Home | History | Annotate | Download | only in process
      1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
      2 // Use of this source code is governed by a BSD-style license that can be
      3 // found in the LICENSE file.
      4 
      5 #include "base/process/launch.h"
      6 
      7 #include <dirent.h>
      8 #include <errno.h>
      9 #include <fcntl.h>
     10 #include <sched.h>
     11 #include <setjmp.h>
     12 #include <signal.h>
     13 #include <stddef.h>
     14 #include <stdint.h>
     15 #include <stdlib.h>
     16 #include <sys/resource.h>
     17 #include <sys/syscall.h>
     18 #include <sys/time.h>
     19 #include <sys/types.h>
     20 #include <sys/wait.h>
     21 #include <unistd.h>
     22 
     23 #include <iterator>
     24 #include <limits>
     25 #include <set>
     26 
     27 #include "base/command_line.h"
     28 #include "base/compiler_specific.h"
     29 #include "base/debug/debugger.h"
     30 #include "base/debug/stack_trace.h"
     31 #include "base/files/dir_reader_posix.h"
     32 #include "base/files/file_util.h"
     33 #include "base/files/scoped_file.h"
     34 #include "base/logging.h"
     35 #include "base/memory/scoped_ptr.h"
     36 #include "base/posix/eintr_wrapper.h"
     37 #include "base/process/process.h"
     38 #include "base/process/process_metrics.h"
     39 #include "base/strings/stringprintf.h"
     40 #include "base/synchronization/waitable_event.h"
     41 #include "base/threading/platform_thread.h"
     42 #include "base/threading/thread_restrictions.h"
     43 #include "build/build_config.h"
     44 #include "third_party/valgrind/valgrind.h"
     45 
     46 #if defined(OS_LINUX)
     47 #include <sys/prctl.h>
     48 #endif
     49 
     50 #if defined(OS_CHROMEOS)
     51 #include <sys/ioctl.h>
     52 #endif
     53 
     54 #if defined(OS_FREEBSD)
     55 #include <sys/event.h>
     56 #include <sys/ucontext.h>
     57 #endif
     58 
     59 #if defined(OS_MACOSX)
     60 #include <crt_externs.h>
     61 #include <sys/event.h>
     62 #else
     63 extern char** environ;
     64 #endif
     65 
     66 namespace base {
     67 
     68 #if !defined(OS_NACL_NONSFI)
     69 
     70 namespace {
     71 
     72 // Get the process's "environment" (i.e. the thing that setenv/getenv
     73 // work with).
     74 char** GetEnvironment() {
     75 #if defined(OS_MACOSX)
     76   return *_NSGetEnviron();
     77 #else
     78   return environ;
     79 #endif
     80 }
     81 
     82 // Set the process's "environment" (i.e. the thing that setenv/getenv
     83 // work with).
     84 void SetEnvironment(char** env) {
     85 #if defined(OS_MACOSX)
     86   *_NSGetEnviron() = env;
     87 #else
     88   environ = env;
     89 #endif
     90 }
     91 
     92 // Set the calling thread's signal mask to new_sigmask and return
     93 // the previous signal mask.
     94 sigset_t SetSignalMask(const sigset_t& new_sigmask) {
     95   sigset_t old_sigmask;
     96 #if defined(OS_ANDROID)
     97   // POSIX says pthread_sigmask() must be used in multi-threaded processes,
     98   // but Android's pthread_sigmask() was broken until 4.1:
     99   // https://code.google.com/p/android/issues/detail?id=15337
    100   // http://stackoverflow.com/questions/13777109/pthread-sigmask-on-android-not-working
    101   RAW_CHECK(sigprocmask(SIG_SETMASK, &new_sigmask, &old_sigmask) == 0);
    102 #else
    103   RAW_CHECK(pthread_sigmask(SIG_SETMASK, &new_sigmask, &old_sigmask) == 0);
    104 #endif
    105   return old_sigmask;
    106 }
    107 
    108 #if !defined(OS_LINUX) || \
    109     (!defined(__i386__) && !defined(__x86_64__) && !defined(__arm__))
    110 void ResetChildSignalHandlersToDefaults() {
    111   // The previous signal handlers are likely to be meaningless in the child's
    112   // context so we reset them to the defaults for now. http://crbug.com/44953
    113   // These signal handlers are set up at least in browser_main_posix.cc:
    114   // BrowserMainPartsPosix::PreEarlyInitialization and stack_trace_posix.cc:
    115   // EnableInProcessStackDumping.
    116   signal(SIGHUP, SIG_DFL);
    117   signal(SIGINT, SIG_DFL);
    118   signal(SIGILL, SIG_DFL);
    119   signal(SIGABRT, SIG_DFL);
    120   signal(SIGFPE, SIG_DFL);
    121   signal(SIGBUS, SIG_DFL);
    122   signal(SIGSEGV, SIG_DFL);
    123   signal(SIGSYS, SIG_DFL);
    124   signal(SIGTERM, SIG_DFL);
    125 }
    126 
    127 #else
    128 
    129 // TODO(jln): remove the Linux special case once kernels are fixed.
    130 
    131 // Internally the kernel makes sigset_t an array of long large enough to have
    132 // one bit per signal.
    133 typedef uint64_t kernel_sigset_t;
    134 
    135 // This is what struct sigaction looks like to the kernel at least on X86 and
    136 // ARM. MIPS, for instance, is very different.
    137 struct kernel_sigaction {
    138   void* k_sa_handler;  // For this usage it only needs to be a generic pointer.
    139   unsigned long k_sa_flags;
    140   void* k_sa_restorer;  // For this usage it only needs to be a generic pointer.
    141   kernel_sigset_t k_sa_mask;
    142 };
    143 
    144 // glibc's sigaction() will prevent access to sa_restorer, so we need to roll
    145 // our own.
    146 int sys_rt_sigaction(int sig, const struct kernel_sigaction* act,
    147                      struct kernel_sigaction* oact) {
    148   return syscall(SYS_rt_sigaction, sig, act, oact, sizeof(kernel_sigset_t));
    149 }
    150 
    151 // This function is intended to be used in between fork() and execve() and will
    152 // reset all signal handlers to the default.
    153 // The motivation for going through all of them is that sa_restorer can leak
    154 // from parents and help defeat ASLR on buggy kernels.  We reset it to NULL.
    155 // See crbug.com/177956.
    156 void ResetChildSignalHandlersToDefaults(void) {
    157   for (int signum = 1; ; ++signum) {
    158     struct kernel_sigaction act;
    159     memset(&act, 0, sizeof(act));
    160     int sigaction_get_ret = sys_rt_sigaction(signum, NULL, &act);
    161     if (sigaction_get_ret && errno == EINVAL) {
    162 #if !defined(NDEBUG)
    163       // Linux supports 32 real-time signals from 33 to 64.
    164       // If the number of signals in the Linux kernel changes, someone should
    165       // look at this code.
    166       const int kNumberOfSignals = 64;
    167       RAW_CHECK(signum == kNumberOfSignals + 1);
    168 #endif  // !defined(NDEBUG)
    169       break;
    170     }
    171     // All other failures are fatal.
    172     if (sigaction_get_ret) {
    173       RAW_LOG(FATAL, "sigaction (get) failed.");
    174     }
    175 
    176     // The kernel won't allow to re-set SIGKILL or SIGSTOP.
    177     if (signum != SIGSTOP && signum != SIGKILL) {
    178       act.k_sa_handler = reinterpret_cast<void*>(SIG_DFL);
    179       act.k_sa_restorer = NULL;
    180       if (sys_rt_sigaction(signum, &act, NULL)) {
    181         RAW_LOG(FATAL, "sigaction (set) failed.");
    182       }
    183     }
    184 #if !defined(NDEBUG)
    185     // Now ask the kernel again and check that no restorer will leak.
    186     if (sys_rt_sigaction(signum, NULL, &act) || act.k_sa_restorer) {
    187       RAW_LOG(FATAL, "Cound not fix sa_restorer.");
    188     }
    189 #endif  // !defined(NDEBUG)
    190   }
    191 }
    192 #endif  // !defined(OS_LINUX) ||
    193         // (!defined(__i386__) && !defined(__x86_64__) && !defined(__arm__))
    194 }  // anonymous namespace
    195 
    196 // Functor for |ScopedDIR| (below).
    197 struct ScopedDIRClose {
    198   inline void operator()(DIR* x) const {
    199     if (x)
    200       closedir(x);
    201   }
    202 };
    203 
    204 // Automatically closes |DIR*|s.
