xref: /linux/arch/x86/lib/retpoline.S (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
1/* SPDX-License-Identifier: GPL-2.0 */
2
3#include <linux/export.h>
4#include <linux/stringify.h>
5#include <linux/linkage.h>
6#include <asm/dwarf2.h>
7#include <asm/cpufeatures.h>
8#include <asm/alternative.h>
9#include <asm/asm-offsets.h>
10#include <asm/nospec-branch.h>
11#include <asm/unwind_hints.h>
12#include <asm/percpu.h>
13#include <asm/frame.h>
14#include <asm/nops.h>
15
16	.section .text..__x86.indirect_thunk
17
18
19.macro POLINE reg
20	ANNOTATE_INTRA_FUNCTION_CALL
21	call    .Ldo_rop_\@
22	int3
23.Ldo_rop_\@:
24	mov     %\reg, (%_ASM_SP)
25	UNWIND_HINT_FUNC
26.endm
27
28.macro RETPOLINE reg
29	POLINE \reg
30	RET
31.endm
32
33.macro THUNK reg
34
35	.align RETPOLINE_THUNK_SIZE
36SYM_INNER_LABEL(__x86_indirect_thunk_\reg, SYM_L_GLOBAL)
37	UNWIND_HINT_UNDEFINED
38	ANNOTATE_NOENDBR
39
40	ALTERNATIVE_2 __stringify(RETPOLINE \reg), \
41		      __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg; int3), X86_FEATURE_RETPOLINE_LFENCE, \
42		      __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), ALT_NOT(X86_FEATURE_RETPOLINE)
43
44.endm
45
46/*
47 * Despite being an assembler file we can't just use .irp here
48 * because __KSYM_DEPS__ only uses the C preprocessor and would
49 * only see one instance of "__x86_indirect_thunk_\reg" rather
50 * than one per register with the correct names. So we do it
51 * the simple and nasty way...
52 *
53 * Worse, you can only have a single EXPORT_SYMBOL per line,
54 * and CPP can't insert newlines, so we have to repeat everything
55 * at least twice.
56 */
57
58#define __EXPORT_THUNK(sym)	_ASM_NOKPROBE(sym); EXPORT_SYMBOL(sym)
59
60	.align RETPOLINE_THUNK_SIZE
61SYM_CODE_START(__x86_indirect_thunk_array)
62
63#define GEN(reg) THUNK reg
64#include <asm/GEN-for-each-reg.h>
65#undef GEN
66
67	.align RETPOLINE_THUNK_SIZE
68SYM_CODE_END(__x86_indirect_thunk_array)
69
70#define GEN(reg) __EXPORT_THUNK(__x86_indirect_thunk_ ## reg)
71#include <asm/GEN-for-each-reg.h>
72#undef GEN
73
74#ifdef CONFIG_CALL_DEPTH_TRACKING
75.macro CALL_THUNK reg
76	.align RETPOLINE_THUNK_SIZE
77
78SYM_INNER_LABEL(__x86_indirect_call_thunk_\reg, SYM_L_GLOBAL)
79	UNWIND_HINT_UNDEFINED
80	ANNOTATE_NOENDBR
81
82	CALL_DEPTH_ACCOUNT
83	POLINE \reg
84	ANNOTATE_UNRET_SAFE
85	ret
86	int3
87.endm
88
89	.align RETPOLINE_THUNK_SIZE
90SYM_CODE_START(__x86_indirect_call_thunk_array)
91
92#define GEN(reg) CALL_THUNK reg
93#include <asm/GEN-for-each-reg.h>
94#undef GEN
95
96	.align RETPOLINE_THUNK_SIZE
97SYM_CODE_END(__x86_indirect_call_thunk_array)
98
99#define GEN(reg) __EXPORT_THUNK(__x86_indirect_call_thunk_ ## reg)
100#include <asm/GEN-for-each-reg.h>
101#undef GEN
102
103.macro JUMP_THUNK reg
104	.align RETPOLINE_THUNK_SIZE
105
106SYM_INNER_LABEL(__x86_indirect_jump_thunk_\reg, SYM_L_GLOBAL)
107	UNWIND_HINT_UNDEFINED
108	ANNOTATE_NOENDBR
109	POLINE \reg
110	ANNOTATE_UNRET_SAFE
111	ret
112	int3
113.endm
114
115	.align RETPOLINE_THUNK_SIZE
116SYM_CODE_START(__x86_indirect_jump_thunk_array)
117
118#define GEN(reg) JUMP_THUNK reg
119#include <asm/GEN-for-each-reg.h>
120#undef GEN
121
122	.align RETPOLINE_THUNK_SIZE
123SYM_CODE_END(__x86_indirect_jump_thunk_array)
124
125#define GEN(reg) __EXPORT_THUNK(__x86_indirect_jump_thunk_ ## reg)
126#include <asm/GEN-for-each-reg.h>
127#undef GEN
128#endif
129
130#ifdef CONFIG_RETHUNK
131
132/*
133 * Be careful here: that label cannot really be removed because in
134 * some configurations and toolchains, the JMP __x86_return_thunk the
135 * compiler issues is either a short one or the compiler doesn't use
136 * relocations for same-section JMPs and that breaks the returns
137 * detection logic in apply_returns() and in objtool.
138 */
139	.section .text..__x86.return_thunk
140
141#ifdef CONFIG_CPU_SRSO
142
143/*
144 * srso_alias_untrain_ret() and srso_alias_safe_ret() are placed at
145 * special addresses:
146 *
147 * - srso_alias_untrain_ret() is 2M aligned
148 * - srso_alias_safe_ret() is also in the same 2M page but bits 2, 8, 14
149 * and 20 in its virtual address are set (while those bits in the
150 * srso_alias_untrain_ret() function are cleared).
151 *
152 * This guarantees that those two addresses will alias in the branch
153 * target buffer of Zen3/4 generations, leading to any potential
154 * poisoned entries at that BTB slot to get evicted.
155 *
156 * As a result, srso_alias_safe_ret() becomes a safe return.
157 */
158	.pushsection .text..__x86.rethunk_untrain
159SYM_CODE_START_NOALIGN(srso_alias_untrain_ret)
160	UNWIND_HINT_FUNC
161	ANNOTATE_NOENDBR
162	ASM_NOP2
163	lfence
164	jmp srso_alias_return_thunk
165SYM_FUNC_END(srso_alias_untrain_ret)
166	.popsection
167
168	.pushsection .text..__x86.rethunk_safe
169SYM_CODE_START_NOALIGN(srso_alias_safe_ret)
170	lea 8(%_ASM_SP), %_ASM_SP
171	UNWIND_HINT_FUNC
172	ANNOTATE_UNRET_SAFE
173	ret
174	int3
175SYM_FUNC_END(srso_alias_safe_ret)
176
177SYM_CODE_START_NOALIGN(srso_alias_return_thunk)
178	UNWIND_HINT_FUNC
179	ANNOTATE_NOENDBR
180	call srso_alias_safe_ret
181	ud2
182SYM_CODE_END(srso_alias_return_thunk)
183	.popsection
184
185/*
186 * SRSO untraining sequence for Zen1/2, similar to retbleed_untrain_ret()
187 * above. On kernel entry, srso_untrain_ret() is executed which is a
188 *
189 * movabs $0xccccc30824648d48,%rax
190 *
191 * and when the return thunk executes the inner label srso_safe_ret()
192 * later, it is a stack manipulation and a RET which is mispredicted and
193 * thus a "safe" one to use.
