 is for SLS protection.
208 jmp retbleed_return_thunk
210 SYM_FUNC_END(retbleed_untrain_ret)
211 __EXPORT_THUNK(retbleed_untrain_ret)
214 * SRSO untraining sequence for Zen1/2, similar to retbleed_untrain_ret()
215 * above. On kernel entry, srso_untrain_ret() is executed which is a
217 * movabs $0xccccc30824648d48,%rax
219 * and when the return thunk executes the inner label srso_safe_ret()
220 * later, it is a stack manipulation and a RET which is mispredicted and
221 * thus a "safe" one to use.
224 .skip 64 - (srso_safe_ret - srso_untrain_ret), 0xcc
225 SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
230 * This forces the function return instruction to speculate into a trap
231 * (UD2 in srso_return_thunk() below). This RET will then mispredict
232 * and execution will continue at the return site read from the top of
235 SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
236 lea 8(%_ASM_SP), %_ASM_SP
244 SYM_CODE_END(srso_safe_ret)
245 SYM_FUNC_END(srso_untrain_ret)
246 __EXPORT_THUNK(srso_untrain_ret)
248 SYM_CODE_START(srso_return_thunk)
253 SYM_CODE_END(srso_return_thunk)
255 SYM_FUNC_START(entry_untrain_ret)
256 ALTERNATIVE "jmp retbleed_untrain_ret", "jmp srso_untrain_ret", X86_FEATURE_SRSO
257 SYM_FUNC_END(entry_untrain_ret)
258 __EXPORT_THUNK(entry_untrain_ret)
260 SYM_CODE_START(__x86_return_thunk)
267 SYM_CODE_END(__x86_return_thunk)
268 EXPORT_SYMBOL(__x86_return_thunk)
270 #endif /* CONFIG_RETHUNK */