1 //===-- interception_linux.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 equilavent 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 
134 namespace __interception {
135 
136 static const int kAddressLength = FIRST_32_SECOND_64(4, 8);
137 static const int kJumpInstructionLength = 5;
138 static const int kShortJumpInstructionLength = 2;
139 static const int kIndirectJumpInstructionLength = 6;
140 static const int kBranchLength =
141     FIRST_32_SECOND_64(kJumpInstructionLength, kIndirectJumpInstructionLength);
142 static const int kDirectBranchLength = kBranchLength + kAddressLength;
143 
InterceptionFailed()144 static void InterceptionFailed() {
145   // Do we have a good way to abort with an error message here?
146   __debugbreak();
147 }
148 
DistanceIsWithin2Gig(uptr from,uptr target)149 static bool DistanceIsWithin2Gig(uptr from, uptr target) {
150 #if SANITIZER_WINDOWS64
151   if (from < target)
152     return target - from <= (uptr)0x7FFFFFFFU;
153   else
154     return from - target <= (uptr)0x80000000U;
155 #else
156   // In a 32-bit address space, the address calculation will wrap, so this check
157   // is unnecessary.
158   return true;
159 #endif
160 }
161 
GetMmapGranularity()162 static uptr GetMmapGranularity() {
163   SYSTEM_INFO si;
164   GetSystemInfo(&si);
165   return si.dwAllocationGranularity;
166 }
167 
RoundUpTo(uptr size,uptr boundary)168 static uptr RoundUpTo(uptr size, uptr boundary) {
169   return (size + boundary - 1) & ~(boundary - 1);
170 }
171 
172 // FIXME: internal_str* and internal_mem* functions should be moved from the
173 // ASan sources into interception/.
174 
_strlen(const char * str)175 static size_t _strlen(const char *str) {
176   const char* p = str;
177   while (*p != '\0') ++p;
178   return p - str;
179 }
180 
_strchr(char * str,char c)181 static char* _strchr(char* str, char c) {
182   while (*str) {
183     if (*str == c)
184       return str;
185     ++str;
186   }
187   return nullptr;
188 }
189 
_memset(void * p,int value,size_t sz)190 static void _memset(void *p, int value, size_t sz) {
191   for (size_t i = 0; i < sz; ++i)
192     ((char*)p)[i] = (char)value;
193 }
194 
_memcpy(void * dst,void * src,size_t sz)195 static void _memcpy(void *dst, void *src, size_t sz) {
196   char *dst_c = (char*)dst,
197        *src_c = (char*)src;
198   for (size_t i = 0; i < sz; ++i)
199     dst_c[i] = src_c[i];
200 }
201 
ChangeMemoryProtection(uptr address,uptr size,DWORD * old_protection)202 static bool ChangeMemoryProtection(
203     uptr address, uptr size, DWORD *old_protection) {
204   return ::VirtualProtect((void*)address, size,
205                           PAGE_EXECUTE_READWRITE,
206                           old_protection) != FALSE;
207 }
208 
RestoreMemoryProtection(uptr address,uptr size,DWORD old_protection)209 static bool RestoreMemoryProtection(
210     uptr address, uptr size, DWORD old_protection) {
211   DWORD unused;
212   return ::VirtualProtect((void*)address, size,
213                           old_protection,
214                           &unused) != FALSE;
215 }
216 
IsMemoryPadding(uptr address,uptr size)217 static bool IsMemoryPadding(uptr address, uptr size) {
218   u8* function = (u8*)address;
219   for (size_t i = 0; i < size; ++i)
220     if (function[i] != 0x90 && function[i] != 0xCC)
221       return false;
222   return true;
223 }
224 
225 static const u8 kHintNop8Bytes[] = {
226   0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00
227 };
228 
229 template<class T>
FunctionHasPrefix(uptr address,const T & pattern)230 static bool FunctionHasPrefix(uptr address, const T &pattern) {
231   u8* function = (u8*)address - sizeof(pattern);
232   for (size_t i = 0; i < sizeof(pattern); ++i)
233     if (function[i] != pattern[i])
234       return false;
235   return true;
236 }
237 
FunctionHasPadding(uptr address,uptr size)238 static bool FunctionHasPadding(uptr address, uptr size) {
239   if (IsMemoryPadding(address - size, size))
240     return true;
241   if (size <= sizeof(kHintNop8Bytes) &&
242       FunctionHasPrefix(address, kHintNop8Bytes))
243     return true;
244   return false;
245 }
246 
WritePadding(uptr from,uptr size)247 static void WritePadding(uptr from, uptr size) {
248   _memset((void*)from, 0xCC, (size_t)size);
249 }
250 
WriteJumpInstruction(uptr from,uptr target)251 static void WriteJumpInstruction(uptr from, uptr target) {
252   if (!DistanceIsWithin2Gig(from + kJumpInstructionLength, target))
253     InterceptionFailed();
254   ptrdiff_t offset = target - from - kJumpInstructionLength;
255   *(u8*)from = 0xE9;
256   *(u32*)(from + 1) = offset;
257 }
258 
WriteShortJumpInstruction(uptr from,uptr target)259 static void WriteShortJumpInstruction(uptr from, uptr target) {
260   sptr offset = target - from - kShortJumpInstructionLength;
261   if (offset < -128 || offset > 127)
262     InterceptionFailed();
263   *(u8*)from = 0xEB;
264   *(u8*)(from + 1) = (u8)offset;
265 }
266 
267 #if SANITIZER_WINDOWS64
WriteIndirectJumpInstruction(uptr from,uptr indirect_target)268 static void WriteIndirectJumpInstruction(uptr from, uptr indirect_target) {
269   // jmp [rip + <offset>] = FF 25 <offset> where <offset> is a relative
270   // offset.
271   // The offset is the distance from then end of the jump instruction to the
272   // memory location containing the targeted address. The displacement is still
273   // 32-bit in x64, so indirect_target must be located within +/- 2GB range.
274   int offset = indirect_target - from - kIndirectJumpInstructionLength;
275   if (!DistanceIsWithin2Gig(from + kIndirectJumpInstructionLength,
276                             indirect_target)) {
277     InterceptionFailed();
278   }
279   *(u16*)from = 0x25FF;
280   *(u32*)(from + 2) = offset;
281 }
282 #endif
283 
WriteBranch(uptr from,uptr indirect_target,uptr target)284 static void WriteBranch(
285     uptr from, uptr indirect_target, uptr target) {
286 #if SANITIZER_WINDOWS64
287   WriteIndirectJumpInstruction(from, indirect_target);
288   *(u64*)indirect_target = target;
289 #else
290   (void)indirect_target;
291   WriteJumpInstruction(from, target);
292 #endif
293 }
294 
WriteDirectBranch(uptr from,uptr target)295 static void WriteDirectBranch(uptr from, uptr target) {
296 #if SANITIZER_WINDOWS64
297   // Emit an indirect jump through immediately following bytes:
298   //   jmp [rip + kBranchLength]
299   //   .quad <target>
300   WriteBranch(from, from + kBranchLength, target);
301 #else
302   WriteJumpInstruction(from, target);
303 #endif
304 }
305 
306 struct TrampolineMemoryRegion {
307   uptr content;
308   uptr allocated_size;
309   uptr max_size;
310 };
311 
312 static const uptr kTrampolineScanLimitRange = 1 << 31;  // 2 gig
313 static const int kMaxTrampolineRegion = 1024;
314 static TrampolineMemoryRegion TrampolineRegions[kMaxTrampolineRegion];
315 
AllocateTrampolineRegion(uptr image_address,size_t granularity)316 static void *AllocateTrampolineRegion(uptr image_address, size_t granularity) {
317 #if SANITIZER_WINDOWS64
318   uptr address = image_address;
319   uptr scanned = 0;
320   while (scanned < kTrampolineScanLimitRange) {
321     MEMORY_BASIC_INFORMATION info;
322     if (!::VirtualQuery((void*)address, &info, sizeof(info)))
323       return nullptr;
324 
325     // Check whether a region can be allocated at |address|.
326     if (info.State == MEM_FREE && info.RegionSize >= granularity) {
327       void *page = ::VirtualAlloc((void*)RoundUpTo(address, granularity),
328                                   granularity,
329                                   MEM_RESERVE | MEM_COMMIT,
330                                   PAGE_EXECUTE_READWRITE);
331       return page;
332     }
333 
334     // Move to the next region.
335     address = (uptr)info.BaseAddress + info.RegionSize;
336     scanned += info.RegionSize;
337   }
338   return nullptr;
339 #else
340   return ::VirtualAlloc(nullptr,
341                         granularity,
342                         MEM_RESERVE | MEM_COMMIT,
343                         PAGE_EXECUTE_READWRITE);
344 #endif
345 }
346 
347 // Used by unittests to release mapped memory space.
TestOnlyReleaseTrampolineRegions()348 void TestOnlyReleaseTrampolineRegions() {
349   for (size_t bucket = 0; bucket < kMaxTrampolineRegion; ++bucket) {
350     TrampolineMemoryRegion *current = &TrampolineRegions[bucket];
351     if (current->content == 0)
352       return;
353     ::VirtualFree((void*)current->content, 0, MEM_RELEASE);
354     current->content = 0;
355   }
356 }
357 
AllocateMemoryForTrampoline(uptr image_address,size_t size)358 static uptr AllocateMemoryForTrampoline(uptr image_address, size_t size) {
359   // Find a region within 2G with enough space to allocate |size| bytes.
360   TrampolineMemoryRegion *region = nullptr;
361   for (size_t bucket = 0; bucket < kMaxTrampolineRegion; ++bucket) {
362     TrampolineMemoryRegion* current = &TrampolineRegions[bucket];
363     if (current->content == 0) {
364       // No valid region found, allocate a new region.
365       size_t bucket_size = GetMmapGranularity();
366       void *content = AllocateTrampolineRegion(image_address, bucket_size);
367       if (content == nullptr)
368         return 0U;
369 
370       current->content = (uptr)content;
371       current->allocated_size = 0;
372       current->max_size = bucket_size;
373       region = current;
374       break;
375     } else if (current->max_size - current->allocated_size > size) {
376 #if SANITIZER_WINDOWS64
377         // In 64-bits, the memory space must be allocated within 2G boundary.
378         uptr next_address = current->content + current->allocated_size;
379         if (next_address < image_address ||
380             next_address - image_address >= 0x7FFF0000)
381           continue;
382 #endif
383       // The space can be allocated in the current region.
384       region = current;
385       break;
386     }
387   }
388 
389   // Failed to find a region.
390   if (region == nullptr)
391     return 0U;
392 
393   // Allocate the space in the current region.
394   uptr allocated_space = region->content + region->allocated_size;
395   region->allocated_size += size;
396   WritePadding(allocated_space, size);
397 
398   return allocated_space;
399 }
400 
401 // Returns 0 on error.
GetInstructionSize(uptr address,size_t * rel_offset=nullptr)402 static size_t GetInstructionSize(uptr address, size_t* rel_offset = nullptr) {
403   switch (*(u64*)address) {
404     case 0x90909090909006EB:  // stub: jmp over 6 x nop.
405       return 8;
406   }
407 
408   switch (*(u8*)address) {
409     case 0x90:  // 90 : nop
410       return 1;
411 
412     case 0x50:  // push eax / rax
413     case 0x51:  // push ecx / rcx
414     case 0x52:  // push edx / rdx
415     case 0x53:  // push ebx / rbx
416     case 0x54:  // push esp / rsp
417     case 0x55:  // push ebp / rbp
418     case 0x56:  // push esi / rsi
419     case 0x57:  // push edi / rdi
420     case 0x5D:  // pop ebp / rbp
421       return 1;
422 
423     case 0x6A:  // 6A XX = push XX
424       return 2;
425 
426     case 0xb8:  // b8 XX XX XX XX : mov eax, XX XX XX XX
427     case 0xB9:  // b9 XX XX XX XX : mov ecx, XX XX XX XX
428       return 5;
429 
430     // Cannot overwrite control-instruction. Return 0 to indicate failure.
431     case 0xE9:  // E9 XX XX XX XX : jmp <label>
432     case 0xE8:  // E8 XX XX XX XX : call <func>
433     case 0xC3:  // C3 : ret
434     case 0xEB:  // EB XX : jmp XX (short jump)
435     case 0x70:  // 7Y YY : jy XX (short conditional jump)
436     case 0x71:
437     case 0x72:
438     case 0x73:
439     case 0x74:
440     case 0x75:
441     case 0x76:
442     case 0x77:
443     case 0x78:
444     case 0x79:
445     case 0x7A:
446     case 0x7B:
447     case 0x7C:
448     case 0x7D:
449     case 0x7E:
450     case 0x7F:
451       return 0;
452   }
453 
454   switch (*(u16*)(address)) {
455     case 0x018A:  // 8A 01 : mov al, byte ptr [ecx]
456     case 0xFF8B:  // 8B FF : mov edi, edi
457     case 0xEC8B:  // 8B EC : mov ebp, esp
458     case 0xc889:  // 89 C8 : mov eax, ecx
459     case 0xC18B:  // 8B C1 : mov eax, ecx
460     case 0xC033:  // 33 C0 : xor eax, eax
461     case 0xC933:  // 33 C9 : xor ecx, ecx
462     case 0xD233:  // 33 D2 : xor edx, edx
463       return 2;
464 
465     // Cannot overwrite control-instruction. Return 0 to indicate failure.
