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 // A passport for a CMP site. We want to keep track of where the given CMP is
170 // and how many times it is evaluated to true or false.
171 struct CmpSitePassport {
175 bool IsInterestingCmpTarget() {
176 static const size_t kRareEnough = 50;
177 size_t C0 = Counter[0];
178 size_t C1 = Counter[1];
179 return C0 > kRareEnough * (C1 + 1) || C1 > kRareEnough * (C0 + 1);
183 // For now, just keep a simple imprecise hash table PC => CmpSitePassport.
184 // Potentially, will need to have a compiler support to have a precise mapping
185 // and also thread-safety.
186 struct CmpSitePassportTable {
187 static const size_t kSize = 99991; // Prime.
188 CmpSitePassport Passports[kSize];
190 CmpSitePassport *GetPassport(uintptr_t PC) {
191 uintptr_t Idx = PC & kSize;
192 CmpSitePassport *Res = &Passports[Idx];
193 if (Res->PC == 0) // Not thread safe.
195 return Res->PC == PC ? Res : nullptr;
199 static CmpSitePassportTable CSPTable; // Zero initialized.
201 // For now, very simple: put Size bytes of Data at position Pos.
202 struct TraceBasedMutation {
210 TraceState(const Fuzzer::FuzzingOptions &Options, const Unit &CurrentUnit)
211 : Options(Options), CurrentUnit(CurrentUnit) {}
213 LabelRange GetLabelRange(dfsan_label L);
214 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
215 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
217 void DFSanSwitchCallback(uint64_t PC, size_t ValSizeInBits, uint64_t Val,
218 size_t NumCases, uint64_t *Cases, dfsan_label L);
219 void TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType, uint64_t Arg1,
222 void TraceSwitchCallback(uintptr_t PC, size_t ValSizeInBits, uint64_t Val,
223 size_t NumCases, uint64_t *Cases);
224 int TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
227 void StartTraceRecording() {
228 if (!Options.UseTraces) return;
229 RecordingTraces = true;
233 size_t StopTraceRecording(FuzzerRandomBase &Rand) {
234 RecordingTraces = false;
235 return Mutations.size();
238 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
241 bool IsTwoByteData(uint64_t Data) {
242 int64_t Signed = static_cast<int64_t>(Data);
244 return Signed == 0 || Signed == -1L;
246 bool RecordingTraces = false;
247 std::vector<TraceBasedMutation> Mutations;
248 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
249 const Fuzzer::FuzzingOptions &Options;
250 const Unit &CurrentUnit;
253 LabelRange TraceState::GetLabelRange(dfsan_label L) {
254 LabelRange &LR = LabelRanges[L];
255 if (LR.Beg < LR.End || L == 0)
257 const dfsan_label_info *LI = dfsan_get_label_info(L);
258 if (LI->l1 || LI->l2)
259 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
260 return LR = LabelRange::Singleton(LI);
263 void TraceState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
264 assert(Idx < Mutations.size());
265 auto &M = Mutations[Idx];
266 if (Options.Verbosity >= 3)
267 Printf("TBM %zd %zd %zd\n", M.Pos, M.Size, M.Data);
268 if (M.Pos + M.Size > U->size()) return;
269 memcpy(U->data() + M.Pos, &M.Data, M.Size);
272 void TraceState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
273 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
275 assert(ReallyHaveDFSan());
276 if (!RecordingTraces) return;
277 if (L1 == 0 && L2 == 0)
278 return; // Not actionable.
279 if (L1 != 0 && L2 != 0)
280 return; // Probably still actionable.
281 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
282 uint64_t Data = L1 ? Arg2 : Arg1;
283 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
285 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
286 Mutations.push_back({Pos, CmpSize, Data});
287 Mutations.push_back({Pos, CmpSize, Data + 1});
288 Mutations.push_back({Pos, CmpSize, Data - 1});
291 if (CmpSize > LR.End - LR.Beg)
292 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
295 if (Options.Verbosity >= 3)
296 Printf("DFSanCmpCallback: PC %lx S %zd T %zd A1 %llx A2 %llx R %d L1 %d L2 "
298 PC, CmpSize, CmpType, Arg1, Arg2, Res, L1, L2, Mutations.size());
301 void TraceState::DFSanSwitchCallback(uint64_t PC, size_t ValSizeInBits,
302 uint64_t Val, size_t NumCases,
303 uint64_t *Cases, dfsan_label L) {
304 assert(ReallyHaveDFSan());
305 if (!RecordingTraces) return;
306 if (!L) return; // Not actionable.
307 LabelRange LR = GetLabelRange(L);
308 size_t ValSize = ValSizeInBits / 8;
309 bool TryShort = IsTwoByteData(Val);
310 for (size_t i = 0; i < NumCases; i++)
311 TryShort &= IsTwoByteData(Cases[i]);
313 for (size_t Pos = LR.Beg; Pos + ValSize <= LR.End; Pos++)
314 for (size_t i = 0; i < NumCases; i++)
315 Mutations.push_back({Pos, ValSize, Cases[i]});
318 for (size_t Pos = LR.Beg; Pos + 2 <= LR.End; Pos++)
319 for (size_t i = 0; i < NumCases; i++)
320 Mutations.push_back({Pos, 2, Cases[i]});
322 if (Options.Verbosity >= 3)
323 Printf("DFSanSwitchCallback: PC %lx Val %zd SZ %zd # %zd L %d: {%d, %d} "
325 PC, Val, ValSize, NumCases, L, LR.Beg, LR.End, TryShort);
328 int TraceState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
331 const uint8_t *Beg = CurrentUnit.data();
332 const uint8_t *End = Beg + CurrentUnit.size();
333 for (const uint8_t *Cur = Beg; Cur < End; Cur += DataSize) {
334 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
337 size_t Pos = Cur - Beg;
338 assert(Pos < CurrentUnit.size());
339 if (Mutations.size() > 100000U) return Res; // Just in case.
340 Mutations.push_back({Pos, DataSize, DesiredData});
341 Mutations.push_back({Pos, DataSize, DesiredData + 1});
342 Mutations.push_back({Pos, DataSize, DesiredData - 1});
349 void TraceState::TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType, uint64_t Arg1,
351 if (!RecordingTraces) return;
353 CmpSitePassport *CSP = CSPTable.GetPassport(PC);
355 CSP->Counter[ComputeCmp(CmpSize, CmpType, Arg1, Arg2)]++;
356 size_t C0 = CSP->Counter[0];
357 size_t C1 = CSP->Counter[1];
358 // FIXME: is this a good idea or a bad?
359 // if (!CSP->IsInterestingCmpTarget())
361 if (Options.Verbosity >= 3)
362 Printf("TraceCmp: %p %zd/%zd; %zd %zd\n", CSP->PC, C0, C1, Arg1, Arg2);
363 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
364 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
365 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
366 Added += TryToAddDesiredData(Arg1, Arg2, 2);
367 Added += TryToAddDesiredData(Arg2, Arg1, 2);
371 void TraceState::TraceSwitchCallback(uintptr_t PC, size_t ValSizeInBits,
372 uint64_t Val, size_t NumCases,
374 if (!RecordingTraces) return;
375 size_t ValSize = ValSizeInBits / 8;
376 bool TryShort = IsTwoByteData(Val);
377 for (size_t i = 0; i < NumCases; i++)
378 TryShort &= IsTwoByteData(Cases[i]);
380 if (Options.Verbosity >= 3)
381 Printf("TraceSwitch: %p %zd # %zd; TryShort %d\n", PC, Val, NumCases,
384 for (size_t i = 0; i < NumCases; i++) {
385 TryToAddDesiredData(Val, Cases[i], ValSize);
387 TryToAddDesiredData(Val, Cases[i], 2);
392 static TraceState *TS;
394 void Fuzzer::StartTraceRecording() {
396 if (ReallyHaveDFSan())
397 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++)
398 dfsan_set_label(i + 1, &CurrentUnit[i], 1);
399 TS->StartTraceRecording();
402 size_t Fuzzer::StopTraceRecording() {
404 return TS->StopTraceRecording(USF.GetRand());
407 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
409 TS->ApplyTraceBasedMutation(Idx, U);
412 void Fuzzer::InitializeTraceState() {
413 if (!Options.UseTraces) return;
414 TS = new TraceState(Options, CurrentUnit);
415 CurrentUnit.resize(Options.MaxLen);
416 // The rest really requires DFSan.
417 if (!ReallyHaveDFSan()) return;
418 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
419 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
420 // We assume that no one else has called dfsan_create_label before.
422 Printf("DFSan labels are not starting from 1, exiting\n");
428 static size_t InternalStrnlen(const char *S, size_t MaxLen) {
430 for (; Len < MaxLen && S[Len]; Len++) {}
434 } // namespace fuzzer
439 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
440 uint64_t Arg2, dfsan_label L0,
441 dfsan_label L1, dfsan_label L2) {
444 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
445 uint64_t CmpSize = (SizeAndType >> 32) / 8;
446 uint64_t Type = (SizeAndType << 32) >> 32;
447 TS->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
450 void __dfsw___sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases,
451 dfsan_label L1, dfsan_label L2) {
453 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
454 TS->DFSanSwitchCallback(PC, Cases[1], Val, Cases[0], Cases+2, L1);
457 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
458 size_t n, dfsan_label s1_label,
459 dfsan_label s2_label, dfsan_label n_label) {
461 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
462 uint64_t S1 = 0, S2 = 0;
463 // Simplification: handle only first 8 bytes.
464 memcpy(&S1, s1, std::min(n, sizeof(S1)));
465 memcpy(&S2, s2, std::min(n, sizeof(S2)));
466 dfsan_label L1 = dfsan_read_label(s1, n);
467 dfsan_label L2 = dfsan_read_label(s2, n);
468 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
471 void dfsan_weak_hook_strncmp(void *caller_pc, const char *s1, const char *s2,
472 size_t n, dfsan_label s1_label,
473 dfsan_label s2_label, dfsan_label n_label) {
475 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
476 uint64_t S1 = 0, S2 = 0;
477 n = std::min(n, fuzzer::InternalStrnlen(s1, n));
478 n = std::min(n, fuzzer::InternalStrnlen(s2, n));
479 // Simplification: handle only first 8 bytes.
480 memcpy(&S1, s1, std::min(n, sizeof(S1)));
481 memcpy(&S2, s2, std::min(n, sizeof(S2)));
482 dfsan_label L1 = dfsan_read_label(s1, n);
483 dfsan_label L2 = dfsan_read_label(s2, n);
484 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
487 void dfsan_weak_hook_strcmp(void *caller_pc, const char *s1, const char *s2,
488 dfsan_label s1_label, dfsan_label s2_label) {
490 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
491 uint64_t S1 = 0, S2 = 0;
492 size_t Len1 = strlen(s1);
493 size_t Len2 = strlen(s2);
494 size_t N = std::min(Len1, Len2);
495 if (N <= 1) return; // Not interesting.
496 // Simplification: handle only first 8 bytes.
497 memcpy(&S1, s1, std::min(N, sizeof(S1)));
498 memcpy(&S2, s2, std::min(N, sizeof(S2)));
499 dfsan_label L1 = dfsan_read_label(s1, Len1);
500 dfsan_label L2 = dfsan_read_label(s2, Len2);
501 TS->DFSanCmpCallback(PC, N, fuzzer::ICMP_EQ, S1, S2, L1, L2);
504 void __sanitizer_weak_hook_memcmp(void *caller_pc, const void *s1,
505 const void *s2, size_t n) {
507 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
508 uint64_t S1 = 0, S2 = 0;
509 // Simplification: handle only first 8 bytes.
510 memcpy(&S1, s1, std::min(n, sizeof(S1)));
511 memcpy(&S2, s2, std::min(n, sizeof(S2)));
512 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
515 void __sanitizer_weak_hook_strncmp(void *caller_pc, const char *s1,
516 const char *s2, size_t n) {
518 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
519 uint64_t S1 = 0, S2 = 0;
520 size_t Len1 = fuzzer::InternalStrnlen(s1, n);
521 size_t Len2 = fuzzer::InternalStrnlen(s2, n);
522 n = std::min(n, Len1);
523 n = std::min(n, Len2);
524 if (n <= 1) return; // Not interesting.
525 // Simplification: handle only first 8 bytes.
526 memcpy(&S1, s1, std::min(n, sizeof(S1)));
527 memcpy(&S2, s2, std::min(n, sizeof(S2)));
528 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
531 void __sanitizer_weak_hook_strcmp(void *caller_pc, const char *s1,
534 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
535 uint64_t S1 = 0, S2 = 0;
536 size_t Len1 = strlen(s1);
537 size_t Len2 = strlen(s2);
538 size_t N = std::min(Len1, Len2);
539 if (N <= 1) return; // Not interesting.
540 // Simplification: handle only first 8 bytes.
541 memcpy(&S1, s1, std::min(N, sizeof(S1)));
542 memcpy(&S2, s2, std::min(N, sizeof(S2)));
543 TS->TraceCmpCallback(PC, N, fuzzer::ICMP_EQ, S1, S2);
547 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
550 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
551 uint64_t CmpSize = (SizeAndType >> 32) / 8;
552 uint64_t Type = (SizeAndType << 32) >> 32;
553 TS->TraceCmpCallback(PC, CmpSize, Type, Arg1, Arg2);
556 void __sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases) {
558 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
559 TS->TraceSwitchCallback(PC, Cases[1], Val, Cases[0], Cases + 2);