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=3 \
61 -mllvm -sanitizer-coverage-experimental-trace-compares=1 \
63 test/dfsan/DFSanSimpleCmpTest.cpp Fuzzer*.o
68 #include "FuzzerInternal.h"
69 #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 // These values are copied from include/llvm/IR/InstrTypes.h.
91 // We do not include the LLVM headers here to remain independent.
92 // If these values ever change, an assertion in ComputeCmp will fail.
94 ICMP_EQ = 32, ///< equal
95 ICMP_NE = 33, ///< not equal
96 ICMP_UGT = 34, ///< unsigned greater than
97 ICMP_UGE = 35, ///< unsigned greater or equal
98 ICMP_ULT = 36, ///< unsigned less than
99 ICMP_ULE = 37, ///< unsigned less or equal
100 ICMP_SGT = 38, ///< signed greater than
101 ICMP_SGE = 39, ///< signed greater or equal
102 ICMP_SLT = 40, ///< signed less than
103 ICMP_SLE = 41, ///< signed less or equal
106 template <class U, class S>
107 bool ComputeCmp(size_t CmpType, U Arg1, U Arg2) {
109 case ICMP_EQ : return Arg1 == Arg2;
110 case ICMP_NE : return Arg1 != Arg2;
111 case ICMP_UGT: return Arg1 > Arg2;
112 case ICMP_UGE: return Arg1 >= Arg2;
113 case ICMP_ULT: return Arg1 < Arg2;
114 case ICMP_ULE: return Arg1 <= Arg2;
115 case ICMP_SGT: return (S)Arg1 > (S)Arg2;
116 case ICMP_SGE: return (S)Arg1 >= (S)Arg2;
117 case ICMP_SLT: return (S)Arg1 < (S)Arg2;
118 case ICMP_SLE: return (S)Arg1 <= (S)Arg2;
119 default: assert(0 && "unsupported CmpType");
124 static bool ComputeCmp(size_t CmpSize, size_t CmpType, uint64_t Arg1,
126 if (CmpSize == 8) return ComputeCmp<uint64_t, int64_t>(CmpType, Arg1, Arg2);
127 if (CmpSize == 4) return ComputeCmp<uint32_t, int32_t>(CmpType, Arg1, Arg2);
128 if (CmpSize == 2) return ComputeCmp<uint16_t, int16_t>(CmpType, Arg1, Arg2);
129 if (CmpSize == 1) return ComputeCmp<uint8_t, int8_t>(CmpType, Arg1, Arg2);
130 assert(0 && "unsupported type size");
134 // As a simplification we use the range of input bytes instead of a set of input
137 uint16_t Beg, End; // Range is [Beg, End), thus Beg==End is an empty range.
139 LabelRange(uint16_t Beg = 0, uint16_t End = 0) : Beg(Beg), End(End) {}
141 static LabelRange Join(LabelRange LR1, LabelRange LR2) {
142 if (LR1.Beg == LR1.End) return LR2;
143 if (LR2.Beg == LR2.End) return LR1;
144 return {std::min(LR1.Beg, LR2.Beg), std::max(LR1.End, LR2.End)};
146 LabelRange &Join(LabelRange LR) {
147 return *this = Join(*this, LR);
149 static LabelRange Singleton(const dfsan_label_info *LI) {
150 uint16_t Idx = (uint16_t)(uintptr_t)LI->userdata;
152 return {(uint16_t)(Idx - 1), Idx};
156 std::ostream &operator<<(std::ostream &os, const LabelRange &LR) {
157 return os << "[" << LR.Beg << "," << LR.End << ")";
162 DFSanState(const fuzzer::Fuzzer::FuzzingOptions &Options)
163 : Options(Options) {}
166 size_t ResCounters[2] = {0, 0};
169 std::unordered_map<uint64_t, size_t> CountedConstants;
172 LabelRange GetLabelRange(dfsan_label L);
173 void DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
174 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
176 bool Mutate(fuzzer::Unit *U);
179 std::unordered_map<uintptr_t, CmpSiteInfo> PcToCmpSiteInfoMap;
180 LabelRange LabelRanges[1 << (sizeof(dfsan_label) * 8)] = {};
181 const fuzzer::Fuzzer::FuzzingOptions &Options;
184 LabelRange DFSanState::GetLabelRange(dfsan_label L) {
185 LabelRange &LR = LabelRanges[L];
186 if (LR.Beg < LR.End || L == 0)
188 const dfsan_label_info *LI = dfsan_get_label_info(L);
189 if (LI->l1 || LI->l2)
190 return LR = LabelRange::Join(GetLabelRange(LI->l1), GetLabelRange(LI->l2));
191 return LR = LabelRange::Singleton(LI);
194 void DFSanState::DFSanCmpCallback(uintptr_t PC, size_t CmpSize, size_t CmpType,
195 uint64_t Arg1, uint64_t Arg2, dfsan_label L1,
197 if (L1 == 0 && L2 == 0)
198 return; // Not actionable.
199 if (L1 != 0 && L2 != 0)
200 return; // Probably still actionable.
201 bool Res = ComputeCmp(CmpSize, CmpType, Arg1, Arg2);
202 CmpSiteInfo &CSI = PcToCmpSiteInfoMap[PC];
203 CSI.CmpSize = CmpSize;
204 CSI.LR.Join(GetLabelRange(L1)).Join(GetLabelRange(L2));
205 if (!L1) CSI.CountedConstants[Arg1]++;
206 if (!L2) CSI.CountedConstants[Arg2]++;
207 size_t Counter = CSI.ResCounters[Res]++;
209 if (Options.Verbosity >= 2 &&
210 (Counter & (Counter - 1)) == 0 &&
211 CSI.ResCounters[!Res] == 0)
212 std::cerr << "DFSAN:"
213 << " PC " << std::hex << PC << std::dec
216 << " A1 " << Arg1 << " A2 " << Arg2 << " R " << Res
217 << " L" << L1 << GetLabelRange(L1)
218 << " L" << L2 << GetLabelRange(L2)
223 bool DFSanState::Mutate(fuzzer::Unit *U) {
224 for (auto &PCToCmp : PcToCmpSiteInfoMap) {
225 auto &CSI = PCToCmp.second;
226 if (CSI.ResCounters[0] * CSI.ResCounters[1] != 0) continue;
227 if (CSI.ResCounters[0] + CSI.ResCounters[1] < 1000) continue;
228 if (CSI.CountedConstants.size() != 1) continue;
229 uintptr_t C = CSI.CountedConstants.begin()->first;
230 if (U->size() >= CSI.CmpSize) {
231 size_t RangeSize = CSI.LR.End - CSI.LR.Beg;
232 size_t Idx = CSI.LR.Beg + rand() % RangeSize;
233 if (Idx + CSI.CmpSize > U->size()) continue;
235 memcpy(U->data() + Idx, &C, CSI.CmpSize);
242 static DFSanState *DFSan;
248 bool Fuzzer::MutateWithDFSan(Unit *U) {
249 if (!&dfsan_create_label || !DFSan) return false;
250 return DFSan->Mutate(U);
253 void Fuzzer::InitializeDFSan() {
254 if (!&dfsan_create_label || !Options.UseDFSan) return;
255 DFSan = new DFSanState(Options);
256 CurrentUnit.resize(Options.MaxLen);
257 for (size_t i = 0; i < static_cast<size_t>(Options.MaxLen); i++) {
258 dfsan_label L = dfsan_create_label("input", (void*)(i + 1));
259 // We assume that no one else has called dfsan_create_label before.
261 dfsan_set_label(L, &CurrentUnit[i], 1);
265 } // namespace fuzzer
268 void __dfsw___sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
269 uint64_t Arg2, dfsan_label L0,
270 dfsan_label L1, dfsan_label L2) {
272 uintptr_t PC = reinterpret_cast<uintptr_t>(__builtin_return_address(0));
273 uint64_t CmpSize = (SizeAndType >> 32) / 8;
274 uint64_t Type = (SizeAndType << 32) >> 32;
275 DFSan->DFSanCmpCallback(PC, CmpSize, Type, Arg1, Arg2, L1, L2);
278 void dfsan_weak_hook_memcmp(void *caller_pc, const void *s1, const void *s2,
279 size_t n, dfsan_label s1_label,
280 dfsan_label s2_label, dfsan_label n_label) {
281 uintptr_t PC = reinterpret_cast<uintptr_t>(caller_pc);
283 // Simplification: handle only first 8 bytes.
284 memcpy(&S1, s1, std::min(n, sizeof(S1)));
285 memcpy(&S2, s2, std::min(n, sizeof(S2)));
286 dfsan_label L1 = dfsan_read_label(s1, n);
287 dfsan_label L2 = dfsan_read_label(s2, n);
288 DFSan->DFSanCmpCallback(PC, n, ICMP_EQ, S1, S2, L1, L2);
291 void __sanitizer_cov_trace_cmp(uint64_t SizeAndType, uint64_t Arg1,
293 // This symbol will be present if dfsan is disabled on the given function.
294 // FIXME: implement poor man's taint analysis here (w/o dfsan).