Merge tag 'dma-mapping-5.14' of git://git.infradead.org/users/hch/dma-mapping
[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/cpufeatures.h>
11 #include <asm/msr-index.h>
12 #include <asm/unwind_hints.h>
13
14 /*
15  * Fill the CPU return stack buffer.
16  *
17  * Each entry in the RSB, if used for a speculative 'ret', contains an
18  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
19  *
20  * This is required in various cases for retpoline and IBRS-based
21  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
22  * eliminate potentially bogus entries from the RSB, and sometimes
23  * purely to ensure that it doesn't get empty, which on some CPUs would
24  * allow predictions from other (unwanted!) sources to be used.
25  *
26  * We define a CPP macro such that it can be used from both .S files and
27  * inline assembly. It's possible to do a .macro and then include that
28  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
29  */
30
31 #define RSB_CLEAR_LOOPS         32      /* To forcibly overwrite all entries */
32
33 /*
34  * Google experimented with loop-unrolling and this turned out to be
35  * the optimal version - two calls, each with their own speculation
36  * trap should their return address end up getting used, in a loop.
37  */
38 #define __FILL_RETURN_BUFFER(reg, nr, sp)       \
39         mov     $(nr/2), reg;                   \
40 771:                                            \
41         ANNOTATE_INTRA_FUNCTION_CALL;           \
42         call    772f;                           \
43 773:    /* speculation trap */                  \
44         UNWIND_HINT_EMPTY;                      \
45         pause;                                  \
46         lfence;                                 \
47         jmp     773b;                           \
48 772:                                            \
49         ANNOTATE_INTRA_FUNCTION_CALL;           \
50         call    774f;                           \
51 775:    /* speculation trap */                  \
52         UNWIND_HINT_EMPTY;                      \
53         pause;                                  \
54         lfence;                                 \
55         jmp     775b;                           \
56 774:                                            \
57         add     $(BITS_PER_LONG/8) * 2, sp;     \
58         dec     reg;                            \
59         jnz     771b;
60
61 #ifdef __ASSEMBLY__
62
63 /*
64  * This should be used immediately before an indirect jump/call. It tells
65  * objtool the subsequent indirect jump/call is vouched safe for retpoline
66  * builds.
67  */
68 .macro ANNOTATE_RETPOLINE_SAFE
69         .Lannotate_\@:
70         .pushsection .discard.retpoline_safe
71         _ASM_PTR .Lannotate_\@
72         .popsection
73 .endm
74
75 /*
76  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
77  * indirect jmp/call which may be susceptible to the Spectre variant 2
78  * attack.
79  */
80 .macro JMP_NOSPEC reg:req
81 #ifdef CONFIG_RETPOLINE
82         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), \
83                       __stringify(jmp __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
84                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), X86_FEATURE_RETPOLINE_AMD
85 #else
86         jmp     *%\reg
87 #endif
88 .endm
89
90 .macro CALL_NOSPEC reg:req
91 #ifdef CONFIG_RETPOLINE
92         ALTERNATIVE_2 __stringify(ANNOTATE_RETPOLINE_SAFE; call *%\reg), \
93                       __stringify(call __x86_indirect_thunk_\reg), X86_FEATURE_RETPOLINE, \
94                       __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; call *%\reg), X86_FEATURE_RETPOLINE_AMD
95 #else
96         call    *%\reg
97 #endif
98 .endm
99
100  /*
101   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
102   * monstrosity above, manually.
103   */
104 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
105 #ifdef CONFIG_RETPOLINE
106         ALTERNATIVE "jmp .Lskip_rsb_\@", "", \ftr
107         __FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)
108 .Lskip_rsb_\@:
109 #endif
110 .endm
111
112 #else /* __ASSEMBLY__ */
113
114 #define ANNOTATE_RETPOLINE_SAFE                                 \
115         "999:\n\t"                                              \
116         ".pushsection .discard.retpoline_safe\n\t"              \
117         _ASM_PTR " 999b\n\t"                                    \
118         ".popsection\n\t"
119
120 #ifdef CONFIG_RETPOLINE
121 #ifdef CONFIG_X86_64
122
123 /*
124  * Inline asm uses the %V modifier which is only in newer GCC
125  * which is ensured when CONFIG_RETPOLINE is defined.
