1 /* SPDX-License-Identifier: GPL-2.0 */
3 #ifndef _ASM_X86_NOSPEC_BRANCH_H_
4 #define _ASM_X86_NOSPEC_BRANCH_H_
6 #include <linux/static_key.h>
8 #include <asm/alternative.h>
9 #include <asm/alternative-asm.h>
10 #include <asm/cpufeatures.h>
11 #include <asm/msr-index.h>
14 * Fill the CPU return stack buffer.
16 * Each entry in the RSB, if used for a speculative 'ret', contains an
17 * infinite 'pause; lfence; jmp' loop to capture speculative execution.
19 * This is required in various cases for retpoline and IBRS-based
20 * mitigations for the Spectre variant 2 vulnerability. Sometimes to
21 * eliminate potentially bogus entries from the RSB, and sometimes
22 * purely to ensure that it doesn't get empty, which on some CPUs would
23 * allow predictions from other (unwanted!) sources to be used.
25 * We define a CPP macro such that it can be used from both .S files and
26 * inline assembly. It's possible to do a .macro and then include that
27 * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
30 #define RSB_CLEAR_LOOPS 32 /* To forcibly overwrite all entries */
31 #define RSB_FILL_LOOPS 16 /* To avoid underflow */
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.
38 #define __FILL_RETURN_BUFFER(reg, nr, sp) \
42 773: /* speculation trap */ \
48 775: /* speculation trap */ \
55 add $(BITS_PER_LONG/8) * nr, sp;
60 * These are the bare retpoline primitives for indirect jmp and call.
61 * Do not use these directly; they only exist to make the ALTERNATIVE
62 * invocation below less ugly.
64 .macro RETPOLINE_JMP reg:req
76 * This is a wrapper around RETPOLINE_JMP so the called function in reg
77 * returns to the instruction after the macro.
79 .macro RETPOLINE_CALL reg:req
81 .Ldo_retpoline_jmp_\@:
84 call .Ldo_retpoline_jmp_\@
88 * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
89 * indirect jmp/call which may be susceptible to the Spectre variant 2
92 .macro JMP_NOSPEC reg:req
93 #ifdef CONFIG_RETPOLINE
94 ALTERNATIVE_2 __stringify(jmp *\reg), \
95 __stringify(RETPOLINE_JMP \reg), X86_FEATURE_RETPOLINE, \
96 __stringify(lfence; jmp *\reg), X86_FEATURE_RETPOLINE_AMD
102 .macro CALL_NOSPEC reg:req
103 #ifdef CONFIG_RETPOLINE
104 ALTERNATIVE_2 __stringify(call *\reg), \
105 __stringify(RETPOLINE_CALL \reg), X86_FEATURE_RETPOLINE,\
106 __stringify(lfence; call *\reg), X86_FEATURE_RETPOLINE_AMD
113 * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
114 * monstrosity above, manually.
116 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req
117 #ifdef CONFIG_RETPOLINE
118 ALTERNATIVE "jmp .Lskip_rsb_\@", \
119 __stringify(__FILL_RETURN_BUFFER(\reg,\nr,%_ASM_SP)) \
125 #else /* __ASSEMBLY__ */
127 #if defined(CONFIG_X86_64) && defined(RETPOLINE)
130 * Since the inline asm uses the %V modifier which is only in newer GCC,
131 * the 64-bit one is dependent on RETPOLINE not CONFIG_RETPOLINE.
133 # define CALL_NOSPEC \
135 "call *%[thunk_target]\n", \
136 "call __x86_indirect_thunk_%V[thunk_target]\n", \
137 X86_FEATURE_RETPOLINE)
138 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
140 #elif defined(CONFIG_X86_32) && defined(CONFIG_RETPOLINE)
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
146 # define CALL_NOSPEC ALTERNATIVE("call *%[thunk_target]\n", \
149 "901: call 903f;\n" \
154 "903: addl $4, %%esp;\n" \
155 " pushl %[thunk_target];\n" \
158 "904: call 901b;\n", \
159 X86_FEATURE_RETPOLINE)
161 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
162 #else /* No retpoline for C / inline asm */
163 # define CALL_NOSPEC "call *%[thunk_target]\n"
164 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
167 /* The Spectre V2 mitigation variants */
168 enum spectre_v2_mitigation {
170 SPECTRE_V2_RETPOLINE_MINIMAL,
171 SPECTRE_V2_RETPOLINE_MINIMAL_AMD,
172 SPECTRE_V2_RETPOLINE_GENERIC,
173 SPECTRE_V2_RETPOLINE_AMD,
174 SPECTRE_V2_IBRS_ENHANCED,
177 /* The indirect branch speculation control variants */
178 enum spectre_v2_user_mitigation {
179 SPECTRE_V2_USER_NONE,
180 SPECTRE_V2_USER_STRICT,
181 SPECTRE_V2_USER_PRCTL,
182 SPECTRE_V2_USER_SECCOMP,
185 /* The Speculative Store Bypass disable variants */
186 enum ssb_mitigation {
187 SPEC_STORE_BYPASS_NONE,
188 SPEC_STORE_BYPASS_DISABLE,
189 SPEC_STORE_BYPASS_PRCTL,
190 SPEC_STORE_BYPASS_SECCOMP,
193 extern char __indirect_thunk_start[];
194 extern char __indirect_thunk_end[];
197 * On VMEXIT we must ensure that no RSB predictions learned in the guest
198 * can be followed in the host, by overwriting the RSB completely. Both
199 * retpoline and IBRS mitigations for Spectre v2 need this; only on future
200 * CPUs with IBRS_ALL *might* it be avoided.
202 static inline void vmexit_fill_RSB(void)
204 #ifdef CONFIG_RETPOLINE
207 asm volatile (ALTERNATIVE("jmp 910f",
208 __stringify(__FILL_RETURN_BUFFER(%0, RSB_CLEAR_LOOPS, %1)),
209 X86_FEATURE_RETPOLINE)
211 : "=r" (loops), ASM_CALL_CONSTRAINT
216 static __always_inline
217 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
219 asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
222 "d" ((u32)(val >> 32)),
223 [feature] "i" (feature)
227 static inline void indirect_branch_prediction_barrier(void)
229 u64 val = PRED_CMD_IBPB;
231 alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
234 /* The Intel SPEC CTRL MSR base value cache */
235 extern u64 x86_spec_ctrl_base;
238 * With retpoline, we must use IBRS to restrict branch prediction
239 * before calling into firmware.
241 * (Implemented as CPP macros due to header hell.)
243 #define firmware_restrict_branch_speculation_start() \
245 u64 val = x86_spec_ctrl_base | SPEC_CTRL_IBRS; \
248 alternative_msr_write(MSR_IA32_SPEC_CTRL, val, \
249 X86_FEATURE_USE_IBRS_FW); \
252 #define firmware_restrict_branch_speculation_end() \
254 u64 val = x86_spec_ctrl_base; \
256 alternative_msr_write(MSR_IA32_SPEC_CTRL, val, \
257 X86_FEATURE_USE_IBRS_FW); \
261 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
262 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
263 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
265 #include <asm/segment.h>
268 * mds_clear_cpu_buffers - Mitigation for MDS vulnerability
270 * This uses the otherwise unused and obsolete VERW instruction in
271 * combination with microcode which triggers a CPU buffer flush when the
272 * instruction is executed.
274 static inline void mds_clear_cpu_buffers(void)
276 static const u16 ds = __KERNEL_DS;
279 * Has to be the memory-operand variant because only that
280 * guarantees the CPU buffer flush functionality according to
281 * documentation. The register-operand variant does not.
282 * Works with any segment selector, but a valid writable
283 * data segment is the fastest variant.
285 * "cc" clobber is required because VERW modifies ZF.
287 asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
290 #endif /* __ASSEMBLY__ */
293 * Below is used in the eBPF JIT compiler and emits the byte sequence
294 * for the following assembly:
296 * With retpolines configured:
307 * Without retpolines configured:
311 #ifdef CONFIG_RETPOLINE
312 # define RETPOLINE_RAX_BPF_JIT_SIZE 17
313 # define RETPOLINE_RAX_BPF_JIT() \
314 EMIT1_off32(0xE8, 7); /* callq do_rop */ \
316 EMIT2(0xF3, 0x90); /* pause */ \
317 EMIT3(0x0F, 0xAE, 0xE8); /* lfence */ \
318 EMIT2(0xEB, 0xF9); /* jmp spec_trap */ \
320 EMIT4(0x48, 0x89, 0x04, 0x24); /* mov %rax,(%rsp) */ \
321 EMIT1(0xC3); /* retq */
323 # define RETPOLINE_RAX_BPF_JIT_SIZE 2
324 # define RETPOLINE_RAX_BPF_JIT() \
325 EMIT2(0xFF, 0xE0); /* jmp *%rax */
328 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */