Merge tag 'amd-drm-fixes-5.19-2022-07-20' of https://gitlab.freedesktop.org/agd5f...
[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 #include <asm/percpu.h>
15
16 #define RETPOLINE_THUNK_SIZE    32
17
18 /*
19  * Fill the CPU return stack buffer.
20  *
21  * Each entry in the RSB, if used for a speculative 'ret', contains an
22  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
23  *
24  * This is required in various cases for retpoline and IBRS-based
25  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
26  * eliminate potentially bogus entries from the RSB, and sometimes
27  * purely to ensure that it doesn't get empty, which on some CPUs would
28  * allow predictions from other (unwanted!) sources to be used.
29  *
30  * We define a CPP macro such that it can be used from both .S files and
31  * inline assembly. It's possible to do a .macro and then include that
32  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
33  */
34
35 #define RSB_CLEAR_LOOPS         32      /* To forcibly overwrite all entries */
36
37 /*
38  * Google experimented with loop-unrolling and this turned out to be
39  * the optimal version - two calls, each with their own speculation
40  * trap should their return address end up getting used, in a loop.
41  */
42 #define __FILL_RETURN_BUFFER(reg, nr, sp)       \
43         mov     $(nr/2), reg;                   \
44 771:                                            \
45         ANNOTATE_INTRA_FUNCTION_CALL;           \
46         call    772f;                           \
47 773:    /* speculation trap */                  \
48         UNWIND_HINT_EMPTY;                      \
49         pause;                                  \
50         lfence;                                 \
51         jmp     773b;                           \
52 772:                                            \
53         ANNOTATE_INTRA_FUNCTION_CALL;           \
54         call    774f;                           \
55 775:    /* speculation trap */                  \
56         UNWIND_HINT_EMPTY;                      \
57         pause;                                  \
58         lfence;                                 \
59         jmp     775b;                           \
60 774:                                            \
61         add     $(BITS_PER_LONG/8) * 2, sp;     \
62         dec     reg;                            \
63         jnz     771b;
64
65 #ifdef __ASSEMBLY__
66
67 /*
68  * This should be used immediately before an indirect jump/call. It tells
69  * objtool the subsequent indirect jump/call is vouched safe for retpoline
70  * builds.
71  */
72 .macro ANNOTATE_RETPOLINE_SAFE
73         .Lannotate_\@:
74         .pushsection .discard.retpoline_safe
75         _ASM_PTR .Lannotate_\@
76         .popsection
77 .endm
78
79 /*
80  * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions
81  * vs RETBleed validation.
82  */
83 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE
84
85 /*
86  * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should
87  * eventually turn into it's own annotation.
88  */
89 .macro ANNOTATE_UNRET_END
90 #ifdef CONFIG_DEBUG_ENTRY
91         ANNOTATE_RETPOLINE_SAFE
92         nop
93 #endif
94 .endm
95
96 /*
97  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
98  * indirect jmp/call which may be susceptible to the Spectre variant 2
99  * attack.
100  */
101 .macro JMP_NOSPEC reg:req
102 #ifdef CONFIG_RETPOLINE
103         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), \
104                       __stringify(jmp __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
105                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), X86_FEATURE_RETPOLINE_LFENCE
106 #else
107         jmp     *%\reg
108 #endif
109 .endm
110
111 .macro CALL_NOSPEC reg:req
112 #ifdef CONFIG_RETPOLINE
113         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; call *%\reg), \
114                       __stringify(call __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
115                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; call *%\reg), X86_FEATURE_RETPOLINE_LFENCE
116 #else
117         call    *%\reg
118 #endif
119 .endm
120
121  /*
122   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
123   * monstrosity above, manually.
124   */
125 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
126         ALTERNATIVE "jmp .Lskip_rsb_\@", "", \ftr
127         __FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)
128 .Lskip_rsb_\@:
129 .endm
130
131 #ifdef CONFIG_CPU_UNRET_ENTRY
132 #define CALL_ZEN_UNTRAIN_RET    "call zen_untrain_ret"
133 #else
134 #define CALL_ZEN_UNTRAIN_RET    ""
135 #endif
136
137 /*
138  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
139  * return thunk isn't mapped into the userspace tables (then again, AMD
140  * typically has NO_MELTDOWN).
141  *
142  * While zen_untrain_ret() doesn't clobber anything but requires stack,
143  * entry_ibpb() will clobber AX, CX, DX.
144  *
145  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
146  * where we have a stack but before any RET instruction.
147  */
148 .macro UNTRAIN_RET
149 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY)
150         ANNOTATE_UNRET_END
151         ALTERNATIVE_2 "",                                               \
152                       CALL_ZEN_UNTRAIN_RET, X86_FEATURE_UNRET,          \
153                       "call entry_ibpb", X86_FEATURE_ENTRY_IBPB
154 #endif
155 .endm
156
157 #else /* __ASSEMBLY__ */
158
159 #define ANNOTATE_RETPOLINE_SAFE                                 \
160         "999:\n\t"                                              \
161         ".pushsection .discard.retpoline_safe\n\t"              \
162         _ASM_PTR " 999b\n\t"                                    \
163         ".popsection\n\t"
164
165 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
166 extern retpoline_thunk_t __x86_indirect_thunk_array[];
167
168 extern void __x86_return_thunk(void);
169 extern void zen_untrain_ret(void);
170 extern void entry_ibpb(void);
171
172 #ifdef CONFIG_RETPOLINE
173
174 #define GEN(reg) \
175         extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
176 #include <asm/GEN-for-each-reg.h>
177 #undef GEN
178
179 #ifdef CONFIG_X86_64
180
181 /*
182  * Inline asm uses the %V modifier which is only in newer GCC
183  * which is ensured when CONFIG_RETPOLINE is defined.
184  */
185 # define CALL_NOSPEC                                            \
186         ALTERNATIVE_2(                                          \
187         ANNOTATE_RETPOLINE_SAFE                                 \
188         "call *%[thunk_target]\n",                              \
189         "call __x86_indirect_thunk_%V[thunk_target]\n",         \
190         X86_FEATURE_RETPOLINE,                                  \
191         "lfence;\n"                                             \
192         ANNOTATE_RETPOLINE_SAFE                                 \
193         "call *%[thunk_target]\n",                              \
194         X86_FEATURE_RETPOLINE_LFENCE)
195
196 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
197
198 #else /* CONFIG_X86_32 */
199 /*
200  * For i386 we use the original ret-equivalent retpoline, because
201  * otherwise we'll run out of registers. We don't care about CET
202  * here, anyway.
203  */
204 # define CALL_NOSPEC                                            \
205         ALTERNATIVE_2(                                          \
206         ANNOTATE_RETPOLINE_SAFE                                 \
207         "call *%[thunk_target]\n",                              \
208         "       jmp    904f;\n"                                 \
209         "       .align 16\n"                                    \
210         "901:   call   903f;\n"                                 \
211         "902:   pause;\n"                                       \
212         "       lfence;\n"                                      \
213         "       jmp    902b;\n"                                 \
214         "       .align 16\n"                                    \
215         "903:   lea    4(%%esp), %%esp;\n"                      \
216         "       pushl  %[thunk_target];\n"                      \
217         "       ret;\n"                                         \
218         "       .align 16\n"                                    \
219         "904:   call   901b;\n",                                \
220         X86_FEATURE_RETPOLINE,                                  \
221         "lfence;\n"                                             \
222         ANNOTATE_RETPOLINE_SAFE                                 \
223         "call *%[thunk_target]\n",                              \
224         X86_FEATURE_RETPOLINE_LFENCE)
225
226 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
227 #endif
228 #else /* No retpoline for C / inline asm */
229 # define CALL_NOSPEC "call *%[thunk_target]\n"
230 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
231 #endif
232
233 /* The Spectre V2 mitigation variants */
234 enum spectre_v2_mitigation {
235         SPECTRE_V2_NONE,
236         SPECTRE_V2_RETPOLINE,
237         SPECTRE_V2_LFENCE,
238         SPECTRE_V2_EIBRS,
239         SPECTRE_V2_EIBRS_RETPOLINE,
240         SPECTRE_V2_EIBRS_LFENCE,
241         SPECTRE_V2_IBRS,
242 };
243
244 /* The indirect branch speculation control variants */
245 enum spectre_v2_user_mitigation {
246         SPECTRE_V2_USER_NONE,
247         SPECTRE_V2_USER_STRICT,
248         SPECTRE_V2_USER_STRICT_PREFERRED,
249         SPECTRE_V2_USER_PRCTL,
250         SPECTRE_V2_USER_SECCOMP,
251 };
252
253 /* The Speculative Store Bypass disable variants */
254 enum ssb_mitigation {
255         SPEC_STORE_BYPASS_NONE,
256         SPEC_STORE_BYPASS_DISABLE,
257         SPEC_STORE_BYPASS_PRCTL,
258         SPEC_STORE_BYPASS_SECCOMP,
259 };
260
261 extern char __indirect_thunk_start[];
262 extern char __indirect_thunk_end[];
263
264 static __always_inline
265 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
266 {
267         asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
268                 : : "c" (msr),
269                     "a" ((u32)val),
270                     "d" ((u32)(val >> 32)),
271                     [feature] "i" (feature)
272                 : "memory");
273 }
274
275 static inline void indirect_branch_prediction_barrier(void)
276 {
277         u64 val = PRED_CMD_IBPB;
278
279         alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
280 }
281
282 /* The Intel SPEC CTRL MSR base value cache */
283 extern u64 x86_spec_ctrl_base;
284 DECLARE_PER_CPU(u64, x86_spec_ctrl_current);
285 extern void write_spec_ctrl_current(u64 val, bool force);
286 extern u64 spec_ctrl_current(void);
287
288 /*
289  * With retpoline, we must use IBRS to restrict branch prediction
290  * before calling into firmware.
291  *
292  * (Implemented as CPP macros due to header hell.)
293  */
294 #define firmware_restrict_branch_speculation_start()                    \
295 do {                                                                    \
296         preempt_disable();                                              \
297         alternative_msr_write(MSR_IA32_SPEC_CTRL,                       \
298                               spec_ctrl_current() | SPEC_CTRL_IBRS,     \
299                               X86_FEATURE_USE_IBRS_FW);                 \
300 } while (0)
301
302 #define firmware_restrict_branch_speculation_end()                      \
303 do {                                                                    \
304         alternative_msr_write(MSR_IA32_SPEC_CTRL,                       \
305                               spec_ctrl_current(),                      \
306                               X86_FEATURE_USE_IBRS_FW);                 \
307         preempt_enable();                                               \
308 } while (0)
309
310 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
311 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
312 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
313
314 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
315 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
316
317 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush);
318
319 DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear);
320
321 #include <asm/segment.h>
322
323 /**
324  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
325  *
326  * This uses the otherwise unused and obsolete VERW instruction in
327  * combination with microcode which triggers a CPU buffer flush when the
328  * instruction is executed.
329  */
330 static __always_inline void mds_clear_cpu_buffers(void)
331 {
332         static const u16 ds = __KERNEL_DS;
333
334         /*
335          * Has to be the memory-operand variant because only that
336          * guarantees the CPU buffer flush functionality according to
337          * documentation. The register-operand variant does not.
338          * Works with any segment selector, but a valid writable
339          * data segment is the fastest variant.
340          *
341          * "cc" clobber is required because VERW modifies ZF.
342          */
343         asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
344 }
345
346 /**
347  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
348  *
349  * Clear CPU buffers if the corresponding static key is enabled
350  */
351 static __always_inline void mds_user_clear_cpu_buffers(void)
352 {
353         if (static_branch_likely(&mds_user_clear))
354                 mds_clear_cpu_buffers();
355 }
356
357 /**
358  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
359  *
360  * Clear CPU buffers if the corresponding static key is enabled
361  */
362 static inline void mds_idle_clear_cpu_buffers(void)
363 {
364         if (static_branch_likely(&mds_idle_clear))
365                 mds_clear_cpu_buffers();
366 }
367
368 #endif /* __ASSEMBLY__ */
369
370 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */