x86/bugs: Optimize SPEC_CTRL MSR writes
[platform/kernel/linux-rpi.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  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
80  * indirect jmp/call which may be susceptible to the Spectre variant 2
81  * attack.
82  */
83 .macro JMP_NOSPEC reg:req
84 #ifdef CONFIG_RETPOLINE
85         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), \
86                       __stringify(jmp __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
87                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), X86_FEATURE_RETPOLINE_LFENCE
88 #else
89         jmp     *%\reg
90 #endif
91 .endm
92
93 .macro CALL_NOSPEC reg:req
94 #ifdef CONFIG_RETPOLINE
95         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; call *%\reg), \
96                       __stringify(call __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
97                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; call *%\reg), X86_FEATURE_RETPOLINE_LFENCE
98 #else
99         call    *%\reg
100 #endif
101 .endm
102
103  /*
104   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
105   * monstrosity above, manually.
106   */
107 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
108 #ifdef CONFIG_RETPOLINE
109         ALTERNATIVE "jmp .Lskip_rsb_\@", "", \ftr
110         __FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)
111 .Lskip_rsb_\@:
112 #endif
113 .endm
114
115 /*
116  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
117  * return thunk isn't mapped into the userspace tables (then again, AMD
118  * typically has NO_MELTDOWN).
119  *
120  * Doesn't clobber any registers but does require a stable stack.
121  *
122  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
123  * where we have a stack but before any RET instruction.
124  */
125 .macro UNTRAIN_RET
126 #ifdef CONFIG_RETPOLINE
127         ALTERNATIVE "", "call zen_untrain_ret", X86_FEATURE_UNRET
128 #endif
129 .endm
130
131 #else /* __ASSEMBLY__ */
132
133 #define ANNOTATE_RETPOLINE_SAFE                                 \
134         "999:\n\t"                                              \
135         ".pushsection .discard.retpoline_safe\n\t"              \
136         _ASM_PTR " 999b\n\t"                                    \
137         ".popsection\n\t"
138
139 extern void __x86_return_thunk(void);
140 extern void zen_untrain_ret(void);
141
142 #ifdef CONFIG_RETPOLINE
143
144 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
145
146 #define GEN(reg) \
147         extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
148 #include <asm/GEN-for-each-reg.h>
149 #undef GEN
150
151 extern retpoline_thunk_t __x86_indirect_thunk_array[];
152
153 #ifdef CONFIG_X86_64
154
155 /*
156  * Inline asm uses the %V modifier which is only in newer GCC
157  * which is ensured when CONFIG_RETPOLINE is defined.
158  */
159 # define CALL_NOSPEC                                            \
160         ALTERNATIVE_2(                                          \
161         ANNOTATE_RETPOLINE_SAFE                                 \
162         "call *%[thunk_target]\n",                              \
163         "call __x86_indirect_thunk_%V[thunk_target]\n",         \
164         X86_FEATURE_RETPOLINE,                                  \
165         "lfence;\n"                                             \
166         ANNOTATE_RETPOLINE_SAFE                                 \
167         "call *%[thunk_target]\n",                              \
168         X86_FEATURE_RETPOLINE_LFENCE)
169
170 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
171
172 #else /* CONFIG_X86_32 */
173 /*
174  * For i386 we use the original ret-equivalent retpoline, because
175  * otherwise we'll run out of registers. We don't care about CET
176  * here, anyway.