194 */
195	.align 64
196	.skip 64 - (srso_safe_ret - srso_untrain_ret), 0xcc
197SYM_CODE_START_LOCAL_NOALIGN(srso_untrain_ret)
198	ANNOTATE_NOENDBR
199	.byte 0x48, 0xb8
200
201/*
202 * This forces the function return instruction to speculate into a trap
203 * (UD2 in srso_return_thunk() below).  This RET will then mispredict
204 * and execution will continue at the return site read from the top of
205 * the stack.
206 */
207SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
208	lea 8(%_ASM_SP), %_ASM_SP
209	ret
210	int3
211	int3
212	/* end of movabs */
213	lfence
214	call srso_safe_ret
215	ud2
216SYM_CODE_END(srso_safe_ret)
217SYM_FUNC_END(srso_untrain_ret)
218
219SYM_CODE_START(srso_return_thunk)
220	UNWIND_HINT_FUNC
221	ANNOTATE_NOENDBR
222	call srso_safe_ret
223	ud2
224SYM_CODE_END(srso_return_thunk)
225
226#define JMP_SRSO_UNTRAIN_RET "jmp srso_untrain_ret"
227#define JMP_SRSO_ALIAS_UNTRAIN_RET "jmp srso_alias_untrain_ret"
228#else /* !CONFIG_CPU_SRSO */
229#define JMP_SRSO_UNTRAIN_RET "ud2"
230#define JMP_SRSO_ALIAS_UNTRAIN_RET "ud2"
231#endif /* CONFIG_CPU_SRSO */
232
233#ifdef CONFIG_CPU_UNRET_ENTRY
234
235/*
236 * Some generic notes on the untraining sequences:
237 *
238 * They are interchangeable when it comes to flushing potentially wrong
239 * RET predictions from the BTB.
240 *
241 * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the
242 * Retbleed sequence because the return sequence done there
243 * (srso_safe_ret()) is longer and the return sequence must fully nest
244 * (end before) the untraining sequence. Therefore, the untraining
245 * sequence must fully overlap the return sequence.
246 *
247 * Regarding alignment - the instructions which need to be untrained,
248 * must all start at a cacheline boundary for Zen1/2 generations. That
249 * is, instruction sequences starting at srso_safe_ret() and
250 * the respective instruction sequences at retbleed_return_thunk()
251 * must start at a cacheline boundary.
252 */
253
254/*
255 * Safety details here pertain to the AMD Zen{1,2} microarchitecture:
256 * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for
257 *    alignment within the BTB.
258 * 2) The instruction at retbleed_untrain_ret must contain, and not
259 *    end with, the 0xc3 byte of the RET.
260 * 3) STIBP must be enabled, or SMT disabled, to prevent the sibling thread
261 *    from re-poisioning the BTB prediction.
262 */
263	.align 64
264	.skip 64 - (retbleed_return_thunk - retbleed_untrain_ret), 0xcc
265SYM_CODE_START_LOCAL_NOALIGN(retbleed_untrain_ret)
266	ANNOTATE_NOENDBR
267	/*
268	 * As executed from retbleed_untrain_ret, this is:
269	 *
270	 *   TEST $0xcc, %bl
271	 *   LFENCE
272	 *   JMP retbleed_return_thunk
273	 *
274	 * Executing the TEST instruction has a side effect of evicting any BTB
275	 * prediction (potentially attacker controlled) attached to the RET, as
276	 * retbleed_return_thunk + 1 isn't an instruction boundary at the moment.
277	 */
278	.byte	0xf6
279
280	/*
281	 * As executed from retbleed_return_thunk, this is a plain RET.
282	 *
283	 * As part of the TEST above, RET is the ModRM byte, and INT3 the imm8.
284	 *
285	 * We subsequently jump backwards and architecturally execute the RET.
286	 * This creates a correct BTB prediction (type=ret), but in the
287	 * meantime we suffer Straight Line Speculation (because the type was
288	 * no branch) which is halted by the INT3.
289	 *
290	 * With SMT enabled and STIBP active, a sibling thread cannot poison
291	 * RET's prediction to a type of its choice, but can evict the
292	 * prediction due to competitive sharing. If the prediction is
293	 * evicted, retbleed_return_thunk will suffer Straight Line Speculation
294	 * which will be contained safely by the INT3.
295	 */
296SYM_INNER_LABEL(retbleed_return_thunk, SYM_L_GLOBAL)
297	ret
298	int3
299SYM_CODE_END(retbleed_return_thunk)
300
301	/*
302	 * Ensure the TEST decoding / BTB invalidation is complete.
303	 */
304	lfence
305
306	/*
307	 * Jump back and execute the RET in the middle of the TEST instruction.
308	 * INT3 is for SLS protection.
309	 */
310	jmp retbleed_return_thunk
311	int3
312SYM_FUNC_END(retbleed_untrain_ret)
313
314#define JMP_RETBLEED_UNTRAIN_RET "jmp retbleed_untrain_ret"
315#else /* !CONFIG_CPU_UNRET_ENTRY */
316#define JMP_RETBLEED_UNTRAIN_RET "ud2"
317#endif /* CONFIG_CPU_UNRET_ENTRY */
318
319#if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO)
320
321SYM_FUNC_START(entry_untrain_ret)
322	ALTERNATIVE_2 JMP_RETBLEED_UNTRAIN_RET,				\
323		      JMP_SRSO_UNTRAIN_RET, X86_FEATURE_SRSO,		\
324		      JMP_SRSO_ALIAS_UNTRAIN_RET, X86_FEATURE_SRSO_ALIAS
325SYM_FUNC_END(entry_untrain_ret)
326__EXPORT_THUNK(entry_untrain_ret)
327
328#endif /* CONFIG_CPU_UNRET_ENTRY || CONFIG_CPU_SRSO */
329
330#ifdef CONFIG_CALL_DEPTH_TRACKING
331
332	.align 64
333SYM_FUNC_START(call_depth_return_thunk)
334	ANNOTATE_NOENDBR
335	/*
336	 * Keep the hotpath in a 16byte I-fetch for the non-debug
337	 * case.
338	 */
339	CALL_THUNKS_DEBUG_INC_RETS
340	shlq	$5, PER_CPU_VAR(pcpu_hot + X86_call_depth)
341	jz	1f
342	ANNOTATE_UNRET_SAFE
343	ret
344	int3
3451:
346	CALL_THUNKS_DEBUG_INC_STUFFS
347	.rept	16
348	ANNOTATE_INTRA_FUNCTION_CALL
349	call	2f
350	int3
3512:
352	.endr
353	add	$(8*16), %rsp
354
355	CREDIT_CALL_DEPTH
356
357	ANNOTATE_UNRET_SAFE
358	ret
359	int3
360SYM_FUNC_END(call_depth_return_thunk)
361
362#endif /* CONFIG_CALL_DEPTH_TRACKING */
363
364/*
365 * This function name is magical and is used by -mfunction-return=thunk-extern
366 * for the compiler to generate JMPs to it.
367 *
368 * This code is only used during kernel boot or module init.  All
369 * 'JMP __x86_return_thunk' sites are changed to something else by
370 * apply_returns().
371 *
372 * This should be converted eventually to call a warning function which
373 * should scream loudly when the default return thunk is called after
374 * alternatives have been applied.
375 *
376 * That warning function cannot BUG() because the bug splat cannot be
377 * displayed in all possible configurations, leading to users not really
378 * knowing why the machine froze.
379 */
380SYM_CODE_START(__x86_return_thunk)
381	UNWIND_HINT_FUNC
382	ANNOTATE_NOENDBR
383	ANNOTATE_UNRET_SAFE
384	ret
385	int3
386SYM_CODE_END(__x86_return_thunk)
387EXPORT_SYMBOL(__x86_return_thunk)
388
389#endif /* CONFIG_RETHUNK */
390