1 // SPDX-License-Identifier: GPL-2.0-only 2 #define pr_fmt(fmt) "SMP alternatives: " fmt 3 4 #include <linux/module.h> 5 #include <linux/sched.h> 6 #include <linux/perf_event.h> 7 #include <linux/mutex.h> 8 #include <linux/list.h> 9 #include <linux/stringify.h> 10 #include <linux/highmem.h> 11 #include <linux/mm.h> 12 #include <linux/vmalloc.h> 13 #include <linux/memory.h> 14 #include <linux/stop_machine.h> 15 #include <linux/slab.h> 16 #include <linux/kdebug.h> 17 #include <linux/kprobes.h> 18 #include <linux/mmu_context.h> 19 #include <linux/bsearch.h> 20 #include <linux/sync_core.h> 21 #include <asm/text-patching.h> 22 #include <asm/alternative.h> 23 #include <asm/sections.h> 24 #include <asm/mce.h> 25 #include <asm/nmi.h> 26 #include <asm/cacheflush.h> 27 #include <asm/tlbflush.h> 28 #include <asm/insn.h> 29 #include <asm/io.h> 30 #include <asm/fixmap.h> 31 #include <asm/paravirt.h> 32 #include <asm/asm-prototypes.h> 33 34 int __read_mostly alternatives_patched; 35 36 EXPORT_SYMBOL_GPL(alternatives_patched); 37 38 #define MAX_PATCH_LEN (255-1) 39 40 #define DA_ALL (~0) 41 #define DA_ALT 0x01 42 #define DA_RET 0x02 43 #define DA_RETPOLINE 0x04 44 #define DA_ENDBR 0x08 45 #define DA_SMP 0x10 46 47 static unsigned int __initdata_or_module debug_alternative; 48 49 static int __init debug_alt(char *str) 50 { 51 if (str && *str == '=') 52 str++; 53 54 if (!str || kstrtouint(str, 0, &debug_alternative)) 55 debug_alternative = DA_ALL; 56 57 return 1; 58 } 59 __setup("debug-alternative", debug_alt); 60 61 static int noreplace_smp; 62 63 static int __init setup_noreplace_smp(char *str) 64 { 65 noreplace_smp = 1; 66 return 1; 67 } 68 __setup("noreplace-smp", setup_noreplace_smp); 69 70 #define DPRINTK(type, fmt, args...) \ 71 do { \ 72 if (debug_alternative & DA_##type) \ 73 printk(KERN_DEBUG pr_fmt(fmt) "\n", ##args); \ 74 } while (0) 75 76 #define DUMP_BYTES(type, buf, len, fmt, args...) \ 77 do { \ 78 if (unlikely(debug_alternative & DA_##type)) { \ 79 int j; \ 80 \ 81 if (!(len)) \ 82 break; \ 83 \ 84 printk(KERN_DEBUG pr_fmt(fmt), ##args); \ 85 for (j = 0; j < (len) - 1; j++) \ 86 printk(KERN_CONT "%02hhx ", buf[j]); \ 87 printk(KERN_CONT "%02hhx\n", buf[j]); \ 88 } \ 89 } while (0) 90 91 static const unsigned char x86nops[] = 92 { 93 BYTES_NOP1, 94 BYTES_NOP2, 95 BYTES_NOP3, 96 BYTES_NOP4, 97 BYTES_NOP5, 98 BYTES_NOP6, 99 BYTES_NOP7, 100 BYTES_NOP8, 101 #ifdef CONFIG_64BIT 102 BYTES_NOP9, 103 BYTES_NOP10, 104 BYTES_NOP11, 105 #endif 106 }; 107 108 const unsigned char * const x86_nops[ASM_NOP_MAX+1] = 109 { 110 NULL, 111 x86nops, 112 x86nops + 1, 113 x86nops + 1 + 2, 114 x86nops + 1 + 2 + 3, 115 x86nops + 1 + 2 + 3 + 4, 116 x86nops + 1 + 2 + 3 + 4 + 5, 117 x86nops + 1 + 2 + 3 + 4 + 5 + 6, 118 x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7, 119 #ifdef CONFIG_64BIT 120 x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8, 121 x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9, 122 x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9 + 10, 123 #endif 124 }; 125 126 /* 127 * Fill the buffer with a single effective instruction of size @len. 128 * 129 * In order not to issue an ORC stack depth tracking CFI entry (Call Frame Info) 130 * for every single-byte NOP, try to generate the maximally available NOP of 131 * size <= ASM_NOP_MAX such that only a single CFI entry is generated (vs one for 132 * each single-byte NOPs). If @len to fill out is > ASM_NOP_MAX, pad with INT3 and 133 * *jump* over instead of executing long and daft NOPs. 134 */ 135 static void __init_or_module add_nop(u8 *instr, unsigned int len) 136 { 137 u8 *target = instr + len; 138 139 if (!len) 140 return; 141 142 if (len <= ASM_NOP_MAX) { 143 memcpy(instr, x86_nops[len], len); 144 return; 145 } 146 147 if (len < 128) { 148 __text_gen_insn(instr, JMP8_INSN_OPCODE, instr, target, JMP8_INSN_SIZE); 149 instr += JMP8_INSN_SIZE; 150 } else { 151 __text_gen_insn(instr, JMP32_INSN_OPCODE, instr, target, JMP32_INSN_SIZE); 152 instr += JMP32_INSN_SIZE; 153 } 154 155 for (;instr < target; instr++) 156 *instr = INT3_INSN_OPCODE; 157 } 158 159 extern s32 __retpoline_sites[], __retpoline_sites_end[]; 160 extern s32 __return_sites[], __return_sites_end[]; 161 extern s32 __cfi_sites[], __cfi_sites_end[]; 162 extern s32 __ibt_endbr_seal[], __ibt_endbr_seal_end[]; 163 extern struct alt_instr __alt_instructions[], __alt_instructions_end[]; 164 extern s32 __smp_locks[], __smp_locks_end[]; 165 void text_poke_early(void *addr, const void *opcode, size_t len); 166 167 /* 168 * Matches NOP and NOPL, not any of the other possible NOPs. 169 */ 170 static bool insn_is_nop(struct insn *insn) 171 { 172 /* Anything NOP, but no REP NOP */ 173 if (insn->opcode.bytes[0] == 0x90 && 174 (!insn->prefixes.nbytes || insn->prefixes.bytes[0] != 0xF3)) 175 return true; 176 177 /* NOPL */ 178 if (insn->opcode.bytes[0] == 0x0F && insn->opcode.bytes[1] == 0x1F) 179 return true; 180 181 /* TODO: more nops */ 182 183 return false; 184 } 185 186 /* 187 * Find the offset of the first non-NOP instruction starting at @offset 188 * but no further than @len. 189 */ 190 static int skip_nops(u8 *instr, int offset, int len) 191 { 192 struct insn insn; 193 194 for (; offset < len; offset += insn.length) { 195 if (insn_decode_kernel(&insn, &instr[offset])) 196 break; 197 198 if (!insn_is_nop(&insn)) 199 break; 200 } 201 202 return offset; 203 } 204 205 /* 206 * Optimize a sequence of NOPs, possibly preceded by an unconditional jump 207 * to the end of the NOP sequence into a single NOP. 208 */ 209 static bool __init_or_module 210 __optimize_nops(u8 *instr, size_t len, struct insn *insn, int *next, int *prev, int *target) 211 { 212 int i = *next - insn->length; 213 214 switch (insn->opcode.bytes[0]) { 215 case JMP8_INSN_OPCODE: 216 case JMP32_INSN_OPCODE: 217 *prev = i; 218 *target = *next + insn->immediate.value; 219 return false; 220 } 221 222 if (insn_is_nop(insn)) { 223 int nop = i; 224 225 *next = skip_nops(instr, *next, len); 226 if (*target && *next == *target) 227 nop = *prev; 228 229 add_nop(instr + nop, *next - nop); 230 DUMP_BYTES(ALT, instr, len, "%px: [%d:%d) optimized NOPs: ", instr, nop, *next); 231 return true; 232 } 233 234 *target = 0; 235 return false; 236 } 237 238 /* 239 * "noinline" to cause control flow change and thus invalidate I$ and 240 * cause refetch after modification. 241 */ 242 static void __init_or_module noinline optimize_nops(u8 *instr, size_t len) 243 { 244 int prev, target = 0; 245 246 for (int next, i = 0; i < len; i = next) { 247 struct insn insn; 248 249 if (insn_decode_kernel(&insn, &instr[i])) 250 return; 251 252 next = i + insn.length; 253 254 __optimize_nops(instr, len, &insn, &next, &prev, &target); 255 } 256 } 257 258 static void __init_or_module noinline optimize_nops_inplace(u8 *instr, size_t len) 259 { 260 unsigned long flags; 261 262 local_irq_save(flags); 263 optimize_nops(instr, len); 264 sync_core(); 265 local_irq_restore(flags); 266 } 267 268 /* 269 * In this context, "source" is where the instructions are placed in the 270 * section .altinstr_replacement, for example during kernel build by the 271 * toolchain. 272 * "Destination" is where the instructions are being patched in by this 273 * machinery. 274 * 275 * The source offset is: 276 * 277 * src_imm = target - src_next_ip (1) 278 * 279 * and the target offset is: 280 * 281 * dst_imm = target - dst_next_ip (2) 282 * 283 * so rework (1) as an expression for target like: 284 * 285 * target = src_imm + src_next_ip (1a) 286 * 287 * and substitute in (2) to get: 288 * 289 * dst_imm = (src_imm + src_next_ip) - dst_next_ip (3) 290 * 291 * Now, since the instruction stream is 'identical' at src and dst (it 292 * is being copied after all) it can be stated that: 293 * 294 * src_next_ip = src + ip_offset 295 * dst_next_ip = dst + ip_offset (4) 296 * 297 * Substitute (4) in (3) and observe ip_offset being cancelled out to 298 * obtain: 299 * 300 * dst_imm = src_imm + (src + ip_offset) - (dst + ip_offset) 301 * = src_imm + src - dst + ip_offset - ip_offset 302 * = src_imm + src - dst (5) 303 * 304 * IOW, only the relative displacement of the code block matters. 