177  */
178 # define CALL_NOSPEC                                            \
179         ALTERNATIVE_2(                                          \
180         ANNOTATE_RETPOLINE_SAFE                                 \
181         "call *%[thunk_target]\n",                              \
182         "       jmp    904f;\n"                                 \
183         "       .align 16\n"                                    \
184         "901:   call   903f;\n"                                 \
185         "902:   pause;\n"                                       \
186         "       lfence;\n"                                      \
187         "       jmp    902b;\n"                                 \
188         "       .align 16\n"                                    \
189         "903:   lea    4(%%esp), %%esp;\n"                      \
190         "       pushl  %[thunk_target];\n"                      \
191         "       ret;\n"                                         \
192         "       .align 16\n"                                    \
193         "904:   call   901b;\n",                                \
194         X86_FEATURE_RETPOLINE,                                  \
195         "lfence;\n"                                             \
196         ANNOTATE_RETPOLINE_SAFE                                 \
197         "call *%[thunk_target]\n",                              \
198         X86_FEATURE_RETPOLINE_LFENCE)
199
200 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
201 #endif
202 #else /* No retpoline for C / inline asm */
203 # define CALL_NOSPEC "call *%[thunk_target]\n"
204 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
205 #endif
206
207 /* The Spectre V2 mitigation variants */
208 enum spectre_v2_mitigation {
209         SPECTRE_V2_NONE,
210         SPECTRE_V2_RETPOLINE,
211         SPECTRE_V2_LFENCE,
212         SPECTRE_V2_EIBRS,
213         SPECTRE_V2_EIBRS_RETPOLINE,
214         SPECTRE_V2_EIBRS_LFENCE,
215 };
216
217 /* The indirect branch speculation control variants */
218 enum spectre_v2_user_mitigation {
219         SPECTRE_V2_USER_NONE,
220         SPECTRE_V2_USER_STRICT,
221         SPECTRE_V2_USER_STRICT_PREFERRED,
222         SPECTRE_V2_USER_PRCTL,
223         SPECTRE_V2_USER_SECCOMP,
224 };
225
226 /* The Speculative Store Bypass disable variants */
227 enum ssb_mitigation {
228         SPEC_STORE_BYPASS_NONE,
229         SPEC_STORE_BYPASS_DISABLE,
230         SPEC_STORE_BYPASS_PRCTL,
231         SPEC_STORE_BYPASS_SECCOMP,
232 };
233
234 extern char __indirect_thunk_start[];
235 extern char __indirect_thunk_end[];
236
237 static __always_inline
238 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
239 {
240         asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
241                 : : "c" (msr),
242                     "a" ((u32)val),
243                     "d" ((u32)(val >> 32)),
244                     [feature] "i" (feature)
245                 : "memory");
246 }
247
248 static inline void indirect_branch_prediction_barrier(void)
249 {
250         u64 val = PRED_CMD_IBPB;
251
252         alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
253 }
254
255 /* The Intel SPEC CTRL MSR base value cache */
256 extern u64 x86_spec_ctrl_base;
257 extern void write_spec_ctrl_current(u64 val, bool force);
258
259 /*
260  * With retpoline, we must use IBRS to restrict branch prediction
261  * before calling into firmware.
262  *
263  * (Implemented as CPP macros due to header hell.)
264  */
265 #define firmware_restrict_branch_speculation_start()                    \
266 do {                                                                    \
267         u64 val = x86_spec_ctrl_base | SPEC_CTRL_IBRS;                  \
268                                                                         \
269         preempt_disable();                                              \
270         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
271                               X86_FEATURE_USE_IBRS_FW);                 \
272 } while (0)
273
274 #define firmware_restrict_branch_speculation_end()                      \
275 do {                                                                    \
276         u64 val = x86_spec_ctrl_base;                                   \
277                                                                         \
278         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
279                               X86_FEATURE_USE_IBRS_FW);                 \
280         preempt_enable();                                               \
281 } while (0)
282
283 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
284 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
285 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
286
287 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
288 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
289
290 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush);
291
292 DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear);
293
294 #include <asm/segment.h>
295
296 /**
297  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
298  *
299  * This uses the otherwise unused and obsolete VERW instruction in
300  * combination with microcode which triggers a CPU buffer flush when the
301  * instruction is executed.
302  */
303 static __always_inline void mds_clear_cpu_buffers(void)
304 {
305         static const u16 ds = __KERNEL_DS;
306
307         /*
308          * Has to be the memory-operand variant because only that
309          * guarantees the CPU buffer flush functionality according to
310          * documentation. The register-operand variant does not.
311          * Works with any segment selector, but a valid writable
312          * data segment is the fastest variant.
313          *
314          * "cc" clobber is required because VERW modifies ZF.
315          */
316         asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
317 }
318
319 /**
320  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
321  *
322  * Clear CPU buffers if the corresponding static key is enabled
323  */
324 static __always_inline void mds_user_clear_cpu_buffers(void)
325 {
326         if (static_branch_likely(&mds_user_clear))
327                 mds_clear_cpu_buffers();
328 }
329
330 /**
331  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
332  *
333  * Clear CPU buffers if the corresponding static key is enabled
334  */
335 static inline void mds_idle_clear_cpu_buffers(void)
336 {
337         if (static_branch_likely(&mds_idle_clear))
338                 mds_clear_cpu_buffers();
339 }
340
341 #endif /* __ASSEMBLY__ */
342
343 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */