1//===-- interception_win.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file is a part of AddressSanitizer, an address sanity checker.
10//
11// Windows-specific interception methods.
12//
13// This file is implementing several hooking techniques to intercept calls
14// to functions. The hooks are dynamically installed by modifying the assembly
15// code.
16//
17// The hooking techniques are making assumptions on the way the code is
18// generated and are safe under these assumptions.
19//
20// On 64-bit architecture, there is no direct 64-bit jump instruction. To allow
21// arbitrary branching on the whole memory space, the notion of trampoline
22// region is used. A trampoline region is a memory space withing 2G boundary
23// where it is safe to add custom assembly code to build 64-bit jumps.
24//
25// Hooking techniques
26// ==================
27//
28// 1) Detour
29//
30// The Detour hooking technique is assuming the presence of an header with
31// padding and an overridable 2-bytes nop instruction (mov edi, edi). The
32// nop instruction can safely be replaced by a 2-bytes jump without any need
33// to save the instruction. A jump to the target is encoded in the function
34// header and the nop instruction is replaced by a short jump to the header.
35//
36// head: 5 x nop head: jmp <hook>
37// func: mov edi, edi --> func: jmp short <head>
38// [...] real: [...]
39//
40// This technique is only implemented on 32-bit architecture.
41// Most of the time, Windows API are hookable with the detour technique.
42//
43// 2) Redirect Jump
44//
45// The redirect jump is applicable when the first instruction is a direct
46// jump. The instruction is replaced by jump to the hook.
47//
48// func: jmp <label> --> func: jmp <hook>
49//
50// On an 64-bit architecture, a trampoline is inserted.
51//
52// func: jmp <label> --> func: jmp <tramp>
53// [...]
54//
55// [trampoline]
56// tramp: jmp QWORD [addr]
57// addr: .bytes <hook>
58//
59// Note: <real> is equivalent to <label>.
60//
61// 3) HotPatch
62//
63// The HotPatch hooking is assuming the presence of an header with padding
64// and a first instruction with at least 2-bytes.
65//
66// The reason to enforce the 2-bytes limitation is to provide the minimal
67// space to encode a short jump. HotPatch technique is only rewriting one
68// instruction to avoid breaking a sequence of instructions containing a
69// branching target.
70//
71// Assumptions are enforced by MSVC compiler by using the /HOTPATCH flag.
72// see: https://msdn.microsoft.com/en-us/library/ms173507.aspx
73// Default padding length is 5 bytes in 32-bits and 6 bytes in 64-bits.
74//
75// head: 5 x nop head: jmp <hook>
76// func: <instr> --> func: jmp short <head>
77// [...] body: [...]
78//
79// [trampoline]
80// real: <instr>
81// jmp <body>
82//
83// On an 64-bit architecture:
84//
85// head: 6 x nop head: jmp QWORD [addr1]
86// func: <instr> --> func: jmp short <head>
87// [...] body: [...]
88//
89// [trampoline]
90// addr1: .bytes <hook>
91// real: <instr>
92// jmp QWORD [addr2]
93// addr2: .bytes <body>
94//
95// 4) Trampoline
96//
97// The Trampoline hooking technique is the most aggressive one. It is
98// assuming that there is a sequence of instructions that can be safely
99// replaced by a jump (enough room and no incoming branches).
100//
101// Unfortunately, these assumptions can't be safely presumed and code may
102// be broken after hooking.
103//
104// func: <instr> --> func: jmp <hook>
105// <instr>
106// [...] body: [...]
107//
108// [trampoline]
109// real: <instr>
110// <instr>
111// jmp <body>
112//
113// On an 64-bit architecture:
114//
115// func: <instr> --> func: jmp QWORD [addr1]
116// <instr>
117// [...] body: [...]
118//
119// [trampoline]
120// addr1: .bytes <hook>
121// real: <instr>
122// <instr>
123// jmp QWORD [addr2]
124// addr2: .bytes <body>
125//===----------------------------------------------------------------------===//
126
127#include "interception.h"
128
129#if SANITIZER_WINDOWS
130#include "sanitizer_common/sanitizer_platform.h"
131#define WIN32_LEAN_AND_MEAN
132#include <windows.h>
133
134namespace __interception {
135
136static const int kAddressLength = FIRST_32_SECOND_64(4, 8);
137static const int kJumpInstructionLength = 5;
138static const int kShortJumpInstructionLength = 2;
139UNUSED static const int kIndirectJumpInstructionLength = 6;
140static const int kBranchLength =
141 FIRST_32_SECOND_64(kJumpInstructionLength, kIndirectJumpInstructionLength);
142static const int kDirectBranchLength = kBranchLength + kAddressLength;
143
144# if defined(_MSC_VER)
145# define INTERCEPTION_FORMAT(f, a)
146# else
147# define INTERCEPTION_FORMAT(f, a) __attribute__((format(printf, f, a)))
148# endif
149
150static void (*ErrorReportCallback)(const char *format, ...)
151 INTERCEPTION_FORMAT(1, 2);
152
153void SetErrorReportCallback(void (*callback)(const char *format, ...)) {
154 ErrorReportCallback = callback;
155}
156
157# define ReportError(...) \
158 do { \
159 if (ErrorReportCallback) \
160 ErrorReportCallback(__VA_ARGS__); \
161 } while (0)
162
163static void InterceptionFailed() {
164 ReportError("interception_win: failed due to an unrecoverable error.\n");
165 // This acts like an abort when no debugger is attached. According to an old
166 // comment, calling abort() leads to an infinite recursion in CheckFailed.
167 __debugbreak();
168}
169
170static bool DistanceIsWithin2Gig(uptr from, uptr target) {
171#if SANITIZER_WINDOWS64
172 if (from < target)
173 return target - from <= (uptr)0x7FFFFFFFU;
174 else
175 return from - target <= (uptr)0x80000000U;
176#else
177 // In a 32-bit address space, the address calculation will wrap, so this check
178 // is unnecessary.
