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>
81 #include <unordered_map>
85 dfsan_label dfsan_create_label(const char *desc, void *userdata);
87 void dfsan_set_label(dfsan_label label, void *addr, size_t size);
89 void dfsan_add_label(dfsan_label label, void *addr, size_t size);
91 const struct dfsan_label_info *dfsan_get_label_info(dfsan_label label);
93 dfsan_label dfsan_read_label(const void *addr, size_t size);
98 static bool ReallyHaveDFSan() {
99 return &dfsan_create_label != nullptr;
102 // These values are copied from include/llvm/IR/InstrTypes.h.
103 // We do not include the LLVM headers here to remain independent.
104 // If these values ever change, an assertion in ComputeCmp will fail.
106 ICMP_EQ = 32, ///< equal
107 ICMP_NE = 33, ///< not equal
108 ICMP_UGT = 34, ///< unsigned greater than
109 ICMP_UGE = 35, ///< unsigned greater or equal
110 ICMP_ULT = 36, ///< unsigned less than
111 ICMP_ULE = 37, ///< unsigned less or equal
112 ICMP_SGT = 38, ///< signed greater than
113 ICMP_SGE = 39, ///< signed greater or equal
114 ICMP_SLT = 40, ///< signed less than
115 ICMP_SLE = 41, ///< signed less or equal
118 template <class U, class S>
119 bool ComputeCmp(size_t CmpType, U Arg1, U Arg2) {
121 case ICMP_EQ : return Arg1 == Arg2;
122 case ICMP_NE : return Arg1 != Arg2;
123 case ICMP_UGT: return Arg1 > Arg2;
124 case ICMP_UGE: return Arg1 >= Arg2;
125 case ICMP_ULT: return Arg1 < Arg2;
126 case ICMP_ULE: return Arg1 <= Arg2;
127 case ICMP_SGT: return (S)Arg1 > (S)Arg2;
128 case ICMP_SGE: return (S)Arg1 >= (S)Arg2;
129 case ICMP_SLT: return (S)Arg1 < (S)Arg2;
130 case ICMP_SLE: return (S)Arg1 <= (S)Arg2;
131 default: assert(0 && "unsupported CmpType");
136 static bool ComputeCmp(size_t CmpSize, size_t CmpType, uint64_t Arg1,
138 if (CmpSize == 8) return ComputeCmp<uint64_t, int64_t>(CmpType, Arg1, Arg2);
139 if (CmpSize == 4) return ComputeCmp<uint32_t, int32_t>(CmpType, Arg1, Arg2);
140 if (CmpSize == 2) return ComputeCmp<uint16_t, int16_t>(CmpType, Arg1, Arg2);
141 if (CmpSize == 1) return ComputeCmp<uint8_t, int8_t>(CmpType, Arg1, Arg2);
142 assert(0 && "unsupported type size");
146 // As a simplification we use the range of input bytes instead of a set of input
149 uint16_t Beg, End; // Range is [Beg, End), thus Beg==End is an empty range.
151 LabelRange(uint16_t Beg = 0, uint16_t End = 0) : Beg(Beg), End(End) {}
153 static LabelRange Join(LabelRange LR1, LabelRange LR2) {
154 if (LR1.Beg == LR1.End) return LR2;
155 if (LR2.Beg == LR2.End) return LR1;
156 return {std::min(LR1.Beg, LR2.Beg), std::max(LR1.End, LR2.End)};
158 LabelRange &Join(LabelRange LR) {
159 return *this = Join(*this, LR);
161 static LabelRange Singleton(const dfsan_label_info *LI) {
162 uint16_t Idx = (uint16_t)(uintptr_t)LI->userdata;
164 return {(uint16_t)(Idx - 1), Idx};
168 std::ostream &operator<<(std::ostream &os, const LabelRange &LR) {
169 return os << "[" << LR.Beg << "," << LR.End << ")";
172 // For now, very simple: put Size bytes of Data at position Pos.
173 struct TraceBasedMutation {
181 TraceState(const Fuzzer::FuzzingOptions &Options, const Unit &CurrentUnit)
182 : Options(Options), CurrentUnit(CurrentUnit) {}
184 LabelRange GetLabelRange(dfsan_label L);
185 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
186 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
188 void TraceCmpCallback(size_t CmpSize, size_t CmpType, uint64_t Arg1,
190 int TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
193 void StartTraceRecording() {
194 if (!Options.UseTraces) return;
195 RecordingTraces = true;
199 size_t StopTraceRecording() {
200 RecordingTraces = false;
201 std::random_shuffle(Mutations.begin(), Mutations.end());
202 return Mutations.size();
205 void ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U);
208 bool IsTwoByteData(uint64_t Data) {
209 int64_t Signed = static_cast<int64_t>(Data);
211 return Signed == 0 || Signed == -1L;
213 bool RecordingTraces = false;
214 std::vector<TraceBasedMutation> Mutations;
215 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
216 const Fuzzer::FuzzingOptions &Options;
217 const Unit &CurrentUnit;
220 LabelRange TraceState::GetLabelRange(dfsan_label L) {
221 LabelRange &LR = LabelRanges[L];
222 if (LR.Beg < LR.End || L == 0)
224 const dfsan_label_info *LI = dfsan_get_label_info(L);
225 if (LI->l1 || LI->l2)
226 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
227 return LR = LabelRange::Singleton(LI);
230 void TraceState::ApplyTraceBasedMutation(size_t Idx, fuzzer::Unit *U) {
231 assert(Idx < Mutations.size());
232 auto &M = Mutations[Idx];
233 if (Options.Verbosity >= 3)
234 std::cerr << "TBM " << M.Pos << " " << M.Size << " " << M.Data << "\n";
235 if (M.Pos + M.Size > U->size()) return;
236 memcpy(U->data() + M.Pos, &M.Data, M.Size);
239 void TraceState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
240 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
242 assert(ReallyHaveDFSan());
243 if (!RecordingTraces) return;
244 if (L1 == 0 && L2 == 0)
245 return; // Not actionable.
246 if (L1 != 0 && L2 != 0)
247 return; // Probably still actionable.
248 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
249 uint64_t Data = L1 ? Arg2 : Arg1;
250 LabelRange LR = L1 ? GetLabelRange(L1) : GetLabelRange(L2);
252 for (size_t Pos = LR.Beg; Pos + CmpSize <= LR.End; Pos++) {
253 Mutations.push_back({Pos, CmpSize, Data});
254 Mutations.push_back({Pos, CmpSize, Data + 1});
255 Mutations.push_back({Pos, CmpSize, Data - 1});
258 if (CmpSize > LR.End - LR.Beg)
259 Mutations.push_back({LR.Beg, (unsigned)(LR.End - LR.Beg), Data});
262 if (Options.Verbosity >= 3)
263 std::cerr << "DFSAN:"
264 << " PC " << std::hex << PC << std::dec
267 << " A1 " << Arg1 << " A2 " << Arg2 << " R " << Res
271 << " MU " << Mutations.size()
275 int TraceState::TryToAddDesiredData(uint64_t PresentData, uint64_t DesiredData,
278 const uint8_t *Beg = CurrentUnit.data();
279 const uint8_t *End = Beg + CurrentUnit.size();
280 for (const uint8_t *Cur = Beg; Cur < End; Cur += DataSize) {
281 Cur = (uint8_t *)memmem(Cur, End - Cur, &PresentData, DataSize);
284 // std::cerr << "Cur " << (void*)Cur << "\n";
285 size_t Pos = Cur - Beg;
286 assert(Pos < CurrentUnit.size());
287 Mutations.push_back({Pos, DataSize, DesiredData});
288 Mutations.push_back({Pos, DataSize, DesiredData + 1});
289 Mutations.push_back({Pos, DataSize, DesiredData - 1});
296 void TraceState::TraceCmpCallback(size_t CmpSize, size_t CmpType, uint64_t Arg1,
298 if (!RecordingTraces) return;
300 if (Options.Verbosity >= 3)
301 std::cerr << "TraceCmp: " << Arg1 << " " << Arg2 << "\n";
302 Added += TryToAddDesiredData(Arg1, Arg2, CmpSize);
303 Added += TryToAddDesiredData(Arg2, Arg1, CmpSize);
304 if (!Added && CmpSize == 4 && IsTwoByteData(Arg1) && IsTwoByteData(Arg2)) {
305 Added += TryToAddDesiredData(Arg1, Arg2, 2);
306 Added += TryToAddDesiredData(Arg2, Arg1, 2);
310 static TraceState *TS;
312 void Fuzzer::StartTraceRecording() {
314 TS->StartTraceRecording();
317 size_t Fuzzer::StopTraceRecording() {
319 return TS->StopTraceRecording();
322 void Fuzzer::ApplyTraceBasedMutation(size_t Idx, Unit *U) {
324 TS->ApplyTraceBasedMutation(Idx, U);
327 void Fuzzer::InitializeTraceState() {
328 if (!Options.UseTraces && !Options.UseDFSan) return;
329 TS = new TraceState(Options, CurrentUnit);
330 CurrentUnit.resize(Options.MaxLen);
331 // The rest really requires DFSan.
332 if (!ReallyHaveDFSan() || !Options.UseDFSan) return;
333 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
334 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
335 // We assume that no one else has called dfsan_create_label before.
337 dfsan_set_label(L, &CurrentUnit[i], 1);
341 } // namespace fuzzer
346 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
347 uint64_t Arg2, dfsan_label L0,
348 dfsan_label L1, dfsan_label L2) {
351 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
352 uint64_t CmpSize = (SizeAndType >> 32) / 8;
353 uint64_t Type = (SizeAndType << 32) >> 32;
354 TS->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
357 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
358 size_t n, dfsan_label s1_label,
359 dfsan_label s2_label, dfsan_label n_label) {
361 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
362 uint64_t S1 = 0, S2 = 0;
363 // Simplification: handle only first 8 bytes.
364 memcpy(&S1, s1, std::min(n, sizeof(S1)));
365 memcpy(&S2, s2, std::min(n, sizeof(S2)));
366 dfsan_label L1 = dfsan_read_label(s1, n);
367 dfsan_label L2 = dfsan_read_label(s2, n);
368 TS->DFSanCmpCallback(PC, n, fuzzer::ICMP_EQ, S1, S2, L1, L2);
371 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
374 uint64_t CmpSize = (SizeAndType >> 32) / 8;
375 uint64_t Type = (SizeAndType << 32) >> 32;
376 TS->TraceCmpCallback(CmpSize, Type, Arg1, Arg2);