126  */
127 # define CALL_NOSPEC                                            \
128         ALTERNATIVE_2(                                          \
129         ANNOTATE_RETPOLINE_SAFE                                 \
130         "call *%[thunk_target]\n",                              \
131         "call __x86_indirect_thunk_%V[thunk_target]\n",         \
132         X86_FEATURE_RETPOLINE,                                  \
133         "lfence;\n"                                             \
134         ANNOTATE_RETPOLINE_SAFE                                 \
135         "call *%[thunk_target]\n",                              \
136         X86_FEATURE_RETPOLINE_AMD)
137
138 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
139
140 #else /* CONFIG_X86_32 */
141 /*
142  * For i386 we use the original ret-equivalent retpoline, because
143  * otherwise we'll run out of registers. We don't care about CET
144  * here, anyway.
145  */
146 # define CALL_NOSPEC                                            \
147         ALTERNATIVE_2(                                          \
148         ANNOTATE_RETPOLINE_SAFE                                 \
149         "call *%[thunk_target]\n",                              \
150         "       jmp    904f;\n"                                 \
151         "       .align 16\n"                                    \
152         "901:   call   903f;\n"                                 \
153         "902:   pause;\n"                                       \
154         "       lfence;\n"                                      \
155         "       jmp    902b;\n"                                 \
156         "       .align 16\n"                                    \
157         "903:   lea    4(%%esp), %%esp;\n"                      \
158         "       pushl  %[thunk_target];\n"                      \
159         "       ret;\n"                                         \
160         "       .align 16\n"                                    \
161         "904:   call   901b;\n",                                \
162         X86_FEATURE_RETPOLINE,                                  \
163         "lfence;\n"                                             \
164         ANNOTATE_RETPOLINE_SAFE                                 \
165         "call *%[thunk_target]\n",                              \
166         X86_FEATURE_RETPOLINE_AMD)
167
168 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
169 #endif
170 #else /* No retpoline for C / inline asm */
171 # define CALL_NOSPEC "call *%[thunk_target]\n"
172 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
173 #endif
174
175 /* The Spectre V2 mitigation variants */
176 enum spectre_v2_mitigation {
177         SPECTRE_V2_NONE,
178         SPECTRE_V2_RETPOLINE_GENERIC,
179         SPECTRE_V2_RETPOLINE_AMD,
180         SPECTRE_V2_IBRS_ENHANCED,
181 };
182
183 /* The indirect branch speculation control variants */
184 enum spectre_v2_user_mitigation {
185         SPECTRE_V2_USER_NONE,
186         SPECTRE_V2_USER_STRICT,
187         SPECTRE_V2_USER_STRICT_PREFERRED,
188         SPECTRE_V2_USER_PRCTL,
189         SPECTRE_V2_USER_SECCOMP,
190 };
191
192 /* The Speculative Store Bypass disable variants */
193 enum ssb_mitigation {
194         SPEC_STORE_BYPASS_NONE,
195         SPEC_STORE_BYPASS_DISABLE,
196         SPEC_STORE_BYPASS_PRCTL,
197         SPEC_STORE_BYPASS_SECCOMP,
198 };
199
200 extern char __indirect_thunk_start[];
201 extern char __indirect_thunk_end[];
202
203 static __always_inline
204 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
205 {
206         asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
207                 : : "c" (msr),
208                     "a" ((u32)val),
209                     "d" ((u32)(val >> 32)),
210                     [feature] "i" (feature)
211                 : "memory");
212 }
213
214 static inline void indirect_branch_prediction_barrier(void)
215 {
216         u64 val = PRED_CMD_IBPB;
217
218         alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
219 }
220
221 /* The Intel SPEC CTRL MSR base value cache */
222 extern u64 x86_spec_ctrl_base;
223
224 /*
225  * With retpoline, we must use IBRS to restrict branch prediction
226  * before calling into firmware.
227  *
228  * (Implemented as CPP macros due to header hell.)
229  */
230 #define firmware_restrict_branch_speculation_start()                    \
231 do {                                                                    \
232         u64 val = x86_spec_ctrl_base | SPEC_CTRL_IBRS;                  \
233                                                                         \
234         preempt_disable();                                              \
235         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
236                               X86_FEATURE_USE_IBRS_FW);                 \
237 } while (0)
238
239 #define firmware_restrict_branch_speculation_end()                      \
240 do {                                                                    \
241         u64 val = x86_spec_ctrl_base;                                   \
242                                                                         \
243         alternative_msr_write(MSR_IA32_SPEC_CTRL, val,                  \
244                               X86_FEATURE_USE_IBRS_FW);                 \
245         preempt_enable();                                               \
246 } while (0)
247
248 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
249 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
250 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
251
252 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
253 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
254
255 #include <asm/segment.h>
256
257 /**
258  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
259  *
260  * This uses the otherwise unused and obsolete VERW instruction in
261  * combination with microcode which triggers a CPU buffer flush when the
262  * instruction is executed.
263  */
264 static __always_inline void mds_clear_cpu_buffers(void)
265 {
266         static const u16 ds = __KERNEL_DS;
267
268         /*
269          * Has to be the memory-operand variant because only that
270          * guarantees the CPU buffer flush functionality according to
271          * documentation. The register-operand variant does not.
272          * Works with any segment selector, but a valid writable
273          * data segment is the fastest variant.
274          *
275          * "cc" clobber is required because VERW modifies ZF.
276          */
277         asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
278 }
279
280 /**
281  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
282  *
283  * Clear CPU buffers if the corresponding static key is enabled
284  */
285 static __always_inline void mds_user_clear_cpu_buffers(void)
286 {
287         if (static_branch_likely(&mds_user_clear))
288                 mds_clear_cpu_buffers();
289 }
290
291 /**
292  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
293  *
294  * Clear CPU buffers if the corresponding static key is enabled
295  */
296 static inline void mds_idle_clear_cpu_buffers(void)
297 {
298         if (static_branch_likely(&mds_idle_clear))
299                 mds_clear_cpu_buffers();
300 }
301
302 #endif /* __ASSEMBLY__ */
303
304 /*
305  * Below is used in the eBPF JIT compiler and emits the byte sequence
306  * for the following assembly:
307  *
308  * With retpolines configured:
309  *
310  *    callq do_rop
311  *  spec_trap:
312  *    pause
313  *    lfence
314  *    jmp spec_trap
315  *  do_rop:
316  *    mov %rcx,(%rsp) for x86_64
317  *    mov %edx,(%esp) for x86_32
318  *    retq
319  *
320  * Without retpolines configured:
321  *
322  *    jmp *%rcx for x86_64
323  *    jmp *%edx for x86_32
324  */
325 #ifdef CONFIG_RETPOLINE
326 # ifdef CONFIG_X86_64
327 #  define RETPOLINE_RCX_BPF_JIT_SIZE    17
328 #  define RETPOLINE_RCX_BPF_JIT()                               \
329 do {                                                            \
330         EMIT1_off32(0xE8, 7);    /* callq do_rop */             \
331         /* spec_trap: */                                        \
332         EMIT2(0xF3, 0x90);       /* pause */                    \
333         EMIT3(0x0F, 0xAE, 0xE8); /* lfence */                   \
334         EMIT2(0xEB, 0xF9);       /* jmp spec_trap */            \
335         /* do_rop: */                                           \
336         EMIT4(0x48, 0x89, 0x0C, 0x24); /* mov %rcx,(%rsp) */    \
337         EMIT1(0xC3);             /* retq */                     \
338 } while (0)
339 # else /* !CONFIG_X86_64 */
340 #  define RETPOLINE_EDX_BPF_JIT()                               \
341 do {                                                            \
342         EMIT1_off32(0xE8, 7);    /* call do_rop */              \
343         /* spec_trap: */                                        \
344         EMIT2(0xF3, 0x90);       /* pause */                    \
345         EMIT3(0x0F, 0xAE, 0xE8); /* lfence */                   \
346         EMIT2(0xEB, 0xF9);       /* jmp spec_trap */            \
347         /* do_rop: */                                           \
348         EMIT3(0x89, 0x14, 0x24); /* mov %edx,(%esp) */          \
349         EMIT1(0xC3);             /* ret */                      \
350 } while (0)
351 # endif
352 #else /* !CONFIG_RETPOLINE */
353 # ifdef CONFIG_X86_64
354 #  define RETPOLINE_RCX_BPF_JIT_SIZE    2
355 #  define RETPOLINE_RCX_BPF_JIT()                               \
356         EMIT2(0xFF, 0xE1);       /* jmp *%rcx */
357 # else /* !CONFIG_X86_64 */
358 #  define RETPOLINE_EDX_BPF_JIT()                               \
359         EMIT2(0xFF, 0xE2)        /* jmp *%edx */
360 # endif
361 #endif
362
363 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */