xref: /freebsd/contrib/llvm-project/compiler-rt/lib/fuzzer/FuzzerCorpus.h (revision 6ba2210ee039f2f12878c217bcf058e9c8b26b29)
1 //===- FuzzerCorpus.h - Internal header for the Fuzzer ----------*- C++ -* ===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 // fuzzer::InputCorpus
9 //===----------------------------------------------------------------------===//
10 
11 #ifndef LLVM_FUZZER_CORPUS
12 #define LLVM_FUZZER_CORPUS
13 
14 #include "FuzzerDataFlowTrace.h"
15 #include "FuzzerDefs.h"
16 #include "FuzzerIO.h"
17 #include "FuzzerRandom.h"
18 #include "FuzzerSHA1.h"
19 #include "FuzzerTracePC.h"
20 #include <algorithm>
21 #include <chrono>
22 #include <numeric>
23 #include <random>
24 #include <unordered_set>
25 
26 namespace fuzzer {
27 
28 struct InputInfo {
29   Unit U;  // The actual input data.
30   std::chrono::microseconds TimeOfUnit;
31   uint8_t Sha1[kSHA1NumBytes];  // Checksum.
32   // Number of features that this input has and no smaller input has.
33   size_t NumFeatures = 0;
34   size_t Tmp = 0; // Used by ValidateFeatureSet.
35   // Stats.
36   size_t NumExecutedMutations = 0;
37   size_t NumSuccessfullMutations = 0;
38   bool NeverReduce = false;
39   bool MayDeleteFile = false;
40   bool Reduced = false;
41   bool HasFocusFunction = false;
42   Vector<uint32_t> UniqFeatureSet;
43   Vector<uint8_t> DataFlowTraceForFocusFunction;
44   // Power schedule.
45   bool NeedsEnergyUpdate = false;
46   double Energy = 0.0;
47   size_t SumIncidence = 0;
48   Vector<std::pair<uint32_t, uint16_t>> FeatureFreqs;
49 
50   // Delete feature Idx and its frequency from FeatureFreqs.
51   bool DeleteFeatureFreq(uint32_t Idx) {
52     if (FeatureFreqs.empty())
53       return false;
54 
55     // Binary search over local feature frequencies sorted by index.
56     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
57                                   std::pair<uint32_t, uint16_t>(Idx, 0));
58 
59     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
60       FeatureFreqs.erase(Lower);
61       return true;
62     }
63     return false;
64   }
65 
66   // Assign more energy to a high-entropy seed, i.e., that reveals more
67   // information about the globally rare features in the neighborhood of the
68   // seed. Since we do not know the entropy of a seed that has never been
69   // executed we assign fresh seeds maximum entropy and let II->Energy approach
70   // the true entropy from above. If ScalePerExecTime is true, the computed
71   // entropy is scaled based on how fast this input executes compared to the
72   // average execution time of inputs. The faster an input executes, the more
73   // energy gets assigned to the input.
74   void UpdateEnergy(size_t GlobalNumberOfFeatures, bool ScalePerExecTime,
75                     std::chrono::microseconds AverageUnitExecutionTime) {
76     Energy = 0.0;
77     SumIncidence = 0;
78 
79     // Apply add-one smoothing to locally discovered features.
80     for (auto F : FeatureFreqs) {
81       size_t LocalIncidence = F.second + 1;
82       Energy -= LocalIncidence * logl(LocalIncidence);
83       SumIncidence += LocalIncidence;
84     }
85 
86     // Apply add-one smoothing to locally undiscovered features.
87     //   PreciseEnergy -= 0; // since logl(1.0) == 0)
88     SumIncidence += (GlobalNumberOfFeatures - FeatureFreqs.size());
89 
90     // Add a single locally abundant feature apply add-one smoothing.
91     size_t AbdIncidence = NumExecutedMutations + 1;
92     Energy -= AbdIncidence * logl(AbdIncidence);
93     SumIncidence += AbdIncidence;
94 
95     // Normalize.
96     if (SumIncidence != 0)
97       Energy = (Energy / SumIncidence) + logl(SumIncidence);
98 
99     if (ScalePerExecTime) {
100       // Scaling to favor inputs with lower execution time.
101       uint32_t PerfScore = 100;
102       if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 10)
103         PerfScore = 10;
104       else if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 4)
105         PerfScore = 25;
106       else if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 2)
107         PerfScore = 50;
108       else if (TimeOfUnit.count() * 3 > AverageUnitExecutionTime.count() * 4)
109         PerfScore = 75;
110       else if (TimeOfUnit.count() * 4 < AverageUnitExecutionTime.count())
111         PerfScore = 300;
112       else if (TimeOfUnit.count() * 3 < AverageUnitExecutionTime.count())
113         PerfScore = 200;
114       else if (TimeOfUnit.count() * 2 < AverageUnitExecutionTime.count())
115         PerfScore = 150;
116 
117       Energy *= PerfScore;
118     }
119   }
120 
121   // Increment the frequency of the feature Idx.
122   void UpdateFeatureFrequency(uint32_t Idx) {
123     NeedsEnergyUpdate = true;
124 
125     // The local feature frequencies is an ordered vector of pairs.
126     // If there are no local feature frequencies, push_back preserves order.
127     // Set the feature frequency for feature Idx32 to 1.
128     if (FeatureFreqs.empty()) {
129       FeatureFreqs.push_back(std::pair<uint32_t, uint16_t>(Idx, 1));
130       return;
131     }
132 
133     // Binary search over local feature frequencies sorted by index.
134     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
135                                   std::pair<uint32_t, uint16_t>(Idx, 0));
136 
137     // If feature Idx32 already exists, increment its frequency.
138     // Otherwise, insert a new pair right after the next lower index.
139     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
140       Lower->second++;
141     } else {
142       FeatureFreqs.insert(Lower, std::pair<uint32_t, uint16_t>(Idx, 1));
143     }
144   }
145 };
146 
147 struct EntropicOptions {
148   bool Enabled;
149   size_t NumberOfRarestFeatures;
150   size_t FeatureFrequencyThreshold;
151   bool ScalePerExecTime;
152 };
153 
154 class InputCorpus {
155   static const uint32_t kFeatureSetSize = 1 << 21;
156   static const uint8_t kMaxMutationFactor = 20;
157   static const size_t kSparseEnergyUpdates = 100;
158 
159   size_t NumExecutedMutations = 0;
160 
161   EntropicOptions Entropic;
162 
163 public:
164   InputCorpus(const std::string &OutputCorpus, EntropicOptions Entropic)
165       : Entropic(Entropic), OutputCorpus(OutputCorpus) {
166     memset(InputSizesPerFeature, 0, sizeof(InputSizesPerFeature));
167     memset(SmallestElementPerFeature, 0, sizeof(SmallestElementPerFeature));
168   }
169   ~InputCorpus() {
170     for (auto II : Inputs)
171       delete II;
172   }
173   size_t size() const { return Inputs.size(); }
174   size_t SizeInBytes() const {
175     size_t Res = 0;
176     for (auto II : Inputs)
177       Res += II->U.size();
178     return Res;
179   }
180   size_t NumActiveUnits() const {
181     size_t Res = 0;
182     for (auto II : Inputs)
183       Res += !II->U.empty();
184     return Res;
185   }
186   size_t MaxInputSize() const {
187     size_t Res = 0;
188     for (auto II : Inputs)
189         Res = std::max(Res, II->U.size());
190     return Res;
191   }
192   void IncrementNumExecutedMutations() { NumExecutedMutations++; }
193 
194   size_t NumInputsThatTouchFocusFunction() {
195     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
196       return II->HasFocusFunction;
197     });
198   }
199 
200   size_t NumInputsWithDataFlowTrace() {
201     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
202       return !II->DataFlowTraceForFocusFunction.empty();
203     });
204   }
205 
206   bool empty() const { return Inputs.empty(); }
207   const Unit &operator[] (size_t Idx) const { return Inputs[Idx]->U; }
208   InputInfo *AddToCorpus(const Unit &U, size_t NumFeatures, bool MayDeleteFile,
209                          bool HasFocusFunction, bool NeverReduce,
210                          std::chrono::microseconds TimeOfUnit,
211                          const Vector<uint32_t> &FeatureSet,
212                          const DataFlowTrace &DFT, const InputInfo *BaseII) {
213     assert(!U.empty());
214     if (FeatureDebug)
215       Printf("ADD_TO_CORPUS %zd NF %zd\n", Inputs.size(), NumFeatures);
216     Inputs.push_back(new InputInfo());
217     InputInfo &II = *Inputs.back();
218     II.U = U;
219     II.NumFeatures = NumFeatures;
220     II.NeverReduce = NeverReduce;
221     II.TimeOfUnit = TimeOfUnit;
222     II.MayDeleteFile = MayDeleteFile;
223     II.UniqFeatureSet = FeatureSet;
224     II.HasFocusFunction = HasFocusFunction;
225     // Assign maximal energy to the new seed.
226     II.Energy = RareFeatures.empty() ? 1.0 : log(RareFeatures.size());
227     II.SumIncidence = RareFeatures.size();
228     II.NeedsEnergyUpdate = false;
229     std::sort(II.UniqFeatureSet.begin(), II.UniqFeatureSet.end());
230     ComputeSHA1(U.data(), U.size(), II.Sha1);
231     auto Sha1Str = Sha1ToString(II.Sha1);
232     Hashes.insert(Sha1Str);
233     if (HasFocusFunction)
234       if (auto V = DFT.Get(Sha1Str))
235         II.DataFlowTraceForFocusFunction = *V;
236     // This is a gross heuristic.
237     // Ideally, when we add an element to a corpus we need to know its DFT.
238     // But if we don't, we'll use the DFT of its base input.
239     if (II.DataFlowTraceForFocusFunction.empty() && BaseII)
240       II.DataFlowTraceForFocusFunction = BaseII->DataFlowTraceForFocusFunction;
241     DistributionNeedsUpdate = true;
242     PrintCorpus();
243     // ValidateFeatureSet();
244     return &II;
245   }
246 
247   // Debug-only
248   void PrintUnit(const Unit &U) {
249     if (!FeatureDebug) return;
250     for (uint8_t C : U) {
251       if (C != 'F' && C != 'U' && C != 'Z')
252         C = '.';
253       Printf("%c", C);
254     }
255   }
256 
257   // Debug-only
258   void PrintFeatureSet(const Vector<uint32_t> &FeatureSet) {
259     if (!FeatureDebug) return;
260     Printf("{");
261     for (uint32_t Feature: FeatureSet)
262       Printf("%u,", Feature);
263     Printf("}");
264   }
265 
266   // Debug-only
267   void PrintCorpus() {
268     if (!FeatureDebug) return;
269     Printf("======= CORPUS:\n");
270     int i = 0;
271     for (auto II : Inputs) {
272       if (std::find(II->U.begin(), II->U.end(), 'F') != II->U.end()) {
273         Printf("[%2d] ", i);
274         Printf("%s sz=%zd ", Sha1ToString(II->Sha1).c_str(), II->U.size());
275         PrintUnit(II->U);
276         Printf(" ");
277         PrintFeatureSet(II->UniqFeatureSet);
278         Printf("\n");
279       }
280       i++;
281     }
282   }
283 
284   void Replace(InputInfo *II, const Unit &U) {
285     assert(II->U.size() > U.size());
286     Hashes.erase(Sha1ToString(II->Sha1));
287     DeleteFile(*II);
288     ComputeSHA1(U.data(), U.size(), II->Sha1);
289     Hashes.insert(Sha1ToString(II->Sha1));
290     II->U = U;
291     II->Reduced = true;
292     DistributionNeedsUpdate = true;
293   }
294 
295   bool HasUnit(const Unit &U) { return Hashes.count(Hash(U)); }
296   bool HasUnit(const std::string &H) { return Hashes.count(H); }
297   InputInfo &ChooseUnitToMutate(Random &Rand) {
298     InputInfo &II = *Inputs[ChooseUnitIdxToMutate(Rand)];
299     assert(!II.U.empty());
300     return II;
301   }
302 
303   InputInfo &ChooseUnitToCrossOverWith(Random &Rand, bool UniformDist) {
304     if (!UniformDist) {
305       return ChooseUnitToMutate(Rand);
306     }
307     InputInfo &II = *Inputs[Rand(Inputs.size())];
308     assert(!II.U.empty());
309     return II;
310   }
311 
312   // Returns an index of random unit from the corpus to mutate.
313   size_t ChooseUnitIdxToMutate(Random &Rand) {
314     UpdateCorpusDistribution(Rand);
315     size_t Idx = static_cast<size_t>(CorpusDistribution(Rand));
316     assert(Idx < Inputs.size());
317     return Idx;
318   }
319 
320   void PrintStats() {
321     for (size_t i = 0; i < Inputs.size(); i++) {
322       const auto &II = *Inputs[i];
323       Printf("  [% 3zd %s] sz: % 5zd runs: % 5zd succ: % 5zd focus: %d\n", i,
324              Sha1ToString(II.Sha1).c_str(), II.U.size(),
325              II.NumExecutedMutations, II.NumSuccessfullMutations, II.HasFocusFunction);
326     }
327   }
328 
329   void PrintFeatureSet() {
330     for (size_t i = 0; i < kFeatureSetSize; i++) {
331       if(size_t Sz = GetFeature(i))
332         Printf("[%zd: id %zd sz%zd] ", i, SmallestElementPerFeature[i], Sz);
333     }
334     Printf("\n\t");
335     for (size_t i = 0; i < Inputs.size(); i++)
336       if (size_t N = Inputs[i]->NumFeatures)
337         Printf(" %zd=>%zd ", i, N);
338     Printf("\n");
339   }
340 
341   void DeleteFile(const InputInfo &II) {
342     if (!OutputCorpus.empty() && II.MayDeleteFile)
343       RemoveFile(DirPlusFile(OutputCorpus, Sha1ToString(II.Sha1)));
344   }
345 
346   void DeleteInput(size_t Idx) {
347     InputInfo &II = *Inputs[Idx];
348     DeleteFile(II);
349     Unit().swap(II.U);
350     II.Energy = 0.0;
351     II.NeedsEnergyUpdate = false;
352     DistributionNeedsUpdate = true;
353     if (FeatureDebug)
354       Printf("EVICTED %zd\n", Idx);
355   }
356 
357   void AddRareFeature(uint32_t Idx) {
358     // Maintain *at least* TopXRarestFeatures many rare features
359     // and all features with a frequency below ConsideredRare.
360     // Remove all other features.
361     while (RareFeatures.size() > Entropic.NumberOfRarestFeatures &&
362            FreqOfMostAbundantRareFeature > Entropic.FeatureFrequencyThreshold) {
363 
364       // Find most and second most abbundant feature.
365       uint32_t MostAbundantRareFeatureIndices[2] = {RareFeatures[0],
366                                                     RareFeatures[0]};
367       size_t Delete = 0;
368       for (size_t i = 0; i < RareFeatures.size(); i++) {
369         uint32_t Idx2 = RareFeatures[i];
370         if (GlobalFeatureFreqs[Idx2] >=
371             GlobalFeatureFreqs[MostAbundantRareFeatureIndices[0]]) {
372           MostAbundantRareFeatureIndices[1] = MostAbundantRareFeatureIndices[0];
373           MostAbundantRareFeatureIndices[0] = Idx2;
374           Delete = i;
375         }
376       }
377 
378       // Remove most abundant rare feature.
379       RareFeatures[Delete] = RareFeatures.back();
380       RareFeatures.pop_back();
381 
382       for (auto II : Inputs) {
383         if (II->DeleteFeatureFreq(MostAbundantRareFeatureIndices[0]))
384           II->NeedsEnergyUpdate = true;
385       }
386 
387       // Set 2nd most abundant as the new most abundant feature count.
388       FreqOfMostAbundantRareFeature =
389           GlobalFeatureFreqs[MostAbundantRareFeatureIndices[1]];
390     }
391 
392     // Add rare feature, handle collisions, and update energy.
393     RareFeatures.push_back(Idx);
394     GlobalFeatureFreqs[Idx] = 0;
395     for (auto II : Inputs) {
396       II->DeleteFeatureFreq(Idx);
397 
398       // Apply add-one smoothing to this locally undiscovered feature.
399       // Zero energy seeds will never be fuzzed and remain zero energy.
400       if (II->Energy > 0.0) {
401         II->SumIncidence += 1;
402         II->Energy += logl(II->SumIncidence) / II->SumIncidence;
403       }
404     }
405 
406     DistributionNeedsUpdate = true;
407   }
408 
409   bool AddFeature(size_t Idx, uint32_t NewSize, bool Shrink) {
410     assert(NewSize);
411     Idx = Idx % kFeatureSetSize;
412     uint32_t OldSize = GetFeature(Idx);
413     if (OldSize == 0 || (Shrink && OldSize > NewSize)) {
414       if (OldSize > 0) {
415         size_t OldIdx = SmallestElementPerFeature[Idx];
416         InputInfo &II = *Inputs[OldIdx];
417         assert(II.NumFeatures > 0);
418         II.NumFeatures--;
419         if (II.NumFeatures == 0)
420           DeleteInput(OldIdx);
421       } else {
422         NumAddedFeatures++;
423         if (Entropic.Enabled)
424           AddRareFeature((uint32_t)Idx);
425       }
426       NumUpdatedFeatures++;
427       if (FeatureDebug)
428         Printf("ADD FEATURE %zd sz %d\n", Idx, NewSize);
429       SmallestElementPerFeature[Idx] = Inputs.size();
430       InputSizesPerFeature[Idx] = NewSize;
431       return true;
432     }
433     return false;
434   }
435 
436   // Increment frequency of feature Idx globally and locally.
437   void UpdateFeatureFrequency(InputInfo *II, size_t Idx) {
438     uint32_t Idx32 = Idx % kFeatureSetSize;
439 
440     // Saturated increment.
441     if (GlobalFeatureFreqs[Idx32] == 0xFFFF)
442       return;
443     uint16_t Freq = GlobalFeatureFreqs[Idx32]++;
444 
445     // Skip if abundant.
446     if (Freq > FreqOfMostAbundantRareFeature ||
447         std::find(RareFeatures.begin(), RareFeatures.end(), Idx32) ==
448             RareFeatures.end())
449       return;
450 
451     // Update global frequencies.
452     if (Freq == FreqOfMostAbundantRareFeature)
453       FreqOfMostAbundantRareFeature++;
454 
455     // Update local frequencies.
456     if (II)
457       II->UpdateFeatureFrequency(Idx32);
458   }
459 
460   size_t NumFeatures() const { return NumAddedFeatures; }
461   size_t NumFeatureUpdates() const { return NumUpdatedFeatures; }
462 
463 private:
464 
465   static const bool FeatureDebug = false;
466 
467   size_t GetFeature(size_t Idx) const { return InputSizesPerFeature[Idx]; }
468 
469   void ValidateFeatureSet() {
470     if (FeatureDebug)
471       PrintFeatureSet();
472     for (size_t Idx = 0; Idx < kFeatureSetSize; Idx++)
473       if (GetFeature(Idx))
474         Inputs[SmallestElementPerFeature[Idx]]->Tmp++;
475     for (auto II: Inputs) {
476       if (II->Tmp != II->NumFeatures)
477         Printf("ZZZ %zd %zd\n", II->Tmp, II->NumFeatures);
478       assert(II->Tmp == II->NumFeatures);
479       II->Tmp = 0;
480     }
481   }
482 
483   // Updates the probability distribution for the units in the corpus.
484   // Must be called whenever the corpus or unit weights are changed.
485   //
486   // Hypothesis: inputs that maximize information about globally rare features
487   // are interesting.
488   void UpdateCorpusDistribution(Random &Rand) {
489     // Skip update if no seeds or rare features were added/deleted.
490     // Sparse updates for local change of feature frequencies,
491     // i.e., randomly do not skip.
492     if (!DistributionNeedsUpdate &&
493         (!Entropic.Enabled || Rand(kSparseEnergyUpdates)))
494       return;
495 
496     DistributionNeedsUpdate = false;
497 
498     size_t N = Inputs.size();
499     assert(N);
500     Intervals.resize(N + 1);
501     Weights.resize(N);
502     std::iota(Intervals.begin(), Intervals.end(), 0);
503 
504     std::chrono::microseconds AverageUnitExecutionTime(0);
505     for (auto II : Inputs) {
506       AverageUnitExecutionTime += II->TimeOfUnit;
507     }
508     AverageUnitExecutionTime /= N;
509 
510     bool VanillaSchedule = true;
511     if (Entropic.Enabled) {
512       for (auto II : Inputs) {
513         if (II->NeedsEnergyUpdate && II->Energy != 0.0) {
514           II->NeedsEnergyUpdate = false;
515           II->UpdateEnergy(RareFeatures.size(), Entropic.ScalePerExecTime,
516                            AverageUnitExecutionTime);
517         }
518       }
519 
520       for (size_t i = 0; i < N; i++) {
521 
522         if (Inputs[i]->NumFeatures == 0) {
523           // If the seed doesn't represent any features, assign zero energy.
524           Weights[i] = 0.;
525         } else if (Inputs[i]->NumExecutedMutations / kMaxMutationFactor >
526                    NumExecutedMutations / Inputs.size()) {
527           // If the seed was fuzzed a lot more than average, assign zero energy.
528           Weights[i] = 0.;
529         } else {
530           // Otherwise, simply assign the computed energy.
531           Weights[i] = Inputs[i]->Energy;
532         }
533 
534         // If energy for all seeds is zero, fall back to vanilla schedule.
535         if (Weights[i] > 0.0)
536           VanillaSchedule = false;
537       }
538     }
539 
540     if (VanillaSchedule) {
541       for (size_t i = 0; i < N; i++)
542         Weights[i] = Inputs[i]->NumFeatures
543                          ? (i + 1) * (Inputs[i]->HasFocusFunction ? 1000 : 1)
544                          : 0.;
545     }
546 
547     if (FeatureDebug) {
548       for (size_t i = 0; i < N; i++)
549         Printf("%zd ", Inputs[i]->NumFeatures);
550       Printf("SCORE\n");
551       for (size_t i = 0; i < N; i++)
552         Printf("%f ", Weights[i]);
553       Printf("Weights\n");
554     }
555     CorpusDistribution = std::piecewise_constant_distribution<double>(
556         Intervals.begin(), Intervals.end(), Weights.begin());
557   }
558   std::piecewise_constant_distribution<double> CorpusDistribution;
559 
560   Vector<double> Intervals;
561   Vector<double> Weights;
562 
563   std::unordered_set<std::string> Hashes;
564   Vector<InputInfo*> Inputs;
565 
566   size_t NumAddedFeatures = 0;
567   size_t NumUpdatedFeatures = 0;
568   uint32_t InputSizesPerFeature[kFeatureSetSize];
569   uint32_t SmallestElementPerFeature[kFeatureSetSize];
570 
571   bool DistributionNeedsUpdate = true;
572   uint16_t FreqOfMostAbundantRareFeature = 0;
573   uint16_t GlobalFeatureFreqs[kFeatureSetSize] = {};
574   Vector<uint32_t> RareFeatures;
575 
576   std::string OutputCorpus;
577 };
578 
579 }  // namespace fuzzer
580 
581 #endif  // LLVM_FUZZER_CORPUS
582