xref: /freebsd/contrib/llvm-project/compiler-rt/lib/fuzzer/FuzzerCorpus.h (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
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 <numeric>
22 #include <random>
23 #include <unordered_set>
24 
25 namespace fuzzer {
26 
27 struct InputInfo {
28   Unit U;  // The actual input data.
29   uint8_t Sha1[kSHA1NumBytes];  // Checksum.
30   // Number of features that this input has and no smaller input has.
31   size_t NumFeatures = 0;
32   size_t Tmp = 0; // Used by ValidateFeatureSet.
33   // Stats.
34   size_t NumExecutedMutations = 0;
35   size_t NumSuccessfullMutations = 0;
36   bool MayDeleteFile = false;
37   bool Reduced = false;
38   bool HasFocusFunction = false;
39   Vector<uint32_t> UniqFeatureSet;
40   Vector<uint8_t> DataFlowTraceForFocusFunction;
41   // Power schedule.
42   bool NeedsEnergyUpdate = false;
43   double Energy = 0.0;
44   size_t SumIncidence = 0;
45   Vector<std::pair<uint32_t, uint16_t>> FeatureFreqs;
46 
47   // Delete feature Idx and its frequency from FeatureFreqs.
48   bool DeleteFeatureFreq(uint32_t Idx) {
49     if (FeatureFreqs.empty())
50       return false;
51 
52     // Binary search over local feature frequencies sorted by index.
53     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
54                                   std::pair<uint32_t, uint16_t>(Idx, 0));
55 
56     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
57       FeatureFreqs.erase(Lower);
58       return true;
59     }
60     return false;
61   }
62 
63   // Assign more energy to a high-entropy seed, i.e., that reveals more
64   // information about the globally rare features in the neighborhood
65   // of the seed. Since we do not know the entropy of a seed that has
66   // never been executed we assign fresh seeds maximum entropy and
67   // let II->Energy approach the true entropy from above.
68   void UpdateEnergy(size_t GlobalNumberOfFeatures) {
69     Energy = 0.0;
70     SumIncidence = 0;
71 
72     // Apply add-one smoothing to locally discovered features.
73     for (auto F : FeatureFreqs) {
74       size_t LocalIncidence = F.second + 1;
75       Energy -= LocalIncidence * logl(LocalIncidence);
76       SumIncidence += LocalIncidence;
77     }
78 
79     // Apply add-one smoothing to locally undiscovered features.
80     //   PreciseEnergy -= 0; // since logl(1.0) == 0)
81     SumIncidence += (GlobalNumberOfFeatures - FeatureFreqs.size());
82 
83     // Add a single locally abundant feature apply add-one smoothing.
84     size_t AbdIncidence = NumExecutedMutations + 1;
85     Energy -= AbdIncidence * logl(AbdIncidence);
86     SumIncidence += AbdIncidence;
87 
88     // Normalize.
89     if (SumIncidence != 0)
90       Energy = (Energy / SumIncidence) + logl(SumIncidence);
91   }
92 
93   // Increment the frequency of the feature Idx.
94   void UpdateFeatureFrequency(uint32_t Idx) {
95     NeedsEnergyUpdate = true;
96 
97     // The local feature frequencies is an ordered vector of pairs.
98     // If there are no local feature frequencies, push_back preserves order.
99     // Set the feature frequency for feature Idx32 to 1.
100     if (FeatureFreqs.empty()) {
101       FeatureFreqs.push_back(std::pair<uint32_t, uint16_t>(Idx, 1));
102       return;
103     }
104 
105     // Binary search over local feature frequencies sorted by index.
106     auto Lower = std::lower_bound(FeatureFreqs.begin(), FeatureFreqs.end(),
107                                   std::pair<uint32_t, uint16_t>(Idx, 0));
108 
109     // If feature Idx32 already exists, increment its frequency.
110     // Otherwise, insert a new pair right after the next lower index.
