1 //===- FuzzerTraceState.cpp - Trace-based fuzzer mutator ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // This file implements a mutation algorithm based on instruction traces and
10 // on taint analysis feedback from DFSan.
12 // Instruction traces are special hooks inserted by the compiler around
13 // interesting instructions. Currently supported traces:
14 // * __sanitizer_cov_trace_cmp -- inserted before every ICMP instruction,
15 // receives the type, size and arguments of ICMP.
17 // Every time a traced event is intercepted we analyse the data involved
18 // in the event and suggest a mutation for future executions.
19 // For example if 4 bytes of data that derive from input bytes {4,5,6,7}
20 // are compared with a constant 12345,
21 // we try to insert 12345, 12344, 12346 into bytes
22 // {4,5,6,7} of the next fuzzed inputs.
24 // The fuzzer can work only with the traces, or with both traces and DFSan.
26 // DataFlowSanitizer (DFSan) is a tool for
27 // generalised dynamic data flow (taint) analysis:
28 // http://clang.llvm.org/docs/DataFlowSanitizer.html .
30 // The approach with DFSan-based fuzzing has some similarity to
31 // "Taint-based Directed Whitebox Fuzzing"
32 // by Vijay Ganesh & Tim Leek & Martin Rinard:
33 // http://dspace.mit.edu/openaccess-disseminate/1721.1/59320,
34 // but it uses a full blown LLVM IR taint analysis and separate instrumentation
35 // to analyze all of the "attack points" at once.
37 // Workflow with DFSan:
38 // * lib/Fuzzer/Fuzzer*.cpp is compiled w/o any instrumentation.
39 // * The code under test is compiled with DFSan *and* with instruction traces.
40 // * Every call to HOOK(a,b) is replaced by DFSan with
41 // __dfsw_HOOK(a, b, label(a), label(b)) so that __dfsw_HOOK
42 // gets all the taint labels for the arguments.
43 // * At the Fuzzer startup we assign a unique DFSan label
44 // to every byte of the input string (Fuzzer::CurrentUnit) so that for any
45 // chunk of data we know which input bytes it has derived from.
46 // * The __dfsw_* functions (implemented in this file) record the
47 // parameters (i.e. the application data and the corresponding taint labels)
49 // * Fuzzer::ApplyTraceBasedMutation() tries to use the data recorded
50 // by __dfsw_* hooks to guide the fuzzing towards new application states.
52 // Parts of this code will not function when DFSan is not linked in.
53 // Instead of using ifdefs and thus requiring a separate build of lib/Fuzzer
54 // we redeclare the dfsan_* interface functions as weak and check if they
55 // are nullptr before calling.
56 // If this approach proves to be useful we may add attribute(weak) to the
57 // dfsan declarations in dfsan_interface.h
59 // This module is in the "proof of concept" stage.
60 // It is capable of solving only the simplest puzzles
61 // like test/dfsan/DFSanSimpleCmpTest.cpp.
62 //===----------------------------------------------------------------------===//
64 /* Example of manual usage (-fsanitize=dataflow is optional):
67 clang -fPIC -c -g -O2 -std=c++11 Fuzzer*.cpp
68 clang++ -O0 -std=c++11 -fsanitize-coverage=edge,trace-cmp \
70 test/dfsan/DFSanSimpleCmpTest.cpp Fuzzer*.o
75 #include "FuzzerInternal.h"
76 #include <sanitizer/dfsan_interface.h>
80 #include <unordered_map>
84 dfsan_label dfsan_create_label(const char *desc, void *userdata);
86 void dfsan_set_label(dfsan_label label, void *addr, size_t size);
88 void dfsan_add_label(dfsan_label label, void *addr, size_t size);
90 const struct dfsan_label_info *dfsan_get_label_info(dfsan_label label);
92 dfsan_label dfsan_read_label(const void *addr, size_t size);
97 static bool ReallyHaveDFSan() {
98 return &dfsan_create_label != nullptr;
101 // These values are copied from include/llvm/IR/InstrTypes.h.
