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);
141 assert(0 && "unsupported type size");
145 // As a simplification we use the range of input bytes instead of a set of input
148 uint16_t Beg, End; // Range is [Beg, End), thus Beg==End is an empty range.
150 LabelRange(uint16_t Beg = 0, uint16_t End = 0) : Beg(Beg), End(End) {}
152 static LabelRange Join(LabelRange LR1, LabelRange LR2) {
153 if (LR1.Beg == LR1.End) return LR2;
154 if (LR2.Beg == LR2.End) return LR1;
155 return {std::min(LR1.Beg, LR2.Beg), std::max(LR1.End, LR2.End)};
157 LabelRange &Join(LabelRange LR) {
158 return *this = Join(*this, LR);
160 static LabelRange Singleton(const dfsan_label_info *LI) {
161 uint16_t Idx = (uint16_t)(uintptr_t)LI->userdata;
163 return {(uint16_t)(Idx - 1), Idx};
167 // A passport for a CMP site. We want to keep track of where the given CMP is
168 // and how many times it is evaluated to true or false.
169 struct CmpSitePassport {
173 bool IsInterestingCmpTarget() {
174 static const size_t kRareEnough = 50;
175 size_t C0 = Counter[0];
176 size_t C1 = Counter[1];
177 return C0 > kRareEnough * (C1 + 1) || C1 > kRareEnough * (C0 + 1);
181 // For now, just keep a simple imprecise hash table PC => CmpSitePassport.
182 // Potentially, will need to have a compiler support to have a precise mapping
183 // and also thread-safety.
184 struct CmpSitePassportTable {
185 static const size_t kSize = 99991; // Prime.
186 CmpSitePassport Passports[kSize];
188 CmpSitePassport *GetPassport(uintptr_t PC) {
189 uintptr_t Idx = PC & kSize;
190 CmpSitePassport *Res = &Passports[Idx];
191 if (Res->PC == 0) // Not thread safe.
193 return Res->PC == PC ? Res : nullptr;
197 static CmpSitePassportTable CSPTable; // Zero initialized.
199 // For now, very simple: put Size bytes of Data at position Pos.
200 struct TraceBasedMutation {
208 TraceState(const Fuzzer::FuzzingOptions &Options, const Unit &CurrentUnit)
209 : Options(Options), CurrentUnit(CurrentUnit) {}
211 LabelRange GetLabelRange(dfsan_label L);
212 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
213 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
215 void TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType, uint64_t Arg1,
217 int TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
220 void StartTraceRecording() {
221 if (!Options.UseTraces) return;
222 RecordingTraces = true;
226 size_t StopTraceRecording(FuzzerRandomBase &Rand) {
227 RecordingTraces = false;
228 std::random_shuffle(Mutations.begin(), Mutations.end(), Rand);
229 return Mutations.size();
232 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
235 bool IsTwoByteData(uint64_t Data) {
236 int64_t Signed = static_cast<int64_t>(Data);
238 return Signed == 0 || Signed == -1L;
240 bool RecordingTraces = false;
241 std::vector<TraceBasedMutation> Mutations;
242 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
243 const Fuzzer::FuzzingOptions &Options;
244 const Unit &CurrentUnit;
247 LabelRange TraceState::GetLabelRange(dfsan_label L) {
248 LabelRange &LR = LabelRanges[L];
249 if (LR.Beg < LR.End || L == 0)
251 const dfsan_label_info *LI = dfsan_get_label_info(L);
252 if (LI->l1 || LI->l2)
253 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
254 return LR = LabelRange::Singleton(LI);
257 void TraceState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
258 assert(Idx < Mutations.size());
259 auto &M = Mutations[Idx];
260 if (Options.Verbosity >= 3)
261 Printf("TBM %zd %zd %zd\n", M.Pos, M.Size, M.Data);
262 if (M.Pos + M.Size > U->size()) return;
263 memcpy(U->data() + M.Pos, &M.Data, M.Size);
266 void TraceState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
267 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
269 assert(ReallyHaveDFSan());
270 if (!RecordingTraces) return;
271 if (L1 == 0 && L2 == 0)
272 return; // Not actionable.
273 if (L1 != 0 && L2 != 0)
274 return; // Probably still actionable.
275 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
276 uint64_t Data = L1 ? Arg2 : Arg1;
277 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
279 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
280 Mutations.push_back({Pos, CmpSize, Data});
281 Mutations.push_back({Pos, CmpSize, Data + 1});
282 Mutations.push_back({Pos, CmpSize, Data - 1});
285 if (CmpSize > LR.End - LR.Beg)
286 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
289 if (Options.Verbosity >= 3)
290 Printf("DFSanCmpCallback: PC %lx S %zd T %zd A1 %llx A2 %llx R %d L1 %d L2 "
292 PC, CmpSize, CmpType, Arg1, Arg2, Res, L1, L2, Mutations.size());
295 int TraceState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
298 const uint8_t *Beg = CurrentUnit.data();
299 const uint8_t *End = Beg + CurrentUnit.size();
300 for (const uint8_t *Cur = Beg; Cur < End; Cur += DataSize) {
301 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
304 size_t Pos = Cur - Beg;
305 assert(Pos < CurrentUnit.size());
306 Mutations.push_back({Pos, DataSize, DesiredData});
307 Mutations.push_back({Pos, DataSize, DesiredData + 1});
308 Mutations.push_back({Pos, DataSize, DesiredData - 1});
315 void TraceState::TraceCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType, uint64_t Arg1,
317 if (!RecordingTraces) return;
319 CmpSitePassport *CSP = CSPTable.GetPassport(PC);
321 CSP->Counter[ComputeCmp(CmpSize, CmpType, Arg1, Arg2)]++;
322 size_t C0 = CSP->Counter[0];
323 size_t C1 = CSP->Counter[1];
324 if (!CSP->IsInterestingCmpTarget())
326 if (Options.Verbosity >= 3)
327 Printf("TraceCmp: %p %zd/%zd; %zd %zd\n", CSP->PC, C0, C1, Arg1, Arg2);
328 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
329 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
330 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
331 Added += TryToAddDesiredData(Arg1, Arg2, 2);
332 Added += TryToAddDesiredData(Arg2, Arg1, 2);
336 static TraceState *TS;
338 void Fuzzer::StartTraceRecording() {
340 TS->StartTraceRecording();
343 size_t Fuzzer::StopTraceRecording() {
345 return TS->StopTraceRecording(USF.GetRand());
348 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
350 TS->ApplyTraceBasedMutation(Idx, U);
353 void Fuzzer::InitializeTraceState() {
354 if (!Options.UseTraces) return;
355 TS = new TraceState(Options, CurrentUnit);
356 CurrentUnit.resize(Options.MaxLen);
357 // The rest really requires DFSan.
358 if (!ReallyHaveDFSan()) return;
359 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
360 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
361 // We assume that no one else has called dfsan_create_label before.
363 dfsan_set_label(L, &CurrentUnit[i], 1);
367 } // namespace fuzzer
372 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
373 uint64_t Arg2, dfsan_label L0,
374 dfsan_label L1, dfsan_label L2) {
377 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
378 uint64_t CmpSize = (SizeAndType >> 32) / 8;
379 uint64_t Type = (SizeAndType << 32) >> 32;
380 TS->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
383 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
384 size_t n, dfsan_label s1_label,
385 dfsan_label s2_label, dfsan_label n_label) {
387 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
388 uint64_t S1 = 0, S2 = 0;
389 // Simplification: handle only first 8 bytes.
390 memcpy(&S1, s1, std::min(n, sizeof(S1)));
391 memcpy(&S2, s2, std::min(n, sizeof(S2)));
392 dfsan_label L1 = dfsan_read_label(s1, n);
393 dfsan_label L2 = dfsan_read_label(s2, n);
394 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
397 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
400 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
401 uint64_t CmpSize = (SizeAndType >> 32) / 8;
402 uint64_t Type = (SizeAndType << 32) >> 32;
403 TS->TraceCmpCallback(PC, CmpSize, Type, Arg1, Arg2);