305 */ 306 307 #define apply_reloc_n(n_, p_, d_) \ 308 do { \ 309 s32 v = *(s##n_ *)(p_); \ 310 v += (d_); \ 311 BUG_ON((v >> 31) != (v >> (n_-1))); \ 312 *(s##n_ *)(p_) = (s##n_)v; \ 313 } while (0) 314 315 316 static __always_inline 317 void apply_reloc(int n, void *ptr, uintptr_t diff) 318 { 319 switch (n) { 320 case 1: apply_reloc_n(8, ptr, diff); break; 321 case 2: apply_reloc_n(16, ptr, diff); break; 322 case 4: apply_reloc_n(32, ptr, diff); break; 323 default: BUG(); 324 } 325 } 326 327 static __always_inline 328 bool need_reloc(unsigned long offset, u8 *src, size_t src_len) 329 { 330 u8 *target = src + offset; 331 /* 332 * If the target is inside the patched block, it's relative to the 333 * block itself and does not need relocation. 334 */ 335 return (target < src || target > src + src_len); 336 } 337 338 static void __init_or_module noinline 339 apply_relocation(u8 *buf, size_t len, u8 *dest, u8 *src, size_t src_len) 340 { 341 int prev, target = 0; 342 343 for (int next, i = 0; i < len; i = next) { 344 struct insn insn; 345 346 if (WARN_ON_ONCE(insn_decode_kernel(&insn, &buf[i]))) 347 return; 348 349 next = i + insn.length; 350 351 if (__optimize_nops(buf, len, &insn, &next, &prev, &target)) 352 continue; 353 354 switch (insn.opcode.bytes[0]) { 355 case 0x0f: 356 if (insn.opcode.bytes[1] < 0x80 || 357 insn.opcode.bytes[1] > 0x8f) 358 break; 359 360 fallthrough; /* Jcc.d32 */ 361 case 0x70 ... 0x7f: /* Jcc.d8 */ 362 case JMP8_INSN_OPCODE: 363 case JMP32_INSN_OPCODE: 364 case CALL_INSN_OPCODE: 365 if (need_reloc(next + insn.immediate.value, src, src_len)) { 366 apply_reloc(insn.immediate.nbytes, 367 buf + i + insn_offset_immediate(&insn), 368 src - dest); 369 } 370 371 /* 372 * Where possible, convert JMP.d32 into JMP.d8. 373 */ 374 if (insn.opcode.bytes[0] == JMP32_INSN_OPCODE) { 375 s32 imm = insn.immediate.value; 376 imm += src - dest; 377 imm += JMP32_INSN_SIZE - JMP8_INSN_SIZE; 378 if ((imm >> 31) == (imm >> 7)) { 379 buf[i+0] = JMP8_INSN_OPCODE; 380 buf[i+1] = (s8)imm; 381 382 memset(&buf[i+2], INT3_INSN_OPCODE, insn.length - 2); 383 } 384 } 385 break; 386 } 387 388 if (insn_rip_relative(&insn)) { 389 if (need_reloc(next + insn.displacement.value, src, src_len)) { 390 apply_reloc(insn.displacement.nbytes, 391 buf + i + insn_offset_displacement(&insn), 392 src - dest); 393 } 394 } 395 } 396 } 397 398 /* 399 * Replace instructions with better alternatives for this CPU type. This runs 400 * before SMP is initialized to avoid SMP problems with self modifying code. 401 * This implies that asymmetric systems where APs have less capabilities than 402 * the boot processor are not handled. Tough. Make sure you disable such 403 * features by hand. 404 * 405 * Marked "noinline" to cause control flow change and thus insn cache 406 * to refetch changed I$ lines. 407 */ 408 void __init_or_module noinline apply_alternatives(struct alt_instr *start, 409 struct alt_instr *end) 410 { 411 struct alt_instr *a; 412 u8 *instr, *replacement; 413 u8 insn_buff[MAX_PATCH_LEN]; 414 415 DPRINTK(ALT, "alt table %px, -> %px", start, end); 416 417 /* 418 * In the case CONFIG_X86_5LEVEL=y, KASAN_SHADOW_START is defined using 419 * cpu_feature_enabled(X86_FEATURE_LA57) and is therefore patched here. 420 * During the process, KASAN becomes confused seeing partial LA57 421 * conversion and triggers a false-positive out-of-bound report. 422 * 423 * Disable KASAN until the patching is complete. 424 */ 425 kasan_disable_current(); 426 427 /* 428 * The scan order should be from start to end. A later scanned 429 * alternative code can overwrite previously scanned alternative code. 430 * Some kernel functions (e.g. memcpy, memset, etc) use this order to 431 * patch code. 432 * 433 * So be careful if you want to change the scan order to any other 434 * order. 435 */ 436 for (a = start; a < end; a++) { 437 int insn_buff_sz = 0; 438 439 instr = (u8 *)&a->instr_offset + a->instr_offset; 440 replacement = (u8 *)&a->repl_offset + a->repl_offset; 441 BUG_ON(a->instrlen > sizeof(insn_buff)); 442 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32); 443 444 /* 445 * Patch if either: 446 * - feature is present 447 * - feature not present but ALT_FLAG_NOT is set to mean, 448 * patch if feature is *NOT* present. 449 */ 450 if (!boot_cpu_has(a->cpuid) == !(a->flags & ALT_FLAG_NOT)) { 451 optimize_nops_inplace(instr, a->instrlen); 452 continue; 453 } 454 455 DPRINTK(ALT, "feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)", 456 (a->flags & ALT_FLAG_NOT) ? "!" : "", 457 a->cpuid >> 5, 458 a->cpuid & 0x1f, 459 instr, instr, a->instrlen, 460 replacement, a->replacementlen); 461 462 memcpy(insn_buff, replacement, a->replacementlen); 463 insn_buff_sz = a->replacementlen; 464 465 for (; insn_buff_sz < a->instrlen; insn_buff_sz++) 466 insn_buff[insn_buff_sz] = 0x90; 467 468 apply_relocation(insn_buff, a->instrlen, instr, replacement, a->replacementlen); 469 470 DUMP_BYTES(ALT, instr, a->instrlen, "%px: old_insn: ", instr); 471 DUMP_BYTES(ALT, replacement, a->replacementlen, "%px: rpl_insn: ", replacement); 472 DUMP_BYTES(ALT, insn_buff, insn_buff_sz, "%px: final_insn: ", instr); 473 474 text_poke_early(instr, insn_buff, insn_buff_sz); 475 } 476 477 kasan_enable_current(); 478 } 479 480 static inline bool is_jcc32(struct insn *insn) 481 { 482 /* Jcc.d32 second opcode byte is in the range: 0x80-0x8f */ 483 return insn->opcode.bytes[0] == 0x0f && (insn->opcode.bytes[1] & 0xf0) == 0x80; 484 } 485 486 #if defined(CONFIG_RETPOLINE) && defined(CONFIG_OBJTOOL) 487 488 /* 489 * CALL/JMP *%\reg 490 */ 491 static int emit_indirect(int op, int reg, u8 *bytes) 492 { 493 int i = 0; 494 u8 modrm; 495 496 switch (op) { 497 case CALL_INSN_OPCODE: 498 modrm = 0x10; /* Reg = 2; CALL r/m */ 499 break; 500 501 case JMP32_INSN_OPCODE: 502 modrm = 0x20; /* Reg = 4; JMP r/m */ 503 break; 504 505 default: 506 WARN_ON_ONCE(1); 507 return -1; 508 } 509 510 if (reg >= 8) { 511 bytes[i++] = 0x41; /* REX.B prefix */ 512 reg -= 8; 513 } 514 515 modrm |= 0xc0; /* Mod = 3 */ 516 modrm += reg; 517 518 bytes[i++] = 0xff; /* opcode */ 519 bytes[i++] = modrm; 520 521 return i; 522 } 523 524 static int emit_call_track_retpoline(void *addr, struct insn *insn, int reg, u8 *bytes) 525 { 526 u8 op = insn->opcode.bytes[0]; 527 int i = 0; 528 529 /* 530 * Clang does 'weird' Jcc __x86_indirect_thunk_r11 conditional 531 * tail-calls. Deal with them. 532 */ 533 if (is_jcc32(insn)) { 534 bytes[i++] = op; 535 op = insn->opcode.bytes[1]; 536 goto clang_jcc; 537 } 538 539 if (insn->length == 6) 540 bytes[i++] = 0x2e; /* CS-prefix */ 541 542 switch (op) { 543 case CALL_INSN_OPCODE: 544 __text_gen_insn(bytes+i, op, addr+i, 545 __x86_indirect_call_thunk_array[reg], 546 CALL_INSN_SIZE); 547 i += CALL_INSN_SIZE; 548 break; 549 550 case JMP32_INSN_OPCODE: 551 clang_jcc: 552 __text_gen_insn(bytes+i, op, addr+i, 553 __x86_indirect_jump_thunk_array[reg], 554 JMP32_INSN_SIZE); 555 i += JMP32_INSN_SIZE; 556 break; 557 558 default: 559 WARN(1, "%pS %px %*ph\n", addr, addr, 6, addr); 560 return -1; 561 } 562 563 WARN_ON_ONCE(i != insn->length); 564 565 return i; 566 } 567 568 /* 569 * Rewrite the compiler generated retpoline thunk calls. 570 * 571 * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate 572 * indirect instructions, avoiding the extra indirection. 573 * 574 * For example, convert: 575 * 576 * CALL __x86_indirect_thunk_\reg 577 * 578 * into: 579 * 580 * CALL *%\reg 581 * 582 * It also tries to inline spectre_v2=retpoline,lfence when size permits. 583 */ 584 static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes) 585 { 586 retpoline_thunk_t *target; 587 int reg, ret, i = 0; 588 u8 op, cc; 589 590 target = addr + insn->length + insn->immediate.value; 591 reg = target - __x86_indirect_thunk_array; 592 593 if (WARN_ON_ONCE(reg & ~0xf)) 594 return -1; 595 596 /* If anyone ever does: CALL/JMP *%rsp, we're in deep trouble. */ 597 BUG_ON(reg == 4); 598 599 if (cpu_feature_enabled(X86_FEATURE_RETPOLINE) && 600 !cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) { 601 if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH)) 602 return emit_call_track_retpoline(addr, insn, reg, bytes); 603 604 return -1; 605 } 606 607 op = insn->opcode.bytes[0]; 608 609 /* 610 * Convert: 611 * 612 * Jcc.d32 __x86_indirect_thunk_\reg 613 * 614 * into: 615 * 616 * Jncc.d8 1f 617 * [ LFENCE ] 618 * JMP *%\reg 619 * [ NOP ] 620 * 1: 621 */ 622 if (is_jcc32(insn)) { 623 cc = insn->opcode.bytes[1] & 0xf; 624 cc ^= 1; /* invert condition */ 625 626 bytes[i++] = 0x70 + cc; /* Jcc.d8 */ 627 bytes[i++] = insn->length - 2; /* sizeof(Jcc.d8) == 2 */ 628 629 /* Continue as if: JMP.d32 __x86_indirect_thunk_\reg */ 630 op = JMP32_INSN_OPCODE; 631 } 632 633 /* 634 * For RETPOLINE_LFENCE: prepend the indirect CALL/JMP with an LFENCE. 635 */ 636 if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) { 637 bytes[i++] = 0x0f; 638 bytes[i++] = 0xae; 639 bytes[i++] = 0xe8; /* LFENCE */ 640 } 641 642 ret = emit_indirect(op, reg, bytes + i); 643 if (ret < 0) 644 return ret; 645 i += ret; 646 647 /* 648 * The compiler is supposed to EMIT an INT3 after every unconditional 649 * JMP instruction due to AMD BTC. However, if the compiler is too old 650 * or SLS isn't enabled, we still need an INT3 after indirect JMPs 651 * even on Intel. 652 */ 653 if (op == JMP32_INSN_OPCODE && i < insn->length) 654 bytes[i++] = INT3_INSN_OPCODE; 655 656 for (; i < insn->length;) 657 bytes[i++] = BYTES_NOP1; 658 659 return i; 660 } 661 662 /* 663 * Generated by 'objtool --retpoline'. 664 */ 665 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) 666 { 667 s32 *s; 668 669 for (s = start; s < end; s++) { 670 void *addr = (void *)s + *s; 671 struct insn insn; 672 int len, ret; 673 u8 bytes[16]; 674 u8 op1, op2; 675 676 ret = insn_decode_kernel(&insn, addr); 677 if (WARN_ON_ONCE(ret < 0)) 678 continue; 679 680 op1 = insn.opcode.bytes[0]; 681 op2 = insn.opcode.bytes[1]; 682 683 switch (op1) { 684 case CALL_INSN_OPCODE: 685 case JMP32_INSN_OPCODE: 686 break; 687 688 case 0x0f: /* escape */ 689 if (op2 >= 0x80 && op2 <= 0x8f) 690 break; 691 fallthrough; 692 default: 693 WARN_ON_ONCE(1); 694 continue; 695 } 696 697 DPRINTK(RETPOLINE, "retpoline at: %pS (%px) len: %d to: %pS", 698 addr, addr, insn.length, 699 addr + insn.length + insn.immediate.value); 700 701 len = patch_retpoline(addr, &insn, bytes); 702 if (len == insn.length) { 703 optimize_nops(bytes, len); 704 DUMP_BYTES(RETPOLINE, ((u8*)addr), len, "%px: orig: ", addr); 705 DUMP_BYTES(RETPOLINE, ((u8*)bytes), len, "%px: repl: ", addr); 706 text_poke_early(addr, bytes, len); 707 } 708 } 709 } 710 711 #ifdef CONFIG_RETHUNK 712 713 /* 714 * Rewrite the compiler generated return thunk tail-calls. 715 * 716 * For example, convert: 717 * 718 * JMP __x86_return_thunk 719 * 720 * into: 721 * 722 * RET 723 */ 724 static int patch_return(void *addr, struct insn *insn, u8 *bytes) 725 { 726 int i = 0; 727 728 /* Patch the custom return thunks... */ 729 if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) { 730 i = JMP32_INSN_SIZE; 731 __text_gen_insn(bytes, JMP32_INSN_OPCODE, addr, x86_return_thunk, i); 732 } else { 733 /* ... or patch them out if not needed. */ 734 bytes[i++] = RET_INSN_OPCODE; 735 } 736 737 for (; i < insn->length;) 738 bytes[i++] = INT3_INSN_OPCODE; 739 return i; 740 } 741 742 void __init_or_module noinline apply_returns(s32 *start, s32 *end) 743 { 744 s32 *s; 745 746 if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) 747 static_call_force_reinit(); 748 749 for (s = start; s < end; s++) { 750 void *dest = NULL, *addr = (void *)s + *s; 751 struct insn insn; 752 int len, ret; 753 u8 bytes[16]; 754 u8 op; 755 756 ret = insn_decode_kernel(&insn, addr); 757 if (WARN_ON_ONCE(ret < 0)) 758 continue; 759 760 op = insn.opcode.bytes[0]; 761 if (op == JMP32_INSN_OPCODE) 762 dest = addr + insn.length + insn.immediate.value; 763 764 if (__static_call_fixup(addr, op, dest) || 765 WARN_ONCE(dest != &__x86_return_thunk, 766 "missing return thunk: %pS-%pS: %*ph", 767 addr, dest, 5, addr)) 768 continue; 769 770 DPRINTK(RET, "return thunk at: %pS (%px) len: %d to: %pS", 771 addr, addr, insn.length, 772 addr + insn.length + insn.immediate.value); 773 774 len = patch_return(addr, &insn, bytes); 775 if (len == insn.length) { 776 DUMP_BYTES(RET, ((u8*)addr), len, "%px: orig: ", addr); 777 DUMP_BYTES(RET, ((u8*)bytes), len, "%px: repl: ", addr); 778 text_poke_early(addr, bytes, len); 779 } 780 } 781 } 782 #else 783 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { } 784 #endif /* CONFIG_RETHUNK */ 785 786 #else /* !CONFIG_RETPOLINE || !CONFIG_OBJTOOL */ 787 788 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) { } 789 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { } 790 791 #endif /* CONFIG_RETPOLINE && CONFIG_OBJTOOL */ 792 793 #ifdef CONFIG_X86_KERNEL_IBT 794 795 static void poison_cfi(void *addr); 796 797 static void __init_or_module poison_endbr(void *addr, bool warn) 798 { 799 u32 endbr, poison = gen_endbr_poison(); 800 801 if (WARN_ON_ONCE(get_kernel_nofault(endbr, addr))) 802 return; 803 804 if (!is_endbr(endbr)) { 805 WARN_ON_ONCE(warn); 806 return; 807 } 808 809 DPRINTK(ENDBR, "ENDBR at: %pS (%px)", addr, addr); 810 811 /* 812 * When we have IBT, the lack of ENDBR will trigger #CP 813 */ 814 DUMP_BYTES(ENDBR, ((u8*)addr), 4, "%px: orig: ", addr); 815 DUMP_BYTES(ENDBR, ((u8*)&poison), 4, "%px: repl: ", addr); 816 text_poke_early(addr, &poison, 4); 817 } 818 819 /* 820 * Generated by: objtool --ibt 821 * 822 * Seal the functions for indirect calls by clobbering the ENDBR instructions 823 * and the kCFI hash value. 824 */ 825 void __init_or_module noinline apply_seal_endbr(s32 *start, s32 *end) 826 { 827 s32 *s; 828 829 for (s = start; s < end; s++) { 830 void *addr = (void *)s + *s; 831 832 poison_endbr(addr, true); 833 if (IS_ENABLED(CONFIG_FINEIBT)) 834 poison_cfi(addr - 16); 835 } 836 } 837 838 #else 839 840 void __init_or_module apply_seal_endbr(s32 *start, s32 *end) { } 841 842 #endif /* CONFIG_X86_KERNEL_IBT */ 843 844 #ifdef CONFIG_FINEIBT 845 846 enum cfi_mode { 847 CFI_DEFAULT, 848 CFI_OFF, 849 CFI_KCFI, 850 CFI_FINEIBT, 851 }; 852 853 static enum cfi_mode cfi_mode __ro_after_init = CFI_DEFAULT; 854 static bool cfi_rand __ro_after_init = true; 855 static u32 cfi_seed __ro_after_init; 856 857 /* 858 * Re-hash the CFI hash with a boot-time seed while making sure the result is 859 * not a valid ENDBR instruction. 860 */ 861 static u32 cfi_rehash(u32 hash) 862 { 863 hash ^= cfi_seed; 864 while (unlikely(is_endbr(hash) || is_endbr(-hash))) { 865 bool lsb = hash & 1; 866 hash >>= 1; 867 if (lsb) 868 hash ^= 0x80200003; 869 } 870 return hash; 871 } 872 873 static __init int cfi_parse_cmdline(char *str) 874 { 875 if (!str) 876 return -EINVAL; 877 878 while (str) { 879 char *next = strchr(str, ','); 880 if (next) { 881 *next = 0; 882 next++; 883 } 884 885 if (!strcmp(str, "auto")) { 886 cfi_mode = CFI_DEFAULT; 887 } else if (!strcmp(str, "off")) { 888 cfi_mode = CFI_OFF; 889 cfi_rand = false; 890 } else if (!strcmp(str, "kcfi")) { 891 cfi_mode = CFI_KCFI; 892 } else if (!strcmp(str, "fineibt")) { 893 cfi_mode = CFI_FINEIBT; 894 } else if (!strcmp(str, "norand")) { 895 cfi_rand = false; 896 } else { 897 pr_err("Ignoring unknown cfi option (%s).", str); 898 } 899 900 str = next; 901 } 902 903 return 0; 904 } 905 early_param("cfi", cfi_parse_cmdline); 906 907 /* 908 * kCFI FineIBT 909 * 910 * __cfi_\func: __cfi_\func: 911 * movl $0x12345678,%eax // 5 endbr64 // 4 912 * nop subl $0x12345678,%r10d // 7 913 * nop jz 1f // 2 914 * nop ud2 // 2 915 * nop 1: nop // 1 916 * nop 917 * nop 918 * nop 919 * nop 920 * nop 921 * nop 922 * nop 923 * 924 * 925 * caller: caller: 926 * movl $(-0x12345678),%r10d // 6 movl $0x12345678,%r10d // 6 927 * addl $-15(%r11),%r10d // 4 sub $16,%r11 // 4 928 * je 1f // 2 nop4 // 4 929 * ud2 // 2 930 * 1: call __x86_indirect_thunk_r11 // 5 call *%r11; nop2; // 5 931 * 932 */ 933 934 asm( ".pushsection .rodata \n" 935 "fineibt_preamble_start: \n" 936 " endbr64 \n" 937 " subl $0x12345678, %r10d \n" 938 " je fineibt_preamble_end \n" 939 " ud2 \n" 940 " nop \n" 941 "fineibt_preamble_end: \n" 942 ".popsection\n" 943 ); 944 945 extern u8 fineibt_preamble_start[]; 946 extern u8 fineibt_preamble_end[]; 947 948 #define fineibt_preamble_size (fineibt_preamble_end - fineibt_preamble_start) 949 #define fineibt_preamble_hash 7 950 951 asm( ".pushsection .rodata \n" 952 "fineibt_caller_start: \n" 953 " movl $0x12345678, %r10d \n" 954 " sub $16, %r11 \n" 955 ASM_NOP4 956 "fineibt_caller_end: \n" 957 ".popsection \n" 958 ); 959 960 extern u8 fineibt_caller_start[]; 961 extern u8 fineibt_caller_end[]; 962 963 #define fineibt_caller_size (fineibt_caller_end - fineibt_caller_start) 964 #define fineibt_caller_hash 2 965 966 #define fineibt_caller_jmp (fineibt_caller_size - 2) 967 968 static u32 decode_preamble_hash(void *addr) 969 { 970 u8 *p = addr; 971 972 /* b8 78 56 34 12 mov $0x12345678,%eax */ 973 if (p[0] == 0xb8) 974 return *(u32 *)(addr + 1); 975 976 return 0; /* invalid hash value */ 977 } 978 979 static u32 decode_caller_hash(void *addr) 980 { 981 u8 *p = addr; 982 983 /* 41 ba 78 56 34 12 mov $0x12345678,%r10d */ 984 if (p[0] == 0x41 && p[1] == 0xba) 985 return -*(u32 *)(addr + 2); 986 987 /* e8 0c 78 56 34 12 jmp.d8 +12 */ 988 if (p[0] == JMP8_INSN_OPCODE && p[1] == fineibt_caller_jmp) 989 return -*(u32 *)(addr + 2); 990 991 return 0; /* invalid hash value */ 992 } 993 994 /* .retpoline_sites */ 995 static int cfi_disable_callers(s32 *start, s32 *end) 996 { 997 /* 998 * Disable kCFI by patching in a JMP.d8, this leaves the hash immediate 999 * in tact for later usage. Also see decode_caller_hash() and 1000 * cfi_rewrite_callers(). 1001 */ 1002 const u8 jmp[] = { JMP8_INSN_OPCODE, fineibt_caller_jmp }; 1003 s32 *s; 1004 1005 for (s = start; s < end; s++) { 1006 void *addr = (void *)s + *s; 1007 u32 hash; 1008 1009 addr -= fineibt_caller_size; 1010 hash = decode_caller_hash(addr); 1011 if (!hash) /* nocfi callers */ 1012 continue; 1013 1014 text_poke_early(addr, jmp, 2); 1015 } 1016 1017 return 0; 1018 } 1019 1020 static int cfi_enable_callers(s32 *start, s32 *end) 1021 { 1022 /* 1023 * Re-enable kCFI, undo what cfi_disable_callers() did. 1024 */ 1025 const u8 mov[] = { 0x41, 0xba }; 1026 s32 *s; 1027 1028 for (s = start; s < end; s++) { 1029 void *addr = (void *)s + *s; 1030 u32 hash; 1031 1032 addr -= fineibt_caller_size; 1033 hash = decode_caller_hash(addr); 1034 if (!hash) /* nocfi callers */ 1035 continue; 1036 1037 text_poke_early(addr, mov, 2); 1038 } 1039 1040 return 0; 1041 } 1042 1043 /* .cfi_sites */ 1044 static int cfi_rand_preamble(s32 *start, s32 *end) 1045 { 1046 s32 *s; 1047 1048 for (s = start; s < end; s++) { 1049 void *addr = (void *)s + *s; 1050 u32 hash; 1051 1052 hash = decode_preamble_hash(addr); 1053 if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n", 1054 addr, addr, 5, addr)) 1055 return -EINVAL; 1056 1057 hash = cfi_rehash(hash); 1058 text_poke_early(addr + 1, &hash, 4); 1059 } 1060 1061 return 0; 1062 } 1063 1064 static int cfi_rewrite_preamble(s32 *start, s32 *end) 1065 { 1066 s32 *s; 1067 1068 for (s = start; s < end; s++) { 1069 void *addr = (void *)s + *s; 1070 u32 hash; 1071 1072 hash = decode_preamble_hash(addr); 1073 if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n", 1074 addr, addr, 5, addr)) 1075 return -EINVAL; 1076 1077 text_poke_early(addr, fineibt_preamble_start, fineibt_preamble_size); 1078 WARN_ON(*(u32 *)(addr + fineibt_preamble_hash) != 0x12345678); 1079 text_poke_early(addr + fineibt_preamble_hash, &hash, 4); 1080 } 1081 1082 return 0; 1083 } 1084 1085 static void cfi_rewrite_endbr(s32 *start, s32 *end) 1086 { 1087 s32 *s; 1088 1089 for (s = start; s < end; s++) { 1090 void *addr = (void *)s + *s; 1091 1092 poison_endbr(addr+16, false); 1093 } 1094 } 1095 1096 /* .retpoline_sites */ 1097 static int cfi_rand_callers(s32 *start, s32 *end) 1098 { 1099 s32 *s; 1100 1101 for (s = start; s < end; s++) { 1102 void *addr = (void *)s + *s; 1103 u32 hash; 1104 1105 addr -= fineibt_caller_size; 1106 hash = decode_caller_hash(addr); 1107 if (hash) { 1108 hash = -cfi_rehash(hash); 1109 text_poke_early(addr + 2, &hash, 4); 1110 } 1111 } 1112 1113 return 0; 1114 } 1115 1116 static int cfi_rewrite_callers(s32 *start, s32 *end) 1117 { 1118 s32 *s; 1119 1120 for (s = start; s < end; s++) { 1121 void *addr = (void *)s + *s; 1122 u32 hash; 1123 1124 addr -= fineibt_caller_size; 1125 hash = decode_caller_hash(addr); 1126 if (hash) { 1127 text_poke_early(addr, fineibt_caller_start, fineibt_caller_size); 1128 WARN_ON(*(u32 *)(addr + fineibt_caller_hash) != 0x12345678); 1129 text_poke_early(addr + fineibt_caller_hash, &hash, 4); 1130 } 1131 /* rely on apply_retpolines() */ 1132 } 1133 1134 return 0; 1135 } 1136 1137 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline, 1138 s32 *start_cfi, s32 *end_cfi, bool builtin) 1139 { 1140 int ret; 1141 1142 if (WARN_ONCE(fineibt_preamble_size != 16, 1143 "FineIBT preamble wrong size: %ld", fineibt_preamble_size)) 1144 return; 1145 1146 if (cfi_mode == CFI_DEFAULT) { 1147 cfi_mode = CFI_KCFI; 1148 if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT)) 1149 cfi_mode = CFI_FINEIBT; 1150 } 1151 1152 /* 1153 * Rewrite the callers to not use the __cfi_ stubs, such that we might 1154 * rewrite them. This disables all CFI. If this succeeds but any of the 1155 * later stages fails, we're without CFI. 1156 */ 1157 ret = cfi_disable_callers(start_retpoline, end_retpoline); 1158 if (ret) 1159 goto err; 1160 1161 if (cfi_rand) { 1162 if (builtin) 1163 cfi_seed = get_random_u32(); 1164 1165 ret = cfi_rand_preamble(start_cfi, end_cfi); 1166 if (ret) 1167 goto err; 1168 1169 ret = cfi_rand_callers(start_retpoline, end_retpoline); 1170 if (ret) 1171 goto err; 1172 } 1173 1174 switch (cfi_mode) { 1175 case CFI_OFF: 1176 if (builtin) 1177 pr_info("Disabling CFI\n"); 1178 return; 1179 1180 case CFI_KCFI: 1181 ret = cfi_enable_callers(start_retpoline, end_retpoline); 1182 if (ret) 1183 goto err; 1184 1185 if (builtin) 1186 pr_info("Using kCFI\n"); 1187 return; 1188 1189 case CFI_FINEIBT: 1190 /* place the FineIBT preamble at func()-16 */ 1191 ret = cfi_rewrite_preamble(start_cfi, end_cfi); 1192 if (ret) 1193 goto err; 1194 1195 /* rewrite the callers to target func()-16 */ 1196 ret = cfi_rewrite_callers(start_retpoline, end_retpoline); 1197 if (ret) 1198 goto err; 1199 1200 /* now that nobody targets func()+0, remove ENDBR there */ 1201 cfi_rewrite_endbr(start_cfi, end_cfi); 1202 1203 if (builtin) 1204 pr_info("Using FineIBT CFI\n"); 1205 return; 1206 1207 default: 1208 break; 1209 } 1210 1211 err: 1212 pr_err("Something went horribly wrong trying to rewrite the CFI implementation.\n"); 1213 } 1214 1215 static inline void poison_hash(void *addr) 1216 { 1217 *(u32 *)addr = 0; 1218 } 1219 1220 static void poison_cfi(void *addr) 1221 { 1222 switch (cfi_mode) { 1223 case CFI_FINEIBT: 1224 /* 1225 * __cfi_\func: 1226 * osp nopl (%rax) 1227 * subl $0, %r10d 1228 * jz 1f 1229 * ud2 1230 * 1: nop 1231 */ 1232 poison_endbr(addr, false); 1233 poison_hash(addr + fineibt_preamble_hash); 1234 break; 1235 1236 case CFI_KCFI: 1237 /* 1238 * __cfi_\func: 1239 * movl $0, %eax 1240 * .skip 11, 0x90 1241 */ 1242 poison_hash(addr + 1); 1243 break; 1244 1245 default: 1246 break; 1247 } 1248 } 1249 1250 #else 1251 1252 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline, 1253 s32 *start_cfi, s32 *end_cfi, bool builtin) 1254 { 1255 } 1256 1257 #ifdef CONFIG_X86_KERNEL_IBT 1258 static void poison_cfi(void *addr) { } 1259 #endif 1260 1261 #endif 1262 1263 void apply_fineibt(s32 *start_retpoline, s32 *end_retpoline, 1264 s32 *start_cfi, s32 *end_cfi) 1265 { 1266 return __apply_fineibt(start_retpoline, end_retpoline, 1267 start_cfi, end_cfi, 1268 /* .builtin = */ false); 1269 } 1270 1271 #ifdef CONFIG_SMP 1272 static void alternatives_smp_lock(const s32 *start, const s32 *end, 1273 u8 *text, u8 *text_end) 1274 { 1275 const s32 *poff; 1276 1277 for (poff = start; poff < end; poff++) { 1278 u8 *ptr = (u8 *)poff + *poff; 1279 1280 if (!*poff || ptr < text || ptr >= text_end) 1281 continue; 1282 /* turn DS segment override prefix into lock prefix */ 1283 if (*ptr == 0x3e) 1284 text_poke(ptr, ((unsigned char []){0xf0}), 1); 1285 } 1286 } 1287 1288 static void alternatives_smp_unlock(const s32 *start, const s32 *end, 1289 u8 *text, u8 *text_end) 1290 { 1291 const s32 *poff; 1292 1293 for (poff = start; poff < end; poff++) { 1294 u8 *ptr = (u8 *)poff + *poff; 1295 1296 if (!*poff || ptr < text || ptr >= text_end) 1297 continue; 1298 /* turn lock prefix into DS segment override prefix */ 1299 if (*ptr == 0xf0) 1300 text_poke(ptr, ((unsigned char []){0x3E}), 1); 1301 } 1302 } 1303 1304 struct smp_alt_module { 1305 /* what is this ??? */ 1306 struct module *mod; 1307 char *name; 1308 1309 /* ptrs to lock prefixes */ 1310 const s32 *locks; 1311 const s32 *locks_end; 1312 1313 /* .text segment, needed to avoid patching init code ;) */ 1314 u8 *text; 1315 u8 *text_end; 1316 1317 struct list_head next; 1318 }; 1319 static LIST_HEAD(smp_alt_modules); 1320 static bool uniproc_patched = false; /* protected by text_mutex */ 1321 1322 void __init_or_module alternatives_smp_module_add(struct module *mod, 1323 char *name, 1324 void *locks, void *locks_end, 1325 void *text, void *text_end) 1326 { 1327 struct smp_alt_module *smp; 1328 1329 mutex_lock(&text_mutex); 1330 if (!uniproc_patched) 1331 goto unlock; 1332 1333 if (num_possible_cpus() == 1) 1334 /* Don't bother remembering, we'll never have to undo it. */ 1335 goto smp_unlock; 1336 1337 smp = kzalloc(sizeof(*smp), GFP_KERNEL); 1338 if (NULL == smp) 1339 /* we'll run the (safe but slow) SMP code then ... */ 1340 goto unlock; 1341 1342 smp->mod = mod; 1343 smp->name = name; 1344 smp->locks = locks; 1345 smp->locks_end = locks_end; 1346 smp->text = text; 1347 smp->text_end = text_end; 1348 DPRINTK(SMP, "locks %p -> %p, text %p -> %p, name %s\n", 1349 smp->locks, smp->locks_end, 1350 smp->text, smp->text_end, smp->name); 1351 1352 list_add_tail(&smp->next, &smp_alt_modules); 1353 smp_unlock: 1354 alternatives_smp_unlock(locks, locks_end, text, text_end); 1355 unlock: 1356 mutex_unlock(&text_mutex); 1357 } 1358 1359 void __init_or_module alternatives_smp_module_del(struct module *mod) 1360 { 1361 struct smp_alt_module *item; 1362 1363 mutex_lock(&text_mutex); 1364 list_for_each_entry(item, &smp_alt_modules, next) { 1365 if (mod != item->mod) 1366 continue; 1367 list_del(&item->next); 1368 kfree(item); 1369 break; 1370 } 1371 mutex_unlock(&text_mutex); 1372 } 1373 1374 void alternatives_enable_smp(void) 1375 { 1376 struct smp_alt_module *mod; 1377 1378 /* Why bother if there are no other CPUs? */ 1379 BUG_ON(num_possible_cpus() == 1); 1380 1381 mutex_lock(&text_mutex); 1382 1383 if (uniproc_patched) { 1384 pr_info("switching to SMP code\n"); 1385 BUG_ON(num_online_cpus() != 1); 1386 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP); 1387 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP); 1388 list_for_each_entry(mod, &smp_alt_modules, next) 1389 alternatives_smp_lock(mod->locks, mod->locks_end, 1390 mod->text, mod->text_end); 1391 uniproc_patched = false; 1392 } 1393 mutex_unlock(&text_mutex); 1394 } 1395 1396 /* 1397 * Return 1 if the address range is reserved for SMP-alternatives. 1398 * Must hold text_mutex. 1399 */ 1400 int alternatives_text_reserved(void *start, void *end) 1401 { 1402 struct smp_alt_module *mod; 1403 const s32 *poff; 1404 u8 *text_start = start; 1405 u8 *text_end = end; 1406 1407 lockdep_assert_held(&text_mutex); 1408 1409 list_for_each_entry(mod, &smp_alt_modules, next) { 1410 if (mod->text > text_end || mod->text_end < text_start) 1411 continue; 1412 for (poff = mod->locks; poff < mod->locks_end; poff++) { 1413 const u8 *ptr = (const u8 *)poff + *poff; 1414 1415 if (text_start <= ptr && text_end > ptr) 1416 return 1; 1417 } 1418 } 1419 1420 return 0; 1421 } 1422 #endif /* CONFIG_SMP */ 1423 1424 #ifdef CONFIG_PARAVIRT 1425 1426 /* Use this to add nops to a buffer, then text_poke the whole buffer. */ 1427 static void __init_or_module add_nops(void *insns, unsigned int len) 1428 { 1429 while (len > 0) { 1430 unsigned int noplen = len; 1431 if (noplen > ASM_NOP_MAX) 1432 noplen = ASM_NOP_MAX; 1433 memcpy(insns, x86_nops[noplen], noplen); 1434 insns += noplen; 1435 len -= noplen; 1436 } 1437 } 1438 1439 void __init_or_module apply_paravirt(struct paravirt_patch_site *start, 1440 struct paravirt_patch_site *end) 1441 { 1442 struct paravirt_patch_site *p; 1443 char insn_buff[MAX_PATCH_LEN]; 1444 1445 for (p = start; p < end; p++) { 1446 unsigned int used; 1447 1448 BUG_ON(p->len > MAX_PATCH_LEN); 1449 /* prep the buffer with the original instructions */ 1450 memcpy(insn_buff, p->instr, p->len); 1451 used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len); 1452 1453 BUG_ON(used > p->len); 1454 1455 /* Pad the rest with nops */ 1456 add_nops(insn_buff + used, p->len - used); 1457 text_poke_early(p->instr, insn_buff, p->len); 1458 } 1459 } 1460 extern struct paravirt_patch_site __start_parainstructions[], 1461 __stop_parainstructions[]; 1462 #endif /* CONFIG_PARAVIRT */ 1463 1464 /* 1465 * Self-test for the INT3 based CALL emulation code. 1466 * 1467 * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up 1468 * properly and that there is a stack gap between the INT3 frame and the 1469 * previous context. Without this gap doing a virtual PUSH on the interrupted 1470 * stack would corrupt the INT3 IRET frame. 1471 * 1472 * See entry_{32,64}.S for more details. 1473 */ 1474 1475 /* 1476 * We define the int3_magic() function in assembly to control the calling 1477 * convention such that we can 'call' it from assembly. 1478 */ 1479 1480 extern void int3_magic(unsigned int *ptr); /* defined in asm */ 1481 1482 asm ( 1483 " .pushsection .init.text, \"ax\", @progbits\n" 1484 " .type int3_magic, @function\n" 1485 "int3_magic:\n" 1486 ANNOTATE_NOENDBR 1487 " movl $1, (%" _ASM_ARG1 ")\n" 1488 ASM_RET 1489 " .size int3_magic, .-int3_magic\n" 1490 " .popsection\n" 1491 ); 1492 1493 extern void int3_selftest_ip(void); /* defined in asm below */ 1494 1495 static int __init 1496 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data) 1497 { 1498 unsigned long selftest = (unsigned long)&int3_selftest_ip; 1499 struct die_args *args = data; 1500 struct pt_regs *regs = args->regs; 1501 1502 OPTIMIZER_HIDE_VAR(selftest); 1503 1504 if (!regs || user_mode(regs)) 1505 return NOTIFY_DONE; 1506 1507 if (val != DIE_INT3) 1508 return NOTIFY_DONE; 1509 1510 if (regs->ip - INT3_INSN_SIZE != selftest) 1511 return NOTIFY_DONE; 1512 1513 int3_emulate_call(regs, (unsigned long)&int3_magic); 1514 return NOTIFY_STOP; 1515 } 1516 1517 /* Must be noinline to ensure uniqueness of int3_selftest_ip. */ 1518 static noinline void __init int3_selftest(void) 1519 { 1520 static __initdata struct notifier_block int3_exception_nb = { 1521 .notifier_call = int3_exception_notify, 1522 .priority = INT_MAX-1, /* last */ 1523 }; 1524 unsigned int val = 0; 1525 1526 BUG_ON(register_die_notifier(&int3_exception_nb)); 1527 1528 /* 1529 * Basically: int3_magic(&val); but really complicated :-) 1530 * 1531 * INT3 padded with NOP to CALL_INSN_SIZE. The int3_exception_nb 1532 * notifier above will emulate CALL for us. 1533 */ 1534 asm volatile ("int3_selftest_ip:\n\t" 1535 ANNOTATE_NOENDBR 1536 " int3; nop; nop; nop; nop\n\t" 1537 : ASM_CALL_CONSTRAINT 1538 : __ASM_SEL_RAW(a, D) (&val) 1539 : "memory"); 1540 1541 BUG_ON(val != 1); 1542 1543 unregister_die_notifier(&int3_exception_nb); 1544 } 1545 1546 static __initdata int __alt_reloc_selftest_addr; 1547 1548 extern void __init __alt_reloc_selftest(void *arg); 1549 __visible noinline void __init __alt_reloc_selftest(void *arg) 1550 { 1551 WARN_ON(arg != &__alt_reloc_selftest_addr); 1552 } 1553 1554 static noinline void __init alt_reloc_selftest(void) 1555 { 1556 /* 1557 * Tests apply_relocation(). 1558 * 1559 * This has a relative immediate (CALL) in a place other than the first 1560 * instruction and additionally on x86_64 we get a RIP-relative LEA: 1561 * 1562 * lea 0x0(%rip),%rdi # 5d0: R_X86_64_PC32 .init.data+0x5566c 1563 * call +0 # 5d5: R_X86_64_PLT32 __alt_reloc_selftest-0x4 1564 * 1565 * Getting this wrong will either crash and burn or tickle the WARN 1566 * above. 1567 */ 1568 asm_inline volatile ( 1569 ALTERNATIVE("", "lea %[mem], %%" _ASM_ARG1 "; call __alt_reloc_selftest;", X86_FEATURE_ALWAYS) 1570 : /* output */ 1571 : [mem] "m" (__alt_reloc_selftest_addr) 1572 : _ASM_ARG1 1573 ); 1574 } 1575 1576 void __init alternative_instructions(void) 1577 { 1578 int3_selftest(); 1579 1580 /* 1581 * The patching is not fully atomic, so try to avoid local 1582 * interruptions that might execute the to be patched code. 1583 * Other CPUs are not running. 1584 */ 1585 stop_nmi(); 1586 1587 /* 1588 * Don't stop machine check exceptions while patching. 1589 * MCEs only happen when something got corrupted and in this 1590 * case we must do something about the corruption. 1591 * Ignoring it is worse than an unlikely patching race. 1592 * Also machine checks tend to be broadcast and if one CPU 1593 * goes into machine check the others follow quickly, so we don't 1594 * expect a machine check to cause undue problems during to code 1595 * patching. 1596 */ 1597 1598 /* 1599 * Paravirt patching and alternative patching can be combined to 1600 * replace a function call with a short direct code sequence (e.g. 1601 * by setting a constant return value instead of doing that in an 1602 * external function). 