xref: /freebsd/contrib/llvm-project/compiler-rt/lib/fuzzer/FuzzerCorpus.h (revision 5f757f3ff9144b609b3c433dfd370cc6bdc191ad)
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 <bitset>
22 #include <chrono>
23 #include <numeric>
24 #include <random>
25 #include <unordered_set>
26 
27 namespace fuzzer {
28 
29 struct InputInfo {
30   Unit U;  // The actual input data.
31   std::chrono::microseconds TimeOfUnit;
32   uint8_t Sha1[kSHA1NumBytes];  // Checksum.
33   // Number of features that this input has and no smaller input has.
34   size_t NumFeatures = 0;
35   size_t Tmp = 0; // Used by ValidateFeatureSet.
36   // Stats.
37   size_t NumExecutedMutations = 0;
38   size_t NumSuccessfullMutations = 0;
39   bool NeverReduce = false;
40   bool MayDeleteFile = false;
41   bool Reduced = false;
42   bool HasFocusFunction = false;
43   std::vector<uint32_t> UniqFeatureSet;
44   std::vector<uint8_t> DataFlowTraceForFocusFunction;
45   // Power schedule.
46   bool NeedsEnergyUpdate = false;
47   double Energy = 0.0;
48   double SumIncidence = 0.0;
49   std::vector<std::pair<uint32_t, uint16_t>> FeatureFreqs;
50 
51   // Delete feature Idx and its frequency from FeatureFreqs.
52   bool DeleteFeatureFreq(uint32_t Idx) {
53     if (FeatureFreqs.empty())
54       return false;
55 
56     // Binary search over local feature frequencies sorted by index.
57     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
58                                   std::pair<uint32_t, uint16_t>(Idx, 0));
59 
60     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
61       FeatureFreqs.erase(Lower);
62       return true;
63     }
64     return false;
65   }
66 
67   // Assign more energy to a high-entropy seed, i.e., that reveals more
68   // information about the globally rare features in the neighborhood of the
69   // seed. Since we do not know the entropy of a seed that has never been
70   // executed we assign fresh seeds maximum entropy and let II->Energy approach
71   // the true entropy from above. If ScalePerExecTime is true, the computed
72   // entropy is scaled based on how fast this input executes compared to the
73   // average execution time of inputs. The faster an input executes, the more
74   // energy gets assigned to the input.
75   void UpdateEnergy(size_t GlobalNumberOfFeatures, bool ScalePerExecTime,
76                     std::chrono::microseconds AverageUnitExecutionTime) {
77     Energy = 0.0;
78     SumIncidence = 0.0;
79 
80     // Apply add-one smoothing to locally discovered features.
81     for (const auto &F : FeatureFreqs) {
82       double LocalIncidence = F.second + 1;
83       Energy -= LocalIncidence * log(LocalIncidence);
84       SumIncidence += LocalIncidence;
85     }
86 
87     // Apply add-one smoothing to locally undiscovered features.
88     //   PreciseEnergy -= 0; // since log(1.0) == 0)
89     SumIncidence +=
90         static_cast<double>(GlobalNumberOfFeatures - FeatureFreqs.size());
91 
92     // Add a single locally abundant feature apply add-one smoothing.
93     double AbdIncidence = static_cast<double>(NumExecutedMutations + 1);
94     Energy -= AbdIncidence * log(AbdIncidence);
95     SumIncidence += AbdIncidence;
96 
97     // Normalize.
98     if (SumIncidence != 0)
99       Energy = Energy / SumIncidence + log(SumIncidence);
100 
101     if (ScalePerExecTime) {
102       // Scaling to favor inputs with lower execution time.
103       uint32_t PerfScore = 100;
104       if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 10)
105         PerfScore = 10;
106       else if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 4)
107         PerfScore = 25;
108       else if (TimeOfUnit.count() > AverageUnitExecutionTime.count() * 2)
109         PerfScore = 50;
110       else if (TimeOfUnit.count() * 3 > AverageUnitExecutionTime.count() * 4)
111         PerfScore = 75;
112       else if (TimeOfUnit.count() * 4 < AverageUnitExecutionTime.count())
113         PerfScore = 300;
114       else if (TimeOfUnit.count() * 3 < AverageUnitExecutionTime.count())
115         PerfScore = 200;
116       else if (TimeOfUnit.count() * 2 < AverageUnitExecutionTime.count())
117         PerfScore = 150;
118 
119       Energy *= PerfScore;
120     }
121   }
122 
123   // Increment the frequency of the feature Idx.
124   void UpdateFeatureFrequency(uint32_t Idx) {
125     NeedsEnergyUpdate = true;
126 
127     // The local feature frequencies is an ordered vector of pairs.
128     // If there are no local feature frequencies, push_back preserves order.
129     // Set the feature frequency for feature Idx32 to 1.
130     if (FeatureFreqs.empty()) {
131       FeatureFreqs.push_back(std::pair<uint32_t, uint16_t>(Idx, 1));
132       return;
133     }
134 
135     // Binary search over local feature frequencies sorted by index.
136     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
137                                   std::pair<uint32_t, uint16_t>(Idx, 0));
138 
139     // If feature Idx32 already exists, increment its frequency.
140     // Otherwise, insert a new pair right after the next lower index.
141     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
142       Lower->second++;
143     } else {
144       FeatureFreqs.insert(Lower, std::pair<uint32_t, uint16_t>(Idx, 1));
145     }
146   }
147 };
148 
149 struct EntropicOptions {
150   bool Enabled;
151   size_t NumberOfRarestFeatures;
152   size_t FeatureFrequencyThreshold;
153   bool ScalePerExecTime;
154 };
155 
156 class InputCorpus {
157   static const uint32_t kFeatureSetSize = 1 << 21;
158   static const uint8_t kMaxMutationFactor = 20;
159   static const size_t kSparseEnergyUpdates = 100;
160 
161   size_t NumExecutedMutations = 0;
162 
163   EntropicOptions Entropic;
164 
165 public:
166   InputCorpus(const std::string &OutputCorpus, EntropicOptions Entropic)
167       : Entropic(Entropic), OutputCorpus(OutputCorpus) {
168     memset(InputSizesPerFeature, 0, sizeof(InputSizesPerFeature));
169     memset(SmallestElementPerFeature, 0, sizeof(SmallestElementPerFeature));
170   }
171   ~InputCorpus() {
172     for (auto II : Inputs)
173       delete II;
174   }
175   size_t size() const { return Inputs.size(); }
176   size_t SizeInBytes() const {
177     size_t Res = 0;
178     for (auto II : Inputs)
179       Res += II->U.size();
180     return Res;
181   }
182   size_t NumActiveUnits() const {
183     size_t Res = 0;
184     for (auto II : Inputs)
185       Res += !II->U.empty();
186     return Res;
187   }
188   size_t MaxInputSize() const {
189     size_t Res = 0;
190     for (auto II : Inputs)
191         Res = std::max(Res, II->U.size());
192     return Res;
193   }
194   void IncrementNumExecutedMutations() { NumExecutedMutations++; }
195 
196   size_t NumInputsThatTouchFocusFunction() {
197     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
198       return II->HasFocusFunction;
199     });
200   }
201 
202   size_t NumInputsWithDataFlowTrace() {
203     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
204       return !II->DataFlowTraceForFocusFunction.empty();
205     });
206   }
207 
208   bool empty() const { return Inputs.empty(); }
209   const Unit &operator[] (size_t Idx) const { return Inputs[Idx]->U; }
210   InputInfo *AddToCorpus(const Unit &U, size_t NumFeatures, bool MayDeleteFile,
211                          bool HasFocusFunction, bool NeverReduce,
212                          std::chrono::microseconds TimeOfUnit,
213                          const std::vector<uint32_t> &FeatureSet,
214                          const DataFlowTrace &DFT, const InputInfo *BaseII) {
215     assert(!U.empty());
216     if (FeatureDebug)
217       Printf("ADD_TO_CORPUS %zd NF %zd\n", Inputs.size(), NumFeatures);
218     // Inputs.size() is cast to uint32_t below.
219     assert(Inputs.size() < std::numeric_limits<uint32_t>::max());
220     Inputs.push_back(new InputInfo());
221     InputInfo &II = *Inputs.back();
222     II.U = U;
223     II.NumFeatures = NumFeatures;
224     II.NeverReduce = NeverReduce;
225     II.TimeOfUnit = TimeOfUnit;
226     II.MayDeleteFile = MayDeleteFile;
227     II.UniqFeatureSet = FeatureSet;
228     II.HasFocusFunction = HasFocusFunction;
229     // Assign maximal energy to the new seed.
230     II.Energy = RareFeatures.empty() ? 1.0 : log(RareFeatures.size());
231     II.SumIncidence = static_cast<double>(RareFeatures.size());
232     II.NeedsEnergyUpdate = false;
233     std::sort(II.UniqFeatureSet.begin(), II.UniqFeatureSet.end());
234     ComputeSHA1(U.data(), U.size(), II.Sha1);
235     auto Sha1Str = Sha1ToString(II.Sha1);
236     Hashes.insert(Sha1Str);
237     if (HasFocusFunction)
238       if (auto V = DFT.Get(Sha1Str))
239         II.DataFlowTraceForFocusFunction = *V;
240     // This is a gross heuristic.
241     // Ideally, when we add an element to a corpus we need to know its DFT.
242     // But if we don't, we'll use the DFT of its base input.
243     if (II.DataFlowTraceForFocusFunction.empty() && BaseII)
244       II.DataFlowTraceForFocusFunction = BaseII->DataFlowTraceForFocusFunction;
245     DistributionNeedsUpdate = true;
246     PrintCorpus();
247     // ValidateFeatureSet();
248     return &II;
249   }
250 
251   // Debug-only
252   void PrintUnit(const Unit &U) {
253     if (!FeatureDebug) return;
254     for (uint8_t C : U) {
255       if (C != 'F' && C != 'U' && C != 'Z')
256         C = '.';
257       Printf("%c", C);
258     }
259   }
260 
261   // Debug-only
262   void PrintFeatureSet(const std::vector<uint32_t> &FeatureSet) {
263     if (!FeatureDebug) return;
264     Printf("{");
265     for (uint32_t Feature: FeatureSet)
266       Printf("%u,", Feature);
267     Printf("}");
268   }
269 
270   // Debug-only
271   void PrintCorpus() {
272     if (!FeatureDebug) return;
273     Printf("======= CORPUS:\n");
274     int i = 0;
275     for (auto II : Inputs) {
276       if (std::find(II->U.begin(), II->U.end(), 'F') != II->U.end()) {
277         Printf("[%2d] ", i);
278         Printf("%s sz=%zd ", Sha1ToString(II->Sha1).c_str(), II->U.size());
279         PrintUnit(II->U);
280         Printf(" ");
281         PrintFeatureSet(II->UniqFeatureSet);
282         Printf("\n");
283       }
284       i++;
285     }
286   }
287 
288   void Replace(InputInfo *II, const Unit &U,
289                std::chrono::microseconds TimeOfUnit) {
290     assert(II->U.size() > U.size());
291     Hashes.erase(Sha1ToString(II->Sha1));
292     DeleteFile(*II);
293     ComputeSHA1(U.data(), U.size(), II->Sha1);
294     Hashes.insert(Sha1ToString(II->Sha1));
295     II->U = U;
296     II->Reduced = true;
297     II->TimeOfUnit = TimeOfUnit;
298     DistributionNeedsUpdate = true;
299   }
300 
301   bool HasUnit(const Unit &U) { return Hashes.count(Hash(U)); }
302   bool HasUnit(const std::string &H) { return Hashes.count(H); }
303   InputInfo &ChooseUnitToMutate(Random &Rand) {
304     InputInfo &II = *Inputs[ChooseUnitIdxToMutate(Rand)];
305     assert(!II.U.empty());
306     return II;
307   }
308 
309   InputInfo &ChooseUnitToCrossOverWith(Random &Rand, bool UniformDist) {
310     if (!UniformDist) {
311       return ChooseUnitToMutate(Rand);
312     }
313     InputInfo &II = *Inputs[Rand(Inputs.size())];
314     assert(!II.U.empty());
315     return II;
316   }
317 
318   // Returns an index of random unit from the corpus to mutate.
319   size_t ChooseUnitIdxToMutate(Random &Rand) {
320     UpdateCorpusDistribution(Rand);
321     size_t Idx = static_cast<size_t>(CorpusDistribution(Rand));
322     assert(Idx < Inputs.size());
323     return Idx;
324   }
325 
326   void PrintStats() {
327     for (size_t i = 0; i < Inputs.size(); i++) {
328       const auto &II = *Inputs[i];
329       Printf("  [% 3zd %s] sz: % 5zd runs: % 5zd succ: % 5zd focus: %d\n", i,
330              Sha1ToString(II.Sha1).c_str(), II.U.size(),
331              II.NumExecutedMutations, II.NumSuccessfullMutations,
332              II.HasFocusFunction);
333     }
334   }
335 
336   void PrintFeatureSet() {
337     for (size_t i = 0; i < kFeatureSetSize; i++) {
338       if(size_t Sz = GetFeature(i))
339         Printf("[%zd: id %zd sz%zd] ", i, SmallestElementPerFeature[i], Sz);
340     }
341     Printf("\n\t");
342     for (size_t i = 0; i < Inputs.size(); i++)
343       if (size_t N = Inputs[i]->NumFeatures)
344         Printf(" %zd=>%zd ", i, N);
345     Printf("\n");
346   }
347 
348   void DeleteFile(const InputInfo &II) {
349     if (!OutputCorpus.empty() && II.MayDeleteFile)
350       RemoveFile(DirPlusFile(OutputCorpus, Sha1ToString(II.Sha1)));
351   }
352 
353   void DeleteInput(size_t Idx) {
354     InputInfo &II = *Inputs[Idx];
355     DeleteFile(II);
356     Unit().swap(II.U);
357     II.Energy = 0.0;
358     II.NeedsEnergyUpdate = false;
359     DistributionNeedsUpdate = true;
360     if (FeatureDebug)
361       Printf("EVICTED %zd\n", Idx);
362   }
363 
364   void AddRareFeature(uint32_t Idx) {
365     // Maintain *at least* TopXRarestFeatures many rare features
366     // and all features with a frequency below ConsideredRare.
367     // Remove all other features.
368     while (RareFeatures.size() > Entropic.NumberOfRarestFeatures &&
369            FreqOfMostAbundantRareFeature > Entropic.FeatureFrequencyThreshold) {
370 
371       // Find most and second most abbundant feature.
372       uint32_t MostAbundantRareFeatureIndices[2] = {RareFeatures[0],
373                                                     RareFeatures[0]};
374       size_t Delete = 0;
375       for (size_t i = 0; i < RareFeatures.size(); i++) {
376         uint32_t Idx2 = RareFeatures[i];
377         if (GlobalFeatureFreqs[Idx2] >=
378             GlobalFeatureFreqs[MostAbundantRareFeatureIndices[0]]) {
379           MostAbundantRareFeatureIndices[1] = MostAbundantRareFeatureIndices[0];
380           MostAbundantRareFeatureIndices[0] = Idx2;
381           Delete = i;
382         }
383       }
384 
385       // Remove most abundant rare feature.
386       IsRareFeature[Delete] = false;
387       RareFeatures[Delete] = RareFeatures.back();
388       RareFeatures.pop_back();
389 
390       for (auto II : Inputs) {
391         if (II->DeleteFeatureFreq(MostAbundantRareFeatureIndices[0]))
392           II->NeedsEnergyUpdate = true;
393       }
394 
395       // Set 2nd most abundant as the new most abundant feature count.
396       FreqOfMostAbundantRareFeature =
397           GlobalFeatureFreqs[MostAbundantRareFeatureIndices[1]];
398     }
399 
400     // Add rare feature, handle collisions, and update energy.
401     RareFeatures.push_back(Idx);
402     IsRareFeature[Idx] = true;
403     GlobalFeatureFreqs[Idx] = 0;
404     for (auto II : Inputs) {
405       II->DeleteFeatureFreq(Idx);
406 
407       // Apply add-one smoothing to this locally undiscovered feature.
408       // Zero energy seeds will never be fuzzed and remain zero energy.
409       if (II->Energy > 0.0) {
410         II->SumIncidence += 1;
411         II->Energy += log(II->SumIncidence) / II->SumIncidence;
412       }
413     }
414 
415     DistributionNeedsUpdate = true;
416   }
417 
418   bool AddFeature(size_t Idx, uint32_t NewSize, bool Shrink) {
419     assert(NewSize);
420     Idx = Idx % kFeatureSetSize;
421     uint32_t OldSize = GetFeature(Idx);
422     if (OldSize == 0 || (Shrink && OldSize > NewSize)) {
423       if (OldSize > 0) {
424         size_t OldIdx = SmallestElementPerFeature[Idx];
425         InputInfo &II = *Inputs[OldIdx];
426         assert(II.NumFeatures > 0);
427         II.NumFeatures--;
428         if (II.NumFeatures == 0)
429           DeleteInput(OldIdx);
430       } else {
431         NumAddedFeatures++;
432         if (Entropic.Enabled)
433           AddRareFeature((uint32_t)Idx);
434       }
435       NumUpdatedFeatures++;
436       if (FeatureDebug)
437         Printf("ADD FEATURE %zd sz %d\n", Idx, NewSize);
438       // Inputs.size() is guaranteed to be less than UINT32_MAX by AddToCorpus.
439       SmallestElementPerFeature[Idx] = static_cast<uint32_t>(Inputs.size());
440       InputSizesPerFeature[Idx] = NewSize;
441       return true;
442     }
443     return false;
444   }
445 
446   // Increment frequency of feature Idx globally and locally.
447   void UpdateFeatureFrequency(InputInfo *II, size_t Idx) {
448     uint32_t Idx32 = Idx % kFeatureSetSize;
449 
450     // Saturated increment.
451     if (GlobalFeatureFreqs[Idx32] == 0xFFFF)
452       return;
453     uint16_t Freq = GlobalFeatureFreqs[Idx32]++;
454 
455     // Skip if abundant.
456     if (Freq > FreqOfMostAbundantRareFeature || !IsRareFeature[Idx32])
457       return;
458 
459     // Update global frequencies.
460     if (Freq == FreqOfMostAbundantRareFeature)
461       FreqOfMostAbundantRareFeature++;
462 
463     // Update local frequencies.
464     if (II)
465       II->UpdateFeatureFrequency(Idx32);
466   }
467 
468   size_t NumFeatures() const { return NumAddedFeatures; }
469   size_t NumFeatureUpdates() const { return NumUpdatedFeatures; }
470 
471 private:
472 
473   static const bool FeatureDebug = false;
474 
475   uint32_t GetFeature(size_t Idx) const { return InputSizesPerFeature[Idx]; }
476 
477   void ValidateFeatureSet() {
478     if (FeatureDebug)
479       PrintFeatureSet();
480     for (size_t Idx = 0; Idx < kFeatureSetSize; Idx++)
481       if (GetFeature(Idx))
482         Inputs[SmallestElementPerFeature[Idx]]->Tmp++;
483     for (auto II: Inputs) {
484       if (II->Tmp != II->NumFeatures)
485         Printf("ZZZ %zd %zd\n", II->Tmp, II->NumFeatures);
486       assert(II->Tmp == II->NumFeatures);
487       II->Tmp = 0;
488     }
489   }
490 
491   // Updates the probability distribution for the units in the corpus.
492   // Must be called whenever the corpus or unit weights are changed.
493   //
494   // Hypothesis: inputs that maximize information about globally rare features
495   // are interesting.
496   void UpdateCorpusDistribution(Random &Rand) {
497     // Skip update if no seeds or rare features were added/deleted.
498     // Sparse updates for local change of feature frequencies,
499     // i.e., randomly do not skip.
500     if (!DistributionNeedsUpdate &&
501         (!Entropic.Enabled || Rand(kSparseEnergyUpdates)))
502       return;
503 
504     DistributionNeedsUpdate = false;
505 
506     size_t N = Inputs.size();
507     assert(N);
508     Intervals.resize(N + 1);
509     Weights.resize(N);
510     std::iota(Intervals.begin(), Intervals.end(), 0);
511 
512     std::chrono::microseconds AverageUnitExecutionTime(0);
513     for (auto II : Inputs) {
514       AverageUnitExecutionTime += II->TimeOfUnit;
515     }
516     AverageUnitExecutionTime /= N;
517 
518     bool VanillaSchedule = true;
519     if (Entropic.Enabled) {
520       for (auto II : Inputs) {
521         if (II->NeedsEnergyUpdate && II->Energy != 0.0) {
522           II->NeedsEnergyUpdate = false;
523           II->UpdateEnergy(RareFeatures.size(), Entropic.ScalePerExecTime,
524                            AverageUnitExecutionTime);
525         }
526       }
527 
528       for (size_t i = 0; i < N; i++) {
529 
530         if (Inputs[i]->NumFeatures == 0) {
531           // If the seed doesn't represent any features, assign zero energy.
532           Weights[i] = 0.;
533         } else if (Inputs[i]->NumExecutedMutations / kMaxMutationFactor >
534                    NumExecutedMutations / Inputs.size()) {
535           // If the seed was fuzzed a lot more than average, assign zero energy.
536           Weights[i] = 0.;
537         } else {
538           // Otherwise, simply assign the computed energy.
539           Weights[i] = Inputs[i]->Energy;
540         }
541 
542         // If energy for all seeds is zero, fall back to vanilla schedule.
543         if (Weights[i] > 0.0)
544           VanillaSchedule = false;
545       }
546     }
547 
548     if (VanillaSchedule) {
549       for (size_t i = 0; i < N; i++)
550         Weights[i] =
551             Inputs[i]->NumFeatures
552                 ? static_cast<double>((i + 1) *
553                                       (Inputs[i]->HasFocusFunction ? 1000 : 1))
554                 : 0.;
555     }
556 
557     if (FeatureDebug) {
558       for (size_t i = 0; i < N; i++)
559         Printf("%zd ", Inputs[i]->NumFeatures);
560       Printf("SCORE\n");
561       for (size_t i = 0; i < N; i++)
562         Printf("%f ", Weights[i]);
563       Printf("Weights\n");
564     }
565     CorpusDistribution = std::piecewise_constant_distribution<double>(
566         Intervals.begin(), Intervals.end(), Weights.begin());
567   }
568   std::piecewise_constant_distribution<double> CorpusDistribution;
569 
570   std::vector<double> Intervals;
571   std::vector<double> Weights;
572 
573   std::unordered_set<std::string> Hashes;
574   std::vector<InputInfo *> Inputs;
575 
576   size_t NumAddedFeatures = 0;
577   size_t NumUpdatedFeatures = 0;
578   uint32_t InputSizesPerFeature[kFeatureSetSize];
579   uint32_t SmallestElementPerFeature[kFeatureSetSize];
580 
581   bool DistributionNeedsUpdate = true;
582   uint16_t FreqOfMostAbundantRareFeature = 0;
583   uint16_t GlobalFeatureFreqs[kFeatureSetSize] = {};
584   std::vector<uint32_t> RareFeatures;
585   std::bitset<kFeatureSetSize> IsRareFeature;
586 
587   std::string OutputCorpus;
588 };
589 
590 }  // namespace fuzzer
591 
592 #endif  // LLVM_FUZZER_CORPUS
593