1//===- llvm/Support/Unix/Program.cpp -----------------------------*- 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//
9// This file implements the Unix specific portion of the Program class.
10//
11//===----------------------------------------------------------------------===//
12
13//===----------------------------------------------------------------------===//
14//=== WARNING: Implementation here must contain only generic UNIX code that
15//===          is guaranteed to work on *all* UNIX variants.
16//===----------------------------------------------------------------------===//
17
18#include "llvm/Support/Program.h"
19
20#include "Unix.h"
21#include "llvm/ADT/StringExtras.h"
22#include "llvm/Config/config.h"
23#include "llvm/Support/Compiler.h"
24#include "llvm/Support/Errc.h"
25#include "llvm/Support/FileSystem.h"
26#include "llvm/Support/Path.h"
27#include "llvm/Support/StringSaver.h"
28#include "llvm/Support/raw_ostream.h"
29#if HAVE_SYS_STAT_H
30#include <sys/stat.h>
31#endif
32#if HAVE_SYS_RESOURCE_H
33#include <sys/resource.h>
34#endif
35#if HAVE_SIGNAL_H
36#include <signal.h>
37#endif
38#if HAVE_FCNTL_H
39#include <fcntl.h>
40#endif
41#if HAVE_UNISTD_H
42#include <unistd.h>
43#endif
44#ifdef HAVE_POSIX_SPAWN
45#include <spawn.h>
46
47#if defined(__APPLE__)
48#include <TargetConditionals.h>
49#endif
50
51#if defined(__APPLE__) && !(defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE)
52#define USE_NSGETENVIRON 1
53#else
54#define USE_NSGETENVIRON 0
55#endif
56
57#if !USE_NSGETENVIRON
58  extern char **environ;
59#else
60#include <crt_externs.h> // _NSGetEnviron
61#endif
62#endif
63
64using namespace llvm;
65using namespace sys;
66
67ProcessInfo::ProcessInfo() : Pid(0), ReturnCode(0) {}
68
69ErrorOr<std::string> sys::findProgramByName(StringRef Name,
70                                            ArrayRef<StringRef> Paths) {
71  assert(!Name.empty() && "Must have a name!");
72  // Use the given path verbatim if it contains any slashes; this matches
73  // the behavior of sh(1) and friends.
74  if (Name.find('/') != StringRef::npos) return std::string(Name);
75
76  SmallVector<StringRef, 16> EnvironmentPaths;
77  if (Paths.empty())
78    if (const char *PathEnv = std::getenv("PATH")) {
79      SplitString(PathEnv, EnvironmentPaths, ":");
80      Paths = EnvironmentPaths;
81    }
82
83  for (auto Path : Paths) {
84    if (Path.empty())
85      continue;
86
87    // Check to see if this first directory contains the executable...
88    SmallString<128> FilePath(Path);
89    sys::path::append(FilePath, Name);
90    if (sys::fs::can_execute(FilePath.c_str()))
91      return std::string(FilePath.str());  // Found the executable!
92  }
93  return errc::no_such_file_or_directory;
94}
95
96static bool RedirectIO(Optional<StringRef> Path, int FD, std::string* ErrMsg) {
97  if (!Path) // Noop
98    return false;
99  std::string File;
100  if (Path->empty())
101    // Redirect empty paths to /dev/null
102    File = "/dev/null";
103  else
104    File = std::string(*Path);
105
106  // Open the file
107  int InFD = open(File.c_str(), FD == 0 ? O_RDONLY : O_WRONLY|O_CREAT, 0666);
108  if (InFD == -1) {
109    MakeErrMsg(ErrMsg, "Cannot open file '" + File + "' for "
110              + (FD == 0 ? "input" : "output"));
111    return true;
112  }
113
114  // Install it as the requested FD
115  if (dup2(InFD, FD) == -1) {
116    MakeErrMsg(ErrMsg, "Cannot dup2");
117    close(InFD);
118    return true;
119  }
120  close(InFD);      // Close the original FD
121  return false;
122}
123
124#ifdef HAVE_POSIX_SPAWN
125static bool RedirectIO_PS(const std::string *Path, int FD, std::string *ErrMsg,
126                          posix_spawn_file_actions_t *FileActions) {
127  if (!Path) // Noop
128    return false;
129  const char *File;
130  if (Path->empty())
131    // Redirect empty paths to /dev/null
132    File = "/dev/null";
133  else
134    File = Path->c_str();
135
136  if (int Err = posix_spawn_file_actions_addopen(
137          FileActions, FD, File,
138          FD == 0 ? O_RDONLY : O_WRONLY | O_CREAT, 0666))
139    return MakeErrMsg(ErrMsg, "Cannot posix_spawn_file_actions_addopen", Err);
140  return false;
141}
142#endif
143
144static void TimeOutHandler(int Sig) {
145}
146
147static void SetMemoryLimits(unsigned size) {
148#if HAVE_SYS_RESOURCE_H && HAVE_GETRLIMIT && HAVE_SETRLIMIT
149  struct rlimit r;
150  __typeof__ (r.rlim_cur) limit = (__typeof__ (r.rlim_cur)) (size) * 1048576;
151
152  // Heap size
153  getrlimit (RLIMIT_DATA, &r);
154  r.rlim_cur = limit;
155  setrlimit (RLIMIT_DATA, &r);
156#ifdef RLIMIT_RSS
157  // Resident set size.
158  getrlimit (RLIMIT_RSS, &r);
159  r.rlim_cur = limit;
160  setrlimit (RLIMIT_RSS, &r);
161#endif
162#endif
163}
164
165static std::vector<const char *>
166toNullTerminatedCStringArray(ArrayRef<StringRef> Strings, StringSaver &Saver) {
167  std::vector<const char *> Result;
168  for (StringRef S : Strings)
169    Result.push_back(Saver.save(S).data());
170  Result.push_back(nullptr);
171  return Result;
172}
173
174static bool Execute(ProcessInfo &PI, StringRef Program,
175                    ArrayRef<StringRef> Args, Optional<ArrayRef<StringRef>> Env,
176                    ArrayRef<Optional<StringRef>> Redirects,
177                    unsigned MemoryLimit, std::string *ErrMsg,
178                    BitVector *AffinityMask) {
179  if (!llvm::sys::fs::exists(Program)) {
180    if (ErrMsg)
181      *ErrMsg = std::string("Executable \"") + Program.str() +
182                std::string("\" doesn't exist!");
183    return false;
184  }
185
186  assert(!AffinityMask && "Starting a process with an affinity mask is "
187                          "currently not supported on Unix!");
188
189  BumpPtrAllocator Allocator;
190  StringSaver Saver(Allocator);
191  std::vector<const char *> ArgVector, EnvVector;
192  const char **Argv = nullptr;
193  const char **Envp = nullptr;
194  ArgVector = toNullTerminatedCStringArray(Args, Saver);
195  Argv = ArgVector.data();
196  if (Env) {
197    EnvVector = toNullTerminatedCStringArray(*Env, Saver);
198    Envp = EnvVector.data();
199  }
200
201  // If this OS has posix_spawn and there is no memory limit being implied, use
202  // posix_spawn.  It is more efficient than fork/exec.
203#ifdef HAVE_POSIX_SPAWN
204  if (MemoryLimit == 0) {
205    posix_spawn_file_actions_t FileActionsStore;
206    posix_spawn_file_actions_t *FileActions = nullptr;
207
208    // If we call posix_spawn_file_actions_addopen we have to make sure the
209    // c strings we pass to it stay alive until the call to posix_spawn,
210    // so we copy any StringRefs into this variable.
211    std::string RedirectsStorage[3];
212
213    if (!Redirects.empty()) {
214      assert(Redirects.size() == 3);
215      std::string *RedirectsStr[3] = {nullptr, nullptr, nullptr};
216      for (int I = 0; I < 3; ++I) {
217        if (Redirects[I]) {
218          RedirectsStorage[I] = std::string(*Redirects[I]);
219          RedirectsStr[I] = &RedirectsStorage[I];
220        }
221      }
222
223      FileActions = &FileActionsStore;
224      posix_spawn_file_actions_init(FileActions);
225
226      // Redirect stdin/stdout.
227      if (RedirectIO_PS(RedirectsStr[0], 0, ErrMsg, FileActions) ||
228          RedirectIO_PS(RedirectsStr[1], 1, ErrMsg, FileActions))
229        return false;
230      if (!Redirects[1] || !Redirects[2] || *Redirects[1] != *Redirects[2]) {
231        // Just redirect stderr
232        if (RedirectIO_PS(RedirectsStr[2], 2, ErrMsg, FileActions))
233          return false;
234      } else {
235        // If stdout and stderr should go to the same place, redirect stderr
236        // to the FD already open for stdout.
237        if (int Err = posix_spawn_file_actions_adddup2(FileActions, 1, 2))
238          return !MakeErrMsg(ErrMsg, "Can't redirect stderr to stdout", Err);
239      }
240    }
241
242    if (!Envp)
243#if !USE_NSGETENVIRON
244      Envp = const_cast<const char **>(environ);
245#else
246      // environ is missing in dylibs.
247      Envp = const_cast<const char **>(*_NSGetEnviron());
248#endif
249
250    constexpr int maxRetries = 8;
251    int retries = 0;
252    pid_t PID;
253    int Err;
254    do {
255      PID = 0; // Make Valgrind happy.
256      Err = posix_spawn(&PID, Program.str().c_str(), FileActions,
257                        /*attrp*/ nullptr, const_cast<char **>(Argv),
258                        const_cast<char **>(Envp));
259    } while (Err == EINTR && ++retries < maxRetries);
260
261    if (FileActions)
262      posix_spawn_file_actions_destroy(FileActions);
263
264    if (Err)
265     return !MakeErrMsg(ErrMsg, "posix_spawn failed", Err);
266
267    PI.Pid = PID;
268    PI.Process = PID;
269
270    return true;
271  }
272#endif
273
274  // Create a child process.
275  int child = fork();
276  switch (child) {
277    // An error occurred:  Return to the caller.
278    case -1:
279      MakeErrMsg(ErrMsg, "Couldn't fork");
280      return false;
281
282    // Child process: Execute the program.
283    case 0: {
284      // Redirect file descriptors...
285      if (!Redirects.empty()) {
286        // Redirect stdin
287        if (RedirectIO(Redirects[0], 0, ErrMsg)) { return false; }
288        // Redirect stdout
289        if (RedirectIO(Redirects[1], 1, ErrMsg)) { return false; }
290        if (Redirects[1] && Redirects[2] && *Redirects[1] == *Redirects[2]) {
291          // If stdout and stderr should go to the same place, redirect stderr
292          // to the FD already open for stdout.
293          if (-1 == dup2(1,2)) {
294            MakeErrMsg(ErrMsg, "Can't redirect stderr to stdout");
295            return false;
296          }
297        } else {
298          // Just redirect stderr
299          if (RedirectIO(Redirects[2], 2, ErrMsg)) { return false; }
300        }
301      }
302
303      // Set memory limits
304      if (MemoryLimit!=0) {
305        SetMemoryLimits(MemoryLimit);
306      }
307
308      // Execute!
309      std::string PathStr = std::string(Program);
310      if (Envp != nullptr)
311        execve(PathStr.c_str(), const_cast<char **>(Argv),
312               const_cast<char **>(Envp));
313      else
314        execv(PathStr.c_str(), const_cast<char **>(Argv));
315      // If the execve() failed, we should exit. Follow Unix protocol and
316      // return 127 if the executable was not found, and 126 otherwise.
317      // Use _exit rather than exit so that atexit functions and static
318      // object destructors cloned from the parent process aren't
319      // redundantly run, and so that any data buffered in stdio buffers
320      // cloned from the parent aren't redundantly written out.
321      _exit(errno == ENOENT ? 127 : 126);
322    }
323
324    // Parent process: Break out of the switch to do our processing.
