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 std::random_shuffle(Mutations.begin(), Mutations.end(), Rand);
236 return Mutations.size();
239 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
242 bool IsTwoByteData(uint64_t Data) {
243 int64_t Signed = static_cast<int64_t>(Data);
245 return Signed == 0 || Signed == -1L;
247 bool RecordingTraces = false;
248 std::vector<TraceBasedMutation> Mutations;
249 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
250 const Fuzzer::FuzzingOptions &Options;
251 const Unit &CurrentUnit;
254 LabelRange TraceState::GetLabelRange(dfsan_label L) {
255 LabelRange &LR = LabelRanges[L];
256 if (LR.Beg < LR.End || L == 0)
258 const dfsan_label_info *LI = dfsan_get_label_info(L);
259 if (LI->l1 || LI->l2)
260 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
261 return LR = LabelRange::Singleton(LI);
264 void TraceState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
265 assert(Idx < Mutations.size());
266 auto &M = Mutations[Idx];
267 if (Options.Verbosity >= 3)
268 Printf("TBM %zd %zd %zd\n", M.Pos, M.Size, M.Data);
269 if (M.Pos + M.Size > U->size()) return;
270 memcpy(U->data() + M.Pos, &M.Data, M.Size);
273 void TraceState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
274 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
276 assert(ReallyHaveDFSan());
277 if (!RecordingTraces) return;
278 if (L1 == 0 && L2 == 0)
279 return; // Not actionable.
280 if (L1 != 0 && L2 != 0)
281 return; // Probably still actionable.
282 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
283 uint64_t Data = L1 ? Arg2 : Arg1;
284 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
286 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
287 Mutations.push_back({Pos, CmpSize, Data});
288 Mutations.push_back({Pos, CmpSize, Data + 1});
289 Mutations.push_back({Pos, CmpSize, Data - 1});
292 if (CmpSize > LR.End - LR.Beg)
293 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
296 if (Options.Verbosity >= 3)
297 Printf("DFSanCmpCallback: PC %lx S %zd T %zd A1 %llx A2 %llx R %d L1 %d L2 "
299 PC, CmpSize, CmpType, Arg1, Arg2, Res, L1, L2, Mutations.size());
302 void TraceState::DFSanSwitchCallback(uint64_t PC, size_t ValSizeInBits,
303 uint64_t Val, size_t NumCases,
304 uint64_t *Cases, dfsan_label L) {
305 assert(ReallyHaveDFSan());
306 if (!RecordingTraces) return;
307 if (!L) return; // Not actionable.
308 LabelRange LR = GetLabelRange(L);
309 size_t ValSize = ValSizeInBits / 8;
310 for (size_t Pos = LR.Beg; Pos + ValSize <= LR.End; Pos++) {
311 for (size_t i = 0; i < NumCases; i++) {
312 Mutations.push_back({Pos, ValSize, Cases[i]});
313 Mutations.push_back({Pos, ValSize, Cases[i] + 1});
314 Mutations.push_back({Pos, ValSize, Cases[i] - 1});
317 if (Options.Verbosity >= 3)
318 Printf("DFSanSwitchCallback: PC %lx Val %zd # %zd L %d\n", PC, Val,
322 int TraceState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
325 const uint8_t *Beg = CurrentUnit.data();
326 const uint8_t *End = Beg + CurrentUnit.size();
327 for (const uint8_t *Cur = Beg; Cur < End; Cur += DataSize) {
328 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
331 size_t Pos = Cur - Beg;
332 assert(Pos < CurrentUnit.size());
333 if (Mutations.size() > 100000U) return Res; // Just in case.
334 Mutations.push_back({Pos, DataSize, DesiredData});
335 Mutations.push_back({Pos, DataSize, DesiredData + 1});
336 Mutations.push_back({Pos, DataSize, DesiredData - 1});
343 void TraceState::TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType, uint64_t Arg1,
345 if (!RecordingTraces) return;
347 CmpSitePassport *CSP = CSPTable.GetPassport(PC);
349 CSP->Counter[ComputeCmp(CmpSize, CmpType, Arg1, Arg2)]++;
350 size_t C0 = CSP->Counter[0];
351 size_t C1 = CSP->Counter[1];
352 if (!CSP->IsInterestingCmpTarget())
354 if (Options.Verbosity >= 3)
355 Printf("TraceCmp: %p %zd/%zd; %zd %zd\n", CSP->PC, C0, C1, Arg1, Arg2);
356 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
357 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
358 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
359 Added += TryToAddDesiredData(Arg1, Arg2, 2);
360 Added += TryToAddDesiredData(Arg2, Arg1, 2);
364 void TraceState::TraceSwitchCallback(uintptr_t PC, size_t ValSizeInBits,
365 uint64_t Val, size_t NumCases,
367 if (!RecordingTraces) return;
368 for (size_t i = 0; i < NumCases; i++)
369 TryToAddDesiredData(Val, Cases[i], ValSizeInBits / 8);
372 static TraceState *TS;
374 void Fuzzer::StartTraceRecording() {
376 TS->StartTraceRecording();
379 size_t Fuzzer::StopTraceRecording() {
381 return TS->StopTraceRecording(USF.GetRand());
384 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
386 TS->ApplyTraceBasedMutation(Idx, U);
389 void Fuzzer::InitializeTraceState() {
390 if (!Options.UseTraces) return;
391 TS = new TraceState(Options, CurrentUnit);
392 CurrentUnit.resize(Options.MaxLen);
393 // The rest really requires DFSan.
394 if (!ReallyHaveDFSan()) return;
395 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
396 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
397 // We assume that no one else has called dfsan_create_label before.
399 dfsan_set_label(L, &CurrentUnit[i], 1);
403 static size_t InternalStrnlen(const char *S, size_t MaxLen) {
405 for (; Len < MaxLen && S[Len]; Len++) {}
409 } // namespace fuzzer
414 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
415 uint64_t Arg2, dfsan_label L0,
416 dfsan_label L1, dfsan_label L2) {
419 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
420 uint64_t CmpSize = (SizeAndType >> 32) / 8;
421 uint64_t Type = (SizeAndType << 32) >> 32;
422 TS->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
425 void __dfsw___sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases,
426 dfsan_label L1, dfsan_label L2) {
428 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
429 TS->DFSanSwitchCallback(PC, Cases[1], Val, Cases[0], Cases+2, L1);
432 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
433 size_t n, dfsan_label s1_label,
434 dfsan_label s2_label, dfsan_label n_label) {
436 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
437 uint64_t S1 = 0, S2 = 0;
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_strncmp(void *caller_pc, const char *s1, const char *s2,
447 size_t n, dfsan_label s1_label,
448 dfsan_label s2_label, dfsan_label n_label) {
450 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
451 uint64_t S1 = 0, S2 = 0;
452 n = std::min(n, fuzzer::InternalStrnlen(s1, n));
453 n = std::min(n, fuzzer::InternalStrnlen(s2, n));
454 // Simplification: handle only first 8 bytes.
455 memcpy(&S1, s1, std::min(n, sizeof(S1)));
456 memcpy(&S2, s2, std::min(n, sizeof(S2)));
457 dfsan_label L1 = dfsan_read_label(s1, n);
458 dfsan_label L2 = dfsan_read_label(s2, n);
459 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
462 void __sanitizer_weak_hook_memcmp(void *caller_pc, const void *s1,
463 const void *s2, size_t n) {
465 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
466 uint64_t S1 = 0, S2 = 0;
467 // Simplification: handle only first 8 bytes.
468 memcpy(&S1, s1, std::min(n, sizeof(S1)));
469 memcpy(&S2, s2, std::min(n, sizeof(S2)));
470 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
473 void __sanitizer_weak_hook_strncmp(void *caller_pc, const char *s1,
474 const char *s2, size_t n) {
476 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
477 uint64_t S1 = 0, S2 = 0;
478 size_t Len1 = fuzzer::InternalStrnlen(s1, n);
479 size_t Len2 = fuzzer::InternalStrnlen(s2, n);
480 n = std::min(n, Len1);
481 n = std::min(n, Len2);
482 if (n <= 1) return; // Not interesting.
483 // Simplification: handle only first 8 bytes.
484 memcpy(&S1, s1, std::min(n, sizeof(S1)));
485 memcpy(&S2, s2, std::min(n, sizeof(S2)));
486 TS->TraceCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2);
489 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
492 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
493 uint64_t CmpSize = (SizeAndType >> 32) / 8;
494 uint64_t Type = (SizeAndType << 32) >> 32;
495 TS->TraceCmpCallback(PC, CmpSize, Type, Arg1, Arg2);
498 void __sanitizer_cov_trace_switch(uint64_t Val, uint64_t *Cases) {
500 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
501 TS->TraceSwitchCallback(PC, Cases[1], Val, Cases[0], Cases + 2);