111     if (Lower != FeatureFreqs.end() && Lower->first == Idx) {
112       Lower->second++;
113     } else {
114       FeatureFreqs.insert(Lower, std::pair<uint32_t, uint16_t>(Idx, 1));
115     }
116   }
117 };
118 
119 struct EntropicOptions {
120   bool Enabled;
121   size_t NumberOfRarestFeatures;
122   size_t FeatureFrequencyThreshold;
123 };
124 
125 class InputCorpus {
126   static const uint32_t kFeatureSetSize = 1 << 21;
127   static const uint8_t kMaxMutationFactor = 20;
128   static const size_t kSparseEnergyUpdates = 100;
129 
130   size_t NumExecutedMutations = 0;
131 
132   EntropicOptions Entropic;
133 
134 public:
135   InputCorpus(const std::string &OutputCorpus, EntropicOptions Entropic)
136       : Entropic(Entropic), OutputCorpus(OutputCorpus) {
137     memset(InputSizesPerFeature, 0, sizeof(InputSizesPerFeature));
138     memset(SmallestElementPerFeature, 0, sizeof(SmallestElementPerFeature));
139   }
140   ~InputCorpus() {
141     for (auto II : Inputs)
142       delete II;
143   }
144   size_t size() const { return Inputs.size(); }
145   size_t SizeInBytes() const {
146     size_t Res = 0;
147     for (auto II : Inputs)
148       Res += II->U.size();
149     return Res;
150   }
151   size_t NumActiveUnits() const {
152     size_t Res = 0;
153     for (auto II : Inputs)
154       Res += !II->U.empty();
155     return Res;
156   }
157   size_t MaxInputSize() const {
158     size_t Res = 0;
159     for (auto II : Inputs)
160         Res = std::max(Res, II->U.size());
161     return Res;
162   }
163   void IncrementNumExecutedMutations() { NumExecutedMutations++; }
164 
165   size_t NumInputsThatTouchFocusFunction() {
166     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
167       return II->HasFocusFunction;
168     });
169   }
170 
171   size_t NumInputsWithDataFlowTrace() {
172     return std::count_if(Inputs.begin(), Inputs.end(), [](const InputInfo *II) {
173       return !II->DataFlowTraceForFocusFunction.empty();
174     });
175   }
176 
177   bool empty() const { return Inputs.empty(); }
178   const Unit &operator[] (size_t Idx) const { return Inputs[Idx]->U; }
179   InputInfo *AddToCorpus(const Unit &U, size_t NumFeatures, bool MayDeleteFile,
180                          bool HasFocusFunction,
181                          const Vector<uint32_t> &FeatureSet,
182                          const DataFlowTrace &DFT, const InputInfo *BaseII) {
183     assert(!U.empty());
184     if (FeatureDebug)
185       Printf("ADD_TO_CORPUS %zd NF %zd\n", Inputs.size(), NumFeatures);
186     Inputs.push_back(new InputInfo());
187     InputInfo &II = *Inputs.back();
188     II.U = U;
189     II.NumFeatures = NumFeatures;
190     II.MayDeleteFile = MayDeleteFile;
191     II.UniqFeatureSet = FeatureSet;
192     II.HasFocusFunction = HasFocusFunction;
193     // Assign maximal energy to the new seed.
194     II.Energy = RareFeatures.empty() ? 1.0 : log(RareFeatures.size());
195     II.SumIncidence = RareFeatures.size();
196     II.NeedsEnergyUpdate = false;
197     std::sort(II.UniqFeatureSet.begin(), II.UniqFeatureSet.end());
198     ComputeSHA1(U.data(), U.size(), II.Sha1);
199     auto Sha1Str = Sha1ToString(II.Sha1);
200     Hashes.insert(Sha1Str);
201     if (HasFocusFunction)
202       if (auto V = DFT.Get(Sha1Str))
203         II.DataFlowTraceForFocusFunction = *V;
204     // This is a gross heuristic.
205     // Ideally, when we add an element to a corpus we need to know its DFT.
206     // But if we don't, we'll use the DFT of its base input.
207     if (II.DataFlowTraceForFocusFunction.empty() && BaseII)
208       II.DataFlowTraceForFocusFunction = BaseII->DataFlowTraceForFocusFunction;
209     DistributionNeedsUpdate = true;
210     PrintCorpus();
211     // ValidateFeatureSet();
212     return &II;
213   }
214 
215   // Debug-only
216   void PrintUnit(const Unit &U) {
217     if (!FeatureDebug) return;
218     for (uint8_t C : U) {
219       if (C != 'F' && C != 'U' && C != 'Z')
220         C = '.';
221       Printf("%c", C);
222     }
223   }
224 
225   // Debug-only
226   void PrintFeatureSet(const Vector<uint32_t> &FeatureSet) {
227     if (!FeatureDebug) return;
228     Printf("{");
229     for (uint32_t Feature: FeatureSet)
230       Printf("%u,", Feature);
231     Printf("}");
232   }
233 
234   // Debug-only
235   void PrintCorpus() {
236     if (!FeatureDebug) return;
237     Printf("======= CORPUS:\n");
238     int i = 0;
239     for (auto II : Inputs) {
240       if (std::find(II->U.begin(), II->U.end(), 'F') != II->U.end()) {
241         Printf("[%2d] ", i);
242         Printf("%s sz=%zd ", Sha1ToString(II->Sha1).c_str(), II->U.size());
243         PrintUnit(II->U);
244         Printf(" ");
245         PrintFeatureSet(II->UniqFeatureSet);
246         Printf("\n");
247       }
248       i++;
249     }
250   }
251 
252   void Replace(InputInfo *II, const Unit &U) {
253     assert(II->U.size() > U.size());
254     Hashes.erase(Sha1ToString(II->Sha1));
255     DeleteFile(*II);
256     ComputeSHA1(U.data(), U.size(), II->Sha1);
257     Hashes.insert(Sha1ToString(II->Sha1));
258     II->U = U;
259     II->Reduced = true;
260     DistributionNeedsUpdate = true;
261   }
262 
263   bool HasUnit(const Unit &U) { return Hashes.count(Hash(U)); }
264   bool HasUnit(const std::string &H) { return Hashes.count(H); }
265   InputInfo &ChooseUnitToMutate(Random &Rand) {
266     InputInfo &II = *Inputs[ChooseUnitIdxToMutate(Rand)];
267     assert(!II.U.empty());
268     return II;
269   }
270 
271   // Returns an index of random unit from the corpus to mutate.
272   size_t ChooseUnitIdxToMutate(Random &Rand) {
273     UpdateCorpusDistribution(Rand);
274     size_t Idx = static_cast<size_t>(CorpusDistribution(Rand));
275     assert(Idx < Inputs.size());
276     return Idx;
277   }
278 
279   void PrintStats() {
280     for (size_t i = 0; i < Inputs.size(); i++) {
281       const auto &II = *Inputs[i];
282       Printf("  [% 3zd %s] sz: % 5zd runs: % 5zd succ: % 5zd focus: %d\n", i,
283              Sha1ToString(II.Sha1).c_str(), II.U.size(),
284              II.NumExecutedMutations, II.NumSuccessfullMutations, II.HasFocusFunction);
285     }
286   }
287 
288   void PrintFeatureSet() {
289     for (size_t i = 0; i < kFeatureSetSize; i++) {
290       if(size_t Sz = GetFeature(i))
291         Printf("[%zd: id %zd sz%zd] ", i, SmallestElementPerFeature[i], Sz);
292     }
293     Printf("\n\t");
294     for (size_t i = 0; i < Inputs.size(); i++)
295       if (size_t N = Inputs[i]->NumFeatures)
296         Printf(" %zd=>%zd ", i, N);
297     Printf("\n");
298   }
299 
300   void DeleteFile(const InputInfo &II) {
301     if (!OutputCorpus.empty() && II.MayDeleteFile)
302       RemoveFile(DirPlusFile(OutputCorpus, Sha1ToString(II.Sha1)));
303   }
304 
305   void DeleteInput(size_t Idx) {
306     InputInfo &II = *Inputs[Idx];
307     DeleteFile(II);
308     Unit().swap(II.U);
309     II.Energy = 0.0;
310     II.NeedsEnergyUpdate = false;
311     DistributionNeedsUpdate = true;
312     if (FeatureDebug)
313       Printf("EVICTED %zd\n", Idx);
314   }
315 
316   void AddRareFeature(uint32_t Idx) {
317     // Maintain *at least* TopXRarestFeatures many rare features
318     // and all features with a frequency below ConsideredRare.
319     // Remove all other features.
320     while (RareFeatures.size() > Entropic.NumberOfRarestFeatures &&
321            FreqOfMostAbundantRareFeature > Entropic.FeatureFrequencyThreshold) {
322 
323       // Find most and second most abbundant feature.
324       uint32_t MostAbundantRareFeatureIndices[2] = {RareFeatures[0],
325                                                     RareFeatures[0]};
326       size_t Delete = 0;
327       for (size_t i = 0; i < RareFeatures.size(); i++) {
328         uint32_t Idx2 = RareFeatures[i];
329         if (GlobalFeatureFreqs[Idx2] >=
330             GlobalFeatureFreqs[MostAbundantRareFeatureIndices[0]]) {
331           MostAbundantRareFeatureIndices[1] = MostAbundantRareFeatureIndices[0];
332           MostAbundantRareFeatureIndices[0] = Idx2;
333           Delete = i;
334         }
335       }
336 
337       // Remove most abundant rare feature.
338       RareFeatures[Delete] = RareFeatures.back();
339       RareFeatures.pop_back();
340 
341       for (auto II : Inputs) {
342         if (II->DeleteFeatureFreq(MostAbundantRareFeatureIndices[0]))
343           II->NeedsEnergyUpdate = true;
344       }
345 
346       // Set 2nd most abundant as the new most abundant feature count.
347       FreqOfMostAbundantRareFeature =
348           GlobalFeatureFreqs[MostAbundantRareFeatureIndices[1]];
349     }
350 
351     // Add rare feature, handle collisions, and update energy.
352     RareFeatures.push_back(Idx);
353     GlobalFeatureFreqs[Idx] = 0;
354     for (auto II : Inputs) {
355       II->DeleteFeatureFreq(Idx);
356 
357       // Apply add-one smoothing to this locally undiscovered feature.
358       // Zero energy seeds will never be fuzzed and remain zero energy.
359       if (II->Energy > 0.0) {
360         II->SumIncidence += 1;
361         II->Energy += logl(II->SumIncidence) / II->SumIncidence;
362       }
363     }
364 
365     DistributionNeedsUpdate = true;
366   }
367 
368   bool AddFeature(size_t Idx, uint32_t NewSize, bool Shrink) {
369     assert(NewSize);
370     Idx = Idx % kFeatureSetSize;
371     uint32_t OldSize = GetFeature(Idx);
372     if (OldSize == 0 || (Shrink && OldSize > NewSize)) {
373       if (OldSize > 0) {
374         size_t OldIdx = SmallestElementPerFeature[Idx];
375         InputInfo &II = *Inputs[OldIdx];
376         assert(II.NumFeatures > 0);
377         II.NumFeatures--;
378         if (II.NumFeatures == 0)
379           DeleteInput(OldIdx);
380       } else {
381         NumAddedFeatures++;
382         if (Entropic.Enabled)
383           AddRareFeature((uint32_t)Idx);
384       }
385       NumUpdatedFeatures++;
386       if (FeatureDebug)
387         Printf("ADD FEATURE %zd sz %d\n", Idx, NewSize);
388       SmallestElementPerFeature[Idx] = Inputs.size();
389       InputSizesPerFeature[Idx] = NewSize;
390       return true;
391     }
392     return false;
393   }
394 
395   // Increment frequency of feature Idx globally and locally.
396   void UpdateFeatureFrequency(InputInfo *II, size_t Idx) {
397     uint32_t Idx32 = Idx % kFeatureSetSize;
398 
399     // Saturated increment.
400     if (GlobalFeatureFreqs[Idx32] == 0xFFFF)
401       return;
402     uint16_t Freq = GlobalFeatureFreqs[Idx32]++;
403 
404     // Skip if abundant.
405     if (Freq > FreqOfMostAbundantRareFeature ||
406         std::find(RareFeatures.begin(), RareFeatures.end(), Idx32) ==
407             RareFeatures.end())
408       return;
409 
410     // Update global frequencies.
411     if (Freq == FreqOfMostAbundantRareFeature)
412       FreqOfMostAbundantRareFeature++;
413 
414     // Update local frequencies.
415     if (II)
416       II->UpdateFeatureFrequency(Idx32);
417   }
418 
419   size_t NumFeatures() const { return NumAddedFeatures; }
420   size_t NumFeatureUpdates() const { return NumUpdatedFeatures; }
421 
422 private:
423 
424   static const bool FeatureDebug = false;
425 
426   size_t GetFeature(size_t Idx) const { return InputSizesPerFeature[Idx]; }
427 
428   void ValidateFeatureSet() {
429     if (FeatureDebug)
430       PrintFeatureSet();
431     for (size_t Idx = 0; Idx < kFeatureSetSize; Idx++)
432       if (GetFeature(Idx))
433         Inputs[SmallestElementPerFeature[Idx]]->Tmp++;
434     for (auto II: Inputs) {
435       if (II->Tmp != II->NumFeatures)
436         Printf("ZZZ %zd %zd\n", II->Tmp, II->NumFeatures);
437       assert(II->Tmp == II->NumFeatures);
438       II->Tmp = 0;
439     }
440   }
441 
442   // Updates the probability distribution for the units in the corpus.
443   // Must be called whenever the corpus or unit weights are changed.
444   //
445   // Hypothesis: inputs that maximize information about globally rare features
446   // are interesting.
447   void UpdateCorpusDistribution(Random &Rand) {
448     // Skip update if no seeds or rare features were added/deleted.
449     // Sparse updates for local change of feature frequencies,
450     // i.e., randomly do not skip.
451     if (!DistributionNeedsUpdate &&
452         (!Entropic.Enabled || Rand(kSparseEnergyUpdates)))
453       return;
454 
455     DistributionNeedsUpdate = false;
456 
457     size_t N = Inputs.size();
458     assert(N);
459     Intervals.resize(N + 1);
460     Weights.resize(N);
461     std::iota(Intervals.begin(), Intervals.end(), 0);
462 
463     bool VanillaSchedule = true;
464     if (Entropic.Enabled) {
465       for (auto II : Inputs) {
466         if (II->NeedsEnergyUpdate && II->Energy != 0.0) {
467           II->NeedsEnergyUpdate = false;
468           II->UpdateEnergy(RareFeatures.size());
469         }
470       }
471 
472       for (size_t i = 0; i < N; i++) {
473 
474         if (Inputs[i]->NumFeatures == 0) {
475           // If the seed doesn't represent any features, assign zero energy.
476           Weights[i] = 0.;
477         } else if (Inputs[i]->NumExecutedMutations / kMaxMutationFactor >
478                    NumExecutedMutations / Inputs.size()) {
479           // If the seed was fuzzed a lot more than average, assign zero energy.
480           Weights[i] = 0.;
481         } else {
482           // Otherwise, simply assign the computed energy.
483           Weights[i] = Inputs[i]->Energy;
484         }
485 
486         // If energy for all seeds is zero, fall back to vanilla schedule.
487         if (Weights[i] > 0.0)
488           VanillaSchedule = false;
489       }
490     }
491 
492     if (VanillaSchedule) {
493       for (size_t i = 0; i < N; i++)
494         Weights[i] = Inputs[i]->NumFeatures
495                          ? (i + 1) * (Inputs[i]->HasFocusFunction ? 1000 : 1)
496                          : 0.;
497     }
498 
499     if (FeatureDebug) {
500       for (size_t i = 0; i < N; i++)
501         Printf("%zd ", Inputs[i]->NumFeatures);
502       Printf("SCORE\n");
503       for (size_t i = 0; i < N; i++)
504         Printf("%f ", Weights[i]);
505       Printf("Weights\n");
506     }
507     CorpusDistribution = std::piecewise_constant_distribution<double>(
508         Intervals.begin(), Intervals.end(), Weights.begin());
509   }
510   std::piecewise_constant_distribution<double> CorpusDistribution;
511 
512   Vector<double> Intervals;
513   Vector<double> Weights;
514 
515   std::unordered_set<std::string> Hashes;
516   Vector<InputInfo*> Inputs;
517 
518   size_t NumAddedFeatures = 0;
519   size_t NumUpdatedFeatures = 0;
520   uint32_t InputSizesPerFeature[kFeatureSetSize];
521   uint32_t SmallestElementPerFeature[kFeatureSetSize];
522 
523   bool DistributionNeedsUpdate = true;
524   uint16_t FreqOfMostAbundantRareFeature = 0;
525   uint16_t GlobalFeatureFreqs[kFeatureSetSize] = {};
526   Vector<uint32_t> RareFeatures;
527 
528   std::string OutputCorpus;
529 };
530 
531 }  // namespace fuzzer
532 
533 #endif  // LLVM_FUZZER_CORPUS
534