179 return true;
180#endif
181}
182
183static uptr GetMmapGranularity() {
184 SYSTEM_INFO si;
185 GetSystemInfo(&si);
186 return si.dwAllocationGranularity;
187}
188
189UNUSED static uptr RoundUpTo(uptr size, uptr boundary) {
190 return (size + boundary - 1) & ~(boundary - 1);
191}
192
193// FIXME: internal_str* and internal_mem* functions should be moved from the
194// ASan sources into interception/.
195
196static size_t _strlen(const char *str) {
197 const char* p = str;
198 while (*p != '\0') ++p;
199 return p - str;
200}
201
202static char* _strchr(char* str, char c) {
203 while (*str) {
204 if (*str == c)
205 return str;
206 ++str;
207 }
208 return nullptr;
209}
210
211static void _memset(void *p, int value, size_t sz) {
212 for (size_t i = 0; i < sz; ++i)
213 ((char*)p)[i] = (char)value;
214}
215
216static void _memcpy(void *dst, void *src, size_t sz) {
217 char *dst_c = (char*)dst,
218 *src_c = (char*)src;
219 for (size_t i = 0; i < sz; ++i)
220 dst_c[i] = src_c[i];
221}
222
223static bool ChangeMemoryProtection(
224 uptr address, uptr size, DWORD *old_protection) {
225 return ::VirtualProtect((void*)address, size,
226 PAGE_EXECUTE_READWRITE,
227 old_protection) != FALSE;
228}
229
230static bool RestoreMemoryProtection(
231 uptr address, uptr size, DWORD old_protection) {
232 DWORD unused;
233 return ::VirtualProtect((void*)address, size,
234 old_protection,
235 &unused) != FALSE;
236}
237
238static bool IsMemoryPadding(uptr address, uptr size) {
239 u8* function = (u8*)address;
240 for (size_t i = 0; i < size; ++i)
241 if (function[i] != 0x90 && function[i] != 0xCC)
242 return false;
243 return true;
244}
245
246static const u8 kHintNop8Bytes[] = {
247 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00
248};
249
250template<class T>
251static bool FunctionHasPrefix(uptr address, const T &pattern) {
252 u8* function = (u8*)address - sizeof(pattern);
253 for (size_t i = 0; i < sizeof(pattern); ++i)
254 if (function[i] != pattern[i])
255 return false;
256 return true;
257}
258
259static bool FunctionHasPadding(uptr address, uptr size) {
260 if (IsMemoryPadding(address - size, size))
261 return true;
262 if (size <= sizeof(kHintNop8Bytes) &&
263 FunctionHasPrefix(address, kHintNop8Bytes))
264 return true;
265 return false;
266}
267
268static void WritePadding(uptr from, uptr size) {
269 _memset((void*)from, 0xCC, (size_t)size);
270}
271
272static void WriteJumpInstruction(uptr from, uptr target) {
273 if (!DistanceIsWithin2Gig(from + kJumpInstructionLength, target)) {
274 ReportError(
275 "interception_win: cannot write jmp further than 2GB away, from %p to "
276 "%p.\n",
277 (void *)from, (void *)target);
278 InterceptionFailed();
279 }
280 ptrdiff_t offset = target - from - kJumpInstructionLength;
281 *(u8*)from = 0xE9;
282 *(u32*)(from + 1) = offset;
283}
284
285static void WriteShortJumpInstruction(uptr from, uptr target) {
286 sptr offset = target - from - kShortJumpInstructionLength;
287 if (offset < -128 || offset > 127)
288 InterceptionFailed();
289 *(u8*)from = 0xEB;
290 *(u8*)(from + 1) = (u8)offset;
291}
292
293#if SANITIZER_WINDOWS64
294static void WriteIndirectJumpInstruction(uptr from, uptr indirect_target) {
295 // jmp [rip + <offset>] = FF 25 <offset> where <offset> is a relative
296 // offset.
297 // The offset is the distance from then end of the jump instruction to the
298 // memory location containing the targeted address. The displacement is still
299 // 32-bit in x64, so indirect_target must be located within +/- 2GB range.
300 int offset = indirect_target - from - kIndirectJumpInstructionLength;
301 if (!DistanceIsWithin2Gig(from + kIndirectJumpInstructionLength,
302 indirect_target)) {
303 ReportError(
304 "interception_win: cannot write indirect jmp with target further than "
305 "2GB away, from %p to %p.\n",
306 (void *)from, (void *)indirect_target);
307 InterceptionFailed();
308 }
309 *(u16*)from = 0x25FF;
310 *(u32*)(from + 2) = offset;
311}
312#endif
313
314static void WriteBranch(
315 uptr from, uptr indirect_target, uptr target) {
316#if SANITIZER_WINDOWS64
317 WriteIndirectJumpInstruction(from, indirect_target);
318 *(u64*)indirect_target = target;
319#else
320 (void)indirect_target;
321 WriteJumpInstruction(from, target);
322#endif
323}
324
325static void WriteDirectBranch(uptr from, uptr target) {
326#if SANITIZER_WINDOWS64
327 // Emit an indirect jump through immediately following bytes:
328 // jmp [rip + kBranchLength]
329 // .quad <target>
330 WriteBranch(from, from + kBranchLength, target);
331#else
332 WriteJumpInstruction(from, target);
333#endif
334}
335
336struct TrampolineMemoryRegion {
337 uptr content;
338 uptr allocated_size;
339 uptr max_size;
340};
341
342UNUSED static const uptr kTrampolineScanLimitRange = 1 << 31; // 2 gig
343static const int kMaxTrampolineRegion = 1024;
344static TrampolineMemoryRegion TrampolineRegions[kMaxTrampolineRegion];
345
346static void *AllocateTrampolineRegion(uptr image_address, size_t granularity) {
347#if SANITIZER_WINDOWS64
348 uptr address = image_address;
349 uptr scanned = 0;
350 while (scanned < kTrampolineScanLimitRange) {
351 MEMORY_BASIC_INFORMATION info;
352 if (!::VirtualQuery((void*)address, &info, sizeof(info)))
353 return nullptr;
354
355 // Check whether a region can be allocated at |address|.
356 if (info.State == MEM_FREE && info.RegionSize >= granularity) {
357 void *page = ::VirtualAlloc((void*)RoundUpTo(address, granularity),
358 granularity,
359 MEM_RESERVE | MEM_COMMIT,
360 PAGE_EXECUTE_READWRITE);
361 return page;
362 }
363
364 // Move to the next region.
365 address = (uptr)info.BaseAddress + info.RegionSize;
366 scanned += info.RegionSize;
367 }
368 return nullptr;
369#else
370 return ::VirtualAlloc(nullptr,
371 granularity,
372 MEM_RESERVE | MEM_COMMIT,
373 PAGE_EXECUTE_READWRITE);
374#endif
375}
376
377// Used by unittests to release mapped memory space.
378void TestOnlyReleaseTrampolineRegions() {
379 for (size_t bucket = 0; bucket < kMaxTrampolineRegion; ++bucket) {
380 TrampolineMemoryRegion *current = &TrampolineRegions[bucket];
381 if (current->content == 0)
382 return;
383 ::VirtualFree((void*)current->content, 0, MEM_RELEASE);
384 current->content = 0;
385 }
386}
387
388static uptr AllocateMemoryForTrampoline(uptr image_address, size_t size) {
389 // Find a region within 2G with enough space to allocate |size| bytes.
390 TrampolineMemoryRegion *region = nullptr;
391 for (size_t bucket = 0; bucket < kMaxTrampolineRegion; ++bucket) {
392 TrampolineMemoryRegion* current = &TrampolineRegions[bucket];
393 if (current->content == 0) {
394 // No valid region found, allocate a new region.
395 size_t bucket_size = GetMmapGranularity();
396 void *content = AllocateTrampolineRegion(image_address, bucket_size);
397 if (content == nullptr)
398 return 0U;
399
400 current->content = (uptr)content;
401 current->allocated_size = 0;
402 current->max_size = bucket_size;
403 region = current;
404 break;
405 } else if (current->max_size - current->allocated_size > size) {
406#if SANITIZER_WINDOWS64
407 // In 64-bits, the memory space must be allocated within 2G boundary.
408 uptr next_address = current->content + current->allocated_size;
409 if (next_address < image_address ||
410 next_address - image_address >= 0x7FFF0000)
411 continue;
412#endif
413 // The space can be allocated in the current region.
414 region = current;
415 break;
416 }
417 }
418
419 // Failed to find a region.
420 if (region == nullptr)
421 return 0U;
422
423 // Allocate the space in the current region.
424 uptr allocated_space = region->content + region->allocated_size;
425 region->allocated_size += size;
426 WritePadding(allocated_space, size);
427
428 return allocated_space;
429}
430
431// The following prologues cannot be patched because of the short jump
432// jumping to the patching region.
433
434// Short jump patterns below are only for x86_64.
435# if SANITIZER_WINDOWS_x64
436// ntdll!wcslen in Win11
437// 488bc1 mov rax,rcx
438// 0fb710 movzx edx,word ptr [rax]
439// 4883c002 add rax,2
440// 6685d2 test dx,dx
441// 75f4 jne -12
442static const u8 kPrologueWithShortJump1[] = {
443 0x48, 0x8b, 0xc1, 0x0f, 0xb7, 0x10, 0x48, 0x83,
444 0xc0, 0x02, 0x66, 0x85, 0xd2, 0x75, 0xf4,
445};
446
447// ntdll!strrchr in Win11
448// 4c8bc1 mov r8,rcx
449// 8a01 mov al,byte ptr [rcx]
450// 48ffc1 inc rcx
451// 84c0 test al,al
452// 75f7 jne -9
453static const u8 kPrologueWithShortJump2[] = {
454 0x4c, 0x8b, 0xc1, 0x8a, 0x01, 0x48, 0xff, 0xc1,
455 0x84, 0xc0, 0x75, 0xf7,
456};
457#endif
458
459// Returns 0 on error.
460static size_t GetInstructionSize(uptr address, size_t* rel_offset = nullptr) {
461#if SANITIZER_ARM64
462 // An ARM64 instruction is 4 bytes long.
463 return 4;
464#endif
465
466# if SANITIZER_WINDOWS_x64
467 if (memcmp((u8*)address, kPrologueWithShortJump1,
468 sizeof(kPrologueWithShortJump1)) == 0 ||
469 memcmp((u8*)address, kPrologueWithShortJump2,
470 sizeof(kPrologueWithShortJump2)) == 0) {
471 return 0;
472 }
473#endif
474
475 switch (*(u64*)address) {
476 case 0x90909090909006EB: // stub: jmp over 6 x nop.
477 return 8;
478 }
479
480 switch (*(u8*)address) {
481 case 0x90: // 90 : nop
482 return 1;
483
484 case 0x50: // push eax / rax
485 case 0x51: // push ecx / rcx
486 case 0x52: // push edx / rdx
487 case 0x53: // push ebx / rbx
488 case 0x54: // push esp / rsp
489 case 0x55: // push ebp / rbp
490 case 0x56: // push esi / rsi
491 case 0x57: // push edi / rdi
492 case 0x5D: // pop ebp / rbp
493 return 1;
494
495 case 0x6A: // 6A XX = push XX
496 return 2;
497
498 case 0xb8: // b8 XX XX XX XX : mov eax, XX XX XX XX
499 case 0xB9: // b9 XX XX XX XX : mov ecx, XX XX XX XX
500 return 5;
501
502 // Cannot overwrite control-instruction. Return 0 to indicate failure.
503 case 0xE9: // E9 XX XX XX XX : jmp <label>
504 case 0xE8: // E8 XX XX XX XX : call <func>
505 case 0xC3: // C3 : ret
506 case 0xEB: // EB XX : jmp XX (short jump)
507 case 0x70: // 7Y YY : jy XX (short conditional jump)
508 case 0x71:
509 case 0x72:
510 case 0x73:
511 case 0x74:
512 case 0x75:
513 case 0x76:
514 case 0x77:
515 case 0x78:
516 case 0x79:
517 case 0x7A:
518 case 0x7B:
519 case 0x7C:
520 case 0x7D:
521 case 0x7E:
522 case 0x7F:
523 return 0;
524 }
525
526 switch (*(u16*)(address)) {
527 case 0x018A: // 8A 01 : mov al, byte ptr [ecx]
528 case 0xFF8B: // 8B FF : mov edi, edi
529 case 0xEC8B: // 8B EC : mov ebp, esp
530 case 0xc889: // 89 C8 : mov eax, ecx
531 case 0xE589: // 89 E5 : mov ebp, esp
532 case 0xC18B: // 8B C1 : mov eax, ecx
533 case 0xC033: // 33 C0 : xor eax, eax
534 case 0xC933: // 33 C9 : xor ecx, ecx
535 case 0xD233: // 33 D2 : xor edx, edx
536 return 2;
537
538 // Cannot overwrite control-instruction. Return 0 to indicate failure.
539 case 0x25FF: // FF 25 XX XX XX XX : jmp [XXXXXXXX]
540 return 0;
541 }
542
543 switch (0x00FFFFFF & *(u32*)address) {
544 case 0x24A48D: // 8D A4 24 XX XX XX XX : lea esp, [esp + XX XX XX XX]
545 return 7;
546 }
547
548# if SANITIZER_WINDOWS_x64
549 switch (*(u8*)address) {
550 case 0xA1: // A1 XX XX XX XX XX XX XX XX :
551 // movabs eax, dword ptr ds:[XXXXXXXX]
552 return 9;
553
554 case 0x83:
555 const u8 next_byte = *(u8*)(address + 1);
556 const u8 mod = next_byte >> 6;
557 const u8 rm = next_byte & 7;
558 if (mod == 1 && rm == 4)
559 return 5; // 83 ModR/M SIB Disp8 Imm8
560 // add|or|adc|sbb|and|sub|xor|cmp [r+disp8], imm8
561 }
562
563 switch (*(u16*)address) {
564 case 0x5040: // push rax
565 case 0x5140: // push rcx
566 case 0x5240: // push rdx
567 case 0x5340: // push rbx
568 case 0x5440: // push rsp
569 case 0x5540: // push rbp
570 case 0x5640: // push rsi
571 case 0x5740: // push rdi
572 case 0x5441: // push r12
573 case 0x5541: // push r13
574 case 0x5641: // push r14
575 case 0x5741: // push r15
576 case 0x9066: // Two-byte NOP
577 case 0xc084: // test al, al
578 case 0x018a: // mov al, byte ptr [rcx]
579 return 2;
580
581 case 0x058A: // 8A 05 XX XX XX XX : mov al, byte ptr [XX XX XX XX]
582 case 0x058B: // 8B 05 XX XX XX XX : mov eax, dword ptr [XX XX XX XX]
583 if (rel_offset)
584 *rel_offset = 2;
585 return 6;
586 }
587
588 switch (0x00FFFFFF & *(u32*)address) {
589 case 0xe58948: // 48 8b c4 : mov rbp, rsp
590 case 0xc18b48: // 48 8b c1 : mov rax, rcx
591 case 0xc48b48: // 48 8b c4 : mov rax, rsp
592 case 0xd9f748: // 48 f7 d9 : neg rcx
593 case 0xd12b48: // 48 2b d1 : sub rdx, rcx
594 case 0x07c1f6: // f6 c1 07 : test cl, 0x7
595 case 0xc98548: // 48 85 C9 : test rcx, rcx
596 case 0xd28548: // 48 85 d2 : test rdx, rdx
597 case 0xc0854d: // 4d 85 c0 : test r8, r8
598 case 0xc2b60f: // 0f b6 c2 : movzx eax, dl
599 case 0xc03345: // 45 33 c0 : xor r8d, r8d
600 case 0xc93345: // 45 33 c9 : xor r9d, r9d
601 case 0xdb3345: // 45 33 DB : xor r11d, r11d
602 case 0xd98b4c: // 4c 8b d9 : mov r11, rcx
603 case 0xd28b4c: // 4c 8b d2 : mov r10, rdx
604 case 0xc98b4c: // 4C 8B C9 : mov r9, rcx
605 case 0xc18b4c: // 4C 8B C1 : mov r8, rcx
606 case 0xd2b60f: // 0f b6 d2 : movzx edx, dl
607 case 0xca2b48: // 48 2b ca : sub rcx, rdx
608 case 0x10b70f: // 0f b7 10 : movzx edx, WORD PTR [rax]
609 case 0xc00b4d: // 3d 0b c0 : or r8, r8
610 case 0xc08b41: // 41 8b c0 : mov eax, r8d
611 case 0xd18b48: // 48 8b d1 : mov rdx, rcx
612 case 0xdc8b4c: // 4c 8b dc : mov r11, rsp
613 case 0xd18b4c: // 4c 8b d1 : mov r10, rcx
614 case 0xE0E483: // 83 E4 E0 : and esp, 0xFFFFFFE0
615 return 3;
616
617 case 0xec8348: // 48 83 ec XX : sub rsp, XX
618 case 0xf88349: // 49 83 f8 XX : cmp r8, XX
619 case 0x588948: // 48 89 58 XX : mov QWORD PTR[rax + XX], rbx
620 return 4;
621
622 case 0xec8148: // 48 81 EC XX XX XX XX : sub rsp, XXXXXXXX
623 return 7;
624
625 case 0x058b48: // 48 8b 05 XX XX XX XX :
626 // mov rax, QWORD PTR [rip + XXXXXXXX]
627 case 0x25ff48: // 48 ff 25 XX XX XX XX :
628 // rex.W jmp QWORD PTR [rip + XXXXXXXX]
629 case 0x158D4C: // 4c 8d 15 XX XX XX XX : lea r10, [rip + XX]
630 // Instructions having offset relative to 'rip' need offset adjustment.
631 if (rel_offset)
632 *rel_offset = 3;
633 return 7;
634
635 case 0x2444c7: // C7 44 24 XX YY YY YY YY
636 // mov dword ptr [rsp + XX], YYYYYYYY
637 return 8;
638 }
639
640 switch (*(u32*)(address)) {
641 case 0x24448b48: // 48 8b 44 24 XX : mov rax, QWORD ptr [rsp + XX]
642 case 0x246c8948: // 48 89 6C 24 XX : mov QWORD ptr [rsp + XX], rbp
643 case 0x245c8948: // 48 89 5c 24 XX : mov QWORD PTR [rsp + XX], rbx
644 case 0x24748948: // 48 89 74 24 XX : mov QWORD PTR [rsp + XX], rsi
645 case 0x247c8948: // 48 89 7c 24 XX : mov QWORD PTR [rsp + XX], rdi
646 case 0x244C8948: // 48 89 4C 24 XX : mov QWORD PTR [rsp + XX], rcx
647 case 0x24548948: // 48 89 54 24 XX : mov QWORD PTR [rsp + XX], rdx
648 case 0x244c894c: // 4c 89 4c 24 XX : mov QWORD PTR [rsp + XX], r9
649 case 0x2444894c: // 4c 89 44 24 XX : mov QWORD PTR [rsp + XX], r8
650 return 5;
651 case 0x24648348: // 48 83 64 24 XX : and QWORD PTR [rsp + XX], YY
652 return 6;
653 }
654
655#else
656
657 switch (*(u8*)address) {
658 case 0xA1: // A1 XX XX XX XX : mov eax, dword ptr ds:[XXXXXXXX]
659 return 5;
660 }
661 switch (*(u16*)address) {
662 case 0x458B: // 8B 45 XX : mov eax, dword ptr [ebp + XX]
663 case 0x5D8B: // 8B 5D XX : mov ebx, dword ptr [ebp + XX]
664 case 0x7D8B: // 8B 7D XX : mov edi, dword ptr [ebp + XX]
665 case 0xEC83: // 83 EC XX : sub esp, XX
666 case 0x75FF: // FF 75 XX : push dword ptr [ebp + XX]
667 return 3;
668 case 0xC1F7: // F7 C1 XX YY ZZ WW : test ecx, WWZZYYXX
669 case 0x25FF: // FF 25 XX YY ZZ WW : jmp dword ptr ds:[WWZZYYXX]
670 return 6;
671 case 0x3D83: // 83 3D XX YY ZZ WW TT : cmp TT, WWZZYYXX
672 return 7;
673 case 0x7D83: // 83 7D XX YY : cmp dword ptr [ebp + XX], YY
674 return 4;
675 }
676
677 switch (0x00FFFFFF & *(u32*)address) {
678 case 0x24448A: // 8A 44 24 XX : mov eal, dword ptr [esp + XX]
679 case 0x24448B: // 8B 44 24 XX : mov eax, dword ptr [esp + XX]
680 case 0x244C8B: // 8B 4C 24 XX : mov ecx, dword ptr [esp + XX]
681 case 0x24548B: // 8B 54 24 XX : mov edx, dword ptr [esp + XX]
682 case 0x245C8B: // 8B 5C 24 XX : mov ebx, dword ptr [esp + XX]
683 case 0x246C8B: // 8B 6C 24 XX : mov ebp, dword ptr [esp + XX]
684 case 0x24748B: // 8B 74 24 XX : mov esi, dword ptr [esp + XX]
685 case 0x247C8B: // 8B 7C 24 XX : mov edi, dword ptr [esp + XX]
686 return 4;
687 }
688
689 switch (*(u32*)address) {
690 case 0x2444B60F: // 0F B6 44 24 XX : movzx eax, byte ptr [esp + XX]
691 return 5;
692 }
693#endif
694
695 // Unknown instruction! This might happen when we add a new interceptor, use
696 // a new compiler version, or if Windows changed how some functions are
697 // compiled. In either case, we print the address and 8 bytes of instructions
698 // to notify the user about the error and to help identify the unknown
699 // instruction. Don't treat this as a fatal error, though we can break the
700 // debugger if one has been attached.
701 u8 *bytes = (u8 *)address;
702 ReportError(
703 "interception_win: unhandled instruction at %p: %02x %02x %02x %02x %02x "
704 "%02x %02x %02x\n",
705 (void *)address, bytes[0], bytes[1], bytes[2], bytes[3], bytes[4],
706 bytes[5], bytes[6], bytes[7]);
707 if (::IsDebuggerPresent())
708 __debugbreak();
709 return 0;
710}
711
712// Returns 0 on error.
713static size_t RoundUpToInstrBoundary(size_t size, uptr address) {
714 size_t cursor = 0;
715 while (cursor < size) {
716 size_t instruction_size = GetInstructionSize(address + cursor);
717 if (!instruction_size)
718 return 0;
719 cursor += instruction_size;
720 }
721 return cursor;
722}
723
724static bool CopyInstructions(uptr to, uptr from, size_t size) {
725 size_t cursor = 0;
726 while (cursor != size) {
727 size_t rel_offset = 0;
728 size_t instruction_size = GetInstructionSize(from + cursor, &rel_offset);
729 if (!instruction_size)
730 return false;
731 _memcpy((void *)(to + cursor), (void *)(from + cursor),
732 (size_t)instruction_size);
733 if (rel_offset) {
734# if SANITIZER_WINDOWS64
735 // we want to make sure that the new relative offset still fits in 32-bits
736 // this will be untrue if relocated_offset \notin [-2**31, 2**31)
737 s64 delta = to - from;
738 s64 relocated_offset = *(s32 *)(to + cursor + rel_offset) - delta;
739 if (-0x8000'0000ll > relocated_offset || relocated_offset > 0x7FFF'FFFFll)
740 return false;
741# else
742 // on 32-bit, the relative offset will always be correct
743 s32 delta = to - from;
744 s32 relocated_offset = *(s32 *)(to + cursor + rel_offset) - delta;
745# endif
746 *(s32 *)(to + cursor + rel_offset) = relocated_offset;
747 }
748 cursor += instruction_size;
749 }
750 return true;
751}
752
753
754#if !SANITIZER_WINDOWS64
755bool OverrideFunctionWithDetour(
756 uptr old_func, uptr new_func, uptr *orig_old_func) {
757 const int kDetourHeaderLen = 5;
758 const u16 kDetourInstruction = 0xFF8B;
759
760 uptr header = (uptr)old_func - kDetourHeaderLen;
761 uptr patch_length = kDetourHeaderLen + kShortJumpInstructionLength;
762
763 // Validate that the function is hookable.
764 if (*(u16*)old_func != kDetourInstruction ||
765 !IsMemoryPadding(header, kDetourHeaderLen))
766 return false;
767
768 // Change memory protection to writable.
769 DWORD protection = 0;
770 if (!ChangeMemoryProtection(header, patch_length, &protection))
771 return false;
772
773 // Write a relative jump to the redirected function.
774 WriteJumpInstruction(header, new_func);
775
776 // Write the short jump to the function prefix.
777 WriteShortJumpInstruction(old_func, header);
778
779 // Restore previous memory protection.
780 if (!RestoreMemoryProtection(header, patch_length, protection))
781 return false;
782
783 if (orig_old_func)
784 *orig_old_func = old_func + kShortJumpInstructionLength;
785
786 return true;
787}
788#endif
789
790bool OverrideFunctionWithRedirectJump(
791 uptr old_func, uptr new_func, uptr *orig_old_func) {
792 // Check whether the first instruction is a relative jump.
793 if (*(u8*)old_func != 0xE9)
794 return false;
795
796 if (orig_old_func) {
797 sptr relative_offset = *(s32 *)(old_func + 1);
798 uptr absolute_target = old_func + relative_offset + kJumpInstructionLength;
799 *orig_old_func = absolute_target;
800 }
801
802#if SANITIZER_WINDOWS64
803 // If needed, get memory space for a trampoline jump.
804 uptr trampoline = AllocateMemoryForTrampoline(old_func, kDirectBranchLength);
805 if (!trampoline)
806 return false;
807 WriteDirectBranch(trampoline, new_func);
808#endif
809
810 // Change memory protection to writable.
811 DWORD protection = 0;
812 if (!ChangeMemoryProtection(old_func, kJumpInstructionLength, &protection))
813 return false;
814
815 // Write a relative jump to the redirected function.
816 WriteJumpInstruction(old_func, FIRST_32_SECOND_64(new_func, trampoline));
817
818 // Restore previous memory protection.
819 if (!RestoreMemoryProtection(old_func, kJumpInstructionLength, protection))
820 return false;
821
822 return true;
823}
824
825bool OverrideFunctionWithHotPatch(
826 uptr old_func, uptr new_func, uptr *orig_old_func) {
827 const int kHotPatchHeaderLen = kBranchLength;
828
829 uptr header = (uptr)old_func - kHotPatchHeaderLen;
830 uptr patch_length = kHotPatchHeaderLen + kShortJumpInstructionLength;
831
832 // Validate that the function is hot patchable.
833 size_t instruction_size = GetInstructionSize(old_func);
834 if (instruction_size < kShortJumpInstructionLength ||
835 !FunctionHasPadding(old_func, kHotPatchHeaderLen))
836 return false;
837
838 if (orig_old_func) {
839 // Put the needed instructions into the trampoline bytes.
840 uptr trampoline_length = instruction_size + kDirectBranchLength;
841 uptr trampoline = AllocateMemoryForTrampoline(old_func, trampoline_length);
842 if (!trampoline)
843 return false;
844 if (!CopyInstructions(trampoline, old_func, instruction_size))
845 return false;
846 WriteDirectBranch(trampoline + instruction_size,
847 old_func + instruction_size);
848 *orig_old_func = trampoline;
849 }
850
851 // If needed, get memory space for indirect address.
852 uptr indirect_address = 0;
853#if SANITIZER_WINDOWS64
854 indirect_address = AllocateMemoryForTrampoline(old_func, kAddressLength);
855 if (!indirect_address)
856 return false;
857#endif
858
859 // Change memory protection to writable.
860 DWORD protection = 0;
861 if (!ChangeMemoryProtection(header, patch_length, &protection))
862 return false;
863
864 // Write jumps to the redirected function.
865 WriteBranch(header, indirect_address, new_func);
866 WriteShortJumpInstruction(old_func, header);
867
868 // Restore previous memory protection.