1603 * In order to make this work the following sequence is required: 1604 * 1. set (artificial) features depending on used paravirt 1605 * functions which can later influence alternative patching 1606 * 2. apply paravirt patching (generally replacing an indirect 1607 * function call with a direct one) 1608 * 3. apply alternative patching (e.g. replacing a direct function 1609 * call with a custom code sequence) 1610 * Doing paravirt patching after alternative patching would clobber 1611 * the optimization of the custom code with a function call again. 1612 */ 1613 paravirt_set_cap(); 1614 1615 /* 1616 * First patch paravirt functions, such that we overwrite the indirect 1617 * call with the direct call. 1618 */ 1619 apply_paravirt(__parainstructions, __parainstructions_end); 1620 1621 __apply_fineibt(__retpoline_sites, __retpoline_sites_end, 1622 __cfi_sites, __cfi_sites_end, true); 1623 1624 /* 1625 * Rewrite the retpolines, must be done before alternatives since 1626 * those can rewrite the retpoline thunks. 1627 */ 1628 apply_retpolines(__retpoline_sites, __retpoline_sites_end); 1629 apply_returns(__return_sites, __return_sites_end); 1630 1631 /* 1632 * Then patch alternatives, such that those paravirt calls that are in 1633 * alternatives can be overwritten by their immediate fragments. 1634 */ 1635 apply_alternatives(__alt_instructions, __alt_instructions_end); 1636 1637 /* 1638 * Now all calls are established. Apply the call thunks if 1639 * required. 1640 */ 1641 callthunks_patch_builtin_calls(); 1642 1643 /* 1644 * Seal all functions that do not have their address taken. 1645 */ 1646 apply_seal_endbr(__ibt_endbr_seal, __ibt_endbr_seal_end); 1647 1648 #ifdef CONFIG_SMP 1649 /* Patch to UP if other cpus not imminent. */ 1650 if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) { 1651 uniproc_patched = true; 1652 alternatives_smp_module_add(NULL, "core kernel", 1653 __smp_locks, __smp_locks_end, 1654 _text, _etext); 1655 } 1656 1657 if (!uniproc_patched || num_possible_cpus() == 1) { 1658 free_init_pages("SMP alternatives", 1659 (unsigned long)__smp_locks, 1660 (unsigned long)__smp_locks_end); 1661 } 1662 #endif 1663 1664 restart_nmi(); 1665 alternatives_patched = 1; 1666 1667 alt_reloc_selftest(); 1668 } 1669 1670 /** 1671 * text_poke_early - Update instructions on a live kernel at boot time 1672 * @addr: address to modify 1673 * @opcode: source of the copy 1674 * @len: length to copy 1675 * 1676 * When you use this code to patch more than one byte of an instruction 1677 * you need to make sure that other CPUs cannot execute this code in parallel. 1678 * Also no thread must be currently preempted in the middle of these 1679 * instructions. And on the local CPU you need to be protected against NMI or 1680 * MCE handlers seeing an inconsistent instruction while you patch. 1681 */ 1682 void __init_or_module text_poke_early(void *addr, const void *opcode, 1683 size_t len) 1684 { 1685 unsigned long flags; 1686 1687 if (boot_cpu_has(X86_FEATURE_NX) && 1688 is_module_text_address((unsigned long)addr)) { 1689 /* 1690 * Modules text is marked initially as non-executable, so the 1691 * code cannot be running and speculative code-fetches are 1692 * prevented. Just change the code. 1693 */ 1694 memcpy(addr, opcode, len); 1695 } else { 1696 local_irq_save(flags); 1697 memcpy(addr, opcode, len); 1698 sync_core(); 1699 local_irq_restore(flags); 1700 1701 /* 1702 * Could also do a CLFLUSH here to speed up CPU recovery; but 1703 * that causes hangs on some VIA CPUs. 1704 */ 1705 } 1706 } 1707 1708 typedef struct { 1709 struct mm_struct *mm; 1710 } temp_mm_state_t; 1711 1712 /* 1713 * Using a temporary mm allows to set temporary mappings that are not accessible 1714 * by other CPUs. Such mappings are needed to perform sensitive memory writes 1715 * that override the kernel memory protections (e.g., W^X), without exposing the 1716 * temporary page-table mappings that are required for these write operations to 1717 * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the 1718 * mapping is torn down. 1719 * 1720 * Context: The temporary mm needs to be used exclusively by a single core. To 1721 * harden security IRQs must be disabled while the temporary mm is 1722 * loaded, thereby preventing interrupt handler bugs from overriding 1723 * the kernel memory protection. 1724 */ 1725 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm) 1726 { 1727 temp_mm_state_t temp_state; 1728 1729 lockdep_assert_irqs_disabled(); 1730 1731 /* 1732 * Make sure not to be in TLB lazy mode, as otherwise we'll end up 1733 * with a stale address space WITHOUT being in lazy mode after 1734 * restoring the previous mm. 1735 */ 1736 if (this_cpu_read(cpu_tlbstate_shared.is_lazy)) 1737 leave_mm(smp_processor_id()); 1738 1739 temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm); 1740 switch_mm_irqs_off(NULL, mm, current); 1741 1742 /* 1743 * If breakpoints are enabled, disable them while the temporary mm is 1744 * used. Userspace might set up watchpoints on addresses that are used 1745 * in the temporary mm, which would lead to wrong signals being sent or 1746 * crashes. 1747 * 1748 * Note that breakpoints are not disabled selectively, which also causes 1749 * kernel breakpoints (e.g., perf's) to be disabled. This might be 1750 * undesirable, but still seems reasonable as the code that runs in the 1751 * temporary mm should be short. 1752 */ 1753 if (hw_breakpoint_active()) 1754 hw_breakpoint_disable(); 1755 1756 return temp_state; 1757 } 1758 1759 static inline void unuse_temporary_mm(temp_mm_state_t prev_state) 1760 { 1761 lockdep_assert_irqs_disabled(); 1762 switch_mm_irqs_off(NULL, prev_state.mm, current); 1763 1764 /* 1765 * Restore the breakpoints if they were disabled before the temporary mm 1766 * was loaded. 1767 */ 1768 if (hw_breakpoint_active()) 1769 hw_breakpoint_restore(); 1770 } 1771 1772 __ro_after_init struct mm_struct *poking_mm; 1773 __ro_after_init unsigned long poking_addr; 1774 1775 static void text_poke_memcpy(void *dst, const void *src, size_t len) 1776 { 1777 memcpy(dst, src, len); 1778 } 1779 1780 static void text_poke_memset(void *dst, const void *src, size_t len) 1781 { 1782 int c = *(const int *)src; 1783 1784 memset(dst, c, len); 1785 } 1786 1787 typedef void text_poke_f(void *dst, const void *src, size_t len); 1788 1789 static void *__text_poke(text_poke_f func, void *addr, const void *src, size_t len) 1790 { 1791 bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE; 1792 struct page *pages[2] = {NULL}; 1793 temp_mm_state_t prev; 1794 unsigned long flags; 1795 pte_t pte, *ptep; 1796 spinlock_t *ptl; 1797 pgprot_t pgprot; 1798 1799 /* 1800 * While boot memory allocator is running we cannot use struct pages as 1801 * they are not yet initialized. There is no way to recover. 1802 */ 1803 BUG_ON(!after_bootmem); 1804 1805 if (!core_kernel_text((unsigned long)addr)) { 1806 pages[0] = vmalloc_to_page(addr); 1807 if (cross_page_boundary) 1808 pages[1] = vmalloc_to_page(addr + PAGE_SIZE); 1809 } else { 1810 pages[0] = virt_to_page(addr); 1811 WARN_ON(!PageReserved(pages[0])); 1812 if (cross_page_boundary) 1813 pages[1] = virt_to_page(addr + PAGE_SIZE); 1814 } 1815 /* 1816 * If something went wrong, crash and burn since recovery paths are not 1817 * implemented. 1818 */ 1819 BUG_ON(!pages[0] || (cross_page_boundary && !pages[1])); 1820 1821 /* 1822 * Map the page without the global bit, as TLB flushing is done with 1823 * flush_tlb_mm_range(), which is intended for non-global PTEs. 1824 */ 1825 pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL); 1826 1827 /* 1828 * The lock is not really needed, but this allows to avoid open-coding. 1829 */ 1830 ptep = get_locked_pte(poking_mm, poking_addr, &ptl); 1831 1832 /* 1833 * This must not fail; preallocated in poking_init(). 1834 */ 1835 VM_BUG_ON(!ptep); 1836 1837 local_irq_save(flags); 1838 1839 pte = mk_pte(pages[0], pgprot); 1840 set_pte_at(poking_mm, poking_addr, ptep, pte); 1841 1842 if (cross_page_boundary) { 1843 pte = mk_pte(pages[1], pgprot); 1844 set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte); 1845 } 1846 1847 /* 1848 * Loading the temporary mm behaves as a compiler barrier, which 1849 * guarantees that the PTE will be set at the time memcpy() is done. 1850 */ 1851 prev = use_temporary_mm(poking_mm); 1852 1853 kasan_disable_current(); 1854 func((u8 *)poking_addr + offset_in_page(addr), src, len); 1855 kasan_enable_current(); 1856 1857 /* 1858 * Ensure that the PTE is only cleared after the instructions of memcpy 1859 * were issued by using a compiler barrier. 1860 */ 1861 barrier(); 1862 1863 pte_clear(poking_mm, poking_addr, ptep); 1864 if (cross_page_boundary) 1865 pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1); 1866 1867 /* 1868 * Loading the previous page-table hierarchy requires a serializing 1869 * instruction that already allows the core to see the updated version. 1870 * Xen-PV is assumed to serialize execution in a similar manner. 1871 */ 1872 unuse_temporary_mm(prev); 1873 1874 /* 1875 * Flushing the TLB might involve IPIs, which would require enabled 1876 * IRQs, but not if the mm is not used, as it is in this point. 1877 */ 1878 flush_tlb_mm_range(poking_mm, poking_addr, poking_addr + 1879 (cross_page_boundary ? 2 : 1) * PAGE_SIZE, 1880 PAGE_SHIFT, false); 1881 1882 if (func == text_poke_memcpy) { 1883 /* 1884 * If the text does not match what we just wrote then something is 1885 * fundamentally screwy; there's nothing we can really do about that. 1886 */ 1887 BUG_ON(memcmp(addr, src, len)); 1888 } 1889 1890 local_irq_restore(flags); 1891 pte_unmap_unlock(ptep, ptl); 1892 return addr; 1893 } 1894 1895 /** 1896 * text_poke - Update instructions on a live kernel 1897 * @addr: address to modify 1898 * @opcode: source of the copy 1899 * @len: length to copy 1900 * 1901 * Only atomic text poke/set should be allowed when not doing early patching. 1902 * It means the size must be writable atomically and the address must be aligned 1903 * in a way that permits an atomic write. It also makes sure we fit on a single 1904 * page. 1905 * 1906 * Note that the caller must ensure that if the modified code is part of a 1907 * module, the module would not be removed during poking. This can be achieved 1908 * by registering a module notifier, and ordering module removal and patching 1909 * trough a mutex. 1910 */ 1911 void *text_poke(void *addr, const void *opcode, size_t len) 1912 { 1913 lockdep_assert_held(&text_mutex); 1914 1915 return __text_poke(text_poke_memcpy, addr, opcode, len); 1916 } 1917 1918 /** 1919 * text_poke_kgdb - Update instructions on a live kernel by kgdb 1920 * @addr: address to modify 1921 * @opcode: source of the copy 1922 * @len: length to copy 1923 * 1924 * Only atomic text poke/set should be allowed when not doing early patching. 1925 * It means the size must be writable atomically and the address must be aligned 1926 * in a way that permits an atomic write. It also makes sure we fit on a single 1927 * page. 1928 * 1929 * Context: should only be used by kgdb, which ensures no other core is running, 1930 * despite the fact it does not hold the text_mutex. 1931 */ 1932 void *text_poke_kgdb(void *addr, const void *opcode, size_t len) 1933 { 1934 return __text_poke(text_poke_memcpy, addr, opcode, len); 1935 } 1936 1937 void *text_poke_copy_locked(void *addr, const void *opcode, size_t len, 1938 bool core_ok) 1939 { 1940 unsigned long start = (unsigned long)addr; 1941 size_t patched = 0; 1942 1943 if (WARN_ON_ONCE(!core_ok && core_kernel_text(start))) 1944 return NULL; 1945 1946 while (patched < len) { 1947 unsigned long ptr = start + patched; 1948 size_t s; 1949 1950 s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched); 1951 1952 __text_poke(text_poke_memcpy, (void *)ptr, opcode + patched, s); 1953 patched += s; 1954 } 1955 return addr; 1956 } 1957 1958 /** 1959 * text_poke_copy - Copy instructions into (an unused part of) RX memory 1960 * @addr: address to modify 1961 * @opcode: source of the copy 1962 * @len: length to copy, could be more than 2x PAGE_SIZE 1963 * 1964 * Not safe against concurrent execution; useful for JITs to dump 1965 * new code blocks into unused regions of RX memory. Can be used in 1966 * conjunction with synchronize_rcu_tasks() to wait for existing 1967 * execution to quiesce after having made sure no existing functions 1968 * pointers are live. 1969 */ 1970 void *text_poke_copy(void *addr, const void *opcode, size_t len) 1971 { 1972 mutex_lock(&text_mutex); 1973 addr = text_poke_copy_locked(addr, opcode, len, false); 1974 mutex_unlock(&text_mutex); 1975 return addr; 1976 } 1977 1978 /** 1979 * text_poke_set - memset into (an unused part of) RX memory 1980 * @addr: address to modify 1981 * @c: the byte to fill the area with 1982 * @len: length to copy, could be more than 2x PAGE_SIZE 1983 * 1984 * This is useful to overwrite unused regions of RX memory with illegal 1985 * instructions. 1986 */ 1987 void *text_poke_set(void *addr, int c, size_t len) 1988 { 1989 unsigned long start = (unsigned long)addr; 1990 size_t patched = 0; 1991 1992 if (WARN_ON_ONCE(core_kernel_text(start))) 1993 return NULL; 1994 1995 mutex_lock(&text_mutex); 1996 while (patched < len) { 1997 unsigned long ptr = start + patched; 1998 size_t s; 1999 2000 s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched); 2001 2002 __text_poke(text_poke_memset, (void *)ptr, (void *)&c, s); 2003 patched += s; 2004 } 2005 mutex_unlock(&text_mutex); 2006 return addr; 2007 } 2008 2009 static void do_sync_core(void *info) 2010 { 2011 sync_core(); 2012 } 2013 2014 void text_poke_sync(void) 2015 { 2016 on_each_cpu(do_sync_core, NULL, 1); 2017 } 2018 2019 /* 2020 * NOTE: crazy scheme to allow patching Jcc.d32 but not increase the size of 2021 * this thing. When len == 6 everything is prefixed with 0x0f and we map 2022 * opcode to Jcc.d8, using len to distinguish. 2023 */ 2024 struct text_poke_loc { 2025 /* addr := _stext + rel_addr */ 2026 s32 rel_addr; 2027 s32 disp; 2028 u8 len; 2029 u8 opcode; 2030 const u8 text[POKE_MAX_OPCODE_SIZE]; 2031 /* see text_poke_bp_batch() */ 2032 u8 old; 2033 }; 2034 2035 struct bp_patching_desc { 2036 struct text_poke_loc *vec; 2037 int nr_entries; 2038 atomic_t refs; 2039 }; 2040 2041 static struct bp_patching_desc bp_desc; 2042 2043 static __always_inline 2044 struct bp_patching_desc *try_get_desc(void) 2045 { 2046 struct bp_patching_desc *desc = &bp_desc; 2047 2048 if (!raw_atomic_inc_not_zero(&desc->refs)) 2049 return NULL; 2050 2051 return desc; 2052 } 2053 2054 static __always_inline void put_desc(void) 2055 { 2056 struct bp_patching_desc *desc = &bp_desc; 2057 2058 smp_mb__before_atomic(); 2059 raw_atomic_dec(&desc->refs); 2060 } 2061 2062 static __always_inline void *text_poke_addr(struct text_poke_loc *tp) 2063 { 2064 return _stext + tp->rel_addr; 2065 } 2066 2067 static __always_inline int patch_cmp(const void *key, const void *elt) 2068 { 2069 struct text_poke_loc *tp = (struct text_poke_loc *) elt; 2070 2071 if (key < text_poke_addr(tp)) 2072 return -1; 2073 if (key > text_poke_addr(tp)) 2074 return 1; 2075 return 0; 2076 } 2077 2078 noinstr int poke_int3_handler(struct pt_regs *regs) 2079 { 2080 struct bp_patching_desc *desc; 2081 struct text_poke_loc *tp; 2082 int ret = 0; 2083 void *ip; 2084 2085 if (user_mode(regs)) 2086 return 0; 2087 2088 /* 2089 * Having observed our INT3 instruction, we now must observe 2090 * bp_desc with non-zero refcount: 2091 * 2092 * bp_desc.refs = 1 INT3 2093 * WMB RMB 2094 * write INT3 if (bp_desc.refs != 0) 2095 */ 2096 smp_rmb(); 2097 2098 desc = try_get_desc(); 2099 if (!desc) 2100 return 0; 2101 2102 /* 2103 * Discount the INT3. See text_poke_bp_batch(). 2104 */ 2105 ip = (void *) regs->ip - INT3_INSN_SIZE; 2106 2107 /* 2108 * Skip the binary search if there is a single member in the vector. 2109 */ 2110 if (unlikely(desc->nr_entries > 1)) { 2111 tp = __inline_bsearch(ip, desc->vec, desc->nr_entries, 2112 sizeof(struct text_poke_loc), 2113 patch_cmp); 2114 if (!tp) 2115 goto out_put; 2116 } else { 2117 tp = desc->vec; 2118 if (text_poke_addr(tp) != ip) 2119 goto out_put; 2120 } 2121 2122 ip += tp->len; 2123 2124 switch (tp->opcode) { 2125 case INT3_INSN_OPCODE: 2126 /* 2127 * Someone poked an explicit INT3, they'll want to handle it, 2128 * do not consume. 2129 */ 2130 goto out_put; 2131 2132 case RET_INSN_OPCODE: 2133 int3_emulate_ret(regs); 2134 break; 2135 2136 case CALL_INSN_OPCODE: 2137 int3_emulate_call(regs, (long)ip + tp->disp); 2138 break; 2139 2140 case JMP32_INSN_OPCODE: 2141 case JMP8_INSN_OPCODE: 2142 int3_emulate_jmp(regs, (long)ip + tp->disp); 2143 break; 2144 2145 case 0x70 ... 0x7f: /* Jcc */ 2146 int3_emulate_jcc(regs, tp->opcode & 0xf, (long)ip, tp->disp); 2147 break; 2148 2149 default: 2150 BUG(); 2151 } 2152 2153 ret = 1; 2154 2155 out_put: 2156 put_desc(); 2157 return ret; 2158 } 2159 2160 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc)) 2161 static struct text_poke_loc tp_vec[TP_VEC_MAX]; 2162 static int tp_vec_nr; 2163 2164 /** 2165 * text_poke_bp_batch() -- update instructions on live kernel on SMP 2166 * @tp: vector of instructions to patch 2167 * @nr_entries: number of entries in the vector 2168 * 2169 * Modify multi-byte instruction by using int3 breakpoint on SMP. 2170 * We completely avoid stop_machine() here, and achieve the 2171 * synchronization using int3 breakpoint. 2172 * 2173 * The way it is done: 2174 * - For each entry in the vector: 2175 * - add a int3 trap to the address that will be patched 2176 * - sync cores 2177 * - For each entry in the vector: 2178 * - update all but the first byte of the patched range 2179 * - sync cores 2180 * - For each entry in the vector: 2181 * - replace the first byte (int3) by the first byte of 2182 * replacing opcode 2183 * - sync cores 2184 */ 2185 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries) 2186 { 2187 unsigned char int3 = INT3_INSN_OPCODE; 2188 unsigned int i; 2189 int do_sync; 2190 2191 lockdep_assert_held(&text_mutex); 2192 2193 bp_desc.vec = tp; 2194 bp_desc.nr_entries = nr_entries; 2195 2196 /* 2197 * Corresponds to the implicit memory barrier in try_get_desc() to 2198 * ensure reading a non-zero refcount provides up to date bp_desc data. 2199 */ 2200 atomic_set_release(&bp_desc.refs, 1); 2201 2202 /* 2203 * Function tracing can enable thousands of places that need to be 2204 * updated. This can take quite some time, and with full kernel debugging 2205 * enabled, this could cause the softlockup watchdog to trigger. 2206 * This function gets called every 256 entries added to be patched. 2207 * Call cond_resched() here to make sure that other tasks can get scheduled 2208 * while processing all the functions being patched. 2209 */ 2210 cond_resched(); 2211 2212 /* 2213 * Corresponding read barrier in int3 notifier for making sure the 2214 * nr_entries and handler are correctly ordered wrt. patching. 2215 */ 2216 smp_wmb(); 2217 2218 /* 2219 * First step: add a int3 trap to the address that will be patched. 2220 */ 2221 for (i = 0; i < nr_entries; i++) { 2222 tp[i].old = *(u8 *)text_poke_addr(&tp[i]); 2223 text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE); 2224 } 2225 2226 text_poke_sync(); 2227 2228 /* 2229 * Second step: update all but the first byte of the patched range. 2230 */ 2231 for (do_sync = 0, i = 0; i < nr_entries; i++) { 2232 u8 old[POKE_MAX_OPCODE_SIZE+1] = { tp[i].old, }; 2233 u8 _new[POKE_MAX_OPCODE_SIZE+1]; 2234 const u8 *new = tp[i].text; 2235 int len = tp[i].len; 2236 2237 if (len - INT3_INSN_SIZE > 0) { 2238 memcpy(old + INT3_INSN_SIZE, 2239 text_poke_addr(&tp[i]) + INT3_INSN_SIZE, 2240 len - INT3_INSN_SIZE); 2241 2242 if (len == 6) { 2243 _new[0] = 0x0f; 2244 memcpy(_new + 1, new, 5); 2245 new = _new; 2246 } 2247 2248 text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE, 2249 new + INT3_INSN_SIZE, 2250 len - INT3_INSN_SIZE); 2251 2252 do_sync++; 2253 } 2254 2255 /* 2256 * Emit a perf event to record the text poke, primarily to 2257 * support Intel PT decoding which must walk the executable code 2258 * to reconstruct the trace. The flow up to here is: 2259 * - write INT3 byte 2260 * - IPI-SYNC 2261 * - write instruction tail 2262 * At this point the actual control flow will be through the 2263 * INT3 and handler and not hit the old or new instruction. 2264 * Intel PT outputs FUP/TIP packets for the INT3, so the flow 2265 * can still be decoded. Subsequently: 2266 * - emit RECORD_TEXT_POKE with the new instruction 2267 * - IPI-SYNC 2268 * - write first byte 2269 * - IPI-SYNC 2270 * So before the text poke event timestamp, the decoder will see 2271 * either the old instruction flow or FUP/TIP of INT3. After the 2272 * text poke event timestamp, the decoder will see either the 2273 * new instruction flow or FUP/TIP of INT3. Thus decoders can 2274 * use the timestamp as the point at which to modify the 2275 * executable code. 2276 * The old instruction is recorded so that the event can be 2277 * processed forwards or backwards. 2278 */ 2279 perf_event_text_poke(text_poke_addr(&tp[i]), old, len, new, len); 2280 } 2281 2282 if (do_sync) { 2283 /* 2284 * According to Intel, this core syncing is very likely 2285 * not necessary and we'd be safe even without it. But 2286 * better safe than sorry (plus there's not only Intel). 2287 */ 2288 text_poke_sync(); 2289 } 2290 2291 /* 2292 * Third step: replace the first byte (int3) by the first byte of 2293 * replacing opcode. 2294 */ 2295 for (do_sync = 0, i = 0; i < nr_entries; i++) { 2296 u8 byte = tp[i].text[0]; 2297 2298 if (tp[i].len == 6) 2299 byte = 0x0f; 2300 2301 if (byte == INT3_INSN_OPCODE) 2302 continue; 2303 2304 text_poke(text_poke_addr(&tp[i]), &byte, INT3_INSN_SIZE); 2305 do_sync++; 2306 } 2307 2308 if (do_sync) 2309 text_poke_sync(); 2310 2311 /* 2312 * Remove and wait for refs to be zero. 2313 */ 2314 if (!atomic_dec_and_test(&bp_desc.refs)) 2315 atomic_cond_read_acquire(&bp_desc.refs, !VAL); 2316 } 2317 2318 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr, 2319 const void *opcode, size_t len, const void *emulate) 2320 { 2321 struct insn insn; 2322 int ret, i = 0; 2323 2324 if (len == 6) 2325 i = 1; 2326 memcpy((void *)tp->text, opcode+i, len-i); 2327 if (!emulate) 2328 emulate = opcode; 2329 2330 ret = insn_decode_kernel(&insn, emulate); 2331 BUG_ON(ret < 0); 2332 2333 tp->rel_addr = addr - (void *)_stext; 2334 tp->len = len; 2335 tp->opcode = insn.opcode.bytes[0]; 2336 2337 if (is_jcc32(&insn)) { 2338 /* 2339 * Map Jcc.d32 onto Jcc.d8 and use len to distinguish. 2340 */ 2341 tp->opcode = insn.opcode.bytes[1] - 0x10; 2342 } 2343 2344 switch (tp->opcode) { 2345 case RET_INSN_OPCODE: 2346 case JMP32_INSN_OPCODE: 2347 case JMP8_INSN_OPCODE: 2348 /* 2349 * Control flow instructions without implied execution of the 2350 * next instruction can be padded with INT3. 2351 */ 2352 for (i = insn.length; i < len; i++) 2353 BUG_ON(tp->text[i] != INT3_INSN_OPCODE); 2354 break; 2355 2356 default: 2357 BUG_ON(len != insn.length); 2358 } 2359 2360 switch (tp->opcode) { 2361 case INT3_INSN_OPCODE: 2362 case RET_INSN_OPCODE: 2363 break; 2364 2365 case CALL_INSN_OPCODE: 2366 case JMP32_INSN_OPCODE: 2367 case JMP8_INSN_OPCODE: 2368 case 0x70 ... 0x7f: /* Jcc */ 2369 tp->disp = insn.immediate.value; 2370 break; 2371 2372 default: /* assume NOP */ 2373 switch (len) { 2374 case 2: /* NOP2 -- emulate as JMP8+0 */ 2375 BUG_ON(memcmp(emulate, x86_nops[len], len)); 2376 tp->opcode = JMP8_INSN_OPCODE; 2377 tp->disp = 0; 2378 break; 2379 2380 case 5: /* NOP5 -- emulate as JMP32+0 */ 2381 BUG_ON(memcmp(emulate, x86_nops[len], len)); 2382 tp->opcode = JMP32_INSN_OPCODE; 2383 tp->disp = 0; 2384 break; 2385 2386 default: /* unknown instruction */ 2387 BUG(); 2388 } 2389 break; 2390 } 2391 } 2392 2393 /* 2394 * We hard rely on the tp_vec being ordered; ensure this is so by flushing 2395 * early if needed. 2396 */ 2397 static bool tp_order_fail(void *addr) 2398 { 2399 struct text_poke_loc *tp; 2400 2401 if (!tp_vec_nr) 2402 return false; 2403 2404 if (!addr) /* force */ 2405 return true; 2406 2407 tp = &tp_vec[tp_vec_nr - 1]; 2408 if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr) 2409 return true; 2410 2411 return false; 2412 } 2413 2414 static void text_poke_flush(void *addr) 2415 { 2416 if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) { 2417 text_poke_bp_batch(tp_vec, tp_vec_nr); 2418 tp_vec_nr = 0; 2419 } 2420 } 2421 2422 void text_poke_finish(void) 2423 { 2424 text_poke_flush(NULL); 2425 } 2426 2427 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate) 2428 { 2429 struct text_poke_loc *tp; 2430 2431 text_poke_flush(addr); 2432 2433 tp = &tp_vec[tp_vec_nr++]; 2434 text_poke_loc_init(tp, addr, opcode, len, emulate); 2435 } 2436 2437 /** 2438 * text_poke_bp() -- update instructions on live kernel on SMP 2439 * @addr: address to patch 2440 * @opcode: opcode of new instruction 2441 * @len: length to copy 2442 * @emulate: instruction to be emulated 2443 * 2444 * Update a single instruction with the vector in the stack, avoiding 2445 * dynamically allocated memory. This function should be used when it is 2446 * not possible to allocate memory. 2447 */ 2448 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate) 2449 { 2450 struct text_poke_loc tp; 2451 2452 text_poke_loc_init(&tp, addr, opcode, len, emulate); 2453 text_poke_bp_batch(&tp, 1); 2454 } 2455