466     case 0x25FF:  // FF 25 XX XX XX XX : jmp [XXXXXXXX]
467       return 0;
468   }
469 
470   switch (0x00FFFFFF & *(u32*)address) {
471     case 0x24A48D:  // 8D A4 24 XX XX XX XX : lea esp, [esp + XX XX XX XX]
472       return 7;
473   }
474 
475 #if SANITIZER_WINDOWS64
476   switch (*(u8*)address) {
477     case 0xA1:  // A1 XX XX XX XX XX XX XX XX :
478                 //   movabs eax, dword ptr ds:[XXXXXXXX]
479       return 9;
480   }
481 
482   switch (*(u16*)address) {
483     case 0x5040:  // push rax
484     case 0x5140:  // push rcx
485     case 0x5240:  // push rdx
486     case 0x5340:  // push rbx
487     case 0x5440:  // push rsp
488     case 0x5540:  // push rbp
489     case 0x5640:  // push rsi
490     case 0x5740:  // push rdi
491     case 0x5441:  // push r12
492     case 0x5541:  // push r13
493     case 0x5641:  // push r14
494     case 0x5741:  // push r15
495     case 0x9066:  // Two-byte NOP
496       return 2;
497 
498     case 0x058B:  // 8B 05 XX XX XX XX : mov eax, dword ptr [XX XX XX XX]
499       if (rel_offset)
500         *rel_offset = 2;
501       return 6;
502   }
503 
504   switch (0x00FFFFFF & *(u32*)address) {
505     case 0xe58948:    // 48 8b c4 : mov rbp, rsp
506     case 0xc18b48:    // 48 8b c1 : mov rax, rcx
507     case 0xc48b48:    // 48 8b c4 : mov rax, rsp
508     case 0xd9f748:    // 48 f7 d9 : neg rcx
509     case 0xd12b48:    // 48 2b d1 : sub rdx, rcx
510     case 0x07c1f6:    // f6 c1 07 : test cl, 0x7
511     case 0xc98548:    // 48 85 C9 : test rcx, rcx
512     case 0xc0854d:    // 4d 85 c0 : test r8, r8
513     case 0xc2b60f:    // 0f b6 c2 : movzx eax, dl
514     case 0xc03345:    // 45 33 c0 : xor r8d, r8d
515     case 0xc93345:    // 45 33 c9 : xor r9d, r9d
516     case 0xdb3345:    // 45 33 DB : xor r11d, r11d
517     case 0xd98b4c:    // 4c 8b d9 : mov r11, rcx
518     case 0xd28b4c:    // 4c 8b d2 : mov r10, rdx
519     case 0xc98b4c:    // 4C 8B C9 : mov r9, rcx
520     case 0xc18b4c:    // 4C 8B C1 : mov r8, rcx
521     case 0xd2b60f:    // 0f b6 d2 : movzx edx, dl
522     case 0xca2b48:    // 48 2b ca : sub rcx, rdx
523     case 0x10b70f:    // 0f b7 10 : movzx edx, WORD PTR [rax]
524     case 0xc00b4d:    // 3d 0b c0 : or r8, r8
525     case 0xd18b48:    // 48 8b d1 : mov rdx, rcx
526     case 0xdc8b4c:    // 4c 8b dc : mov r11, rsp
527     case 0xd18b4c:    // 4c 8b d1 : mov r10, rcx
528     case 0xE0E483:    // 83 E4 E0 : and esp, 0xFFFFFFE0
529       return 3;
530 
531     case 0xec8348:    // 48 83 ec XX : sub rsp, XX
532     case 0xf88349:    // 49 83 f8 XX : cmp r8, XX
533     case 0x588948:    // 48 89 58 XX : mov QWORD PTR[rax + XX], rbx
534       return 4;
535 
536     case 0xec8148:    // 48 81 EC XX XX XX XX : sub rsp, XXXXXXXX
537       return 7;
538 
539     case 0x058b48:    // 48 8b 05 XX XX XX XX :
540                       //   mov rax, QWORD PTR [rip + XXXXXXXX]
541     case 0x25ff48:    // 48 ff 25 XX XX XX XX :
542                       //   rex.W jmp QWORD PTR [rip + XXXXXXXX]
543 
544       // Instructions having offset relative to 'rip' need offset adjustment.
545       if (rel_offset)
546         *rel_offset = 3;
547       return 7;
548 
549     case 0x2444c7:    // C7 44 24 XX YY YY YY YY
550                       //   mov dword ptr [rsp + XX], YYYYYYYY
551       return 8;
552   }
553 
554   switch (*(u32*)(address)) {
555     case 0x24448b48:  // 48 8b 44 24 XX : mov rax, QWORD ptr [rsp + XX]
556     case 0x246c8948:  // 48 89 6C 24 XX : mov QWORD ptr [rsp + XX], rbp
557     case 0x245c8948:  // 48 89 5c 24 XX : mov QWORD PTR [rsp + XX], rbx
558     case 0x24748948:  // 48 89 74 24 XX : mov QWORD PTR [rsp + XX], rsi
559     case 0x244C8948:  // 48 89 4C 24 XX : mov QWORD PTR [rsp + XX], rcx
560     case 0x24548948:  // 48 89 54 24 XX : mov QWORD PTR [rsp + XX], rdx
561     case 0x244c894c:  // 4c 89 4c 24 XX : mov QWORD PTR [rsp + XX], r9
562     case 0x2444894c:  // 4c 89 44 24 XX : mov QWORD PTR [rsp + XX], r8
563       return 5;
564     case 0x24648348:  // 48 83 64 24 XX : and QWORD PTR [rsp + XX], YY
565       return 6;
566   }
567 
568 #else
569 
570   switch (*(u8*)address) {
571     case 0xA1:  // A1 XX XX XX XX :  mov eax, dword ptr ds:[XXXXXXXX]
572       return 5;
573   }
574   switch (*(u16*)address) {
575     case 0x458B:  // 8B 45 XX : mov eax, dword ptr [ebp + XX]
576     case 0x5D8B:  // 8B 5D XX : mov ebx, dword ptr [ebp + XX]
577     case 0x7D8B:  // 8B 7D XX : mov edi, dword ptr [ebp + XX]
578     case 0xEC83:  // 83 EC XX : sub esp, XX
579     case 0x75FF:  // FF 75 XX : push dword ptr [ebp + XX]
580       return 3;
581     case 0xC1F7:  // F7 C1 XX YY ZZ WW : test ecx, WWZZYYXX
582     case 0x25FF:  // FF 25 XX YY ZZ WW : jmp dword ptr ds:[WWZZYYXX]
583       return 6;
584     case 0x3D83:  // 83 3D XX YY ZZ WW TT : cmp TT, WWZZYYXX
585       return 7;
586     case 0x7D83:  // 83 7D XX YY : cmp dword ptr [ebp + XX], YY
587       return 4;
588   }
589 
590   switch (0x00FFFFFF & *(u32*)address) {
591     case 0x24448A:  // 8A 44 24 XX : mov eal, dword ptr [esp + XX]
592     case 0x24448B:  // 8B 44 24 XX : mov eax, dword ptr [esp + XX]
593     case 0x244C8B:  // 8B 4C 24 XX : mov ecx, dword ptr [esp + XX]
594     case 0x24548B:  // 8B 54 24 XX : mov edx, dword ptr [esp + XX]
595     case 0x24748B:  // 8B 74 24 XX : mov esi, dword ptr [esp + XX]
596     case 0x247C8B:  // 8B 7C 24 XX : mov edi, dword ptr [esp + XX]
597       return 4;
598   }
599 
600   switch (*(u32*)address) {
601     case 0x2444B60F:  // 0F B6 44 24 XX : movzx eax, byte ptr [esp + XX]
602       return 5;
603   }
604 #endif
605 
606   // Unknown instruction!
607   // FIXME: Unknown instruction failures might happen when we add a new
608   // interceptor or a new compiler version. In either case, they should result
609   // in visible and readable error messages. However, merely calling abort()
610   // leads to an infinite recursion in CheckFailed.
611   InterceptionFailed();
612   return 0;
613 }
614 
615 // Returns 0 on error.
RoundUpToInstrBoundary(size_t size,uptr address)616 static size_t RoundUpToInstrBoundary(size_t size, uptr address) {
617   size_t cursor = 0;
618   while (cursor < size) {
619     size_t instruction_size = GetInstructionSize(address + cursor);
620     if (!instruction_size)
621       return 0;
622     cursor += instruction_size;
623   }
624   return cursor;
625 }
626 
CopyInstructions(uptr to,uptr from,size_t size)627 static bool CopyInstructions(uptr to, uptr from, size_t size) {
628   size_t cursor = 0;
629   while (cursor != size) {
630     size_t rel_offset = 0;
631     size_t instruction_size = GetInstructionSize(from + cursor, &rel_offset);
632     _memcpy((void*)(to + cursor), (void*)(from + cursor),
633             (size_t)instruction_size);
634     if (rel_offset) {
635       uptr delta = to - from;
636       uptr relocated_offset = *(u32*)(to + cursor + rel_offset) - delta;
637 #if SANITIZER_WINDOWS64
638       if (relocated_offset + 0x80000000U >= 0xFFFFFFFFU)
639         return false;
640 #endif
641       *(u32*)(to + cursor + rel_offset) = relocated_offset;
642     }
643     cursor += instruction_size;
644   }
645   return true;
646 }
647 
648 
649 #if !SANITIZER_WINDOWS64
OverrideFunctionWithDetour(uptr old_func,uptr new_func,uptr * orig_old_func)650 bool OverrideFunctionWithDetour(
651     uptr old_func, uptr new_func, uptr *orig_old_func) {
652   const int kDetourHeaderLen = 5;
653   const u16 kDetourInstruction = 0xFF8B;
654 
655   uptr header = (uptr)old_func - kDetourHeaderLen;
656   uptr patch_length = kDetourHeaderLen + kShortJumpInstructionLength;
657 
658   // Validate that the function is hookable.
659   if (*(u16*)old_func != kDetourInstruction ||
660       !IsMemoryPadding(header, kDetourHeaderLen))
661     return false;
662 
663   // Change memory protection to writable.
664   DWORD protection = 0;
665   if (!ChangeMemoryProtection(header, patch_length, &protection))
666     return false;
667 
668   // Write a relative jump to the redirected function.
669   WriteJumpInstruction(header, new_func);
670 
671   // Write the short jump to the function prefix.
672   WriteShortJumpInstruction(old_func, header);
673 
674   // Restore previous memory protection.
675   if (!RestoreMemoryProtection(header, patch_length, protection))
676     return false;
677 
678   if (orig_old_func)
679     *orig_old_func = old_func + kShortJumpInstructionLength;
680 
681   return true;
682 }
683 #endif
684 
OverrideFunctionWithRedirectJump(uptr old_func,uptr new_func,uptr * orig_old_func)685 bool OverrideFunctionWithRedirectJump(
686     uptr old_func, uptr new_func, uptr *orig_old_func) {
687   // Check whether the first instruction is a relative jump.