869 if (!RestoreMemoryProtection(header, patch_length, protection))
870 return false;
871
872 return true;
873}
874
875bool OverrideFunctionWithTrampoline(
876 uptr old_func, uptr new_func, uptr *orig_old_func) {
877
878 size_t instructions_length = kBranchLength;
879 size_t padding_length = 0;
880 uptr indirect_address = 0;
881
882 if (orig_old_func) {
883 // Find out the number of bytes of the instructions we need to copy
884 // to the trampoline.
885 instructions_length = RoundUpToInstrBoundary(kBranchLength, old_func);
886 if (!instructions_length)
887 return false;
888
889 // Put the needed instructions into the trampoline bytes.
890 uptr trampoline_length = instructions_length + kDirectBranchLength;
891 uptr trampoline = AllocateMemoryForTrampoline(old_func, trampoline_length);
892 if (!trampoline)
893 return false;
894 if (!CopyInstructions(trampoline, old_func, instructions_length))
895 return false;
896 WriteDirectBranch(trampoline + instructions_length,
897 old_func + instructions_length);
898 *orig_old_func = trampoline;
899 }
900
901#if SANITIZER_WINDOWS64
902 // Check if the targeted address can be encoded in the function padding.
903 // Otherwise, allocate it in the trampoline region.
904 if (IsMemoryPadding(old_func - kAddressLength, kAddressLength)) {
905 indirect_address = old_func - kAddressLength;
906 padding_length = kAddressLength;
907 } else {
908 indirect_address = AllocateMemoryForTrampoline(old_func, kAddressLength);
909 if (!indirect_address)
910 return false;
911 }
912#endif
913
914 // Change memory protection to writable.
915 uptr patch_address = old_func - padding_length;
916 uptr patch_length = instructions_length + padding_length;
917 DWORD protection = 0;
918 if (!ChangeMemoryProtection(patch_address, patch_length, &protection))
919 return false;
920
921 // Patch the original function.
922 WriteBranch(old_func, indirect_address, new_func);
923
924 // Restore previous memory protection.
925 if (!RestoreMemoryProtection(patch_address, patch_length, protection))
926 return false;
927
928 return true;
929}
930
931bool OverrideFunction(
932 uptr old_func, uptr new_func, uptr *orig_old_func) {
933#if !SANITIZER_WINDOWS64
934 if (OverrideFunctionWithDetour(old_func, new_func, orig_old_func))
935 return true;
936#endif
937 if (OverrideFunctionWithRedirectJump(old_func, new_func, orig_old_func))
938 return true;
939 if (OverrideFunctionWithHotPatch(old_func, new_func, orig_old_func))
940 return true;
941 if (OverrideFunctionWithTrampoline(old_func, new_func, orig_old_func))
942 return true;
943 return false;
944}
945
946static void **InterestingDLLsAvailable() {
947 static const char *InterestingDLLs[] = {
948 "kernel32.dll",
949 "msvcr100d.dll", // VS2010
950 "msvcr110d.dll", // VS2012
951 "msvcr120d.dll", // VS2013
952 "vcruntime140d.dll", // VS2015
953 "ucrtbased.dll", // Universal CRT
954 "msvcr100.dll", // VS2010
955 "msvcr110.dll", // VS2012
956 "msvcr120.dll", // VS2013
957 "vcruntime140.dll", // VS2015
958 "ucrtbase.dll", // Universal CRT
959# if (defined(__MINGW32__) && defined(__i386__))
960 "libc++.dll", // libc++
961 "libunwind.dll", // libunwind
962# endif
963 // NTDLL should go last as it exports some functions that we should
964 // override in the CRT [presumably only used internally].
965 "ntdll.dll",
966 NULL
967 };
968 static void *result[ARRAY_SIZE(InterestingDLLs)] = { 0 };
969 if (!result[0]) {
970 for (size_t i = 0, j = 0; InterestingDLLs[i]; ++i) {
971 if (HMODULE h = GetModuleHandleA(InterestingDLLs[i]))
972 result[j++] = (void *)h;
973 }
974 }
975 return &result[0];
976}
977
978namespace {
979// Utility for reading loaded PE images.
980template <typename T> class RVAPtr {
981 public:
982 RVAPtr(void *module, uptr rva)
983 : ptr_(reinterpret_cast<T *>(reinterpret_cast<char *>(module) + rva)) {}
984 operator T *() { return ptr_; }
985 T *operator->() { return ptr_; }
986 T *operator++() { return ++ptr_; }
987
988 private:
989 T *ptr_;
990};
991} // namespace
992
993// Internal implementation of GetProcAddress. At least since Windows 8,
994// GetProcAddress appears to initialize DLLs before returning function pointers
995// into them. This is problematic for the sanitizers, because they typically
996// want to intercept malloc *before* MSVCRT initializes. Our internal
997// implementation walks the export list manually without doing initialization.
998uptr InternalGetProcAddress(void *module, const char *func_name) {
999 // Check that the module header is full and present.
1000 RVAPtr<IMAGE_DOS_HEADER> dos_stub(module, 0);
1001 RVAPtr<IMAGE_NT_HEADERS> headers(module, dos_stub->e_lfanew);
1002 if (!module || dos_stub->e_magic != IMAGE_DOS_SIGNATURE || // "MZ"
1003 headers->Signature != IMAGE_NT_SIGNATURE || // "PE\0\0"
1004 headers->FileHeader.SizeOfOptionalHeader <
1005 sizeof(IMAGE_OPTIONAL_HEADER)) {
1006 return 0;
1007 }
1008
1009 IMAGE_DATA_DIRECTORY *export_directory =
1010 &headers->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_EXPORT];
1011 if (export_directory->Size == 0)
1012 return 0;
1013 RVAPtr<IMAGE_EXPORT_DIRECTORY> exports(module,
1014 export_directory->VirtualAddress);
1015 RVAPtr<DWORD> functions(module, exports->AddressOfFunctions);
1016 RVAPtr<DWORD> names(module, exports->AddressOfNames);
1017 RVAPtr<WORD> ordinals(module, exports->AddressOfNameOrdinals);
1018
1019 for (DWORD i = 0; i < exports->NumberOfNames; i++) {
1020 RVAPtr<char> name(module, names[i]);
1021 if (!strcmp(func_name, name)) {
1022 DWORD index = ordinals[i];
1023 RVAPtr<char> func(module, functions[index]);
1024
1025 // Handle forwarded functions.
1026 DWORD offset = functions[index];
1027 if (offset >= export_directory->VirtualAddress &&
1028 offset < export_directory->VirtualAddress + export_directory->Size) {
1029 // An entry for a forwarded function is a string with the following
1030 // format: "<module> . <function_name>" that is stored into the
1031 // exported directory.
1032 char function_name[256];
1033 size_t funtion_name_length = _strlen(func);
1034 if (funtion_name_length >= sizeof(function_name) - 1)
1035 InterceptionFailed();
1036
1037 _memcpy(function_name, func, funtion_name_length);
1038 function_name[funtion_name_length] = '\0';
1039 char* separator = _strchr(function_name, '.');
1040 if (!separator)
1041 InterceptionFailed();
1042 *separator = '\0';
1043
1044 void* redirected_module = GetModuleHandleA(function_name);
1045 if (!redirected_module)
1046 InterceptionFailed();
1047 return InternalGetProcAddress(redirected_module, separator + 1);
1048 }
1049
1050 return (uptr)(char *)func;
1051 }
1052 }
1053
1054 return 0;
1055}
1056
1057bool OverrideFunction(
1058 const char *func_name, uptr new_func, uptr *orig_old_func) {
1059 bool hooked = false;
1060 void **DLLs = InterestingDLLsAvailable();
1061 for (size_t i = 0; DLLs[i]; ++i) {
1062 uptr func_addr = InternalGetProcAddress(DLLs[i], func_name);
1063 if (func_addr &&
1064 OverrideFunction(func_addr, new_func, orig_old_func)) {
1065 hooked = true;
1066 }
1067 }
1068 return hooked;
1069}
1070
1071bool OverrideImportedFunction(const char *module_to_patch,
1072 const char *imported_module,
1073 const char *function_name, uptr new_function,
1074 uptr *orig_old_func) {
1075 HMODULE module = GetModuleHandleA(module_to_patch);
1076 if (!module)
1077 return false;
1078
1079 // Check that the module header is full and present.
1080 RVAPtr<IMAGE_DOS_HEADER> dos_stub(module, 0);
1081 RVAPtr<IMAGE_NT_HEADERS> headers(module, dos_stub->e_lfanew);
1082 if (!module || dos_stub->e_magic != IMAGE_DOS_SIGNATURE || // "MZ"
1083 headers->Signature != IMAGE_NT_SIGNATURE || // "PE\0\0"
1084 headers->FileHeader.SizeOfOptionalHeader <
1085 sizeof(IMAGE_OPTIONAL_HEADER)) {
1086 return false;
1087 }
1088
1089 IMAGE_DATA_DIRECTORY *import_directory =
1090 &headers->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_IMPORT];
1091
1092 // Iterate the list of imported DLLs. FirstThunk will be null for the last
1093 // entry.
1094 RVAPtr<IMAGE_IMPORT_DESCRIPTOR> imports(module,
1095 import_directory->VirtualAddress);
1096 for (; imports->FirstThunk != 0; ++imports) {
1097 RVAPtr<const char> modname(module, imports->Name);
1098 if (_stricmp(&*modname, imported_module) == 0)
1099 break;
1100 }
1101 if (imports->FirstThunk == 0)
1102 return false;
1103
1104 // We have two parallel arrays: the import address table (IAT) and the table
1105 // of names. They start out containing the same data, but the loader rewrites
1106 // the IAT to hold imported addresses and leaves the name table in
1107 // OriginalFirstThunk alone.
1108 RVAPtr<IMAGE_THUNK_DATA> name_table(module, imports->OriginalFirstThunk);
1109 RVAPtr<IMAGE_THUNK_DATA> iat(module, imports->FirstThunk);
1110 for (; name_table->u1.Ordinal != 0; ++name_table, ++iat) {
1111 if (!IMAGE_SNAP_BY_ORDINAL(name_table->u1.Ordinal)) {
1112 RVAPtr<IMAGE_IMPORT_BY_NAME> import_by_name(
1113 module, name_table->u1.ForwarderString);
1114 const char *funcname = &import_by_name->Name[0];
1115 if (strcmp(funcname, function_name) == 0)
1116 break;
1117 }
1118 }
1119 if (name_table->u1.Ordinal == 0)
1120 return false;
1121
1122 // Now we have the correct IAT entry. Do the swap. We have to make the page
1123 // read/write first.
1124 if (orig_old_func)
1125 *orig_old_func = iat->u1.AddressOfData;
1126 DWORD old_prot, unused_prot;
1127 if (!VirtualProtect(&iat->u1.AddressOfData, 4, PAGE_EXECUTE_READWRITE,
1128 &old_prot))
1129 return false;
1130 iat->u1.AddressOfData = new_function;
1131 if (!VirtualProtect(&iat->u1.AddressOfData, 4, old_prot, &unused_prot))
1132 return false; // Not clear if this failure bothers us.
1133 return true;
1134}
1135
1136} // namespace __interception
1137
1138#endif // SANITIZER_APPLE
1139

source code of compiler-rt/lib/interception/interception_win.cpp