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_AMD 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_AMD 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 #else /* __ASSEMBLY__ */ 116 117 #define ANNOTATE_RETPOLINE_SAFE \ 118 "999:\n\t" \ 119 ".pushsection .discard.retpoline_safe\n\t" \ 120 _ASM_PTR " 999b\n\t" \ 121 ".popsection\n\t" 122 123 #ifdef CONFIG_RETPOLINE 124 125 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE]; 126 127 #define GEN(reg) \ 128 extern retpoline_thunk_t __x86_indirect_thunk_ ## reg; 129 #include <asm/GEN-for-each-reg.h> 130 #undef GEN 131 132 extern retpoline_thunk_t __x86_indirect_thunk_array[]; 133 134 #ifdef CONFIG_X86_64 135 136 /* 137 * Inline asm uses the %V modifier which is only in newer GCC 138 * which is ensured when CONFIG_RETPOLINE is defined. 139 */ 140 # define CALL_NOSPEC \ 141 ALTERNATIVE_2( \ 142 ANNOTATE_RETPOLINE_SAFE \ 143 "call *%[thunk_target]\n", \ 144 "call __x86_indirect_thunk_%V[thunk_target]\n", \ 145 X86_FEATURE_RETPOLINE, \ 146 "lfence;\n" \ 147 ANNOTATE_RETPOLINE_SAFE \ 148 "call *%[thunk_target]\n", \ 149 X86_FEATURE_RETPOLINE_AMD) 150 151 # define THUNK_TARGET(addr) [thunk_target] "r" (addr) 152 153 #else /* CONFIG_X86_32 */ 154 /* 155 * For i386 we use the original ret-equivalent retpoline, because 156 * otherwise we'll run out of registers. We don't care about CET 157 * here, anyway. 158 */ 159 # define CALL_NOSPEC \ 160 ALTERNATIVE_2( \ 161 ANNOTATE_RETPOLINE_SAFE \ 162 "call *%[thunk_target]\n", \ 163 " jmp 904f;\n" \ 164 " .align 16\n" \ 165 "901: call 903f;\n" \ 166 "902: pause;\n" \ 167 " lfence;\n" \ 168 " jmp 902b;\n" \ 169 " .align 16\n" \ 170 "903: lea 4(%%esp), %%esp;\n" \ 171 " pushl %[thunk_target];\n" \ 172 " ret;\n" \ 173 " .align 16\n" \ 174 "904: call 901b;\n", \ 175 X86_FEATURE_RETPOLINE, \ 176 "lfence;\n" \ 177 ANNOTATE_RETPOLINE_SAFE \ 178 "call *%[thunk_target]\n", \ 179 X86_FEATURE_RETPOLINE_AMD) 180 181 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 182 #endif 183 #else /* No retpoline for C / inline asm */ 184 # define CALL_NOSPEC "call *%[thunk_target]\n" 185 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 186 #endif 187 188 /* The Spectre V2 mitigation variants */ 189 enum spectre_v2_mitigation { 190 SPECTRE_V2_NONE, 191 SPECTRE_V2_RETPOLINE_GENERIC, 192 SPECTRE_V2_RETPOLINE_AMD, 193 SPECTRE_V2_IBRS_ENHANCED, 194 }; 195 196 /* The indirect branch speculation control variants */ 197 enum spectre_v2_user_mitigation { 198 SPECTRE_V2_USER_NONE, 199 SPECTRE_V2_USER_STRICT, 200 SPECTRE_V2_USER_STRICT_PREFERRED, 201 SPECTRE_V2_USER_PRCTL, 202 SPECTRE_V2_USER_SECCOMP, 203 }; 204 205 /* The Speculative Store Bypass disable variants */ 206 enum ssb_mitigation { 207 SPEC_STORE_BYPASS_NONE, 208 SPEC_STORE_BYPASS_DISABLE, 209 SPEC_STORE_BYPASS_PRCTL, 210 SPEC_STORE_BYPASS_SECCOMP, 211 }; 212 213 extern char __indirect_thunk_start[]; 214 extern char __indirect_thunk_end[]; 215 216 static __always_inline 217 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature) 218 { 219 asm volatile(ALTERNATIVE("", "wrmsr", %c[feature]) 220 : : "c" (msr), 221 "a" ((u32)val), 222 "d" ((u32)(val >> 32)), 223 [feature] "i" (feature) 224 : "memory"); 225 } 226 227 static inline void indirect_branch_prediction_barrier(void) 228 { 229 u64 val = PRED_CMD_IBPB; 230 231 alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB); 232 } 233 234 /* The Intel SPEC CTRL MSR base value cache */ 235 extern u64 x86_spec_ctrl_base; 236 237 /* 238 * With retpoline, we must use IBRS to restrict branch prediction 239 * before calling into firmware. 240 * 241 * (Implemented as CPP macros due to header hell.) 242 */ 243 #define firmware_restrict_branch_speculation_start() \ 244 do { \ 245 u64 val = x86_spec_ctrl_base | SPEC_CTRL_IBRS; \ 246 \ 247 preempt_disable(); \ 248 alternative_msr_write(MSR_IA32_SPEC_CTRL, val, \ 249 X86_FEATURE_USE_IBRS_FW); \ 250 } while (0) 251 252 #define firmware_restrict_branch_speculation_end() \ 253 do { \ 254 u64 val = x86_spec_ctrl_base; \ 255 \ 256 alternative_msr_write(MSR_IA32_SPEC_CTRL, val, \ 257 X86_FEATURE_USE_IBRS_FW); \ 258 preempt_enable(); \ 259 } while (0) 260 261 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); 264 265 DECLARE_STATIC_KEY_FALSE(mds_user_clear); 266 DECLARE_STATIC_KEY_FALSE(mds_idle_clear); 267 268 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush); 269 270 #include <asm/segment.h> 271 272 /** 273 * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability 274 * 275 * This uses the otherwise unused and obsolete VERW instruction in 276 * combination with microcode which triggers a CPU buffer flush when the 277 * instruction is executed. 278 */ 279 static __always_inline void mds_clear_cpu_buffers(void) 280 { 281 static const u16 ds = __KERNEL_DS; 282 283 /* 284 * Has to be the memory-operand variant because only that 285 * guarantees the CPU buffer flush functionality according to 286 * documentation. The register-operand variant does not. 287 * Works with any segment selector, but a valid writable 288 * data segment is the fastest variant. 289 * 290 * "cc" clobber is required because VERW modifies ZF. 291 */ 292 asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc"); 293 } 294 295 /** 296 * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability 297 * 298 * Clear CPU buffers if the corresponding static key is enabled 299 */ 300 static __always_inline void mds_user_clear_cpu_buffers(void) 301 { 302 if (static_branch_likely(&mds_user_clear)) 303 mds_clear_cpu_buffers(); 304 } 305 306 /** 307 * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability 308 * 309 * Clear CPU buffers if the corresponding static key is enabled 310 */ 311 static inline void mds_idle_clear_cpu_buffers(void) 312 { 313 if (static_branch_likely(&mds_idle_clear)) 314 mds_clear_cpu_buffers(); 315 } 316 317 #endif /* __ASSEMBLY__ */ 318 319 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */ 320