Merge tag 'cgroup-for-5.19-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel...
[platform/kernel/linux-starfive.git] / arch / x86 / include / asm / nospec-branch.h
1 /* SPDX-License-Identifier: GPL-2.0 */
2
3 #ifndef _ASM_X86_NOSPEC_BRANCH_H_
4 #define _ASM_X86_NOSPEC_BRANCH_H_
5
6 #include <linux/static_key.h>
7 #include <linux/objtool.h>
8 #include <linux/linkage.h>
9
10 #include <asm/alternative.h>
11 #include <asm/cpufeatures.h>
12 #include <asm/msr-index.h>
13 #include <asm/unwind_hints.h>
14
15 #define RETPOLINE_THUNK_SIZE    32
16
17 /*
18  * Fill the CPU return stack buffer.
19  *
20  * Each entry in the RSB, if used for a speculative 'ret', contains an
21  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
22  *
23  * This is required in various cases for retpoline and IBRS-based
24  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
25  * eliminate potentially bogus entries from the RSB, and sometimes
26  * purely to ensure that it doesn't get empty, which on some CPUs would
27  * allow predictions from other (unwanted!) sources to be used.
28  *
29  * We define a CPP macro such that it can be used from both .S files and
30  * inline assembly. It's possible to do a .macro and then include that
31  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
32  */
33
34 #define RSB_CLEAR_LOOPS         32      /* To forcibly overwrite all entries */
35
36 /*
37  * Google experimented with loop-unrolling and this turned out to be
38  * the optimal version - two calls, each with their own speculation
39  * trap should their return address end up getting used, in a loop.
40  */
41 #define __FILL_RETURN_BUFFER(reg, nr, sp)       \
42         mov     $(nr/2), reg;                   \
43 771:                                            \
44         ANNOTATE_INTRA_FUNCTION_CALL;           \
45         call    772f;                           \
46 773:    /* speculation trap */                  \
47         UNWIND_HINT_EMPTY;                      \
48         pause;                                  \
49         lfence;                                 \
50         jmp     773b;                           \
51 772:                                            \
52         ANNOTATE_INTRA_FUNCTION_CALL;           \
53         call    774f;                           \
54 775:    /* speculation trap */                  \
55         UNWIND_HINT_EMPTY;                      \
56         pause;                                  \
57         lfence;                                 \
58         jmp     775b;                           \
59 774:                                            \
60         add     $(BITS_PER_LONG/8) * 2, sp;     \
61         dec     reg;                            \
62         jnz     771b;
63
64 #ifdef __ASSEMBLY__
65
66 /*
67  * This should be used immediately before an indirect jump/call. It tells
68  * objtool the subsequent indirect jump/call is vouched safe for retpoline
69  * builds.
70  */
71 .macro ANNOTATE_RETPOLINE_SAFE
72         .Lannotate_\@:
73         .pushsection .discard.retpoline_safe
74         _ASM_PTR .Lannotate_\@
75         .popsection
76 .endm
77
78 /*
79  * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions
80  * vs RETBleed validation.
81  */
82 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE
83
84 /*
85  * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should
86  * eventually turn into it's own annotation.
87  */
88 .macro ANNOTATE_UNRET_END
89 #ifdef CONFIG_DEBUG_ENTRY
90         ANNOTATE_RETPOLINE_SAFE
91         nop
92 #endif
93 .endm
94
95 /*
96  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
97  * indirect jmp/call which may be susceptible to the Spectre variant 2
98  * attack.
99  */
100 .macro JMP_NOSPEC reg:req
101 #ifdef CONFIG_RETPOLINE
102         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), \
103                       __stringify(jmp __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
104                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), X86_FEATURE_RETPOLINE_LFENCE
105 #else
106         jmp     *%\reg
107 #endif
108 .endm
109
110 .macro CALL_NOSPEC reg:req
111 #ifdef CONFIG_RETPOLINE
112         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; call *%\reg), \
113                       __stringify(call __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
114                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; call *%\reg), X86_FEATURE_RETPOLINE_LFENCE
115 #else
116         call    *%\reg
117 #endif
118 .endm
119
120  /*
121   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
122   * monstrosity above, manually.
123   */
124 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
125         ALTERNATIVE "jmp .Lskip_rsb_\@", "", \ftr
126         __FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)
127 .Lskip_rsb_\@:
128 .endm
129
130 #ifdef CONFIG_CPU_UNRET_ENTRY
131 #define CALL_ZEN_UNTRAIN_RET    "call zen_untrain_ret"
132 #else
133 #define CALL_ZEN_UNTRAIN_RET    ""
134 #endif
135
136 /*
137  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
138  * return thunk isn't mapped into the userspace tables (then again, AMD
139  * typically has NO_MELTDOWN).
140  *
141  * While zen_untrain_ret() doesn't clobber anything but requires stack,
142  * entry_ibpb() will clobber AX, CX, DX.
143  *
144  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
145  * where we have a stack but before any RET instruction.
146  */
147 .macro UNTRAIN_RET
148 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY)
149         ANNOTATE_UNRET_END
150         ALTERNATIVE_2 "",                                               \
151                       CALL_ZEN_UNTRAIN_RET, X86_FEATURE_UNRET,          \
152                       "call entry_ibpb", X86_FEATURE_ENTRY_IBPB
153 #endif
154 .endm
155
156 #else /* __ASSEMBLY__ */
157
158 #define ANNOTATE_RETPOLINE_SAFE                                 \
159         "999:\n\t"                                              \
160         ".pushsection .discard.retpoline_safe\n\t"              \
161         _ASM_PTR " 999b\n\t"                                    \
162         ".popsection\n\t"
163
164 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
165 extern retpoline_thunk_t __x86_indirect_thunk_array[];
166
167 extern void __x86_return_thunk(void);
168 extern void zen_untrain_ret(void);
169 extern void entry_ibpb(void);
170
171 #ifdef CONFIG_RETPOLINE
172
173 #define GEN(reg) \
174         extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
175 #include <asm/GEN-for-each-reg.h>
176 #undef GEN
177
178 #ifdef CONFIG_X86_64
179
180 /*
181  * Inline asm uses the %V modifier which is only in newer GCC
182  * which is ensured when CONFIG_RETPOLINE is defined.
183  */
184 # define CALL_NOSPEC                                            \
185         ALTERNATIVE_2(                                          \
186         ANNOTATE_RETPOLINE_SAFE                                 \
187         "call *%[thunk_target]\n",                              \
188         "call __x86_indirect_thunk_%V[thunk_target]\n",         \
189         X86_FEATURE_RETPOLINE,                                  \
190         "lfence;\n"                                             \
191         ANNOTATE_RETPOLINE_SAFE                                 \
192         "call *%[thunk_target]\n",                              \
193         X86_FEATURE_RETPOLINE_LFENCE)
194
195 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
196
197 #else /* CONFIG_X86_32 */
198 /*
199  * For i386 we use the original ret-equivalent retpoline, because
200  * otherwise we'll run out of registers. We don't care about CET
201  * here, anyway.
202  */
203 # define CALL_NOSPEC                                            \
204         ALTERNATIVE_2(                                          \
205         ANNOTATE_RETPOLINE_SAFE                                 \
206         "call *%[thunk_target]\n",                              \
207         "       jmp    904f;\n"                                 \
208         "       .align 16\n"                                    \
209         "901:   call   903f;\n"                                 \
210         "902:   pause;\n"                                       \
211         "       lfence;\n"                                      \
212         "       jmp    902b;\n"                                 \
213         "       .align 16\n"                                    \
214         "903:   lea    4(%%esp), %%esp;\n"                      \
215         "       pushl  %[thunk_target];\n"                      \
216         "       ret;\n"                                         \
217         "       .align 16\n"                                    \
218         "904:   call   901b;\n",                                \
219         X86_FEATURE_RETPOLINE,                                  \
220         "lfence;\n"                                             \
221         ANNOTATE_RETPOLINE_SAFE                                 \
222         "call *%[thunk_target]\n",                              \
223         X86_FEATURE_RETPOLINE_LFENCE)
224
225 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
226 #endif
227 #else /* No retpoline for C / inline asm */
228 # define CALL_NOSPEC "call *%[thunk_target]\n"
229 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
230 #endif
231
232 /* The Spectre V2 mitigation variants */
233 enum spectre_v2_mitigation {
234         SPECTRE_V2_NONE,
235         SPECTRE_V2_RETPOLINE,
236         SPECTRE_V2_LFENCE,
237         SPECTRE_V2_EIBRS,
238         SPECTRE_V2_EIBRS_RETPOLINE,
239         SPECTRE_V2_EIBRS_LFENCE,
240         SPECTRE_V2_IBRS,
241 };
242
243 /* The indirect branch speculation control variants */
244 enum spectre_v2_user_mitigation {
245         SPECTRE_V2_USER_NONE,
246         SPECTRE_V2_USER_STRICT,
247         SPECTRE_V2_USER_STRICT_PREFERRED,
248         SPECTRE_V2_USER_PRCTL,
249         SPECTRE_V2_USER_SECCOMP,
250 };
251
252 /* The Speculative Store Bypass disable variants */
253 enum ssb_mitigation {
254         SPEC_STORE_BYPASS_NONE,
255         SPEC_STORE_BYPASS_DISABLE,
256         SPEC_STORE_BYPASS_PRCTL,
257         SPEC_STORE_BYPASS_SECCOMP,
258 };
259
260 extern char __indirect_thunk_start[];
261 extern char __indirect_thunk_end[];
262
263 static __always_inline
264 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
265 {
266         asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
267                 : : "c" (msr),
268                     "a" ((u32)val),
269                     "d" ((u32)(val >> 32)),
270                     [feature] "i" (feature)
271                 : "memory");
272 }
273
274 static inline void indirect_branch_prediction_barrier(void)
275 {
276         u64 val = PRED_CMD_IBPB;
277
278         alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
279 }
280
281 /* The Intel SPEC CTRL MSR base value cache */
282 extern u64 x86_spec_ctrl_base;
283 extern u64 x86_spec_ctrl_current;
284 extern void write_spec_ctrl_current(u64 val, bool force);
285 extern u64 spec_ctrl_current(void);
286
287 /*
288  * With retpoline, we must use IBRS to restrict branch prediction
289  * before calling into firmware.
290  *
291  * (Implemented as CPP macros due to header hell.)
292  */
293 #define firmware_restrict_branch_speculation_start()                    \
294 do {                                                                    \
295         preempt_disable();                                              \
296         alternative_msr_write(MSR_IA32_SPEC_CTRL,                       \
297                               spec_ctrl_current() | SPEC_CTRL_IBRS,     \
298                               X86_FEATURE_USE_IBRS_FW);                 \
299 } while (0)
300
301 #define firmware_restrict_branch_speculation_end()                      \
302 do {                                                                    \
303         alternative_msr_write(MSR_IA32_SPEC_CTRL,                       \
304                               spec_ctrl_current(),                      \
305                               X86_FEATURE_USE_IBRS_FW);                 \
306         preempt_enable();                                               \
307 } while (0)
308
309 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
310 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
311 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
312
313 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
314 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
315
316 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush);
317
318 DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear);
319
320 #include <asm/segment.h>
321
322 /**
323  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
324  *
325  * This uses the otherwise unused and obsolete VERW instruction in
326  * combination with microcode which triggers a CPU buffer flush when the
327  * instruction is executed.
328  */
329 static __always_inline void mds_clear_cpu_buffers(void)
330 {
331         static const u16 ds = __KERNEL_DS;
332
333         /*
334          * Has to be the memory-operand variant because only that
335          * guarantees the CPU buffer flush functionality according to
336          * documentation. The register-operand variant does not.
337          * Works with any segment selector, but a valid writable
338          * data segment is the fastest variant.
339          *
340          * "cc" clobber is required because VERW modifies ZF.
341          */
342         asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
343 }
344
345 /**
346  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
347  *
348  * Clear CPU buffers if the corresponding static key is enabled
349  */
350 static __always_inline void mds_user_clear_cpu_buffers(void)
351 {
352         if (static_branch_likely(&mds_user_clear))
353                 mds_clear_cpu_buffers();
354 }
355
356 /**
357  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
358  *
359  * Clear CPU buffers if the corresponding static key is enabled
360  */
361 static inline void mds_idle_clear_cpu_buffers(void)
362 {
363         if (static_branch_likely(&mds_idle_clear))
364                 mds_clear_cpu_buffers();
365 }
366
367 #endif /* __ASSEMBLY__ */
368
369 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */