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#include <asm/percpu.h> 15#include <asm/current.h> 16 17/* 18 * Call depth tracking for Intel SKL CPUs to address the RSB underflow 19 * issue in software. 20 * 21 * The tracking does not use a counter. It uses uses arithmetic shift 22 * right on call entry and logical shift left on return. 23 * 24 * The depth tracking variable is initialized to 0x8000.... when the call 25 * depth is zero. The arithmetic shift right sign extends the MSB and 26 * saturates after the 12th call. The shift count is 5 for both directions 27 * so the tracking covers 12 nested calls. 28 * 29 * Call 30 * 0: 0x8000000000000000 0x0000000000000000 31 * 1: 0xfc00000000000000 0xf000000000000000 32 * ... 33 * 11: 0xfffffffffffffff8 0xfffffffffffffc00 34 * 12: 0xffffffffffffffff 0xffffffffffffffe0 35 * 36 * After a return buffer fill the depth is credited 12 calls before the 37 * next stuffing has to take place. 38 * 39 * There is a inaccuracy for situations like this: 40 * 41 * 10 calls 42 * 5 returns 43 * 3 calls 44 * 4 returns 45 * 3 calls 46 * .... 47 * 48 * The shift count might cause this to be off by one in either direction, 49 * but there is still a cushion vs. the RSB depth. The algorithm does not 50 * claim to be perfect and it can be speculated around by the CPU, but it 51 * is considered that it obfuscates the problem enough to make exploitation 52 * extremly difficult. 53 */ 54#define RET_DEPTH_SHIFT 5 55#define RSB_RET_STUFF_LOOPS 16 56#define RET_DEPTH_INIT 0x8000000000000000ULL 57#define RET_DEPTH_INIT_FROM_CALL 0xfc00000000000000ULL 58#define RET_DEPTH_CREDIT 0xffffffffffffffffULL 59 60#ifdef CONFIG_CALL_THUNKS_DEBUG 61# define CALL_THUNKS_DEBUG_INC_CALLS \ 62 incq %gs:__x86_call_count; 63# define CALL_THUNKS_DEBUG_INC_RETS \ 64 incq %gs:__x86_ret_count; 65# define CALL_THUNKS_DEBUG_INC_STUFFS \ 66 incq %gs:__x86_stuffs_count; 67# define CALL_THUNKS_DEBUG_INC_CTXSW \ 68 incq %gs:__x86_ctxsw_count; 69#else 70# define CALL_THUNKS_DEBUG_INC_CALLS 71# define CALL_THUNKS_DEBUG_INC_RETS 72# define CALL_THUNKS_DEBUG_INC_STUFFS 73# define CALL_THUNKS_DEBUG_INC_CTXSW 74#endif 75 76#if defined(CONFIG_CALL_DEPTH_TRACKING) && !defined(COMPILE_OFFSETS) 77 78#include <asm/asm-offsets.h> 79 80#define CREDIT_CALL_DEPTH \ 81 movq $-1, PER_CPU_VAR(pcpu_hot + X86_call_depth); 82 83#define ASM_CREDIT_CALL_DEPTH \ 84 movq $-1, PER_CPU_VAR(pcpu_hot + X86_call_depth); 85 86#define RESET_CALL_DEPTH \ 87 xor %eax, %eax; \ 88 bts $63, %rax; \ 89 movq %rax, PER_CPU_VAR(pcpu_hot + X86_call_depth); 90 91#define RESET_CALL_DEPTH_FROM_CALL \ 92 movb $0xfc, %al; \ 93 shl $56, %rax; \ 94 movq %rax, PER_CPU_VAR(pcpu_hot + X86_call_depth); \ 95 CALL_THUNKS_DEBUG_INC_CALLS 96 97#define INCREMENT_CALL_DEPTH \ 98 sarq $5, %gs:pcpu_hot + X86_call_depth; \ 99 CALL_THUNKS_DEBUG_INC_CALLS 100 101#define ASM_INCREMENT_CALL_DEPTH \ 102 sarq $5, PER_CPU_VAR(pcpu_hot + X86_call_depth); \ 103 CALL_THUNKS_DEBUG_INC_CALLS 104 105#else 106#define CREDIT_CALL_DEPTH 107#define ASM_CREDIT_CALL_DEPTH 108#define RESET_CALL_DEPTH 109#define INCREMENT_CALL_DEPTH 110#define ASM_INCREMENT_CALL_DEPTH 111#define RESET_CALL_DEPTH_FROM_CALL 112#endif 113 114/* 115 * Fill the CPU return stack buffer. 116 * 117 * Each entry in the RSB, if used for a speculative 'ret', contains an 118 * infinite 'pause; lfence; jmp' loop to capture speculative execution. 119 * 120 * This is required in various cases for retpoline and IBRS-based 121 * mitigations for the Spectre variant 2 vulnerability. Sometimes to 122 * eliminate potentially bogus entries from the RSB, and sometimes 123 * purely to ensure that it doesn't get empty, which on some CPUs would 124 * allow predictions from other (unwanted!) sources to be used. 125 * 126 * We define a CPP macro such that it can be used from both .S files and 127 * inline assembly. It's possible to do a .macro and then include that 128 * from C via asm(".include <asm/nospec-branch.h>") but let's not go there. 129 */ 130 131#define RETPOLINE_THUNK_SIZE 32 132#define RSB_CLEAR_LOOPS 32 /* To forcibly overwrite all entries */ 133 134/* 135 * Common helper for __FILL_RETURN_BUFFER and __FILL_ONE_RETURN. 136 */ 137#define __FILL_RETURN_SLOT \ 138 ANNOTATE_INTRA_FUNCTION_CALL; \ 139 call 772f; \ 140 int3; \ 141772: 142 143/* 144 * Stuff the entire RSB. 145 * 146 * Google experimented with loop-unrolling and this turned out to be 147 * the optimal version - two calls, each with their own speculation 148 * trap should their return address end up getting used, in a loop. 149 */ 150#ifdef CONFIG_X86_64 151#define __FILL_RETURN_BUFFER(reg, nr) \ 152 mov $(nr/2), reg; \ 153771: \ 154 __FILL_RETURN_SLOT \ 155 __FILL_RETURN_SLOT \ 156 add $(BITS_PER_LONG/8) * 2, %_ASM_SP; \ 157 dec reg; \ 158 jnz 771b; \ 159 /* barrier for jnz misprediction */ \ 160 lfence; \ 161 ASM_CREDIT_CALL_DEPTH \ 162 CALL_THUNKS_DEBUG_INC_CTXSW 163#else 164/* 165 * i386 doesn't unconditionally have LFENCE, as such it can't 166 * do a loop. 167 */ 168#define __FILL_RETURN_BUFFER(reg, nr) \ 169 .rept nr; \ 170 __FILL_RETURN_SLOT; \ 171 .endr; \ 172 add $(BITS_PER_LONG/8) * nr, %_ASM_SP; 173#endif 174 175/* 176 * Stuff a single RSB slot. 177 * 178 * To mitigate Post-Barrier RSB speculation, one CALL instruction must be 179 * forced to retire before letting a RET instruction execute. 180 * 181 * On PBRSB-vulnerable CPUs, it is not safe for a RET to be executed 182 * before this point. 183 */ 184#define __FILL_ONE_RETURN \ 185 __FILL_RETURN_SLOT \ 186 add $(BITS_PER_LONG/8), %_ASM_SP; \ 187 lfence; 188 189#ifdef __ASSEMBLY__ 190 191/* 192 * This should be used immediately before an indirect jump/call. It tells 193 * objtool the subsequent indirect jump/call is vouched safe for retpoline 194 * builds. 195 */ 196.macro ANNOTATE_RETPOLINE_SAFE 197.Lhere_\@: 198 .pushsection .discard.retpoline_safe 199 .long .Lhere_\@ 200 .popsection 201.endm 202 203/* 204 * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions 205 * vs RETBleed validation. 206 */ 207#define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE 208 209/* 210 * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should 211 * eventually turn into it's own annotation. 212 */ 213.macro VALIDATE_UNRET_END 214#if defined(CONFIG_NOINSTR_VALIDATION) && \ 215 (defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO)) 216 ANNOTATE_RETPOLINE_SAFE 217 nop 218#endif 219.endm 220 221/* 222 * Equivalent to -mindirect-branch-cs-prefix; emit the 5 byte jmp/call 223 * to the retpoline thunk with a CS prefix when the register requires 224 * a RAX prefix byte to encode. Also see apply_retpolines(). 225 */ 226.macro __CS_PREFIX reg:req 227 .irp rs,r8,r9,r10,r11,r12,r13,r14,r15 228 .ifc \reg,\rs 229 .byte 0x2e 230 .endif 231 .endr 232.endm 233 234/* 235 * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple 236 * indirect jmp/call which may be susceptible to the Spectre variant 2 237 * attack. 238 * 239 * NOTE: these do not take kCFI into account and are thus not comparable to C 240 * indirect calls, take care when using. The target of these should be an ENDBR 241 * instruction irrespective of kCFI. 242 */ 243.macro JMP_NOSPEC reg:req 244#ifdef CONFIG_RETPOLINE 245 __CS_PREFIX \reg 246 jmp __x86_indirect_thunk_\reg 247#else 248 jmp *%\reg 249 int3 250#endif 251.endm 252 253.macro CALL_NOSPEC reg:req 254#ifdef CONFIG_RETPOLINE 255 __CS_PREFIX \reg 256 call __x86_indirect_thunk_\reg 257#else 258 call *%\reg 259#endif 260.endm 261 262 /* 263 * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP 264 * monstrosity above, manually. 265 */ 266.macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS) 267 ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \ 268 __stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \ 269 __stringify(nop;nop;__FILL_ONE_RETURN), \ftr2 270 271.Lskip_rsb_\@: 272.endm 273 274#if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO) 275#define CALL_UNTRAIN_RET "call entry_untrain_ret" 276#else 277#define CALL_UNTRAIN_RET "" 278#endif 279 280/* 281 * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the 282 * return thunk isn't mapped into the userspace tables (then again, AMD 283 * typically has NO_MELTDOWN). 284 * 285 * While retbleed_untrain_ret() doesn't clobber anything but requires stack, 286 * entry_ibpb() will clobber AX, CX, DX. 287 * 288 * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point 289 * where we have a stack but before any RET instruction. 290 */ 291.macro UNTRAIN_RET 292#if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 293 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 294 VALIDATE_UNRET_END 295 ALTERNATIVE_3 "", \ 296 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 297 "call entry_ibpb", X86_FEATURE_ENTRY_IBPB, \ 298 __stringify(RESET_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 299#endif 300.endm 301 302.macro UNTRAIN_RET_VM 303#if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 304 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 305 VALIDATE_UNRET_END 306 ALTERNATIVE_3 "", \ 307 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 308 "call entry_ibpb", X86_FEATURE_IBPB_ON_VMEXIT, \ 309 __stringify(RESET_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 310#endif 311.endm 312 313.macro UNTRAIN_RET_FROM_CALL 314#if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 315 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 316 VALIDATE_UNRET_END 317 ALTERNATIVE_3 "", \ 318 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 319 "call entry_ibpb", X86_FEATURE_ENTRY_IBPB, \ 320 __stringify(RESET_CALL_DEPTH_FROM_CALL), X86_FEATURE_CALL_DEPTH 321#endif 322.endm 323 324 325.macro CALL_DEPTH_ACCOUNT 326#ifdef CONFIG_CALL_DEPTH_TRACKING 327 ALTERNATIVE "", \ 328 __stringify(ASM_INCREMENT_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 329#endif 330.endm 331 332/* 333 * Macro to execute VERW instruction that mitigate transient data sampling 334 * attacks such as MDS. On affected systems a microcode update overloaded VERW 335 * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF. 336 * 337 * Note: Only the memory operand variant of VERW clears the CPU buffers. 338 */ 339.macro CLEAR_CPU_BUFFERS 340 ALTERNATIVE "", __stringify(verw _ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF 341.endm 342 343#else /* __ASSEMBLY__ */ 344 345#define ANNOTATE_RETPOLINE_SAFE \ 346 "999:\n\t" \ 347 ".pushsection .discard.retpoline_safe\n\t" \ 348 ".long 999b\n\t" \ 349 ".popsection\n\t" 350 351typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE]; 352extern retpoline_thunk_t __x86_indirect_thunk_array[]; 353extern retpoline_thunk_t __x86_indirect_call_thunk_array[]; 354extern retpoline_thunk_t __x86_indirect_jump_thunk_array[]; 355 356#ifdef CONFIG_RETHUNK 357extern void __x86_return_thunk(void); 358#else 359static inline void __x86_return_thunk(void) {} 360#endif 361 362extern void retbleed_return_thunk(void); 363extern void srso_return_thunk(void); 364extern void srso_alias_return_thunk(void); 365 366extern void retbleed_untrain_ret(void); 367extern void srso_untrain_ret(void); 368extern void srso_alias_untrain_ret(void); 369 370extern void entry_untrain_ret(void); 371extern void entry_ibpb(void); 372 373extern void (*x86_return_thunk)(void); 374 375#ifdef CONFIG_CALL_DEPTH_TRACKING 376extern void __x86_return_skl(void); 377 378static inline void x86_set_skl_return_thunk(void) 379{ 380 x86_return_thunk = &__x86_return_skl; 381} 382 383#define CALL_DEPTH_ACCOUNT \ 384 ALTERNATIVE("", \ 385 __stringify(INCREMENT_CALL_DEPTH), \ 386 X86_FEATURE_CALL_DEPTH) 387 388#ifdef CONFIG_CALL_THUNKS_DEBUG 389DECLARE_PER_CPU(u64, __x86_call_count); 390DECLARE_PER_CPU(u64, __x86_ret_count); 391DECLARE_PER_CPU(u64, __x86_stuffs_count); 392DECLARE_PER_CPU(u64, __x86_ctxsw_count); 393#endif 394#else 395static inline void x86_set_skl_return_thunk(void) {} 396 397#define CALL_DEPTH_ACCOUNT "" 398 399#endif 400 401#ifdef CONFIG_RETPOLINE 402 403#define GEN(reg) \ 404 extern retpoline_thunk_t __x86_indirect_thunk_ ## reg; 405#include <asm/GEN-for-each-reg.h> 406#undef GEN 407 408#define GEN(reg) \ 409 extern retpoline_thunk_t __x86_indirect_call_thunk_ ## reg; 410#include <asm/GEN-for-each-reg.h> 411#undef GEN 412 413#define GEN(reg) \ 414 extern retpoline_thunk_t __x86_indirect_jump_thunk_ ## reg; 415#include <asm/GEN-for-each-reg.h> 416#undef GEN 417 418#ifdef CONFIG_X86_64 419 420/* 421 * Inline asm uses the %V modifier which is only in newer GCC 422 * which is ensured when CONFIG_RETPOLINE is defined. 423 */ 424# define CALL_NOSPEC \ 425 ALTERNATIVE_2( \ 426 ANNOTATE_RETPOLINE_SAFE \ 427 "call *%[thunk_target]\n", \ 428 "call __x86_indirect_thunk_%V[thunk_target]\n", \ 429 X86_FEATURE_RETPOLINE, \ 430 "lfence;\n" \ 431 ANNOTATE_RETPOLINE_SAFE \ 432 "call *%[thunk_target]\n", \ 433 X86_FEATURE_RETPOLINE_LFENCE) 434 435# define THUNK_TARGET(addr) [thunk_target] "r" (addr) 436 437#else /* CONFIG_X86_32 */ 438/* 439 * For i386 we use the original ret-equivalent retpoline, because 440 * otherwise we'll run out of registers. We don't care about CET 441 * here, anyway. 442 */ 443# define CALL_NOSPEC \ 444 ALTERNATIVE_2( \ 445 ANNOTATE_RETPOLINE_SAFE \ 446 "call *%[thunk_target]\n", \ 447 " jmp 904f;\n" \ 448 " .align 16\n" \ 449 "901: call 903f;\n" \ 450 "902: pause;\n" \ 451 " lfence;\n" \ 452 " jmp 902b;\n" \ 453 " .align 16\n" \ 454 "903: lea 4(%%esp), %%esp;\n" \ 455 " pushl %[thunk_target];\n" \ 456 " ret;\n" \ 457 " .align 16\n" \ 458 "904: call 901b;\n", \ 459 X86_FEATURE_RETPOLINE, \ 460 "lfence;\n" \ 461 ANNOTATE_RETPOLINE_SAFE \ 462 "call *%[thunk_target]\n", \ 463 X86_FEATURE_RETPOLINE_LFENCE) 464 465# define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 466#endif 467#else /* No retpoline for C / inline asm */ 468# define CALL_NOSPEC "call *%[thunk_target]\n" 469# define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 470#endif 471 472/* The Spectre V2 mitigation variants */ 473enum spectre_v2_mitigation { 474 SPECTRE_V2_NONE, 475 SPECTRE_V2_RETPOLINE, 476 SPECTRE_V2_LFENCE, 477 SPECTRE_V2_EIBRS, 478 SPECTRE_V2_EIBRS_RETPOLINE, 479 SPECTRE_V2_EIBRS_LFENCE, 480 SPECTRE_V2_IBRS, 481}; 482 483/* The indirect branch speculation control variants */ 484enum spectre_v2_user_mitigation { 485 SPECTRE_V2_USER_NONE, 486 SPECTRE_V2_USER_STRICT, 487 SPECTRE_V2_USER_STRICT_PREFERRED, 488 SPECTRE_V2_USER_PRCTL, 489 SPECTRE_V2_USER_SECCOMP, 490}; 491 492/* The Speculative Store Bypass disable variants */ 493enum ssb_mitigation { 494 SPEC_STORE_BYPASS_NONE, 495 SPEC_STORE_BYPASS_DISABLE, 496 SPEC_STORE_BYPASS_PRCTL, 497 SPEC_STORE_BYPASS_SECCOMP, 498}; 499 500static __always_inline 501void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature) 502{ 503 asm volatile(ALTERNATIVE("", "wrmsr", %c[feature]) 504 : : "c" (msr), 505 "a" ((u32)val), 506 "d" ((u32)(val >> 32)), 507 [feature] "i" (feature) 508 : "memory"); 509} 510 511extern u64 x86_pred_cmd; 512 513static inline void indirect_branch_prediction_barrier(void) 514{ 515 alternative_msr_write(MSR_IA32_PRED_CMD, x86_pred_cmd, X86_FEATURE_USE_IBPB); 516} 517 518/* The Intel SPEC CTRL MSR base value cache */ 519extern u64 x86_spec_ctrl_base; 520DECLARE_PER_CPU(u64, x86_spec_ctrl_current); 521extern void update_spec_ctrl_cond(u64 val); 522extern u64 spec_ctrl_current(void); 523 524/* 525 * With retpoline, we must use IBRS to restrict branch prediction 526 * before calling into firmware. 527 * 528 * (Implemented as CPP macros due to header hell.) 529 */ 530#define firmware_restrict_branch_speculation_start() \ 531do { \ 532 preempt_disable(); \ 533 alternative_msr_write(MSR_IA32_SPEC_CTRL, \ 534 spec_ctrl_current() | SPEC_CTRL_IBRS, \ 535 X86_FEATURE_USE_IBRS_FW); \ 536 alternative_msr_write(MSR_IA32_PRED_CMD, PRED_CMD_IBPB, \ 537 X86_FEATURE_USE_IBPB_FW); \ 538} while (0) 539 540#define firmware_restrict_branch_speculation_end() \ 541do { \ 542 alternative_msr_write(MSR_IA32_SPEC_CTRL, \ 543 spec_ctrl_current(), \ 544 X86_FEATURE_USE_IBRS_FW); \ 545 preempt_enable(); \ 546} while (0) 547 548DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp); 549DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb); 550DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb); 551 552DECLARE_STATIC_KEY_FALSE(mds_idle_clear); 553 554DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush); 555 556DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear); 557 558extern u16 mds_verw_sel; 559 560#include <asm/segment.h> 561 562/** 563 * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability 564 * 565 * This uses the otherwise unused and obsolete VERW instruction in 566 * combination with microcode which triggers a CPU buffer flush when the 567 * instruction is executed. 568 */ 569static __always_inline void mds_clear_cpu_buffers(void) 570{ 571 static const u16 ds = __KERNEL_DS; 572 573 /* 574 * Has to be the memory-operand variant because only that 575 * guarantees the CPU buffer flush functionality according to 576 * documentation. The register-operand variant does not. 577 * Works with any segment selector, but a valid writable 578 * data segment is the fastest variant. 579 * 580 * "cc" clobber is required because VERW modifies ZF. 581 */ 582 asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc"); 583} 584 585/** 586 * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability 587 * 588 * Clear CPU buffers if the corresponding static key is enabled 589 */ 590static __always_inline void mds_idle_clear_cpu_buffers(void) 591{ 592 if (static_branch_likely(&mds_idle_clear)) 593 mds_clear_cpu_buffers(); 594} 595 596#endif /* __ASSEMBLY__ */ 597 598#endif /* _ASM_X86_NOSPEC_BRANCH_H_ */ 599