1/* SPDX-License-Identifier: GPL-2.0 */ 2 3#include <linux/export.h> 4#include <linux/stringify.h> 5#include <linux/linkage.h> 6#include <asm/dwarf2.h> 7#include <asm/cpufeatures.h> 8#include <asm/alternative.h> 9#include <asm/asm-offsets.h> 10#include <asm/nospec-branch.h> 11#include <asm/unwind_hints.h> 12#include <asm/percpu.h> 13#include <asm/frame.h> 14#include <asm/nops.h> 15 16 .section .text..__x86.indirect_thunk 17 18 19.macro POLINE reg 20 ANNOTATE_INTRA_FUNCTION_CALL 21 call .Ldo_rop_\@ 22 int3 23.Ldo_rop_\@: 24 mov %\reg, (%_ASM_SP) 25 UNWIND_HINT_FUNC 26.endm 27 28.macro RETPOLINE reg 29 POLINE \reg 30 RET 31.endm 32 33.macro THUNK reg 34 35 .align RETPOLINE_THUNK_SIZE 36SYM_INNER_LABEL(__x86_indirect_thunk_\reg, SYM_L_GLOBAL) 37 UNWIND_HINT_UNDEFINED 38 ANNOTATE_NOENDBR 39 40 ALTERNATIVE_2 __stringify(RETPOLINE \reg), \ 41 __stringify(lfence; ANNOTATE_RETPOLINE_SAFE; jmp *%\reg; int3), X86_FEATURE_RETPOLINE_LFENCE, \ 42 __stringify(ANNOTATE_RETPOLINE_SAFE; jmp *%\reg), ALT_NOT(X86_FEATURE_RETPOLINE) 43 44.endm 45 46/* 47 * Despite being an assembler file we can't just use .irp here 48 * because __KSYM_DEPS__ only uses the C preprocessor and would 49 * only see one instance of "__x86_indirect_thunk_\reg" rather 50 * than one per register with the correct names. So we do it 51 * the simple and nasty way... 52 * 53 * Worse, you can only have a single EXPORT_SYMBOL per line, 54 * and CPP can't insert newlines, so we have to repeat everything 55 * at least twice. 56 */ 57 58#define __EXPORT_THUNK(sym) _ASM_NOKPROBE(sym); EXPORT_SYMBOL(sym) 59 60 .align RETPOLINE_THUNK_SIZE 61SYM_CODE_START(__x86_indirect_thunk_array) 62 63#define GEN(reg) THUNK reg 64#include <asm/GEN-for-each-reg.h> 65#undef GEN 66 67 .align RETPOLINE_THUNK_SIZE 68SYM_CODE_END(__x86_indirect_thunk_array) 69 70#define GEN(reg) __EXPORT_THUNK(__x86_indirect_thunk_ ## reg) 71#include <asm/GEN-for-each-reg.h> 72#undef GEN 73 74#ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING 75.macro CALL_THUNK reg 76 .align RETPOLINE_THUNK_SIZE 77 78SYM_INNER_LABEL(__x86_indirect_call_thunk_\reg, SYM_L_GLOBAL) 79 UNWIND_HINT_UNDEFINED 80 ANNOTATE_NOENDBR 81 82 CALL_DEPTH_ACCOUNT 83 POLINE \reg 84 ANNOTATE_UNRET_SAFE 85 ret 86 int3 87.endm 88 89 .align RETPOLINE_THUNK_SIZE 90SYM_CODE_START(__x86_indirect_call_thunk_array) 91 92#define GEN(reg) CALL_THUNK reg 93#include <asm/GEN-for-each-reg.h> 94#undef GEN 95 96 .align RETPOLINE_THUNK_SIZE 97SYM_CODE_END(__x86_indirect_call_thunk_array) 98 99#define GEN(reg) __EXPORT_THUNK(__x86_indirect_call_thunk_ ## reg) 100#include <asm/GEN-for-each-reg.h> 101#undef GEN 102 103.macro JUMP_THUNK reg 104 .align RETPOLINE_THUNK_SIZE 105 106SYM_INNER_LABEL(__x86_indirect_jump_thunk_\reg, SYM_L_GLOBAL) 107 UNWIND_HINT_UNDEFINED 108 ANNOTATE_NOENDBR 109 POLINE \reg 110 ANNOTATE_UNRET_SAFE 111 ret 112 int3 113.endm 114 115 .align RETPOLINE_THUNK_SIZE 116SYM_CODE_START(__x86_indirect_jump_thunk_array) 117 118#define GEN(reg) JUMP_THUNK reg 119#include <asm/GEN-for-each-reg.h> 120#undef GEN 121 122 .align RETPOLINE_THUNK_SIZE 123SYM_CODE_END(__x86_indirect_jump_thunk_array) 124 125#define GEN(reg) __EXPORT_THUNK(__x86_indirect_jump_thunk_ ## reg) 126#include <asm/GEN-for-each-reg.h> 127#undef GEN 128#endif 129 130#ifdef CONFIG_MITIGATION_RETHUNK 131 132/* 133 * Be careful here: that label cannot really be removed because in 134 * some configurations and toolchains, the JMP __x86_return_thunk the 135 * compiler issues is either a short one or the compiler doesn't use 136 * relocations for same-section JMPs and that breaks the returns 137 * detection logic in apply_returns() and in objtool. 138 */ 139 .section .text..__x86.return_thunk 140 141#ifdef CONFIG_MITIGATION_SRSO 142 143/* 144 * srso_alias_untrain_ret() and srso_alias_safe_ret() are placed at 145 * special addresses: 146 * 147 * - srso_alias_untrain_ret() is 2M aligned 148 * - srso_alias_safe_ret() is also in the same 2M page but bits 2, 8, 14 149 * and 20 in its virtual address are set (while those bits in the 150 * srso_alias_untrain_ret() function are cleared). 151 * 152 * This guarantees that those two addresses will alias in the branch 153 * target buffer of Zen3/4 generations, leading to any potential 154 * poisoned entries at that BTB slot to get evicted. 155 * 156 * As a result, srso_alias_safe_ret() becomes a safe return. 157 */ 158 .pushsection .text..__x86.rethunk_untrain 159SYM_CODE_START_NOALIGN(srso_alias_untrain_ret) 160 UNWIND_HINT_FUNC 161 ANNOTATE_NOENDBR 162 ASM_NOP2 163 lfence 164 jmp srso_alias_return_thunk 165SYM_FUNC_END(srso_alias_untrain_ret) 166__EXPORT_THUNK(srso_alias_untrain_ret) 167 .popsection 168 169 .pushsection .text..__x86.rethunk_safe 170SYM_CODE_START_NOALIGN(srso_alias_safe_ret) 171 lea 8(%_ASM_SP), %_ASM_SP 172 UNWIND_HINT_FUNC 173 ANNOTATE_UNRET_SAFE 174 ret 175 int3 176SYM_FUNC_END(srso_alias_safe_ret) 177 178SYM_CODE_START_NOALIGN(srso_alias_return_thunk) 179 UNWIND_HINT_FUNC 180 ANNOTATE_NOENDBR 181 call srso_alias_safe_ret 182 ud2 183SYM_CODE_END(srso_alias_return_thunk) 184 .popsection 185 186/* 187 * SRSO untraining sequence for Zen1/2, similar to retbleed_untrain_ret() 188 * above. On kernel entry, srso_untrain_ret() is executed which is a 189 * 190 * movabs $0xccccc30824648d48,%rax 191 * 192 * and when the return thunk executes the inner label srso_safe_ret() 193 * later, it is a stack manipulation and a RET which is mispredicted and 194 * thus a "safe" one to use. 195 */ 196 .align 64 197 .skip 64 - (srso_safe_ret - srso_untrain_ret), 0xcc 198SYM_CODE_START_LOCAL_NOALIGN(srso_untrain_ret) 199 ANNOTATE_NOENDBR 200 .byte 0x48, 0xb8 201 202/* 203 * This forces the function return instruction to speculate into a trap 204 * (UD2 in srso_return_thunk() below). This RET will then mispredict 205 * and execution will continue at the return site read from the top of 206 * the stack. 207 */ 208SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL) 209 lea 8(%_ASM_SP), %_ASM_SP 210 ret 211 int3 212 int3 213 /* end of movabs */ 214 lfence 215 call srso_safe_ret 216 ud2 217SYM_CODE_END(srso_safe_ret) 218SYM_FUNC_END(srso_untrain_ret) 219 220SYM_CODE_START(srso_return_thunk) 221 UNWIND_HINT_FUNC 222 ANNOTATE_NOENDBR 223 call srso_safe_ret 224 ud2 225SYM_CODE_END(srso_return_thunk) 226 227#define JMP_SRSO_UNTRAIN_RET "jmp srso_untrain_ret" 228#else /* !CONFIG_MITIGATION_SRSO */ 229/* Dummy for the alternative in CALL_UNTRAIN_RET. */ 230SYM_CODE_START(srso_alias_untrain_ret) 231 RET 232SYM_FUNC_END(srso_alias_untrain_ret) 233#define JMP_SRSO_UNTRAIN_RET "ud2" 234#endif /* CONFIG_MITIGATION_SRSO */ 235 236#ifdef CONFIG_MITIGATION_UNRET_ENTRY 237 238/* 239 * Some generic notes on the untraining sequences: 240 * 241 * They are interchangeable when it comes to flushing potentially wrong 242 * RET predictions from the BTB. 243 * 244 * The SRSO Zen1/2 (MOVABS) untraining sequence is longer than the 245 * Retbleed sequence because the return sequence done there 246 * (srso_safe_ret()) is longer and the return sequence must fully nest 247 * (end before) the untraining sequence. Therefore, the untraining 248 * sequence must fully overlap the return sequence. 249 * 250 * Regarding alignment - the instructions which need to be untrained, 251 * must all start at a cacheline boundary for Zen1/2 generations. That 252 * is, instruction sequences starting at srso_safe_ret() and 253 * the respective instruction sequences at retbleed_return_thunk() 254 * must start at a cacheline boundary. 255 */ 256 257/* 258 * Safety details here pertain to the AMD Zen{1,2} microarchitecture: 259 * 1) The RET at retbleed_return_thunk must be on a 64 byte boundary, for 260 * alignment within the BTB. 261 * 2) The instruction at retbleed_untrain_ret must contain, and not 262 * end with, the 0xc3 byte of the RET. 263 * 3) STIBP must be enabled, or SMT disabled, to prevent the sibling thread 264 * from re-poisioning the BTB prediction. 265 */ 266 .align 64 267 .skip 64 - (retbleed_return_thunk - retbleed_untrain_ret), 0xcc 268SYM_CODE_START_LOCAL_NOALIGN(retbleed_untrain_ret) 269 ANNOTATE_NOENDBR 270 /* 271 * As executed from retbleed_untrain_ret, this is: 272 * 273 * TEST $0xcc, %bl 274 * LFENCE 275 * JMP retbleed_return_thunk 276 * 277 * Executing the TEST instruction has a side effect of evicting any BTB 278 * prediction (potentially attacker controlled) attached to the RET, as 279 * retbleed_return_thunk + 1 isn't an instruction boundary at the moment. 280 */ 281 .byte 0xf6 282 283 /* 284 * As executed from retbleed_return_thunk, this is a plain RET. 285 * 286 * As part of the TEST above, RET is the ModRM byte, and INT3 the imm8. 287 * 288 * We subsequently jump backwards and architecturally execute the RET. 289 * This creates a correct BTB prediction (type=ret), but in the 290 * meantime we suffer Straight Line Speculation (because the type was 291 * no branch) which is halted by the INT3. 292 * 293 * With SMT enabled and STIBP active, a sibling thread cannot poison 294 * RET's prediction to a type of its choice, but can evict the 295 * prediction due to competitive sharing. If the prediction is 296 * evicted, retbleed_return_thunk will suffer Straight Line Speculation 297 * which will be contained safely by the INT3. 298 */ 299SYM_INNER_LABEL(retbleed_return_thunk, SYM_L_GLOBAL) 300 ret 301 int3 302SYM_CODE_END(retbleed_return_thunk) 303 304 /* 305 * Ensure the TEST decoding / BTB invalidation is complete. 306 */ 307 lfence 308 309 /* 310 * Jump back and execute the RET in the middle of the TEST instruction. 311 * INT3 is for SLS protection. 312 */ 313 jmp retbleed_return_thunk 314 int3 315SYM_FUNC_END(retbleed_untrain_ret) 316 317#define JMP_RETBLEED_UNTRAIN_RET "jmp retbleed_untrain_ret" 318#else /* !CONFIG_MITIGATION_UNRET_ENTRY */ 319#define JMP_RETBLEED_UNTRAIN_RET "ud2" 320#endif /* CONFIG_MITIGATION_UNRET_ENTRY */ 321 322#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || defined(CONFIG_MITIGATION_SRSO) 323 324SYM_FUNC_START(entry_untrain_ret) 325 ALTERNATIVE JMP_RETBLEED_UNTRAIN_RET, JMP_SRSO_UNTRAIN_RET, X86_FEATURE_SRSO 326SYM_FUNC_END(entry_untrain_ret) 327__EXPORT_THUNK(entry_untrain_ret) 328 329#endif /* CONFIG_MITIGATION_UNRET_ENTRY || CONFIG_MITIGATION_SRSO */ 330 331#ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING 332 333 .align 64 334SYM_FUNC_START(call_depth_return_thunk) 335 ANNOTATE_NOENDBR 336 /* 337 * Keep the hotpath in a 16byte I-fetch for the non-debug 338 * case. 339 */ 340 CALL_THUNKS_DEBUG_INC_RETS 341 shlq $5, PER_CPU_VAR(pcpu_hot + X86_call_depth) 342 jz 1f 343 ANNOTATE_UNRET_SAFE 344 ret 345 int3 3461: 347 CALL_THUNKS_DEBUG_INC_STUFFS 348 .rept 16 349 ANNOTATE_INTRA_FUNCTION_CALL 350 call 2f 351 int3 3522: 353 .endr 354 add $(8*16), %rsp 355 356 CREDIT_CALL_DEPTH 357 358 ANNOTATE_UNRET_SAFE 359 ret 360 int3 361SYM_FUNC_END(call_depth_return_thunk) 362 363#endif /* CONFIG_MITIGATION_CALL_DEPTH_TRACKING */ 364 365/* 366 * This function name is magical and is used by -mfunction-return=thunk-extern 367 * for the compiler to generate JMPs to it. 368 * 369 * This code is only used during kernel boot or module init. All 370 * 'JMP __x86_return_thunk' sites are changed to something else by 371 * apply_returns(). 372 * 373 * The ALTERNATIVE below adds a really loud warning to catch the case 374 * where the insufficient default return thunk ends up getting used for 375 * whatever reason like miscompilation or failure of 376 * objtool/alternatives/etc to patch all the return sites. 377 */ 378SYM_CODE_START(__x86_return_thunk) 379 UNWIND_HINT_FUNC 380 ANNOTATE_NOENDBR 381 ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \ 382 "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS 383 int3 384SYM_CODE_END(__x86_return_thunk) 385EXPORT_SYMBOL(__x86_return_thunk) 386 387#endif /* CONFIG_MITIGATION_RETHUNK */ 388