688   if (*(u8*)old_func != 0xE9)
689     return false;
690 
691   if (orig_old_func) {
692     uptr relative_offset = *(u32*)(old_func + 1);
693     uptr absolute_target = old_func + relative_offset + kJumpInstructionLength;
694     *orig_old_func = absolute_target;
695   }
696 
697 #if SANITIZER_WINDOWS64
698   // If needed, get memory space for a trampoline jump.
699   uptr trampoline = AllocateMemoryForTrampoline(old_func, kDirectBranchLength);
700   if (!trampoline)
701     return false;
702   WriteDirectBranch(trampoline, new_func);
703 #endif
704 
705   // Change memory protection to writable.
706   DWORD protection = 0;
707   if (!ChangeMemoryProtection(old_func, kJumpInstructionLength, &protection))
708     return false;
709 
710   // Write a relative jump to the redirected function.
711   WriteJumpInstruction(old_func, FIRST_32_SECOND_64(new_func, trampoline));
712 
713   // Restore previous memory protection.
714   if (!RestoreMemoryProtection(old_func, kJumpInstructionLength, protection))
715     return false;
716 
717   return true;
718 }
719 
OverrideFunctionWithHotPatch(uptr old_func,uptr new_func,uptr * orig_old_func)720 bool OverrideFunctionWithHotPatch(
721     uptr old_func, uptr new_func, uptr *orig_old_func) {
722   const int kHotPatchHeaderLen = kBranchLength;
723 
724   uptr header = (uptr)old_func - kHotPatchHeaderLen;
725   uptr patch_length = kHotPatchHeaderLen + kShortJumpInstructionLength;
726 
727   // Validate that the function is hot patchable.
728   size_t instruction_size = GetInstructionSize(old_func);
729   if (instruction_size < kShortJumpInstructionLength ||
730       !FunctionHasPadding(old_func, kHotPatchHeaderLen))
731     return false;
732 
733   if (orig_old_func) {
734     // Put the needed instructions into the trampoline bytes.
735     uptr trampoline_length = instruction_size + kDirectBranchLength;
736     uptr trampoline = AllocateMemoryForTrampoline(old_func, trampoline_length);
737     if (!trampoline)
738       return false;
739     if (!CopyInstructions(trampoline, old_func, instruction_size))
740       return false;
741     WriteDirectBranch(trampoline + instruction_size,
742                       old_func + instruction_size);
743     *orig_old_func = trampoline;
744   }
745 
746   // If needed, get memory space for indirect address.
747   uptr indirect_address = 0;
748 #if SANITIZER_WINDOWS64
749   indirect_address = AllocateMemoryForTrampoline(old_func, kAddressLength);
750   if (!indirect_address)
751     return false;
752 #endif
753 
754   // Change memory protection to writable.
755   DWORD protection = 0;
756   if (!ChangeMemoryProtection(header, patch_length, &protection))
757     return false;
758 
759   // Write jumps to the redirected function.
760   WriteBranch(header, indirect_address, new_func);
761   WriteShortJumpInstruction(old_func, header);
762 
763   // Restore previous memory protection.
764   if (!RestoreMemoryProtection(header, patch_length, protection))
765     return false;
766 
767   return true;
768 }
769 
OverrideFunctionWithTrampoline(uptr old_func,uptr new_func,uptr * orig_old_func)770 bool OverrideFunctionWithTrampoline(
771     uptr old_func, uptr new_func, uptr *orig_old_func) {
772 
773   size_t instructions_length = kBranchLength;
774   size_t padding_length = 0;
775   uptr indirect_address = 0;
776 
777   if (orig_old_func) {
778     // Find out the number of bytes of the instructions we need to copy
779     // to the trampoline.
780     instructions_length = RoundUpToInstrBoundary(kBranchLength, old_func);
781     if (!instructions_length)
782       return false;
783 
784     // Put the needed instructions into the trampoline bytes.
785     uptr trampoline_length = instructions_length + kDirectBranchLength;
786     uptr trampoline = AllocateMemoryForTrampoline(old_func, trampoline_length);
787     if (!trampoline)
788       return false;
789     if (!CopyInstructions(trampoline, old_func, instructions_length))
790       return false;
791     WriteDirectBranch(trampoline + instructions_length,
792                       old_func + instructions_length);
793     *orig_old_func = trampoline;
794   }
795 
796 #if SANITIZER_WINDOWS64
797   // Check if the targeted address can be encoded in the function padding.
798   // Otherwise, allocate it in the trampoline region.
799   if (IsMemoryPadding(old_func - kAddressLength, kAddressLength)) {
800     indirect_address = old_func - kAddressLength;
801     padding_length = kAddressLength;
802   } else {
803     indirect_address = AllocateMemoryForTrampoline(old_func, kAddressLength);
804     if (!indirect_address)
805       return false;
806   }
807 #endif
808 
809   // Change memory protection to writable.
810   uptr patch_address = old_func - padding_length;
811   uptr patch_length = instructions_length + padding_length;
812   DWORD protection = 0;
813   if (!ChangeMemoryProtection(patch_address, patch_length, &protection))
814     return false;
815 
816   // Patch the original function.
817   WriteBranch(old_func, indirect_address, new_func);
818 
819   // Restore previous memory protection.
820   if (!RestoreMemoryProtection(patch_address, patch_length, protection))
821     return false;
822 
823   return true;
824 }
825 
OverrideFunction(uptr old_func,uptr new_func,uptr * orig_old_func)826 bool OverrideFunction(
827     uptr old_func, uptr new_func, uptr *orig_old_func) {
828 #if !SANITIZER_WINDOWS64
829   if (OverrideFunctionWithDetour(old_func, new_func, orig_old_func))
830     return true;
831 #endif
832   if (OverrideFunctionWithRedirectJump(old_func, new_func, orig_old_func))
833     return true;
834   if (OverrideFunctionWithHotPatch(old_func, new_func, orig_old_func))
835     return true;
836   if (OverrideFunctionWithTrampoline(old_func, new_func, orig_old_func))
837     return true;
838   return false;
839 }
840 
InterestingDLLsAvailable()841 static void **InterestingDLLsAvailable() {
842   static const char *InterestingDLLs[] = {
843       "kernel32.dll",
844       "msvcr100.dll",      // VS2010
845       "msvcr110.dll",      // VS2012
846       "msvcr120.dll",      // VS2013
847       "vcruntime140.dll",  // VS2015
848       "ucrtbase.dll",      // Universal CRT
849       // NTDLL should go last as it exports some functions that we should
850       // override in the CRT [presumably only used internally].
851       "ntdll.dll", NULL};
852   static void *result[ARRAY_SIZE(InterestingDLLs)] = { 0 };
853   if (!result[0]) {
854     for (size_t i = 0, j = 0; InterestingDLLs[i]; ++i) {
855       if (HMODULE h = GetModuleHandleA(InterestingDLLs[i]))
856         result[j++] = (void *)h;
857     }
858   }
859   return &result[0];
860 }
861 
862 namespace {
863 // Utility for reading loaded PE images.
864 template <typename T> class RVAPtr {
865  public:
RVAPtr(void * module,uptr rva)866   RVAPtr(void *module, uptr rva)
867       : ptr_(reinterpret_cast<T *>(reinterpret_cast<char *>(module) + rva)) {}
operator T*()868   operator T *() { return ptr_; }
operator ->()869   T *operator->() { return ptr_; }
operator ++()870   T *operator++() { return ++ptr_; }
871 
872  private:
873   T *ptr_;
874 };
875 } // namespace
876 
877 // Internal implementation of GetProcAddress. At least since Windows 8,
878 // GetProcAddress appears to initialize DLLs before returning function pointers
879 // into them. This is problematic for the sanitizers, because they typically
880 // want to intercept malloc *before* MSVCRT initializes. Our internal
881 // implementation walks the export list manually without doing initialization.
InternalGetProcAddress(void * module,const char * func_name)882 uptr InternalGetProcAddress(void *module, const char *func_name) {
883   // Check that the module header is full and present.
884   RVAPtr<IMAGE_DOS_HEADER> dos_stub(module, 0);
885   RVAPtr<IMAGE_NT_HEADERS> headers(module, dos_stub->e_lfanew);
886   if (!module || dos_stub->e_magic != IMAGE_DOS_SIGNATURE ||  // "MZ"
887       headers->Signature != IMAGE_NT_SIGNATURE ||             // "PE\0\0"
888       headers->FileHeader.SizeOfOptionalHeader <
889           sizeof(IMAGE_OPTIONAL_HEADER)) {
890     return 0;
891   }
892 
893   IMAGE_DATA_DIRECTORY *export_directory =
894       &headers->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_EXPORT];
895   if (export_directory->Size == 0)
896     return 0;
897   RVAPtr<IMAGE_EXPORT_DIRECTORY> exports(module,
898                                          export_directory->VirtualAddress);
899   RVAPtr<DWORD> functions(module, exports->AddressOfFunctions);
900   RVAPtr<DWORD> names(module, exports->AddressOfNames);
901   RVAPtr<WORD> ordinals(module, exports->AddressOfNameOrdinals);
902 
903   for (DWORD i = 0; i < exports->NumberOfNames; i++) {
904     RVAPtr<char> name(module, names[i]);
905     if (!strcmp(func_name, name)) {
906       DWORD index = ordinals[i];
907       RVAPtr<char> func(module, functions[index]);
908 
909       // Handle forwarded functions.