325    default:
326      break;
327  }
328
329  PI.Pid = child;
330  PI.Process = child;
331
332  return true;
333}
334
335namespace llvm {
336namespace sys {
337
338#ifndef _AIX
339using ::wait4;
340#else
341static pid_t (wait4)(pid_t pid, int *status, int options, struct rusage *usage);
342#endif
343
344} // namespace sys
345} // namespace llvm
346
347#ifdef _AIX
348#ifndef _ALL_SOURCE
349extern "C" pid_t (wait4)(pid_t pid, int *status, int options,
350                         struct rusage *usage);
351#endif
352pid_t (llvm::sys::wait4)(pid_t pid, int *status, int options,
353                         struct rusage *usage) {
354  assert(pid > 0 && "Only expecting to handle actual PID values!");
355  assert((options & ~WNOHANG) == 0 && "Expecting WNOHANG at most!");
356  assert(usage && "Expecting usage collection!");
357
358  // AIX wait4 does not work well with WNOHANG.
359  if (!(options & WNOHANG))
360    return ::wait4(pid, status, options, usage);
361
362  // For WNOHANG, we use waitid (which supports WNOWAIT) until the child process
363  // has terminated.
364  siginfo_t WaitIdInfo;
365  WaitIdInfo.si_pid = 0;
366  int WaitIdRetVal =
367      waitid(P_PID, pid, &WaitIdInfo, WNOWAIT | WEXITED | options);
368
369  if (WaitIdRetVal == -1 || WaitIdInfo.si_pid == 0)
370    return WaitIdRetVal;
371
372  assert(WaitIdInfo.si_pid == pid);
373
374  // The child has already terminated, so a blocking wait on it is okay in the
375  // absence of indiscriminate `wait` calls from the current process (which
376  // would cause the call here to fail with ECHILD).
377  return ::wait4(pid, status, options & ~WNOHANG, usage);
378}
379#endif
380
381ProcessInfo llvm::sys::Wait(const ProcessInfo &PI, unsigned SecondsToWait,
382                            bool WaitUntilTerminates, std::string *ErrMsg,
383                            Optional<ProcessStatistics> *ProcStat) {
384  struct sigaction Act, Old;
385  assert(PI.Pid && "invalid pid to wait on, process not started?");
386
387  int WaitPidOptions = 0;
388  pid_t ChildPid = PI.Pid;
389  if (WaitUntilTerminates) {
390    SecondsToWait = 0;
391  } else if (SecondsToWait) {
392    // Install a timeout handler.  The handler itself does nothing, but the
393    // simple fact of having a handler at all causes the wait below to return
394    // with EINTR, unlike if we used SIG_IGN.
395    memset(&Act, 0, sizeof(Act));
396    Act.sa_handler = TimeOutHandler;
397    sigemptyset(&Act.sa_mask);
398    sigaction(SIGALRM, &Act, &Old);
399    // FIXME The alarm signal may be delivered to another thread.
400    alarm(SecondsToWait);
401  } else if (SecondsToWait == 0)
402    WaitPidOptions = WNOHANG;
403
404  // Parent process: Wait for the child process to terminate.
405  int status;
406  ProcessInfo WaitResult;
407  rusage Info;
408  if (ProcStat)
409    ProcStat->reset();
410
411  do {
412    WaitResult.Pid = sys::wait4(ChildPid, &status, WaitPidOptions, &Info);
413  } while (WaitUntilTerminates && WaitResult.Pid == -1 && errno == EINTR);
414
415  if (WaitResult.Pid != PI.Pid) {
416    if (WaitResult.Pid == 0) {
417      // Non-blocking wait.
418      return WaitResult;
419    } else {
420      if (SecondsToWait && errno == EINTR) {
421        // Kill the child.
422        kill(PI.Pid, SIGKILL);
423
424        // Turn off the alarm and restore the signal handler
425        alarm(0);
426        sigaction(SIGALRM, &Old, nullptr);
427
428        // Wait for child to die
429        // FIXME This could grab some other child process out from another
430        // waiting thread and then leave a zombie anyway.
431        if (wait(&status) != ChildPid)
432          MakeErrMsg(ErrMsg, "Child timed out but wouldn't die");
433        else
434          MakeErrMsg(ErrMsg, "Child timed out", 0);
435
436        WaitResult.ReturnCode = -2; // Timeout detected
437        return WaitResult;
438      } else if (errno != EINTR) {
439        MakeErrMsg(ErrMsg, "Error waiting for child process");
440        WaitResult.ReturnCode = -1;
441        return WaitResult;
442      }
443    }
444  }
445
446  // We exited normally without timeout, so turn off the timer.
447  if (SecondsToWait && !WaitUntilTerminates) {
448    alarm(0);
449    sigaction(SIGALRM, &Old, nullptr);
450  }
451
452  if (ProcStat) {
453    std::chrono::microseconds UserT = toDuration(Info.ru_utime);
454    std::chrono::microseconds KernelT = toDuration(Info.ru_stime);
455    uint64_t PeakMemory = static_cast<uint64_t>(Info.ru_maxrss);
456    *ProcStat = ProcessStatistics{UserT + KernelT, UserT, PeakMemory};
457  }
458
459  // Return the proper exit status. Detect error conditions
460  // so we can return -1 for them and set ErrMsg informatively.
461  int result = 0;
462  if (WIFEXITED(status)) {
463    result = WEXITSTATUS(status);
464    WaitResult.ReturnCode = result;
465
466    if (result == 127) {
467      if (ErrMsg)
468        *ErrMsg = llvm::sys::StrError(ENOENT);
469      WaitResult.ReturnCode = -1;
470      return WaitResult;
471    }
472    if (result == 126) {
473      if (ErrMsg)
474        *ErrMsg = "Program could not be executed";
475      WaitResult.ReturnCode = -1;
476      return WaitResult;
477    }
478  } else if (WIFSIGNALED(status)) {
479    if (ErrMsg) {
480      *ErrMsg = strsignal(WTERMSIG(status));
481#ifdef WCOREDUMP
482      if (WCOREDUMP(status))
483        *ErrMsg += " (core dumped)";
484#endif
485    }
486    // Return a special value to indicate that the process received an unhandled
487    // signal during execution as opposed to failing to execute.
488    WaitResult.ReturnCode = -2;
489  }
490  return WaitResult;
491}
492
493std::error_code llvm::sys::ChangeStdinToBinary() {
494  // Do nothing, as Unix doesn't differentiate between text and binary.
495  return std::error_code();
496}
497
498std::error_code llvm::sys::ChangeStdoutToBinary() {
499  // Do nothing, as Unix doesn't differentiate between text and binary.
500  return std::error_code();
501}
502
503std::error_code
504llvm::sys::writeFileWithEncoding(StringRef FileName, StringRef Contents,
505                                 WindowsEncodingMethod Encoding /*unused*/) {
506  std::error_code EC;
507  llvm::raw_fd_ostream OS(FileName, EC, llvm::sys::fs::OpenFlags::OF_Text);
508
509  if (EC)
510    return EC;
511
512  OS << Contents;
513
514  if (OS.has_error())
515    return make_error_code(errc::io_error);
516
517  return EC;
518}
519
520bool llvm::sys::commandLineFitsWithinSystemLimits(StringRef Program,
521                                                  ArrayRef<StringRef> Args) {
522  static long ArgMax = sysconf(_SC_ARG_MAX);
523  // POSIX requires that _POSIX_ARG_MAX is 4096, which is the lowest possible
524  // value for ARG_MAX on a POSIX compliant system.
525  static long ArgMin = _POSIX_ARG_MAX;
526
527  // This the same baseline used by xargs.
528  long EffectiveArgMax = 128 * 1024;
529
530  if (EffectiveArgMax > ArgMax)
531    EffectiveArgMax = ArgMax;
532  else if (EffectiveArgMax < ArgMin)
533    EffectiveArgMax = ArgMin;
534
535  // System says no practical limit.
536  if (ArgMax == -1)
537    return true;
538
539  // Conservatively account for space required by environment variables.
540  long HalfArgMax = EffectiveArgMax / 2;
541
542  size_t ArgLength = Program.size() + 1;
543  for (StringRef Arg : Args) {
544    // Ensure that we do not exceed the MAX_ARG_STRLEN constant on Linux, which
545    // does not have a constant unlike what the man pages would have you
546    // believe. Since this limit is pretty high, perform the check
547    // unconditionally rather than trying to be aggressive and limiting it to
548    // Linux only.
549    if (Arg.size() >= (32 * 4096))
550      return false;
551
552    ArgLength += Arg.size() + 1;
553    if (ArgLength > size_t(HalfArgMax)) {
554      return false;
555    }
556  }
557
558  return true;
559}
560