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
24namespace fuzzer {
25
26bool Merger::Parse(const std::string &Str, bool ParseCoverage) {
27 std::istringstream SS(Str);
28 return Parse(SS, ParseCoverage);
29}
30
31void 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
52bool 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 bool HaveFtMarker = true;
81 std::vector<uint32_t> TmpFeatures;
82 std::set<uint32_t> PCs;
83 while (std::getline(IS, Line, '\n')) {
84 std::istringstream ISS1(Line);
85 std::string Marker;
86 uint32_t N;
87 if (!(ISS1 >> Marker) || !(ISS1 >> N))
88 return false;
89 if (Marker == "STARTED") {
90 // STARTED FILE_ID FILE_SIZE
91 if (ExpectedStartMarker != N)
92 return false;
93 ISS1 >> Files[ExpectedStartMarker].Size;
94 LastSeenStartMarker = ExpectedStartMarker;
95 assert(ExpectedStartMarker < Files.size());
96 ExpectedStartMarker++;
97 HaveFtMarker = false;
98 } else if (Marker == "FT") {
99 // FT FILE_ID COV1 COV2 COV3 ...
100 size_t CurrentFileIdx = N;
101 if (CurrentFileIdx != LastSeenStartMarker)
102 return false;
103 HaveFtMarker = true;
104 if (ParseCoverage) {
105 TmpFeatures.clear(); // use a vector from outer scope to avoid resizes.
106 while (ISS1 >> N)
107 TmpFeatures.push_back(N);
108 std::sort(TmpFeatures.begin(), TmpFeatures.end());
109 Files[CurrentFileIdx].Features = TmpFeatures;
110 }
111 } else if (Marker == "COV") {
112 size_t CurrentFileIdx = N;
113 if (CurrentFileIdx != LastSeenStartMarker)
114 return false;
115 if (ParseCoverage)
116 while (ISS1 >> N)
117 if (PCs.insert(N).second)
118 Files[CurrentFileIdx].Cov.push_back(N);
119 } else {
120 return false;
121 }
122 }
123 if (!HaveFtMarker && LastSeenStartMarker != kInvalidStartMarker)
124 LastFailure = Files[LastSeenStartMarker].Name;
125
126 FirstNotProcessedFile = ExpectedStartMarker;
127 return true;
128}
129
130size_t Merger::ApproximateMemoryConsumption() const {
131 size_t Res = 0;
132 for (const auto &F: Files)
133 Res += sizeof(F) + F.Features.size() * sizeof(F.Features[0]);
134 return Res;
135}
136
137// Decides which files need to be merged (add those to NewFiles).
138// Returns the number of new features added.
139size_t Merger::Merge(const std::set<uint32_t> &InitialFeatures,
140 std::set<uint32_t> *NewFeatures,
141 const std::set<uint32_t> &InitialCov,
142 std::set<uint32_t> *NewCov,
143 std::vector<std::string> *NewFiles) {
144 NewFiles->clear();
145 NewFeatures->clear();
146 NewCov->clear();
147 assert(NumFilesInFirstCorpus <= Files.size());
148 std::set<uint32_t> AllFeatures = InitialFeatures;
149
150 // What features are in the initial corpus?
151 for (size_t i = 0; i < NumFilesInFirstCorpus; i++) {
152 auto &Cur = Files[i].Features;
153 AllFeatures.insert(Cur.begin(), Cur.end());
154 }
155 // Remove all features that we already know from all other inputs.
156 for (size_t i = NumFilesInFirstCorpus; i < Files.size(); i++) {
157 auto &Cur = Files[i].Features;
158 std::vector<uint32_t> Tmp;
159 std::set_difference(Cur.begin(), Cur.end(), AllFeatures.begin(),
160 AllFeatures.end(), std::inserter(Tmp, Tmp.begin()));
161 Cur.swap(Tmp);
162 }
163
164 // Sort. Give preference to
165 // * smaller files
166 // * files with more features.
167 std::sort(Files.begin() + NumFilesInFirstCorpus, Files.end(),
168 [&](const MergeFileInfo &a, const MergeFileInfo &b) -> bool {
169 if (a.Size != b.Size)
170 return a.Size < b.Size;
171 return a.Features.size() > b.Features.size();
172 });
173
174 // One greedy pass: add the file's features to AllFeatures.
175 // If new features were added, add this file to NewFiles.
176 for (size_t i = NumFilesInFirstCorpus; i < Files.size(); i++) {
177 auto &Cur = Files[i].Features;
178 // Printf("%s -> sz %zd ft %zd\n", Files[i].Name.c_str(),
179 // Files[i].Size, Cur.size());
180 bool FoundNewFeatures = false;
181 for (auto Fe: Cur) {
182 if (AllFeatures.insert(Fe).second) {
183 FoundNewFeatures = true;
184 NewFeatures->insert(Fe);
185 }
186 }
187 if (FoundNewFeatures)
188 NewFiles->push_back(Files[i].Name);
189 for (auto Cov : Files[i].Cov)
190 if (InitialCov.find(Cov) == InitialCov.end())
191 NewCov->insert(Cov);
192 }
193 return NewFeatures->size();
194}
195
196std::set<uint32_t> Merger::AllFeatures() const {
197 std::set<uint32_t> S;
198 for (auto &File : Files)
199 S.insert(File.Features.begin(), File.Features.end());
200 return S;
201}
202
203// Inner process. May crash if the target crashes.
204void Fuzzer::CrashResistantMergeInternalStep(const std::string &CFPath,
205 bool IsSetCoverMerge) {
206 Printf("MERGE-INNER: using the control file '%s'\n", CFPath.c_str());
207 Merger M;
208 std::ifstream IF(CFPath);
209 M.ParseOrExit(IF, false);
210 IF.close();
211 if (!M.LastFailure.empty())
212 Printf("MERGE-INNER: '%s' caused a failure at the previous merge step\n",
213 M.LastFailure.c_str());
214
215 Printf("MERGE-INNER: %zd total files;"
216 " %zd processed earlier; will process %zd files now\n",
217 M.Files.size(), M.FirstNotProcessedFile,
218 M.Files.size() - M.FirstNotProcessedFile);
219
220 std::ofstream OF(CFPath, std::ofstream::out | std::ofstream::app);
221 std::set<size_t> AllFeatures;
222 auto PrintStatsWrapper = [this, &AllFeatures](const char* Where) {
223 this->PrintStats(Where, "\n", 0, AllFeatures.size());
224 };
225 std::set<const TracePC::PCTableEntry *> AllPCs;
226 for (size_t i = M.FirstNotProcessedFile; i < M.Files.size(); i++) {
227 Fuzzer::MaybeExitGracefully();
228 auto U = FileToVector(M.Files[i].Name);
229 if (U.size() > MaxInputLen) {
230 U.resize(MaxInputLen);
231 U.shrink_to_fit();
232 }
233
234 // Write the pre-run marker.
235 OF << "STARTED " << i << " " << U.size() << "\n";
236 OF.flush(); // Flush is important since Command::Execute may crash.
237 // Run.
238 TPC.ResetMaps();
239 ExecuteCallback(U.data(), U.size());
240 // Collect coverage. We are iterating over the files in this order:
241 // * First, files in the initial corpus ordered by size, smallest first.
242 // * Then, all other files, smallest first.
243 std::set<size_t> Features;
244 if (IsSetCoverMerge)
245 TPC.CollectFeatures([&](size_t Feature) { Features.insert(Feature); });
246 else
247 TPC.CollectFeatures([&](size_t Feature) {
248 if (AllFeatures.insert(Feature).second)
249 Features.insert(Feature);
250 });
251 TPC.UpdateObservedPCs();
252 // Show stats.
253 if (!(TotalNumberOfRuns & (TotalNumberOfRuns - 1)))
254 PrintStatsWrapper("pulse ");
255 if (TotalNumberOfRuns == M.NumFilesInFirstCorpus)
256 PrintStatsWrapper("LOADED");
257 // Write the post-run marker and the coverage.
258 OF << "FT " << i;
259 for (size_t F : Features)
260 OF << " " << F;
261 OF << "\n";
262 OF << "COV " << i;
263 TPC.ForEachObservedPC([&](const TracePC::PCTableEntry *TE) {
264 if (AllPCs.insert(TE).second)
265 OF << " " << TPC.PCTableEntryIdx(TE);
266 });
267 OF << "\n";
268 OF.flush();
269 }
270 PrintStatsWrapper("DONE ");
271}
272
273// Merges all corpora into the first corpus. A file is added into
274// the first corpus only if it adds new features. Unlike `Merger::Merge`,
275// this implementation calculates an approximation of the minimum set
276// of corpora files, that cover all known features (set cover problem).
277// Generally, this means that files with more features are preferred for
278// merge into the first corpus. When two files have the same number of
279// features, the smaller one is preferred.
280size_t Merger::SetCoverMerge(const std::set<uint32_t> &InitialFeatures,
281 std::set<uint32_t> *NewFeatures,
282 const std::set<uint32_t> &InitialCov,
283 std::set<uint32_t> *NewCov,
284 std::vector<std::string> *NewFiles) {
285 assert(NumFilesInFirstCorpus <= Files.size());
286 NewFiles->clear();
287 NewFeatures->clear();
288 NewCov->clear();
289 std::set<uint32_t> AllFeatures;
290 // 1 << 21 - 1 is the maximum feature index.
291 // See 'kFeatureSetSize' in 'FuzzerCorpus.h'.
292 const uint32_t kFeatureSetSize = 1 << 21;
293 std::vector<bool> Covered(kFeatureSetSize, false);
294 size_t NumCovered = 0;
295
296 std::set<uint32_t> ExistingFeatures = InitialFeatures;
297 for (size_t i = 0; i < NumFilesInFirstCorpus; ++i)
298 ExistingFeatures.insert(Files[i].Features.begin(), Files[i].Features.end());
299
300 // Mark the existing features as covered.
301 for (const auto &F : ExistingFeatures) {
302 if (!Covered[F % kFeatureSetSize]) {
303 ++NumCovered;
304 Covered[F % kFeatureSetSize] = true;
305 }
306 // Calculate an underestimation of the set of covered features
307 // since the `Covered` bitvector is smaller than the feature range.
308 AllFeatures.insert(F % kFeatureSetSize);
309 }
310
311 std::set<size_t> RemainingFiles;
312 for (size_t i = NumFilesInFirstCorpus; i < Files.size(); ++i) {
313 // Construct an incremental sequence which represent the
314 // indices to all files (excluding those in the initial corpus).
315 // RemainingFiles = range(NumFilesInFirstCorpus..Files.size()).
316 RemainingFiles.insert(i);
317 // Insert this file's unique features to all features.
318 for (const auto &F : Files[i].Features)
319 AllFeatures.insert(F % kFeatureSetSize);
320 }
321
322 // Integrate files into Covered until set is complete.
323 while (NumCovered != AllFeatures.size()) {
324 // Index to file with largest number of unique features.
325 size_t MaxFeaturesIndex = NumFilesInFirstCorpus;
326 // Indices to remove from RemainingFiles.
327 std::set<size_t> RemoveIndices;
328 // Running max unique feature count.
329 // Updated upon finding a file with more features.
330 size_t MaxNumFeatures = 0;
331
332 // Iterate over all files not yet integrated into Covered,
333 // to find the file which has the largest number of
334 // features that are not already in Covered.
335 for (const auto &i : RemainingFiles) {
336 const auto &File = Files[i];
337 size_t CurrentUnique = 0;
338 // Count number of features in this file
339 // which are not yet in Covered.
340 for (const auto &F : File.Features)
341 if (!Covered[F % kFeatureSetSize])
342 ++CurrentUnique;
343
344 if (CurrentUnique == 0) {
345 // All features in this file are already in Covered: skip next time.
346 RemoveIndices.insert(i);
347 } else if (CurrentUnique > MaxNumFeatures ||
348 (CurrentUnique == MaxNumFeatures &&
349 File.Size < Files[MaxFeaturesIndex].Size)) {
350 // Update the max features file based on unique features
351 // Break ties by selecting smaller files.
352 MaxNumFeatures = CurrentUnique;
353 MaxFeaturesIndex = i;
354 }
355 }
356 // Must be a valid index/
357 assert(MaxFeaturesIndex < Files.size());
358 // Remove any feature-less files found.
359 for (const auto &i : RemoveIndices)
360 RemainingFiles.erase(i);
361 if (MaxNumFeatures == 0) {
362 // Did not find a file that adds unique features.
363 // This means that we should have no remaining files.
364 assert(RemainingFiles.size() == 0);
365 assert(NumCovered == AllFeatures.size());
366 break;
367 }
368
369 // MaxFeaturesIndex must be an element of Remaining.
370 assert(RemainingFiles.find(MaxFeaturesIndex) != RemainingFiles.end());
371 // Remove the file with the most features from Remaining.
372 RemainingFiles.erase(MaxFeaturesIndex);
373 const auto &MaxFeatureFile = Files[MaxFeaturesIndex];
374 // Add the features of the max feature file to Covered.
375 for (const auto &F : MaxFeatureFile.Features) {
376 if (!Covered[F % kFeatureSetSize]) {
377 ++NumCovered;
378 Covered[F % kFeatureSetSize] = true;
379 NewFeatures->insert(F);
380 }
381 }
382 // Add the index to this file to the result.
383 NewFiles->push_back(MaxFeatureFile.Name);
384 // Update NewCov with the additional coverage
385 // that MaxFeatureFile provides.
386 for (const auto &C : MaxFeatureFile.Cov)
387 if (InitialCov.find(C) == InitialCov.end())
388 NewCov->insert(C);
389 }
390
391 return NewFeatures->size();
392}
393
394static size_t
395WriteNewControlFile(const std::string &CFPath,
396 const std::vector<SizedFile> &OldCorpus,
397 const std::vector<SizedFile> &NewCorpus,
398 const std::vector<MergeFileInfo> &KnownFiles) {
399 std::unordered_set<std::string> FilesToSkip;
400 for (auto &SF: KnownFiles)
401 FilesToSkip.insert(SF.Name);
402
403 std::vector<std::string> FilesToUse;
404 auto MaybeUseFile = [=, &FilesToUse](std::string Name) {
405 if (FilesToSkip.find(Name) == FilesToSkip.end())
406 FilesToUse.push_back(Name);
407 };
408 for (auto &SF: OldCorpus)
409 MaybeUseFile(SF.File);
410 auto FilesToUseFromOldCorpus = FilesToUse.size();
411 for (auto &SF: NewCorpus)
412 MaybeUseFile(SF.File);
413
414 RemoveFile(CFPath);
415 std::ofstream ControlFile(CFPath);
416 ControlFile << FilesToUse.size() << "\n";
417 ControlFile << FilesToUseFromOldCorpus << "\n";
418 for (auto &FN: FilesToUse)
419 ControlFile << FN << "\n";
420
421 if (!ControlFile) {
422 Printf("MERGE-OUTER: failed to write to the control file: %s\n",
423 CFPath.c_str());
424 exit(1);
425 }
426
427 return FilesToUse.size();
428}
429
430// Outer process. Does not call the target code and thus should not fail.
431void CrashResistantMerge(const std::vector<std::string> &Args,
432 const std::vector<SizedFile> &OldCorpus,
433 const std::vector<SizedFile> &NewCorpus,
434 std::vector<std::string> *NewFiles,
435 const std::set<uint32_t> &InitialFeatures,
436 std::set<uint32_t> *NewFeatures,
437 const std::set<uint32_t> &InitialCov,
438 std::set<uint32_t> *NewCov, const std::string &CFPath,
439 bool V, /*Verbose*/
440 bool IsSetCoverMerge) {
441 if (NewCorpus.empty() && OldCorpus.empty()) return; // Nothing to merge.
442 size_t NumAttempts = 0;
443 std::vector<MergeFileInfo> KnownFiles;
444 if (FileSize(CFPath)) {
445 VPrintf(V, "MERGE-OUTER: non-empty control file provided: '%s'\n",
446 CFPath.c_str());
447 Merger M;
448 std::ifstream IF(CFPath);
449 if (M.Parse(IF, /*ParseCoverage=*/true)) {
450 VPrintf(V, "MERGE-OUTER: control file ok, %zd files total,"
451 " first not processed file %zd\n",
452 M.Files.size(), M.FirstNotProcessedFile);
453 if (!M.LastFailure.empty())
454 VPrintf(V, "MERGE-OUTER: '%s' will be skipped as unlucky "
455 "(merge has stumbled on it the last time)\n",
456 M.LastFailure.c_str());
457 if (M.FirstNotProcessedFile >= M.Files.size()) {
458 // Merge has already been completed with the given merge control file.
459 if (M.Files.size() == OldCorpus.size() + NewCorpus.size()) {
460 VPrintf(
461 V,
462 "MERGE-OUTER: nothing to do, merge has been completed before\n");
463 exit(0);
464 }
465
466 // Number of input files likely changed, start merge from scratch, but
467 // reuse coverage information from the given merge control file.
468 VPrintf(
469 V,
470 "MERGE-OUTER: starting merge from scratch, but reusing coverage "
471 "information from the given control file\n");
472 KnownFiles = M.Files;
473 } else {
474 // There is a merge in progress, continue.
475 NumAttempts = M.Files.size() - M.FirstNotProcessedFile;
476 }
477 } else {
478 VPrintf(V, "MERGE-OUTER: bad control file, will overwrite it\n");
479 }
480 }
481
482 if (!NumAttempts) {
483 // The supplied control file is empty or bad, create a fresh one.
484 VPrintf(V, "MERGE-OUTER: "
485 "%zd files, %zd in the initial corpus, %zd processed earlier\n",
486 OldCorpus.size() + NewCorpus.size(), OldCorpus.size(),
487 KnownFiles.size());
488 NumAttempts = WriteNewControlFile(CFPath, OldCorpus, NewCorpus, KnownFiles);
489 }
490
491 // Execute the inner process until it passes.
492 // Every inner process should execute at least one input.
493 Command BaseCmd(Args);
494 BaseCmd.removeFlag("merge");
495 BaseCmd.removeFlag("set_cover_merge");
496 BaseCmd.removeFlag("fork");
497 BaseCmd.removeFlag("collect_data_flow");
498 for (size_t Attempt = 1; Attempt <= NumAttempts; Attempt++) {
499 Fuzzer::MaybeExitGracefully();
500 VPrintf(V, "MERGE-OUTER: attempt %zd\n", Attempt);
501 Command Cmd(BaseCmd);
502 Cmd.addFlag("merge_control_file", CFPath);
503 // If we are going to use the set cover implementation for
504 // minimization add the merge_inner=2 internal flag.
505 Cmd.addFlag("merge_inner", IsSetCoverMerge ? "2" : "1");
506 if (!V) {
507 Cmd.setOutputFile(getDevNull());
508 Cmd.combineOutAndErr();
509 }
510 auto ExitCode = ExecuteCommand(Cmd);
511 if (!ExitCode) {
512 VPrintf(V, "MERGE-OUTER: successful in %zd attempt(s)\n", Attempt);
513 break;
514 }
515 }
516 // Read the control file and do the merge.
517 Merger M;
518 std::ifstream IF(CFPath);
519 IF.seekg(0, IF.end);
520 VPrintf(V, "MERGE-OUTER: the control file has %zd bytes\n",
521 (size_t)IF.tellg());
522 IF.seekg(0, IF.beg);
523 M.ParseOrExit(IF, true);
524 IF.close();
525 VPrintf(V,
526 "MERGE-OUTER: consumed %zdMb (%zdMb rss) to parse the control file\n",
527 M.ApproximateMemoryConsumption() >> 20, GetPeakRSSMb());
528
529 M.Files.insert(M.Files.end(), KnownFiles.begin(), KnownFiles.end());
530 if (IsSetCoverMerge)
531 M.SetCoverMerge(InitialFeatures, NewFeatures, InitialCov, NewCov, NewFiles);
532 else
533 M.Merge(InitialFeatures, NewFeatures, InitialCov, NewCov, NewFiles);
534 VPrintf(V, "MERGE-OUTER: %zd new files with %zd new features added; "
535 "%zd new coverage edges\n",
536 NewFiles->size(), NewFeatures->size(), NewCov->size());
537}
538
539} // namespace fuzzer
540