1 //===- FuzzerDFSan.cpp - DFSan-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 // DataFlowSanitizer (DFSan) is a tool for
10 // generalised dynamic data flow (taint) analysis:
11 // http://clang.llvm.org/docs/DataFlowSanitizer.html .
13 // This file implements a mutation algorithm based on taint
14 // analysis feedback from DFSan.
16 // The approach has some similarity to "Taint-based Directed Whitebox Fuzzing"
17 // by Vijay Ganesh & Tim Leek & Martin Rinard:
18 // http://dspace.mit.edu/openaccess-disseminate/1721.1/59320,
19 // but it uses a full blown LLVM IR taint analysis and separate instrumentation
20 // to analyze all of the "attack points" at once.
23 // * lib/Fuzzer/Fuzzer*.cpp is compiled w/o any instrumentation.
24 // * The code under test is compiled with DFSan *and* with special extra hooks
25 // that are inserted before dfsan. Currently supported hooks:
26 // - __sanitizer_cov_trace_cmp: inserted before every ICMP instruction,
27 // receives the type, size and arguments of ICMP.
28 // * Every call to HOOK(a,b) is replaced by DFSan with
29 // __dfsw_HOOK(a, b, label(a), label(b)) so that __dfsw_HOOK
30 // gets all the taint labels for the arguments.
31 // * At the Fuzzer startup we assign a unique DFSan label
32 // to every byte of the input string (Fuzzer::CurrentUnit) so that for any
33 // chunk of data we know which input bytes it has derived from.
34 // * The __dfsw_* functions (implemented in this file) record the
35 // parameters (i.e. the application data and the corresponding taint labels)
37 // * Fuzzer::MutateWithDFSan() tries to use the data recorded by __dfsw_*
38 // hooks to guide the fuzzing towards new application states.
39 // For example if 4 bytes of data that derive from input bytes {4,5,6,7}
40 // are compared with a constant 12345 and the comparison always yields
41 // the same result, we try to insert 12345, 12344, 12346 into bytes
42 // {4,5,6,7} of the next fuzzed inputs.
44 // This code does not function when DFSan is not linked in.
45 // Instead of using ifdefs and thus requiring a separate build of lib/Fuzzer
46 // we redeclare the dfsan_* interface functions as weak and check if they
47 // are nullptr before calling.
48 // If this approach proves to be useful we may add attribute(weak) to the
49 // dfsan declarations in dfsan_interface.h
51 // This module is in the "proof of concept" stage.
52 // It is capable of solving only the simplest puzzles
53 // like test/dfsan/DFSanSimpleCmpTest.cpp.
54 //===----------------------------------------------------------------------===//
56 /* Example of manual usage:
59 clang -fPIC -c -g -O2 -std=c++11 Fuzzer*.cpp
60 clang++ -O0 -std=c++11 -fsanitize-coverage=edge,trace-cmp \
62 test/dfsan/DFSanSimpleCmpTest.cpp Fuzzer*.o
67 #include "FuzzerInternal.h"
68 #include <sanitizer/dfsan_interface.h>
73 #include <unordered_map>
77 dfsan_label dfsan_create_label(const char *desc, void *userdata);
79 void dfsan_set_label(dfsan_label label, void *addr, size_t size);
81 void dfsan_add_label(dfsan_label label, void *addr, size_t size);
83 const struct dfsan_label_info *dfsan_get_label_info(dfsan_label label);
85 dfsan_label dfsan_read_label(const void *addr, size_t size);
90 static bool ReallyHaveDFSan() {
91 return &dfsan_create_label != nullptr;
94 // These values are copied from include/llvm/IR/InstrTypes.h.
95 // We do not include the LLVM headers here to remain independent.
96 // If these values ever change, an assertion in ComputeCmp will fail.
98 ICMP_EQ = 32, ///< equal
99 ICMP_NE = 33, ///< not equal
100 ICMP_UGT = 34, ///< unsigned greater than
101 ICMP_UGE = 35, ///< unsigned greater or equal
102 ICMP_ULT = 36, ///< unsigned less than
103 ICMP_ULE = 37, ///< unsigned less or equal
104 ICMP_SGT = 38, ///< signed greater than
105 ICMP_SGE = 39, ///< signed greater or equal
106 ICMP_SLT = 40, ///< signed less than
107 ICMP_SLE = 41, ///< signed less or equal
110 template <class U, class S>
111 bool ComputeCmp(size_t CmpType, U Arg1, U Arg2) {
113 case ICMP_EQ : return Arg1 == Arg2;
114 case ICMP_NE : return Arg1 != Arg2;
115 case ICMP_UGT: return Arg1 > Arg2;
116 case ICMP_UGE: return Arg1 >= Arg2;
117 case ICMP_ULT: return Arg1 < Arg2;
118 case ICMP_ULE: return Arg1 <= Arg2;
119 case ICMP_SGT: return (S)Arg1 > (S)Arg2;
120 case ICMP_SGE: return (S)Arg1 >= (S)Arg2;
121 case ICMP_SLT: return (S)Arg1 < (S)Arg2;
122 case ICMP_SLE: return (S)Arg1 <= (S)Arg2;
123 default: assert(0 && "unsupported CmpType");
128 static bool ComputeCmp(size_t CmpSize, size_t CmpType, uint64_t Arg1,
130 if (CmpSize == 8) return ComputeCmp<uint64_t, int64_t>(CmpType, Arg1, Arg2);
131 if (CmpSize == 4) return ComputeCmp<uint32_t, int32_t>(CmpType, Arg1, Arg2);
132 if (CmpSize == 2) return ComputeCmp<uint16_t, int16_t>(CmpType, Arg1, Arg2);
133 if (CmpSize == 1) return ComputeCmp<uint8_t, int8_t>(CmpType, Arg1, Arg2);
134 assert(0 && "unsupported type size");
138 // As a simplification we use the range of input bytes instead of a set of input
141 uint16_t Beg, End; // Range is [Beg, End), thus Beg==End is an empty range.
143 LabelRange(uint16_t Beg = 0, uint16_t End = 0) : Beg(Beg), End(End) {}
145 static LabelRange Join(LabelRange LR1, LabelRange LR2) {
146 if (LR1.Beg == LR1.End) return LR2;
147 if (LR2.Beg == LR2.End) return LR1;
148 return {std::min(LR1.Beg, LR2.Beg), std::max(LR1.End, LR2.End)};
150 LabelRange &Join(LabelRange LR) {
151 return *this = Join(*this, LR);
153 static LabelRange Singleton(const dfsan_label_info *LI) {
154 uint16_t Idx = (uint16_t)(uintptr_t)LI->userdata;
156 return {(uint16_t)(Idx - 1), Idx};
160 std::ostream &operator<<(std::ostream &os, const LabelRange &LR) {
161 return os << "[" << LR.Beg << "," << LR.End << ")";
164 // For now, very simple: put Size bytes of Data at position Pos.
165 struct TraceBasedMutation {
173 DFSanState(const Fuzzer::FuzzingOptions &Options, const Unit &CurrentUnit)
174 : Options(Options), CurrentUnit(CurrentUnit) {}
176 LabelRange GetLabelRange(dfsan_label L);
177 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
178 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
180 void TraceCmpCallback(size_t CmpSize, size_t CmpType, uint64_t Arg1,
182 int TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
185 void StartTraceRecording() {
186 RecordingTraces = true;
190 size_t StopTraceRecording() {
191 RecordingTraces = false;
192 std::random_shuffle(Mutations.begin(), Mutations.end());
193 return Mutations.size();
196 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
199 bool IsTwoByteData(uint64_t Data) {
200 int64_t Signed = static_cast<int64_t>(Data);
202 return Signed == 0 || Signed == -1L;
204 bool RecordingTraces = false;
205 std::vector<TraceBasedMutation> Mutations;
206 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
207 const Fuzzer::FuzzingOptions &Options;
208 const Unit &CurrentUnit;
211 LabelRange DFSanState::GetLabelRange(dfsan_label L) {
212 LabelRange &LR = LabelRanges[L];
213 if (LR.Beg < LR.End || L == 0)
215 const dfsan_label_info *LI = dfsan_get_label_info(L);
216 if (LI->l1 || LI->l2)
217 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
218 return LR = LabelRange::Singleton(LI);
221 void DFSanState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
222 assert(Idx < Mutations.size());
223 auto &M = Mutations[Idx];
224 if (Options.Verbosity >= 3)
225 std::cerr << "TBM " << M.Pos << " " << M.Size << " " << M.Data << "\n";
226 if (M.Pos + M.Size > U->size()) return;
227 memcpy(U->data() + M.Pos, &M.Data, M.Size);
230 void DFSanState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
231 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
233 assert(ReallyHaveDFSan());
234 if (!RecordingTraces) return;
235 if (L1 == 0 && L2 == 0)
236 return; // Not actionable.
237 if (L1 != 0 && L2 != 0)
238 return; // Probably still actionable.
239 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
240 uint64_t Data = L1 ? Arg2 : Arg1;
241 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
243 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
244 Mutations.push_back({Pos, CmpSize, Data});
245 Mutations.push_back({Pos, CmpSize, Data + 1});
246 Mutations.push_back({Pos, CmpSize, Data - 1});
249 if (CmpSize > LR.End - LR.Beg)
250 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
253 if (Options.Verbosity >= 3)
254 std::cerr << "DFSAN:"
255 << " PC " << std::hex << PC << std::dec
258 << " A1 " << Arg1 << " A2 " << Arg2 << " R " << Res
262 << " MU " << Mutations.size()
266 int DFSanState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
269 const uint8_t *Beg = CurrentUnit.data();
270 const uint8_t *End = Beg + CurrentUnit.size();
271 for (const uint8_t *Cur = Beg; Cur < End; Cur += DataSize) {
272 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
275 // std::cerr << "Cur " << (void*)Cur << "\n";
276 size_t Pos = Cur - Beg;
277 assert(Pos < CurrentUnit.size());
278 Mutations.push_back({Pos, DataSize, DesiredData});
279 Mutations.push_back({Pos, DataSize, DesiredData + 1});
280 Mutations.push_back({Pos, DataSize, DesiredData - 1});
287 void DFSanState::TraceCmpCallback(size_t CmpSize, size_t CmpType, uint64_t Arg1,
289 if (!Options.UseTraces) return;
291 if (Options.Verbosity >= 3)
292 std::cerr << "TraceCmp: " << Arg1 << " " << Arg2 << "\n";
293 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
294 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
295 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
296 Added += TryToAddDesiredData(Arg1, Arg2, 2);
297 Added += TryToAddDesiredData(Arg2, Arg1, 2);
301 static DFSanState *DFSan;
303 void Fuzzer::StartTraceRecording() {
305 DFSan->StartTraceRecording();
308 size_t Fuzzer::StopTraceRecording() {
309 if (!DFSan) return 0;
310 return DFSan->StopTraceRecording();
313 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
315 DFSan->ApplyTraceBasedMutation(Idx, U);
318 void Fuzzer::InitializeDFSan() {
319 if (!Options.UseDFSan) return;
320 DFSan = new DFSanState(Options, CurrentUnit);
321 CurrentUnit.resize(Options.MaxLen);
322 // The rest really requires DFSan.
323 if (!ReallyHaveDFSan()) return;
324 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
325 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
326 // We assume that no one else has called dfsan_create_label before.
328 dfsan_set_label(L, &CurrentUnit[i], 1);
332 } // namespace fuzzer
337 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
338 uint64_t Arg2, dfsan_label L0,
339 dfsan_label L1, dfsan_label L2) {
341 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
342 uint64_t CmpSize = (SizeAndType >> 32) / 8;
343 uint64_t Type = (SizeAndType << 32) >> 32;
344 DFSan->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
347 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
348 size_t n, dfsan_label s1_label,
349 dfsan_label s2_label, dfsan_label n_label) {
350 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
351 uint64_t S1 = 0, S2 = 0;
352 // Simplification: handle only first 8 bytes.
353 memcpy(&S1, s1, std::min(n, sizeof(S1)));
354 memcpy(&S2, s2, std::min(n, sizeof(S2)));
355 dfsan_label L1 = dfsan_read_label(s1, n);
356 dfsan_label L2 = dfsan_read_label(s2, n);
357 DFSan->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
360 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
363 uint64_t CmpSize = (SizeAndType >> 32) / 8;
364 uint64_t Type = (SizeAndType << 32) >> 32;
365 DFSan->TraceCmpCallback(CmpSize, Type, Arg1, Arg2);