1 //===- FuzzerFork.cpp - run fuzzing in separate subprocesses --------------===//
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 // Spawn and orchestrate separate fuzzing processes.
9 //===----------------------------------------------------------------------===//
10 
11 #include "FuzzerCommand.h"
12 #include "FuzzerFork.h"
13 #include "FuzzerIO.h"
14 #include "FuzzerInternal.h"
15 #include "FuzzerMerge.h"
16 #include "FuzzerSHA1.h"
17 #include "FuzzerTracePC.h"
18 #include "FuzzerUtil.h"
19 
20 #include <atomic>
21 #include <chrono>
22 #include <fstream>
23 #include <memory>
24 #include <mutex>
25 #include <queue>
26 #include <sstream>
27 #include <thread>
28 
29 namespace fuzzer {
30 
31 struct Stats {
32   size_t number_of_executed_units = 0;
33   size_t peak_rss_mb = 0;
34   size_t average_exec_per_sec = 0;
35 };
36 
37 static Stats ParseFinalStatsFromLog(const std::string &LogPath) {
38   std::ifstream In(LogPath);
39   std::string Line;
40   Stats Res;
41   struct {
42     const char *Name;
43     size_t *Var;
44   } NameVarPairs[] = {
45       {"stat::number_of_executed_units:", &Res.number_of_executed_units},
46       {"stat::peak_rss_mb:", &Res.peak_rss_mb},
47       {"stat::average_exec_per_sec:", &Res.average_exec_per_sec},
48       {nullptr, nullptr},
49   };
50   while (std::getline(In, Line, '\n')) {
51     if (Line.find("stat::") != 0) continue;
52     std::istringstream ISS(Line);
53     std::string Name;
54     size_t Val;
55     ISS >> Name >> Val;
56     for (size_t i = 0; NameVarPairs[i].Name; i++)
57       if (Name == NameVarPairs[i].Name)
58         *NameVarPairs[i].Var = Val;
59   }
60   return Res;
61 }
62 
63 struct FuzzJob {
64   // Inputs.
65   Command Cmd;
66   std::string CorpusDir;
67   std::string FeaturesDir;
68   std::string LogPath;
69   std::string CFPath;
70 
71   // Fuzzing Outputs.
72   int ExitCode;
73 
74   ~FuzzJob() {
75     RemoveFile(CFPath);
76     RemoveFile(LogPath);
77     RmDirRecursive(CorpusDir);
78     RmDirRecursive(FeaturesDir);
79   }
80 };
81 
82 struct GlobalEnv {
83   Vector<std::string> Args;
84   Vector<std::string> CorpusDirs;
85   std::string MainCorpusDir;
86   std::string TempDir;
87   Set<uint32_t> Features, Cov;
88   Vector<std::string> Files;
89   Random *Rand;
90   std::chrono::system_clock::time_point ProcessStartTime;
91   int Verbosity = 0;
92 
93   size_t NumTimeouts = 0;
94   size_t NumOOMs = 0;
95   size_t NumCrashes = 0;
96 
97 
98   size_t NumRuns = 0;
99 
100   size_t secondsSinceProcessStartUp() const {
101     return std::chrono::duration_cast<std::chrono::seconds>(
102                std::chrono::system_clock::now() - ProcessStartTime)
103         .count();
104   }
105 
106   FuzzJob *CreateNewJob(size_t JobId) {
107     Command Cmd(Args);
108     Cmd.removeFlag("fork");
109     Cmd.removeFlag("runs");
110     for (auto &C : CorpusDirs) // Remove all corpora from the args.
111       Cmd.removeArgument(C);
112     Cmd.addFlag("reload", "0");  // working in an isolated dir, no reload.
113     Cmd.addFlag("print_final_stats", "1");
114     Cmd.addFlag("print_funcs", "0");  // no need to spend time symbolizing.
115     Cmd.addFlag("max_total_time", std::to_string(std::min((size_t)300, JobId)));
116 
117     auto Job = new FuzzJob;
118     std::string Seeds;
119     if (size_t CorpusSubsetSize =
120             std::min(Files.size(), (size_t)sqrt(Files.size() + 2)))
121       for (size_t i = 0; i < CorpusSubsetSize; i++)
122         Seeds += (Seeds.empty() ? "" : ",") +
123                  Files[Rand->SkewTowardsLast(Files.size())];
124     if (!Seeds.empty())
125       Cmd.addFlag("seed_inputs", Seeds);
126     Job->LogPath = DirPlusFile(TempDir, std::to_string(JobId) + ".log");
127     Job->CorpusDir = DirPlusFile(TempDir, "C" + std::to_string(JobId));
128     Job->FeaturesDir = DirPlusFile(TempDir, "F" + std::to_string(JobId));
129     Job->CFPath = DirPlusFile(TempDir, std::to_string(JobId) + ".merge");
130 
131 
132     Cmd.addArgument(Job->CorpusDir);
133     Cmd.addFlag("features_dir", Job->FeaturesDir);
134 
135     for (auto &D : {Job->CorpusDir, Job->FeaturesDir}) {
136       RmDirRecursive(D);
137       MkDir(D);
138     }
139 
140     Cmd.setOutputFile(Job->LogPath);
141     Cmd.combineOutAndErr();
142 
143     Job->Cmd = Cmd;
144 
145     if (Verbosity >= 2)
146       Printf("Job %zd/%p Created: %s\n", JobId, Job,
147              Job->Cmd.toString().c_str());
148     // Start from very short runs and gradually increase them.
149     return Job;
150   }
151 
152   void RunOneMergeJob(FuzzJob *Job) {
153     auto Stats = ParseFinalStatsFromLog(Job->LogPath);
154     NumRuns += Stats.number_of_executed_units;
155 
156     Vector<SizedFile> TempFiles, MergeCandidates;
157     // Read all newly created inputs and their feature sets.
158     // Choose only those inputs that have new features.
159     GetSizedFilesFromDir(Job->CorpusDir, &TempFiles);
160     std::sort(TempFiles.begin(), TempFiles.end());
161     for (auto &F : TempFiles) {
162       auto FeatureFile = F.File;
163       FeatureFile.replace(0, Job->CorpusDir.size(), Job->FeaturesDir);
164       auto FeatureBytes = FileToVector(FeatureFile, 0, false);
165       assert((FeatureBytes.size() % sizeof(uint32_t)) == 0);
166       Vector<uint32_t> NewFeatures(FeatureBytes.size() / sizeof(uint32_t));
167       memcpy(NewFeatures.data(), FeatureBytes.data(), FeatureBytes.size());
168       for (auto Ft : NewFeatures) {
169         if (!Features.count(Ft)) {
170           MergeCandidates.push_back(F);
171           break;
172         }
173       }
174     }
175     if (MergeCandidates.empty()) return;
176 
177     Vector<std::string> FilesToAdd;
178     Set<uint32_t> NewFeatures, NewCov;
179     CrashResistantMerge(Args, {}, MergeCandidates, &FilesToAdd, Features,
180                         &NewFeatures, Cov, &NewCov, Job->CFPath, false);
181     for (auto &Path : FilesToAdd) {
182       auto U = FileToVector(Path);
183       auto NewPath = DirPlusFile(MainCorpusDir, Hash(U));
184       WriteToFile(U, NewPath);
185       Files.push_back(NewPath);
186     }
187     Features.insert(NewFeatures.begin(), NewFeatures.end());
188     Cov.insert(NewCov.begin(), NewCov.end());
189     for (auto Idx : NewCov)
190       if (auto *TE = TPC.PCTableEntryByIdx(Idx))
191         if (TPC.PcIsFuncEntry(TE))
192           PrintPC("  NEW_FUNC: %p %F %L\n", "",
193                   TPC.GetNextInstructionPc(TE->PC));
194 
195     if (!FilesToAdd.empty() || Job->ExitCode != 0)
196       Printf("#%zd: cov: %zd ft: %zd corp: %zd exec/s %zd "
197              "oom/timeout/crash: %zd/%zd/%zd time: %zds\n", NumRuns,
198              Cov.size(), Features.size(), Files.size(),
199              Stats.average_exec_per_sec,
200              NumOOMs, NumTimeouts, NumCrashes, secondsSinceProcessStartUp());
201   }
202 };
203 
204 struct JobQueue {
205   std::queue<FuzzJob *> Qu;
206   std::mutex Mu;
207 
208   void Push(FuzzJob *Job) {
209     std::lock_guard<std::mutex> Lock(Mu);
210     Qu.push(Job);
211   }
212   FuzzJob *Pop() {
213     std::lock_guard<std::mutex> Lock(Mu);
214     if (Qu.empty()) return nullptr;
215     auto Job = Qu.front();
216     Qu.pop();
217     return Job;
218   }
219 };
220 
221 void WorkerThread(std::atomic<bool> *Stop, JobQueue *FuzzQ, JobQueue *MergeQ) {
222   while (!Stop->load()) {
223     auto Job = FuzzQ->Pop();
224     // Printf("WorkerThread: job %p\n", Job);
225     if (!Job) {
226       SleepSeconds(1);
227       continue;
228     }
229     Job->ExitCode = ExecuteCommand(Job->Cmd);
230     MergeQ->Push(Job);
231   }
232 }
233 
234 // This is just a skeleton of an experimental -fork=1 feature.
235 void FuzzWithFork(Random &Rand, const FuzzingOptions &Options,
236                   const Vector<std::string> &Args,
237                   const Vector<std::string> &CorpusDirs, int NumJobs) {
238   Printf("INFO: -fork=%d: fuzzing in separate process(s)\n", NumJobs);
239 
240   GlobalEnv Env;
241   Env.Args = Args;
242   Env.CorpusDirs = CorpusDirs;
243   Env.Rand = &Rand;
244   Env.Verbosity = Options.Verbosity;
245   Env.ProcessStartTime = std::chrono::system_clock::now();
246 
247   Vector<SizedFile> SeedFiles;
248   for (auto &Dir : CorpusDirs)
249     GetSizedFilesFromDir(Dir, &SeedFiles);
250   std::sort(SeedFiles.begin(), SeedFiles.end());
251   Env.TempDir = TempPath(".dir");
252   RmDirRecursive(Env.TempDir);  // in case there is a leftover from old runs.
253   MkDir(Env.TempDir);
254 
255 
256   if (CorpusDirs.empty())
257     MkDir(Env.MainCorpusDir = DirPlusFile(Env.TempDir, "C"));
258   else
259     Env.MainCorpusDir = CorpusDirs[0];
260 
261   auto CFPath = DirPlusFile(Env.TempDir, "merge.txt");
262   CrashResistantMerge(Env.Args, {}, SeedFiles, &Env.Files, {}, &Env.Features,
263                       {}, &Env.Cov,
264                       CFPath, false);
265   RemoveFile(CFPath);
266   Printf("INFO: -fork=%d: %zd seed inputs, starting to fuzz in %s\n", NumJobs,
267          Env.Files.size(), Env.TempDir.c_str());
268 
269   int ExitCode = 0;
270 
271   JobQueue FuzzQ, MergeQ;
272   std::atomic<bool> Stop(false);
273 
274   size_t JobId = 1;
275   Vector<std::thread> Threads;
276   for (int t = 0; t < NumJobs; t++) {
277     Threads.push_back(std::thread(WorkerThread, &Stop, &FuzzQ, &MergeQ));
278     FuzzQ.Push(Env.CreateNewJob(JobId++));
279   }
280 
281   while (true) {
282     std::unique_ptr<FuzzJob> Job(MergeQ.Pop());
283     if (!Job) {
284       if (Stop)
285         break;
286       SleepSeconds(1);
287       continue;
288     }
289     ExitCode = Job->ExitCode;
290     if (ExitCode == Options.InterruptExitCode) {
291       Printf("==%lu== libFuzzer: a child was interrupted; exiting\n", GetPid());
292       Stop = true;
293       break;
294     }
295     Fuzzer::MaybeExitGracefully();
296 
297     Env.RunOneMergeJob(Job.get());
298 
299     // Continue if our crash is one of the ignorred ones.
300     if (Options.IgnoreTimeouts && ExitCode == Options.TimeoutExitCode)
301       Env.NumTimeouts++;
302     else if (Options.IgnoreOOMs && ExitCode == Options.OOMExitCode)
303       Env.NumOOMs++;
304     else if (ExitCode != 0) {
305       Env.NumCrashes++;
306       if (Options.IgnoreCrashes) {
307         std::ifstream In(Job->LogPath);
308         std::string Line;
309         while (std::getline(In, Line, '\n'))
310           if (Line.find("ERROR:") != Line.npos ||
311               Line.find("runtime error:") != Line.npos)
312             Printf("%s\n", Line.c_str());
313       } else {
314         // And exit if we don't ignore this crash.
315         Printf("INFO: log from the inner process:\n%s",
316                FileToString(Job->LogPath).c_str());
317         Stop = true;
318       }
319     }
320 
321     // Stop if we are over the time budget.
322     // This is not precise, since other threads are still running
323     // and we will wait while joining them.
324     // We also don't stop instantly: other jobs need to finish.
325     if (Options.MaxTotalTimeSec > 0 && !Stop &&
326         Env.secondsSinceProcessStartUp() >= (size_t)Options.MaxTotalTimeSec) {
327       Printf("INFO: fuzzed for %zd seconds, wrapping up soon\n",
328              Env.secondsSinceProcessStartUp());
329       Stop = true;
330     }
331     if (!Stop && Env.NumRuns >= Options.MaxNumberOfRuns) {
332       Printf("INFO: fuzzed for %zd iterations, wrapping up soon\n",
333              Env.NumRuns);
334       Stop = true;
335     }
336 
337     if (!Stop)
338       FuzzQ.Push(Env.CreateNewJob(JobId++));
339   }
340   Stop = true;
341 
342   for (auto &T : Threads)
343     T.join();
344 
345   // The workers have terminated. Don't try to remove the directory before they
346   // terminate to avoid a race condition preventing cleanup on Windows.
347   RmDirRecursive(Env.TempDir);
348 
349   // Use the exit code from the last child process.
350   Printf("INFO: exiting: %d time: %zds\n", ExitCode,
351          Env.secondsSinceProcessStartUp());
352   exit(ExitCode);
353 }
354 
355 } // namespace fuzzer
356