Merge tag 'arm64-upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64...
[linux-2.6-microblaze.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
9 #include <asm/alternative.h>
10 #include <asm/alternative-asm.h>
11 #include <asm/cpufeatures.h>
12 #include <asm/msr-index.h>
13 #include <asm/unwind_hints.h>
14
15 /*
16  * Fill the CPU return stack buffer.
17  *
18  * Each entry in the RSB, if used for a speculative 'ret', contains an
19  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
20  *
21  * This is required in various cases for retpoline and IBRS-based
22  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
23  * eliminate potentially bogus entries from the RSB, and sometimes
24  * purely to ensure that it doesn't get empty, which on some CPUs would
25  * allow predictions from other (unwanted!) sources to be used.
26  *
27  * We define a CPP macro such that it can be used from both .S files and
28  * inline assembly. It's possible to do a .macro and then include that
29  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
30  */
31
32 #define RSB_CLEAR_LOOPS         32      /* To forcibly overwrite all entries */
33
34 /*
35  * Google experimented with loop-unrolling and this turned out to be
36  * the optimal version — two calls, each with their own speculation
37  * trap should their return address end up getting used, in a loop.
38  */
39 #define __FILL_RETURN_BUFFER(reg, nr, sp)       \
40         mov     $(nr/2), reg;                   \
41 771:                                            \
42         ANNOTATE_INTRA_FUNCTION_CALL;           \
43         call    772f;                           \
44 773:    /* speculation trap */                  \
45         UNWIND_HINT_EMPTY;                      \
46         pause;                                  \
47         lfence;                                 \
48         jmp     773b;                           \
49 772:                                            \
50         ANNOTATE_INTRA_FUNCTION_CALL;           \
51         call    774f;                           \
52 775:    /* speculation trap */                  \
53         UNWIND_HINT_EMPTY;                      \
54         pause;                                  \
55         lfence;                                 \
56         jmp     775b;                           \
57 774:                                            \
58         add     $(BITS_PER_LONG/8) * 2, sp;     \
59         dec     reg;                            \
60         jnz     771b;
61
62 #ifdef __ASSEMBLY__
63
64 /*
65  * This should be used immediately before an indirect jump/call. It tells
66  * objtool the subsequent indirect jump/call is vouched safe for retpoline
67  * builds.
68  */
69 .macro ANNOTATE_RETPOLINE_SAFE
70         .Lannotate_\@:
71         .pushsection .discard.retpoline_safe
72         _ASM_PTR .Lannotate_\@
73         .popsection
74 .endm
75
76 /*
77  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
78  * indirect jmp/call which may be susceptible to the Spectre variant 2
79  * attack.
80  */
81 .macro JMP_NOSPEC reg:req
82 #ifdef CONFIG_RETPOLINE
83         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), \
84                       __stringify(jmp __x86_retpoline_\reg), X86_FEATURE_RETPOLINE, \
85                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), X86_FEATURE_RETPOLINE_AMD
86 #else
87         jmp     *%\reg
88 #endif
89 .endm
90
91 .macro CALL_NOSPEC reg:req
92 #ifdef CONFIG_RETPOLINE
93         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; call *%\reg), \
94                       __stringify(call __x86_retpoline_\reg), X86_FEATURE_RETPOLINE, \
95                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; call *%\reg), X86_FEATURE_RETPOLINE_AMD
96 #else
97         call    *%\reg
98 #endif
99 .endm
100
101  /*
102   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
103   * monstrosity above, manually.
104   */
105 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
106 #ifdef CONFIG_RETPOLINE
107         ALTERNATIVE "jmp .Lskip_rsb_\@", "", \ftr
108         __FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)
109 .Lskip_rsb_\@:
110 #endif
111 .endm
112
113 #else /* __ASSEMBLY__ */
114
115 #define ANNOTATE_RETPOLINE_SAFE                                 \
116         "999:\n\t"                                              \
117         ".pushsection .discard.retpoline_safe\n\t"              \
118         _ASM_PTR " 999b\n\t"                                    \
119         ".popsection\n\t"
120
121 #ifdef CONFIG_RETPOLINE
122 #ifdef CONFIG_X86_64
123
124 /*
125  * Inline asm uses the %V modifier which is only in newer GCC
126  * which is ensured when CONFIG_RETPOLINE is defined.
127  */
128 # define CALL_NOSPEC                                            \
129         ALTERNATIVE_2(                                          \
130         ANNOTATE_RETPOLINE_SAFE                                 \
131         "call *%[thunk_target]\n",                              \
132         "call __x86_retpoline_%V[thunk_target]\n",              \
133         X86_FEATURE_RETPOLINE,                                  \
134         "lfence;\n"                                             \
135         ANNOTATE_RETPOLINE_SAFE                                 \
136         "call *%[thunk_target]\n",                              \
137         X86_FEATURE_RETPOLINE_AMD)
138
139 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
140
141 #else /* CONFIG_X86_32 */
142 /*
143  * For i386 we use the original ret-equivalent retpoline, because
144  * otherwise we'll run out of registers. We don't care about CET
145  * here, anyway.
146  */
147 # define CALL_NOSPEC                                            \
148         ALTERNATIVE_2(                                          \
149         ANNOTATE_RETPOLINE_SAFE                                 \
150         "call *%[thunk_target]\n",                              \
151         "       jmp    904f;\n"                                 \
152         "       .align 16\n"                                    \
153         "901:   call   903f;\n"                                 \
154         "902:   pause;\n"                                       \
155         "       lfence;\n"                                      \
156         "       jmp    902b;\n"                                 \
157         "       .align 16\n"                                    \
158         "903:   lea    4(%%esp), %%esp;\n"                      \
159         "       pushl  %[thunk_target];\n"                      \
160         "       ret;\n"                                         \
161         "       .align 16\n"                                    \
162         "904:   call   901b;\n",                                \
163         X86_FEATURE_RETPOLINE,                                  \
164         "lfence;\n"                                             \
165         ANNOTATE_RETPOLINE_SAFE                                 \
166         "call *%[thunk_target]\n",                              \
167         X86_FEATURE_RETPOLINE_AMD)
168
169 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
170 #endif
171 #else /* No retpoline for C / inline asm */
172 # define CALL_NOSPEC "call *%[thunk_target]\n"
173 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
174 #endif
175
176 /* The Spectre V2 mitigation variants */
177 enum spectre_v2_mitigation {
178         SPECTRE_V2_NONE,
179         SPECTRE_V2_RETPOLINE_GENERIC,
180         SPECTRE_V2_RETPOLINE_AMD,
181         SPECTRE_V2_IBRS_ENHANCED,
182 };
183
184 /* The indirect branch speculation control variants */
185 enum spectre_v2_user_mitigation {
186         SPECTRE_V2_USER_NONE,
187         SPECTRE_V2_USER_STRICT,
188         SPECTRE_V2_USER_STRICT_PREFERRED,
189         SPECTRE_V2_USER_PRCTL,
190         SPECTRE_V2_USER_SECCOMP,
191 };
192
193 /* The Speculative Store Bypass disable variants */
194 enum ssb_mitigation {
195         SPEC_STORE_BYPASS_NONE,
196         SPEC_STORE_BYPASS_DISABLE,
197         SPEC_STORE_BYPASS_PRCTL,
198         SPEC_STORE_BYPASS_SECCOMP,
199 };
200
201 extern char __indirect_thunk_start[];
202 extern char __indirect_thunk_end[];
203
204 static __always_inline
205 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
206 {
207         asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
208                 : : "c" (msr),
209                     "a" ((u32)val),
210                     "d" ((u32)(val >> 32)),
211                     [feature] "i" (feature)
212                 : "memory");
213 }
214
215 static inline void indirect_branch_prediction_barrier(void)
216 {
217         u64 val = PRED_CMD_IBPB;
218
219         alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
220 }
221
222 /* The Intel SPEC CTRL MSR base value cache */
223 extern u64 x86_spec_ctrl_base;
224
225 /*
226  * With retpoline, we must use IBRS to restrict branch prediction
227  * before calling into firmware.
228  *
229  * (Implemented as CPP macros due to header hell.)
230  */
231 #define firmware_restrict_branch_speculation_start()                    \
232 do {                                                                    \
233         u64 val = x86_spec_ctrl_base | SPEC_CTRL_IBRS;                  \
234                                                                         \
235         preempt_disable();                                              \
236         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
237                               X86_FEATURE_USE_IBRS_FW);                 \
238 } while (0)
239
240 #define firmware_restrict_branch_speculation_end()                      \
241 do {                                                                    \
242         u64 val = x86_spec_ctrl_base;                                   \
243                                                                         \
244         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
245                               X86_FEATURE_USE_IBRS_FW);                 \
246         preempt_enable();                                               \
247 } while (0)
248
249 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
250 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
251 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
252
253 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
254 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
255
256 #include <asm/segment.h>
257
258 /**
259  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
260  *
261  * This uses the otherwise unused and obsolete VERW instruction in
262  * combination with microcode which triggers a CPU buffer flush when the
263  * instruction is executed.
264  */
265 static __always_inline void mds_clear_cpu_buffers(void)
266 {
267         static const u16 ds = __KERNEL_DS;
268
269         /*
270          * Has to be the memory-operand variant because only that
271          * guarantees the CPU buffer flush functionality according to
272          * documentation. The register-operand variant does not.
273          * Works with any segment selector, but a valid writable
274          * data segment is the fastest variant.
275          *
276          * "cc" clobber is required because VERW modifies ZF.
277          */
278         asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
279 }
280
281 /**
282  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
283  *
284  * Clear CPU buffers if the corresponding static key is enabled
285  */
286 static __always_inline void mds_user_clear_cpu_buffers(void)
287 {
288         if (static_branch_likely(&mds_user_clear))
289                 mds_clear_cpu_buffers();
290 }
291
292 /**
293  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
294  *
295  * Clear CPU buffers if the corresponding static key is enabled
296  */
297 static inline void mds_idle_clear_cpu_buffers(void)
298 {
299         if (static_branch_likely(&mds_idle_clear))
300                 mds_clear_cpu_buffers();
301 }
302
303 #endif /* __ASSEMBLY__ */
304
305 /*
306  * Below is used in the eBPF JIT compiler and emits the byte sequence
307  * for the following assembly:
308  *
309  * With retpolines configured:
310  *
311  *    callq do_rop
312  *  spec_trap:
313  *    pause
314  *    lfence
315  *    jmp spec_trap
316  *  do_rop:
317  *    mov %rcx,(%rsp) for x86_64
318  *    mov %edx,(%esp) for x86_32
319  *    retq
320  *
321  * Without retpolines configured:
322  *
323  *    jmp *%rcx for x86_64
324  *    jmp *%edx for x86_32
325  */
326 #ifdef CONFIG_RETPOLINE
327 # ifdef CONFIG_X86_64
328 #  define RETPOLINE_RCX_BPF_JIT_SIZE    17
329 #  define RETPOLINE_RCX_BPF_JIT()                               \
330 do {                                                            \
331         EMIT1_off32(0xE8, 7);    /* callq do_rop */             \
332         /* spec_trap: */                                        \
333         EMIT2(0xF3, 0x90);       /* pause */                    \
334         EMIT3(0x0F, 0xAE, 0xE8); /* lfence */                   \
335         EMIT2(0xEB, 0xF9);       /* jmp spec_trap */            \
336         /* do_rop: */                                           \
337         EMIT4(0x48, 0x89, 0x0C, 0x24); /* mov %rcx,(%rsp) */    \
338         EMIT1(0xC3);             /* retq */                     \
339 } while (0)
340 # else /* !CONFIG_X86_64 */
341 #  define RETPOLINE_EDX_BPF_JIT()                               \
342 do {                                                            \
343         EMIT1_off32(0xE8, 7);    /* call do_rop */              \
344         /* spec_trap: */                                        \
345         EMIT2(0xF3, 0x90);       /* pause */                    \
346         EMIT3(0x0F, 0xAE, 0xE8); /* lfence */                   \
347         EMIT2(0xEB, 0xF9);       /* jmp spec_trap */            \
348         /* do_rop: */                                           \
349         EMIT3(0x89, 0x14, 0x24); /* mov %edx,(%esp) */          \
350         EMIT1(0xC3);             /* ret */                      \
351 } while (0)
352 # endif
353 #else /* !CONFIG_RETPOLINE */
354 # ifdef CONFIG_X86_64
355 #  define RETPOLINE_RCX_BPF_JIT_SIZE    2
356 #  define RETPOLINE_RCX_BPF_JIT()                               \
357         EMIT2(0xFF, 0xE1);       /* jmp *%rcx */
358 # else /* !CONFIG_X86_64 */
359 #  define RETPOLINE_EDX_BPF_JIT()                               \
360         EMIT2(0xFF, 0xE2)        /* jmp *%edx */
361 # endif
362 #endif
363
364 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */