xref: /linux/arch/x86/lib/retpoline.S (revision 3f41368fbfe1b3d5922d317fe1a0a0cab6846802)
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_MITIGATION_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_MITIGATION_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_MITIGATION_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__EXPORT_THUNK(srso_alias_untrain_ret)
167	.popsection
168
169	.pushsection .text..__x86.rethunk_safe
170SYM_CODE_START_NOALIGN(srso_alias_safe_ret)
171	lea 8(%_ASM_SP), %_ASM_SP
172	UNWIND_HINT_FUNC
173	ANNOTATE_UNRET_SAFE
174	ret
175	int3
176SYM_FUNC_END(srso_alias_safe_ret)
177
178SYM_CODE_START_NOALIGN(srso_alias_return_thunk)
179	UNWIND_HINT_FUNC
180	ANNOTATE_NOENDBR
181	call srso_alias_safe_ret
182	ud2
183SYM_CODE_END(srso_alias_return_thunk)
184	.popsection
185
186/*
187 * SRSO untraining sequence for Zen1/2, similar to retbleed_untrain_ret()
188 * above. On kernel entry, srso_untrain_ret() is executed which is a
189 *
190 * movabs $0xccccc30824648d48,%rax
191 *
192 * and when the return thunk executes the inner label srso_safe_ret()
193 * later, it is a stack manipulation and a RET which is mispredicted and
194 * thus a "safe" one to use.
195 */
196	.align 64
197	.skip 64 - (srso_safe_ret - srso_untrain_ret), 0xcc
198SYM_CODE_START_LOCAL_NOALIGN(srso_untrain_ret)
199	ANNOTATE_NOENDBR
200	.byte 0x48, 0xb8
201
202/*
203 * This forces the function return instruction to speculate into a trap
204 * (UD2 in srso_return_thunk() below).  This RET will then mispredict
205 * and execution will continue at the return site read from the top of
206 * the stack.
207 */
208SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
209	lea 8(%_ASM_SP), %_ASM_SP
210	ret
211	int3
212	int3
213	/* end of movabs */
214	lfence
215	call srso_safe_ret
216	ud2
217SYM_CODE_END(srso_safe_ret)
218SYM_FUNC_END(srso_untrain_ret)
219
220SYM_CODE_START(srso_return_thunk)
221	UNWIND_HINT_FUNC
222	ANNOTATE_NOENDBR
223	call srso_safe_ret
224	ud2
225SYM_CODE_END(srso_return_thunk)
226
227#define JMP_SRSO_UNTRAIN_RET "jmp srso_untrain_ret"
228#else /* !CONFIG_MITIGATION_SRSO */
229/* Dummy for the alternative in CALL_UNTRAIN_RET. */
230SYM_CODE_START(srso_alias_untrain_ret)
231	ANNOTATE_UNRET_SAFE
232	ANNOTATE_NOENDBR
233	ret
234	int3
235SYM_FUNC_END(srso_alias_untrain_ret)
236__EXPORT_THUNK(srso_alias_untrain_ret)
237#define JMP_SRSO_UNTRAIN_RET "ud2"
238#endif /* CONFIG_MITIGATION_SRSO */
239
240#ifdef CONFIG_MITIGATION_UNRET_ENTRY
241
242/*
243 * Some generic notes on the untraining sequences:
244 *
245 * They are interchangeable when it comes to flushing potentially wrong
246 * RET predictions from the BTB.
247 *
248 * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the
249 * Retbleed sequence because the return sequence done there
250 * (srso_safe_ret()) is longer and the return sequence must fully nest
251 * (end before) the untraining sequence. Therefore, the untraining
252 * sequence must fully overlap the return sequence.
253 *
254 * Regarding alignment - the instructions which need to be untrained,
255 * must all start at a cacheline boundary for Zen1/2 generations. That
256 * is, instruction sequences starting at srso_safe_ret() and
257 * the respective instruction sequences at retbleed_return_thunk()
258 * must start at a cacheline boundary.
259 */
260
261/*
262 * Safety details here pertain to the AMD Zen{1,2} microarchitecture:
263 * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for
264 *    alignment within the BTB.
265 * 2) The instruction at retbleed_untrain_ret must contain, and not
266 *    end with, the 0xc3 byte of the RET.
267 * 3) STIBP must be enabled, or SMT disabled, to prevent the sibling thread
268 *    from re-poisioning the BTB prediction.
269 */
270	.align 64
271	.skip 64 - (retbleed_return_thunk - retbleed_untrain_ret), 0xcc
272SYM_CODE_START_LOCAL_NOALIGN(retbleed_untrain_ret)
273	ANNOTATE_NOENDBR
274	/*
275	 * As executed from retbleed_untrain_ret, this is:
276	 *
277	 *   TEST $0xcc, %bl
278	 *   LFENCE
279	 *   JMP retbleed_return_thunk
280	 *
281	 * Executing the TEST instruction has a side effect of evicting any BTB
282	 * prediction (potentially attacker controlled) attached to the RET, as
283	 * retbleed_return_thunk + 1 isn't an instruction boundary at the moment.
284	 */
285	.byte	0xf6
286
287	/*
288	 * As executed from retbleed_return_thunk, this is a plain RET.
289	 *
290	 * As part of the TEST above, RET is the ModRM byte, and INT3 the imm8.
291	 *
292	 * We subsequently jump backwards and architecturally execute the RET.
293	 * This creates a correct BTB prediction (type=ret), but in the
294	 * meantime we suffer Straight Line Speculation (because the type was
295	 * no branch) which is halted by the INT3.
296	 *
297	 * With SMT enabled and STIBP active, a sibling thread cannot poison
298	 * RET's prediction to a type of its choice, but can evict the
299	 * prediction due to competitive sharing. If the prediction is
300	 * evicted, retbleed_return_thunk will suffer Straight Line Speculation
301	 * which will be contained safely by the INT3.
302	 */
303SYM_INNER_LABEL(retbleed_return_thunk, SYM_L_GLOBAL)
304	ret
305	int3
306SYM_CODE_END(retbleed_return_thunk)
307
308	/*
309	 * Ensure the TEST decoding / BTB invalidation is complete.
310	 */
311	lfence
312
313	/*
314	 * Jump back and execute the RET in the middle of the TEST instruction.
315	 * INT3 is for SLS protection.
316	 */
317	jmp retbleed_return_thunk
318	int3
319SYM_FUNC_END(retbleed_untrain_ret)
320
321#define JMP_RETBLEED_UNTRAIN_RET "jmp retbleed_untrain_ret"
322#else /* !CONFIG_MITIGATION_UNRET_ENTRY */
323#define JMP_RETBLEED_UNTRAIN_RET "ud2"
324#endif /* CONFIG_MITIGATION_UNRET_ENTRY */
325
326#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || defined(CONFIG_MITIGATION_SRSO)
327
328SYM_FUNC_START(entry_untrain_ret)
329	ALTERNATIVE JMP_RETBLEED_UNTRAIN_RET, JMP_SRSO_UNTRAIN_RET, X86_FEATURE_SRSO
330SYM_FUNC_END(entry_untrain_ret)
331__EXPORT_THUNK(entry_untrain_ret)
332
333#endif /* CONFIG_MITIGATION_UNRET_ENTRY || CONFIG_MITIGATION_SRSO */
334
335#ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING
336
337	.align 64
338SYM_FUNC_START(call_depth_return_thunk)
339	ANNOTATE_NOENDBR
340	/*
341	 * Keep the hotpath in a 16byte I-fetch for the non-debug
342	 * case.
343	 */
344	CALL_THUNKS_DEBUG_INC_RETS
345	shlq	$5, PER_CPU_VAR(pcpu_hot + X86_call_depth)
346	jz	1f
347	ANNOTATE_UNRET_SAFE
348	ret
349	int3
3501:
351	CALL_THUNKS_DEBUG_INC_STUFFS
352	.rept	16
353	ANNOTATE_INTRA_FUNCTION_CALL
354	call	2f
355	int3
3562:
357	.endr
358	add	$(8*16), %rsp
359
360	CREDIT_CALL_DEPTH
361
362	ANNOTATE_UNRET_SAFE
363	ret
364	int3
365SYM_FUNC_END(call_depth_return_thunk)
366
367#endif /* CONFIG_MITIGATION_CALL_DEPTH_TRACKING */
368
369/*
370 * This function name is magical and is used by -mfunction-return=thunk-extern
371 * for the compiler to generate JMPs to it.
372 *
373 * This code is only used during kernel boot or module init.  All
374 * 'JMP __x86_return_thunk' sites are changed to something else by
375 * apply_returns().
376 *
377 * The ALTERNATIVE below adds a really loud warning to catch the case
378 * where the insufficient default return thunk ends up getting used for
379 * whatever reason like miscompilation or failure of
380 * objtool/alternatives/etc to patch all the return sites.
381 */
382SYM_CODE_START(__x86_return_thunk)
383	UNWIND_HINT_FUNC
384	ANNOTATE_NOENDBR
385#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
386    defined(CONFIG_MITIGATION_SRSO) || \
387    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
388	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
389		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
390#else
391	ANNOTATE_UNRET_SAFE
392	ret
393#endif
394	int3
395SYM_CODE_END(__x86_return_thunk)
396EXPORT_SYMBOL(__x86_return_thunk)
397
398#endif /* CONFIG_MITIGATION_RETHUNK */
399