102 // We do not include the LLVM headers here to remain independent.
103 // If these values ever change, an assertion in ComputeCmp will fail.
105 ICMP_EQ = 32, ///< equal
106 ICMP_NE = 33, ///< not equal
107 ICMP_UGT = 34, ///< unsigned greater than
108 ICMP_UGE = 35, ///< unsigned greater or equal
109 ICMP_ULT = 36, ///< unsigned less than
110 ICMP_ULE = 37, ///< unsigned less or equal
111 ICMP_SGT = 38, ///< signed greater than
112 ICMP_SGE = 39, ///< signed greater or equal
113 ICMP_SLT = 40, ///< signed less than
114 ICMP_SLE = 41, ///< signed less or equal
117 template <class U, class S>
118 bool ComputeCmp(size_t CmpType, U Arg1, U Arg2) {
120 case ICMP_EQ : return Arg1 == Arg2;
121 case ICMP_NE : return Arg1 != Arg2;
122 case ICMP_UGT: return Arg1 > Arg2;
123 case ICMP_UGE: return Arg1 >= Arg2;
124 case ICMP_ULT: return Arg1 < Arg2;
125 case ICMP_ULE: return Arg1 <= Arg2;
126 case ICMP_SGT: return (S)Arg1 > (S)Arg2;
127 case ICMP_SGE: return (S)Arg1 >= (S)Arg2;
128 case ICMP_SLT: return (S)Arg1 < (S)Arg2;
129 case ICMP_SLE: return (S)Arg1 <= (S)Arg2;
130 default: assert(0 && "unsupported CmpType");
135 static bool ComputeCmp(size_t CmpSize, size_t CmpType, uint64_t Arg1,
137 if (CmpSize == 8) return ComputeCmp<uint64_t, int64_t>(CmpType, Arg1, Arg2);
138 if (CmpSize == 4) return ComputeCmp<uint32_t, int32_t>(CmpType, Arg1, Arg2);
139 if (CmpSize == 2) return ComputeCmp<uint16_t, int16_t>(CmpType, Arg1, Arg2);
140 if (CmpSize == 1) return ComputeCmp<uint8_t, int8_t>(CmpType, Arg1, Arg2);
142 if (CmpType == ICMP_EQ) return Arg1 == Arg2;
143 // assert(0 && "unsupported cmp and type size combination");
147 // As a simplification we use the range of input bytes instead of a set of input
150 uint16_t Beg, End; // Range is [Beg, End), thus Beg==End is an empty range.
152 LabelRange(uint16_t Beg = 0, uint16_t End = 0) : Beg(Beg), End(End) {}
154 static LabelRange Join(LabelRange LR1, LabelRange LR2) {
155 if (LR1.Beg == LR1.End) return LR2;
156 if (LR2.Beg == LR2.End) return LR1;
157 return {std::min(LR1.Beg, LR2.Beg), std::max(LR1.End, LR2.End)};
159 LabelRange &Join(LabelRange LR) {
160 return *this = Join(*this, LR);
162 static LabelRange Singleton(const dfsan_label_info *LI) {
163 uint16_t Idx = (uint16_t)(uintptr_t)LI->userdata;
165 return {(uint16_t)(Idx - 1), Idx};
169 // For now, very simple: put Size bytes of Data at position Pos.
170 struct TraceBasedMutation {
178 TraceState(const Fuzzer::FuzzingOptions &Options, const Unit &CurrentUnit)
179 : Options(Options), CurrentUnit(CurrentUnit) {}
181 LabelRange GetLabelRange(dfsan_label L);
182 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
183 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
185 void DFSanSwitchCallback(uint64_t PC, size_t ValSizeInBits, uint64_t Val,
186 size_t NumCases, uint64_t *Cases, dfsan_label L);
187 void TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
188 uint64_t Arg1, uint64_t Arg2);
190 void TraceSwitchCallback(uintptr_t PC, size_t ValSizeInBits, uint64_t Val,
191 size_t NumCases, uint64_t *Cases);
192 int TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
195 void StartTraceRecording() {
196 if (!Options.UseTraces) return;
197 RecordingTraces = true;
201 size_t StopTraceRecording(FuzzerRandomBase &Rand) {
202 RecordingTraces = false;
203 return Mutations.size();
206 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
209 bool IsTwoByteData(uint64_t Data) {
210 int64_t Signed = static_cast<int64_t>(Data);
212 return Signed == 0 || Signed == -1L;
214 bool RecordingTraces = false;
215 std::vector<TraceBasedMutation> Mutations;
216 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
217 const Fuzzer::FuzzingOptions &Options;
218 const Unit &CurrentUnit;
221 LabelRange TraceState::GetLabelRange(dfsan_label L) {
222 LabelRange &LR = LabelRanges[L];
223 if (LR.Beg < LR.End || L == 0)
225 const dfsan_label_info *LI = dfsan_get_label_info(L);
226 if (LI->l1 || LI->l2)
227 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
228 return LR = LabelRange::Singleton(LI);
231 void TraceState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
232 assert(Idx < Mutations.size());
233 auto &M = Mutations[Idx];
234 if (Options.Verbosity >= 3)
235 Printf("TBM %zd %zd %zd\n", M.Pos, M.Size, M.Data);
236 if (M.Pos + M.Size > U->size()) return;
237 memcpy(U->data() + M.Pos, &M.Data, M.Size);
240 void TraceState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
241 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
243 assert(ReallyHaveDFSan());
244 if (!RecordingTraces) return;
245 if (L1 == 0 && L2 == 0)
246 return; // Not actionable.
247 if (L1 != 0 && L2 != 0)
248 return; // Probably still actionable.
249 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
250 uint64_t Data = L1 ? Arg2 : Arg1;
251 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
253 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
254 Mutations.push_back({Pos, CmpSize, Data});
255 Mutations.push_back({Pos, CmpSize, Data + 1});
256 Mutations.push_back({Pos, CmpSize, Data - 1});
259 if (CmpSize > LR.End - LR.Beg)
260 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
263 if (Options.Verbosity >= 3)
264 Printf("DFSanCmpCallback: PC %lx S %zd T %zd A1 %llx A2 %llx R %d L1 %d L2 "
266 PC, CmpSize, CmpType, Arg1, Arg2, Res, L1, L2, Mutations.size());
269 void TraceState::DFSanSwitchCallback(uint64_t PC, size_t ValSizeInBits,
270 uint64_t Val, size_t NumCases,
271 uint64_t *Cases, dfsan_label L) {
272 assert(ReallyHaveDFSan());
273 if (!RecordingTraces) return;
274 if (!L) return; // Not actionable.
275 LabelRange LR = GetLabelRange(L);
276 size_t ValSize = ValSizeInBits / 8;
277 bool TryShort = IsTwoByteData(Val);
278 for (size_t i = 0; i < NumCases; i++)
279 TryShort &= IsTwoByteData(Cases[i]);
281 for (size_t Pos = LR.Beg; Pos + ValSize <= LR.End; Pos++)
282 for (size_t i = 0; i < NumCases; i++)
283 Mutations.push_back({Pos, ValSize, Cases[i]});
286 for (size_t Pos = LR.Beg; Pos + 2 <= LR.End; Pos++)
287 for (size_t i = 0; i < NumCases; i++)
288 Mutations.push_back({Pos, 2, Cases[i]});
290 if (Options.Verbosity >= 3)
291 Printf("DFSanSwitchCallback: PC %lx Val %zd SZ %zd # %zd L %d: {%d, %d} "
293 PC, Val, ValSize, NumCases, L, LR.Beg, LR.End, TryShort);
296 int TraceState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
299 const uint8_t *Beg = CurrentUnit.data();
300 const uint8_t *End = Beg + CurrentUnit.size();
301 for (const uint8_t *Cur = Beg; Cur < End; Cur++) {
302 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
305 size_t Pos = Cur - Beg;
306 assert(Pos < CurrentUnit.size());
307 if (Mutations.size() > 100000U) return Res; // Just in case.
308 Mutations.push_back({Pos, DataSize, DesiredData});
309 Mutations.push_back({Pos, DataSize, DesiredData + 1});
310 Mutations.push_back({Pos, DataSize, DesiredData - 1});
316 void TraceState::TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
317 uint64_t Arg1, uint64_t Arg2) {
318 if (!RecordingTraces) return;
320 if (Options.Verbosity >= 3)
321 Printf("TraceCmp %zd/%zd: %p %zd %zd\n", CmpSize, CmpType, PC, Arg1, Arg2);
322 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
323 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
324 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
325 Added += TryToAddDesiredData(Arg1, Arg2, 2);
326 Added += TryToAddDesiredData(Arg2, Arg1, 2);
330 void TraceState::TraceSwitchCallback(uintptr_t PC, size_t ValSizeInBits,
331 uint64_t Val, size_t NumCases,
333 if (!RecordingTraces) return;
334 size_t ValSize = ValSizeInBits / 8;
335 bool TryShort = IsTwoByteData(Val);
336 for (size_t i = 0; i < NumCases; i++)
337 TryShort &= IsTwoByteData(Cases[i]);
339 if (Options.Verbosity >= 3)
340 Printf("TraceSwitch: %p %zd # %zd; TryShort %d\n", PC, Val, NumCases,
343 for (size_t i = 0; i < NumCases; i++) {
344 TryToAddDesiredData(Val, Cases[i], ValSize);
346 TryToAddDesiredData(Val, Cases[i], 2);
351 static TraceState *TS;
353 void Fuzzer::StartTraceRecording() {
355 if (ReallyHaveDFSan())
356 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++)
357 dfsan_set_label(i + 1, &CurrentUnit[i], 1);
358 TS->StartTraceRecording();
361 size_t Fuzzer::StopTraceRecording() {
363 return TS->StopTraceRecording(USF.GetRand());
366 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
368 TS->ApplyTraceBasedMutation(Idx, U);
371 void Fuzzer::InitializeTraceState() {
372 if (!Options.UseTraces) return;
373 TS = new TraceState(Options, CurrentUnit);
374 CurrentUnit.resize(Options.MaxLen);
375 // The rest really requires DFSan.
376 if (!ReallyHaveDFSan()) return;
377 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
378 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
379 // We assume that no one else has called dfsan_create_label before.
381 Printf("DFSan labels are not starting from 1, exiting\n");
387 static size_t InternalStrnlen(const char *S, size_t MaxLen) {
389 for (; Len < MaxLen && S[Len]; Len++) {}
393 } // namespace fuzzer
398 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
399 uint64_t Arg2, dfsan_label L0,
400 dfsan_label L1, dfsan_label L2) {
403 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
404 uint64_t CmpSize = (SizeAndType >> 32) / 8;
405 uint64_t Type = (SizeAndType << 32) >> 32;
406 TS->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
409 void __dfsw___sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases,
410 dfsan_label L1, dfsan_label L2) {
412 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
413 TS->DFSanSwitchCallback(PC, Cases[1], Val, Cases[0], Cases+2, L1);
416 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
417 size_t n, dfsan_label s1_label,
418 dfsan_label s2_label, dfsan_label n_label) {
420 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
421 uint64_t S1 = 0, S2 = 0;
422 // Simplification: handle only first 8 bytes.
423 memcpy(&S1, s1, std::min(n, sizeof(S1)));
424 memcpy(&S2, s2, std::min(n, sizeof(S2)));
425 dfsan_label L1 = dfsan_read_label(s1, n);
426 dfsan_label L2 = dfsan_read_label(s2, n);
427 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
430 void dfsan_weak_hook_strncmp(void *caller_pc, const char *s1, const char *s2,
431 size_t n, dfsan_label s1_label,
432 dfsan_label s2_label, dfsan_label n_label) {
434 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
435 uint64_t S1 = 0, S2 = 0;
436 n = std::min(n, fuzzer::InternalStrnlen(s1, n));
437 n = std::min(n, fuzzer::InternalStrnlen(s2, n));
438 // Simplification: handle only first 8 bytes.
439 memcpy(&S1, s1, std::min(n, sizeof(S1)));
440 memcpy(&S2, s2, std::min(n, sizeof(S2)));
441 dfsan_label L1 = dfsan_read_label(s1, n);
442 dfsan_label L2 = dfsan_read_label(s2, n);
443 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
446 void dfsan_weak_hook_strcmp(void *caller_pc, const char *s1, const char *s2,
447 dfsan_label s1_label, dfsan_label s2_label) {
449 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
450 uint64_t S1 = 0, S2 = 0;
451 size_t Len1 = strlen(s1);
452 size_t Len2 = strlen(s2);
453 size_t N = std::min(Len1, Len2);
454 if (N <= 1) return; // Not interesting.
455 // Simplification: handle only first 8 bytes.
456 memcpy(&S1, s1, std::min(N, sizeof(S1)));
457 memcpy(&S2, s2, std::min(N, sizeof(S2)));
458 dfsan_label L1 = dfsan_read_label(s1, Len1);
459 dfsan_label L2 = dfsan_read_label(s2, Len2);
460 TS->DFSanCmpCallback(PC, N, fuzzer::ICMP_EQ, S1, S2, L1, L2);
463 void __sanitizer_weak_hook_memcmp(void *caller_pc, const void *s1,
464 const void *s2, size_t n) {
466 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
467 uint64_t S1 = 0, S2 = 0;
468 // Simplification: handle only first 8 bytes.
469 memcpy(&S1, s1, std::min(n, sizeof(S1)));
470 memcpy(&S2, s2, std::min(n, sizeof(S2)));
471 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
474 void __sanitizer_weak_hook_strncmp(void *caller_pc, const char *s1,
475 const char *s2, size_t n) {
477 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
478 uint64_t S1 = 0, S2 = 0;
479 size_t Len1 = fuzzer::InternalStrnlen(s1, n);
480 size_t Len2 = fuzzer::InternalStrnlen(s2, n);
481 n = std::min(n, Len1);
482 n = std::min(n, Len2);
483 if (n <= 1) return; // Not interesting.
484 // Simplification: handle only first 8 bytes.
485 memcpy(&S1, s1, std::min(n, sizeof(S1)));
486 memcpy(&S2, s2, std::min(n, sizeof(S2)));
487 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
490 void __sanitizer_weak_hook_strcmp(void *caller_pc, const char *s1,
493 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
494 uint64_t S1 = 0, S2 = 0;
495 size_t Len1 = strlen(s1);
496 size_t Len2 = strlen(s2);
497 size_t N = std::min(Len1, Len2);
498 if (N <= 1) return; // Not interesting.
499 // Simplification: handle only first 8 bytes.
500 memcpy(&S1, s1, std::min(N, sizeof(S1)));
501 memcpy(&S2, s2, std::min(N, sizeof(S2)));
502 TS->TraceCmpCallback(PC, N, fuzzer::ICMP_EQ, S1, S2);
506 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
509 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
510 uint64_t CmpSize = (SizeAndType >> 32) / 8;
511 uint64_t Type = (SizeAndType << 32) >> 32;
512 TS->TraceCmpCallback(PC, CmpSize, Type, Arg1, Arg2);
515 void __sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases) {
517 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
518 TS->TraceSwitchCallback(PC, Cases[1], Val, Cases[0], Cases + 2);