910       DWORD offset = functions[index];
911       if (offset >= export_directory->VirtualAddress &&
912           offset < export_directory->VirtualAddress + export_directory->Size) {
913         // An entry for a forwarded function is a string with the following
914         // format: "<module> . <function_name>" that is stored into the
915         // exported directory.
916         char function_name[256];
917         size_t funtion_name_length = _strlen(func);
918         if (funtion_name_length >= sizeof(function_name) - 1)
919           InterceptionFailed();
920 
921         _memcpy(function_name, func, funtion_name_length);
922         function_name[funtion_name_length] = '\0';
923         char* separator = _strchr(function_name, '.');
924         if (!separator)
925           InterceptionFailed();
926         *separator = '\0';
927 
928         void* redirected_module = GetModuleHandleA(function_name);
929         if (!redirected_module)
930           InterceptionFailed();
931         return InternalGetProcAddress(redirected_module, separator + 1);
932       }
933 
934       return (uptr)(char *)func;
935     }
936   }
937 
938   return 0;
939 }
940 
OverrideFunction(const char * func_name,uptr new_func,uptr * orig_old_func)941 bool OverrideFunction(
942     const char *func_name, uptr new_func, uptr *orig_old_func) {
943   bool hooked = false;
944   void **DLLs = InterestingDLLsAvailable();
945   for (size_t i = 0; DLLs[i]; ++i) {
946     uptr func_addr = InternalGetProcAddress(DLLs[i], func_name);
947     if (func_addr &&
948         OverrideFunction(func_addr, new_func, orig_old_func)) {
949       hooked = true;
950     }
951   }
952   return hooked;
953 }
954 
OverrideImportedFunction(const char * module_to_patch,const char * imported_module,const char * function_name,uptr new_function,uptr * orig_old_func)955 bool OverrideImportedFunction(const char *module_to_patch,
956                               const char *imported_module,
957                               const char *function_name, uptr new_function,
958                               uptr *orig_old_func) {
959   HMODULE module = GetModuleHandleA(module_to_patch);
960   if (!module)
961     return false;
962 
963   // Check that the module header is full and present.
964   RVAPtr<IMAGE_DOS_HEADER> dos_stub(module, 0);
965   RVAPtr<IMAGE_NT_HEADERS> headers(module, dos_stub->e_lfanew);
966   if (!module || dos_stub->e_magic != IMAGE_DOS_SIGNATURE ||  // "MZ"
967       headers->Signature != IMAGE_NT_SIGNATURE ||             // "PE\0\0"
968       headers->FileHeader.SizeOfOptionalHeader <
969           sizeof(IMAGE_OPTIONAL_HEADER)) {
970     return false;
971   }
972 
973   IMAGE_DATA_DIRECTORY *import_directory =
974       &headers->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_IMPORT];
975 
976   // Iterate the list of imported DLLs. FirstThunk will be null for the last
977   // entry.
978   RVAPtr<IMAGE_IMPORT_DESCRIPTOR> imports(module,
979                                           import_directory->VirtualAddress);
980   for (; imports->FirstThunk != 0; ++imports) {
981     RVAPtr<const char> modname(module, imports->Name);
982     if (_stricmp(&*modname, imported_module) == 0)
983       break;
984   }
985   if (imports->FirstThunk == 0)
986     return false;
987 
988   // We have two parallel arrays: the import address table (IAT) and the table
989   // of names. They start out containing the same data, but the loader rewrites
990   // the IAT to hold imported addresses and leaves the name table in
991   // OriginalFirstThunk alone.
992   RVAPtr<IMAGE_THUNK_DATA> name_table(module, imports->OriginalFirstThunk);
993   RVAPtr<IMAGE_THUNK_DATA> iat(module, imports->FirstThunk);
994   for (; name_table->u1.Ordinal != 0; ++name_table, ++iat) {
995     if (!IMAGE_SNAP_BY_ORDINAL(name_table->u1.Ordinal)) {
996       RVAPtr<IMAGE_IMPORT_BY_NAME> import_by_name(
997           module, name_table->u1.ForwarderString);
998       const char *funcname = &import_by_name->Name[0];
999       if (strcmp(funcname, function_name) == 0)
1000         break;
1001     }
1002   }
1003   if (name_table->u1.Ordinal == 0)
1004     return false;
1005 
1006   // Now we have the correct IAT entry. Do the swap. We have to make the page
1007   // read/write first.
1008   if (orig_old_func)
1009     *orig_old_func = iat->u1.AddressOfData;
1010   DWORD old_prot, unused_prot;
1011   if (!VirtualProtect(&iat->u1.AddressOfData, 4, PAGE_EXECUTE_READWRITE,
1012                       &old_prot))
1013     return false;
1014   iat->u1.AddressOfData = new_function;
1015   if (!VirtualProtect(&iat->u1.AddressOfData, 4, old_prot, &unused_prot))
1016     return false;  // Not clear if this failure bothers us.
1017   return true;
1018 }
1019 
1020 }  // namespace __interception
1021 
1022 #endif  // SANITIZER_MAC
1023