compiler-rt/lib/fuzzer/FuzzerMerge.cpp

   1 //===- FuzzerMerge.cpp - merging corpora ----------------------------------===//
   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 // Merging corpora.
   9 //===----------------------------------------------------------------------===//
  10
  11 #include "FuzzerCommand.h"
  12 #include "FuzzerMerge.h"
  13 #include "FuzzerIO.h"
  14 #include "FuzzerInternal.h"
  15 #include "FuzzerTracePC.h"
  16 #include "FuzzerUtil.h"
  17
  18 #include <fstream>
  19 #include <iterator>
  20 #include <set>
  21 #include <sstream>
  22 #include <unordered_set>
  23
  24 namespace fuzzer {
  25
  26 bool Merger::Parse(const std::string &Str, bool ParseCoverage) {
  27   std::istringstream SS(Str);
  28   return Parse(SS, ParseCoverage);
  29 }
  30
  31 void Merger::ParseOrExit(std::istream &IS, bool ParseCoverage) {
  32   if (!Parse(IS, ParseCoverage)) {
  33     Printf("MERGE: failed to parse the control file (unexpected error)\n");
  34     exit(1);
  35   }
  36 }
  37
  38 // The control file example:
  39 //
  40 // 3 # The number of inputs
  41 // 1 # The number of inputs in the first corpus, <= the previous number
  42 // file0
  43 // file1
  44 // file2  # One file name per line.
  45 // STARTED 0 123  # FileID, file size
  46 // FT 0 1 4 6 8  # FileID COV1 COV2 ...
  47 // COV 0 7 8 9 # FileID COV1 COV1
  48 // STARTED 1 456  # If FT is missing, the input crashed while processing.
  49 // STARTED 2 567
  50 // FT 2 8 9
  51 // COV 2 11 12
  52 bool Merger::Parse(std::istream &IS, bool ParseCoverage) {
  53   LastFailure.clear();
  54   std::string Line;
  55
  56   // Parse NumFiles.
  57   if (!std::getline(IS, Line, '\n')) return false;
  58   std::istringstream L1(Line);
  59   size_t NumFiles = 0;
  60   L1 >> NumFiles;
  61   if (NumFiles == 0 || NumFiles > 10000000) return false;
  62
  63   // Parse NumFilesInFirstCorpus.
  64   if (!std::getline(IS, Line, '\n')) return false;
  65   std::istringstream L2(Line);
  66   NumFilesInFirstCorpus = NumFiles + 1;
  67   L2 >> NumFilesInFirstCorpus;
  68   if (NumFilesInFirstCorpus > NumFiles) return false;
  69
  70   // Parse file names.
  71   Files.resize(NumFiles);
  72   for (size_t i = 0; i < NumFiles; i++)
  73     if (!std::getline(IS, Files[i].Name, '\n'))
  74       return false;
  75
  76   // Parse STARTED, FT, and COV lines.
  77   size_t ExpectedStartMarker = 0;
  78   const size_t kInvalidStartMarker = -1;
  79   size_t LastSeenStartMarker = kInvalidStartMarker;
  80   std::vector<uint32_t> TmpFeatures;
  81   std::set<uint32_t> PCs;
  82   while (std::getline(IS, Line, '\n')) {
  83     std::istringstream ISS1(Line);
  84     std::string Marker;
  85     uint32_t N;
  86     if (!(ISS1 >> Marker) || !(ISS1 >> N))
  87       return false;
  88     if (Marker == "STARTED") {
  89       // STARTED FILE_ID FILE_SIZE
  90       if (ExpectedStartMarker != N)
  91         return false;
  92       ISS1 >> Files[ExpectedStartMarker].Size;
  93       LastSeenStartMarker = ExpectedStartMarker;
  94       assert(ExpectedStartMarker < Files.size());
  95       ExpectedStartMarker++;
  96     } else if (Marker == "FT") {
  97       // FT FILE_ID COV1 COV2 COV3 ...
  98       size_t CurrentFileIdx = N;
  99       if (CurrentFileIdx != LastSeenStartMarker)
 100         return false;
 101       LastSeenStartMarker = kInvalidStartMarker;
 102       if (ParseCoverage) {
 103         TmpFeatures.clear();  // use a vector from outer scope to avoid resizes.
 104         while (ISS1 >> N)
 105           TmpFeatures.push_back(N);
 106         std::sort(TmpFeatures.begin(), TmpFeatures.end());
 107         Files[CurrentFileIdx].Features = TmpFeatures;
 108       }
 109     } else if (Marker == "COV") {
 110       size_t CurrentFileIdx = N;
 111       if (ParseCoverage)
 112         while (ISS1 >> N)
 113           if (PCs.insert(N).second)
 114             Files[CurrentFileIdx].Cov.push_back(N);
 115     } else {
 116       return false;
 117     }
 118   }
 119   if (LastSeenStartMarker != kInvalidStartMarker)
 120     LastFailure = Files[LastSeenStartMarker].Name;
 121
 122   FirstNotProcessedFile = ExpectedStartMarker;
 123   return true;
 124 }
 125
 126 size_t Merger::ApproximateMemoryConsumption() const  {
 127   size_t Res = 0;
 128   for (const auto &F: Files)
 129     Res += sizeof(F) + F.Features.size() * sizeof(F.Features[0]);
 130   return Res;
 131 }
 132
 133 // Decides which files need to be merged (add those to NewFiles).
 134 // Returns the number of new features added.
 135 size_t Merger::Merge(const std::set<uint32_t> &InitialFeatures,
 136                      std::set<uint32_t> *NewFeatures,
 137                      const std::set<uint32_t> &InitialCov,
 138                      std::set<uint32_t> *NewCov,
 139                      std::vector<std::string> *NewFiles) {
 140   NewFiles->clear();
 141   NewFeatures->clear();
 142   NewCov->clear();
 143   assert(NumFilesInFirstCorpus <= Files.size());
 144   std::set<uint32_t> AllFeatures = InitialFeatures;
 145
 146   // What features are in the initial corpus?
 147   for (size_t i = 0; i < NumFilesInFirstCorpus; i++) {
 148     auto &Cur = Files[i].Features;
 149     AllFeatures.insert(Cur.begin(), Cur.end());
 150   }
 151   // Remove all features that we already know from all other inputs.
 152   for (size_t i = NumFilesInFirstCorpus; i < Files.size(); i++) {
 153     auto &Cur = Files[i].Features;
 154     std::vector<uint32_t> Tmp;
 155     std::set_difference(Cur.begin(), Cur.end(), AllFeatures.begin(),
 156                         AllFeatures.end(), std::inserter(Tmp, Tmp.begin()));
 157     Cur.swap(Tmp);
 158   }
 159
 160   // Sort. Give preference to
 161   //   * smaller files
 162   //   * files with more features.
 163   std::sort(Files.begin() + NumFilesInFirstCorpus, Files.end(),
 164             [&](const MergeFileInfo &a, const MergeFileInfo &b) -> bool {
 165               if (a.Size != b.Size)
 166                 return a.Size < b.Size;
 167               return a.Features.size() > b.Features.size();
 168             });
 169
 170   // One greedy pass: add the file's features to AllFeatures.
 171   // If new features were added, add this file to NewFiles.
 172   for (size_t i = NumFilesInFirstCorpus; i < Files.size(); i++) {
 173     auto &Cur = Files[i].Features;
 174     // Printf("%s -> sz %zd ft %zd\n", Files[i].Name.c_str(),
 175     //       Files[i].Size, Cur.size());
 176     bool FoundNewFeatures = false;
 177     for (auto Fe: Cur) {
 178       if (AllFeatures.insert(Fe).second) {
 179         FoundNewFeatures = true;
 180         NewFeatures->insert(Fe);
 181       }
 182     }
 183     if (FoundNewFeatures)
 184       NewFiles->push_back(Files[i].Name);
 185     for (auto Cov : Files[i].Cov)
 186       if (InitialCov.find(Cov) == InitialCov.end())
 187         NewCov->insert(Cov);
 188   }
 189   return NewFeatures->size();
 190 }
 191
 192 std::set<uint32_t> Merger::AllFeatures() const {
 193   std::set<uint32_t> S;
 194   for (auto &File : Files)
 195     S.insert(File.Features.begin(), File.Features.end());
 196   return S;
 197 }
 198
 199 // Inner process. May crash if the target crashes.
 200 void Fuzzer::CrashResistantMergeInternalStep(const std::string &CFPath,
 201                                              bool IsSetCoverMerge) {
 202   Printf("MERGE-INNER: using the control file '%s'\n", CFPath.c_str());
 203   Merger M;
 204   std::ifstream IF(CFPath);
 205   M.ParseOrExit(IF, false);
 206   IF.close();
 207   if (!M.LastFailure.empty())
 208     Printf("MERGE-INNER: '%s' caused a failure at the previous merge step\n",
 209            M.LastFailure.c_str());
 210
 211   Printf("MERGE-INNER: %zd total files;"
 212          " %zd processed earlier; will process %zd files now\n",
 213          M.Files.size(), M.FirstNotProcessedFile,
 214          M.Files.size() - M.FirstNotProcessedFile);
 215
 216   std::ofstream OF(CFPath, std::ofstream::out | std::ofstream::app);
 217   std::set<size_t> AllFeatures;
 218   auto PrintStatsWrapper = [this, &AllFeatures](const char* Where) {
 219     this->PrintStats(Where, "\n", 0, AllFeatures.size());
 220   };
 221   std::set<const TracePC::PCTableEntry *> AllPCs;
 222   for (size_t i = M.FirstNotProcessedFile; i < M.Files.size(); i++) {
 223     Fuzzer::MaybeExitGracefully();
 224     auto U = FileToVector(M.Files[i].Name);
 225     if (U.size() > MaxInputLen) {
 226       U.resize(MaxInputLen);
 227       U.shrink_to_fit();
 228     }
 229
 230     // Write the pre-run marker.
 231     OF << "STARTED " << i << " " << U.size() << "\n";
 232     OF.flush();  // Flush is important since Command::Execute may crash.
 233     // Run.
 234     TPC.ResetMaps();
 235     ExecuteCallback(U.data(), U.size());
 236     // Collect coverage. We are iterating over the files in this order:
 237     // * First, files in the initial corpus ordered by size, smallest first.
 238     // * Then, all other files, smallest first.
 239     std::set<size_t> Features;
 240     if (IsSetCoverMerge)
 241       TPC.CollectFeatures([&](size_t Feature) { Features.insert(Feature); });
 242     else
 243       TPC.CollectFeatures([&](size_t Feature) {
 244         if (AllFeatures.insert(Feature).second)
 245           Features.insert(Feature);
 246       });
 247     TPC.UpdateObservedPCs();
 248     // Show stats.
 249     if (!(TotalNumberOfRuns & (TotalNumberOfRuns - 1)))
 250       PrintStatsWrapper("pulse ");
 251     if (TotalNumberOfRuns == M.NumFilesInFirstCorpus)
 252       PrintStatsWrapper("LOADED");
 253     // Write the post-run marker and the coverage.
 254     OF << "FT " << i;
 255     for (size_t F : Features)
 256       OF << " " << F;
 257     OF << "\n";
 258     OF << "COV " << i;
 259     TPC.ForEachObservedPC([&](const TracePC::PCTableEntry *TE) {
 260       if (AllPCs.insert(TE).second)
 261         OF << " " << TPC.PCTableEntryIdx(TE);
 262     });
 263     OF << "\n";
 264     OF.flush();
 265   }
 266   PrintStatsWrapper("DONE  ");
 267 }
 268
 269 // Merges all corpora into the first corpus. A file is added into
 270 // the first corpus only if it adds new features. Unlike `Merger::Merge`,
 271 // this implementation calculates an approximation of the minimum set
 272 // of corpora files, that cover all known features (set cover problem).
 273 // Generally, this means that files with more features are preferred for
 274 // merge into the first corpus. When two files have the same number of
 275 // features, the smaller one is preferred.
 276 size_t Merger::SetCoverMerge(const std::set<uint32_t> &InitialFeatures,
 277                              std::set<uint32_t> *NewFeatures,
 278                              const std::set<uint32_t> &InitialCov,
 279                              std::set<uint32_t> *NewCov,
 280                              std::vector<std::string> *NewFiles) {
 281   assert(NumFilesInFirstCorpus <= Files.size());
 282   NewFiles->clear();
 283   NewFeatures->clear();
 284   NewCov->clear();
 285   std::set<uint32_t> AllFeatures;
 286   // 1 << 21 - 1 is the maximum feature index.
 287   // See 'kFeatureSetSize' in 'FuzzerCorpus.h'.
 288   const uint32_t kFeatureSetSize = 1 << 21;
 289   std::vector<bool> Covered(kFeatureSetSize, false);
 290   size_t NumCovered = 0;
 291
 292   std::set<uint32_t> ExistingFeatures = InitialFeatures;
 293   for (size_t i = 0; i < NumFilesInFirstCorpus; ++i)
 294     ExistingFeatures.insert(Files[i].Features.begin(), Files[i].Features.end());
 295
 296   // Mark the existing features as covered.
 297   for (const auto &F : ExistingFeatures) {
 298     if (!Covered[F % kFeatureSetSize]) {
 299       ++NumCovered;
 300       Covered[F % kFeatureSetSize] = true;
 301     }
 302     // Calculate an underestimation of the set of covered features
 303     // since the `Covered` bitvector is smaller than the feature range.
 304     AllFeatures.insert(F % kFeatureSetSize);
 305   }
 306
 307   std::set<size_t> RemainingFiles;
 308   for (size_t i = NumFilesInFirstCorpus; i < Files.size(); ++i) {
 309     // Construct an incremental sequence which represent the
 310     // indices to all files (excluding those in the initial corpus).
 311     // RemainingFiles = range(NumFilesInFirstCorpus..Files.size()).
 312     RemainingFiles.insert(i);
 313     // Insert this file's unique features to all features.
 314     for (const auto &F : Files[i].Features)
 315       AllFeatures.insert(F % kFeatureSetSize);
 316   }
 317
 318   // Integrate files into Covered until set is complete.
 319   while (NumCovered != AllFeatures.size()) {
 320     // Index to file with largest number of unique features.
 321     size_t MaxFeaturesIndex = NumFilesInFirstCorpus;
 322     // Indices to remove from RemainingFiles.
 323     std::set<size_t> RemoveIndices;
 324     // Running max unique feature count.
 325     // Updated upon finding a file with more features.
 326     size_t MaxNumFeatures = 0;
 327
 328     // Iterate over all files not yet integrated into Covered,
 329     // to find the file which has the largest number of
 330     // features that are not already in Covered.
 331     for (const auto &i : RemainingFiles) {
 332       const auto &File = Files[i];
 333       size_t CurrentUnique = 0;
 334       // Count number of features in this file
 335       // which are not yet in Covered.
 336       for (const auto &F : File.Features)
 337         if (!Covered[F % kFeatureSetSize])
 338           ++CurrentUnique;
 339
 340       if (CurrentUnique == 0) {
 341         // All features in this file are already in Covered: skip next time.
 342         RemoveIndices.insert(i);
 343       } else if (CurrentUnique > MaxNumFeatures ||
 344                  (CurrentUnique == MaxNumFeatures &&
 345                   File.Size < Files[MaxFeaturesIndex].Size)) {
 346         // Update the max features file based on unique features
 347         // Break ties by selecting smaller files.
 348         MaxNumFeatures = CurrentUnique;
 349         MaxFeaturesIndex = i;
 350       }
 351     }
 352     // Must be a valid index/
 353     assert(MaxFeaturesIndex < Files.size());
 354     // Remove any feature-less files found.
 355     for (const auto &i : RemoveIndices)
 356       RemainingFiles.erase(i);
 357     if (MaxNumFeatures == 0) {
 358       // Did not find a file that adds unique features.
 359       // This means that we should have no remaining files.
 360       assert(RemainingFiles.size() == 0);
 361       assert(NumCovered == AllFeatures.size());
 362       break;
 363     }
 364
 365     // MaxFeaturesIndex must be an element of Remaining.
 366     assert(RemainingFiles.find(MaxFeaturesIndex) != RemainingFiles.end());
 367     // Remove the file with the most features from Remaining.
 368     RemainingFiles.erase(MaxFeaturesIndex);
 369     const auto &MaxFeatureFile = Files[MaxFeaturesIndex];
 370     // Add the features of the max feature file to Covered.
 371     for (const auto &F : MaxFeatureFile.Features) {
 372       if (!Covered[F % kFeatureSetSize]) {
 373         ++NumCovered;
 374         Covered[F % kFeatureSetSize] = true;
 375         NewFeatures->insert(F);
 376       }
 377     }
 378     // Add the index to this file to the result.
 379     NewFiles->push_back(MaxFeatureFile.Name);
 380     // Update NewCov with the additional coverage
 381     // that MaxFeatureFile provides.
 382     for (const auto &C : MaxFeatureFile.Cov)
 383       if (InitialCov.find(C) == InitialCov.end())
 384         NewCov->insert(C);
 385   }
 386
 387   return NewFeatures->size();
 388 }
 389
 390 static size_t
 391 WriteNewControlFile(const std::string &CFPath,
 392                     const std::vector<SizedFile> &OldCorpus,
 393                     const std::vector<SizedFile> &NewCorpus,
 394                     const std::vector<MergeFileInfo> &KnownFiles) {
 395   std::unordered_set<std::string> FilesToSkip;
 396   for (auto &SF: KnownFiles)
 397     FilesToSkip.insert(SF.Name);
 398
 399   std::vector<std::string> FilesToUse;
 400   auto MaybeUseFile = [=, &FilesToUse](std::string Name) {
 401     if (FilesToSkip.find(Name) == FilesToSkip.end())
 402       FilesToUse.push_back(Name);
 403   };
 404   for (auto &SF: OldCorpus)
 405     MaybeUseFile(SF.File);
 406   auto FilesToUseFromOldCorpus = FilesToUse.size();
 407   for (auto &SF: NewCorpus)
 408     MaybeUseFile(SF.File);
 409
 410   RemoveFile(CFPath);
 411   std::ofstream ControlFile(CFPath);
 412   ControlFile << FilesToUse.size() << "\n";
 413   ControlFile << FilesToUseFromOldCorpus << "\n";
 414   for (auto &FN: FilesToUse)
 415     ControlFile << FN << "\n";
 416
 417   if (!ControlFile) {
 418     Printf("MERGE-OUTER: failed to write to the control file: %s\n",
 419            CFPath.c_str());
 420     exit(1);
 421   }
 422
 423   return FilesToUse.size();
 424 }
 425
 426 // Outer process. Does not call the target code and thus should not fail.
 427 void CrashResistantMerge(const std::vector<std::string> &Args,
 428                          const std::vector<SizedFile> &OldCorpus,
 429                          const std::vector<SizedFile> &NewCorpus,
 430                          std::vector<std::string> *NewFiles,
 431                          const std::set<uint32_t> &InitialFeatures,
 432                          std::set<uint32_t> *NewFeatures,
 433                          const std::set<uint32_t> &InitialCov,
 434                          std::set<uint32_t> *NewCov, const std::string &CFPath,
 435                          bool V, /*Verbose*/
 436                          bool IsSetCoverMerge) {
 437   if (NewCorpus.empty() && OldCorpus.empty()) return;  // Nothing to merge.
 438   size_t NumAttempts = 0;
 439   std::vector<MergeFileInfo> KnownFiles;
 440   if (FileSize(CFPath)) {
 441     VPrintf(V, "MERGE-OUTER: non-empty control file provided: '%s'\n",
 442            CFPath.c_str());
 443     Merger M;
 444     std::ifstream IF(CFPath);
 445     if (M.Parse(IF, /*ParseCoverage=*/true)) {
 446       VPrintf(V, "MERGE-OUTER: control file ok, %zd files total,"
 447              " first not processed file %zd\n",
 448              M.Files.size(), M.FirstNotProcessedFile);
 449       if (!M.LastFailure.empty())
 450         VPrintf(V, "MERGE-OUTER: '%s' will be skipped as unlucky "
 451                "(merge has stumbled on it the last time)\n",
 452                M.LastFailure.c_str());
 453       if (M.FirstNotProcessedFile >= M.Files.size()) {
 454         // Merge has already been completed with the given merge control file.
 455         if (M.Files.size() == OldCorpus.size() + NewCorpus.size()) {
 456           VPrintf(
 457               V,
 458               "MERGE-OUTER: nothing to do, merge has been completed before\n");
 459           exit(0);
 460         }
 461
 462         // Number of input files likely changed, start merge from scratch, but
 463         // reuse coverage information from the given merge control file.
 464         VPrintf(
 465             V,
 466             "MERGE-OUTER: starting merge from scratch, but reusing coverage "
 467             "information from the given control file\n");
 468         KnownFiles = M.Files;
 469       } else {
 470         // There is a merge in progress, continue.
 471         NumAttempts = M.Files.size() - M.FirstNotProcessedFile;
 472       }
 473     } else {
 474       VPrintf(V, "MERGE-OUTER: bad control file, will overwrite it\n");
 475     }
 476   }
 477
 478   if (!NumAttempts) {
 479     // The supplied control file is empty or bad, create a fresh one.
 480     VPrintf(V, "MERGE-OUTER: "
 481             "%zd files, %zd in the initial corpus, %zd processed earlier\n",
 482             OldCorpus.size() + NewCorpus.size(), OldCorpus.size(),
 483             KnownFiles.size());
 484     NumAttempts = WriteNewControlFile(CFPath, OldCorpus, NewCorpus, KnownFiles);
 485   }
 486
 487   // Execute the inner process until it passes.
 488   // Every inner process should execute at least one input.
 489   Command BaseCmd(Args);
 490   BaseCmd.removeFlag("merge");
 491   BaseCmd.removeFlag("set_cover_merge");
 492   BaseCmd.removeFlag("fork");
 493   BaseCmd.removeFlag("collect_data_flow");
 494   for (size_t Attempt = 1; Attempt <= NumAttempts; Attempt++) {
 495     Fuzzer::MaybeExitGracefully();
 496     VPrintf(V, "MERGE-OUTER: attempt %zd\n", Attempt);
 497     Command Cmd(BaseCmd);
 498     Cmd.addFlag("merge_control_file", CFPath);
 499     // If we are going to use the set cover implementation for
 500     // minimization add the merge_inner=2 internal flag.
 501     Cmd.addFlag("merge_inner", IsSetCoverMerge ? "2" : "1");
 502     if (!V) {
 503       Cmd.setOutputFile(getDevNull());
 504       Cmd.combineOutAndErr();
 505     }
 506     auto ExitCode = ExecuteCommand(Cmd);
 507     if (!ExitCode) {
 508       VPrintf(V, "MERGE-OUTER: successful in %zd attempt(s)\n", Attempt);
 509       break;
 510     }
 511   }
 512   // Read the control file and do the merge.
 513   Merger M;
 514   std::ifstream IF(CFPath);
 515   IF.seekg(0, IF.end);
 516   VPrintf(V, "MERGE-OUTER: the control file has %zd bytes\n",
 517           (size_t)IF.tellg());
 518   IF.seekg(0, IF.beg);
 519   M.ParseOrExit(IF, true);
 520   IF.close();
 521   VPrintf(V,
 522           "MERGE-OUTER: consumed %zdMb (%zdMb rss) to parse the control file\n",
 523           M.ApproximateMemoryConsumption() >> 20, GetPeakRSSMb());
 524
 525   M.Files.insert(M.Files.end(), KnownFiles.begin(), KnownFiles.end());
 526   if (IsSetCoverMerge)
 527     M.SetCoverMerge(InitialFeatures, NewFeatures, InitialCov, NewCov, NewFiles);
 528   else
 529     M.Merge(InitialFeatures, NewFeatures, InitialCov, NewCov, NewFiles);
 530   VPrintf(V, "MERGE-OUTER: %zd new files with %zd new features added; "
 531           "%zd new coverage edges\n",
 532          NewFiles->size(), NewFeatures->size(), NewCov->size());
 533 }
 534
 535 } // namespace fuzzer