    205 typedef scoped_ptr<DIR, ScopedDIRClose> ScopedDIR;
    206 
    207 #if defined(OS_LINUX)
    208 static const char kFDDir[] = "/proc/self/fd";
    209 #elif defined(OS_MACOSX)
    210 static const char kFDDir[] = "/dev/fd";
    211 #elif defined(OS_SOLARIS)
    212 static const char kFDDir[] = "/dev/fd";
    213 #elif defined(OS_FREEBSD)
    214 static const char kFDDir[] = "/dev/fd";
    215 #elif defined(OS_OPENBSD)
    216 static const char kFDDir[] = "/dev/fd";
    217 #elif defined(OS_ANDROID)
    218 static const char kFDDir[] = "/proc/self/fd";
    219 #endif
    220 
    221 void CloseSuperfluousFds(const base::InjectiveMultimap& saved_mapping) {
    222   // DANGER: no calls to malloc or locks are allowed from now on:
    223   // http://crbug.com/36678
    224 
    225   // Get the maximum number of FDs possible.
    226   size_t max_fds = GetMaxFds();
    227 
    228   DirReaderPosix fd_dir(kFDDir);
    229   if (!fd_dir.IsValid()) {
    230     // Fallback case: Try every possible fd.
    231     for (size_t i = 0; i < max_fds; ++i) {
    232       const int fd = static_cast<int>(i);
    233       if (fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO)
    234         continue;
    235       // Cannot use STL iterators here, since debug iterators use locks.
    236       size_t j;
    237       for (j = 0; j < saved_mapping.size(); j++) {
    238         if (fd == saved_mapping[j].dest)
    239           break;
    240       }
    241       if (j < saved_mapping.size())
    242         continue;
    243 
    244       // Since we're just trying to close anything we can find,
    245       // ignore any error return values of close().
    246       close(fd);
    247     }
    248     return;
    249   }
    250 
    251   const int dir_fd = fd_dir.fd();
    252 
    253   for ( ; fd_dir.Next(); ) {
    254     // Skip . and .. entries.
    255     if (fd_dir.name()[0] == '.')
    256       continue;
    257 
    258     char *endptr;
    259     errno = 0;
    260     const long int fd = strtol(fd_dir.name(), &endptr, 10);
    261     if (fd_dir.name()[0] == 0 || *endptr || fd < 0 || errno)
    262       continue;
    263     if (fd == STDIN_FILENO || fd == STDOUT_FILENO || fd == STDERR_FILENO)
    264       continue;
    265     // Cannot use STL iterators here, since debug iterators use locks.
    266     size_t i;
    267     for (i = 0; i < saved_mapping.size(); i++) {
    268       if (fd == saved_mapping[i].dest)
    269         break;
    270     }
    271     if (i < saved_mapping.size())
    272       continue;
    273     if (fd == dir_fd)
    274       continue;
    275 
    276     // When running under Valgrind, Valgrind opens several FDs for its
    277     // own use and will complain if we try to close them.  All of
    278     // these FDs are >= |max_fds|, so we can check against that here
    279     // before closing.  See https://bugs.kde.org/show_bug.cgi?id=191758
    280     if (fd < static_cast<int>(max_fds)) {
    281       int ret = IGNORE_EINTR(close(fd));
    282       DPCHECK(ret == 0);
    283     }
    284   }
    285 }
    286 
    287 Process LaunchProcess(const CommandLine& cmdline,
    288                       const LaunchOptions& options) {
    289   return LaunchProcess(cmdline.argv(), options);
    290 }
    291 
    292 Process LaunchProcess(const std::vector<std::string>& argv,
    293                       const LaunchOptions& options) {
    294   size_t fd_shuffle_size = 0;
    295   if (options.fds_to_remap) {
    296     fd_shuffle_size = options.fds_to_remap->size();
    297   }
    298 
    299   InjectiveMultimap fd_shuffle1;
    300   InjectiveMultimap fd_shuffle2;
    301   fd_shuffle1.reserve(fd_shuffle_size);
    302   fd_shuffle2.reserve(fd_shuffle_size);
    303 
    304   scoped_ptr<char* []> argv_cstr(new char* [argv.size() + 1]);
    305   for (size_t i = 0; i < argv.size(); i++) {
    306     argv_cstr[i] = const_cast<char*>(argv[i].c_str());
    307   }
    308   argv_cstr[argv.size()] = NULL;
    309 
    310   scoped_ptr<char*[]> new_environ;
    311   char* const empty_environ = NULL;
    312   char* const* old_environ = GetEnvironment();
    313   if (options.clear_environ)
    314     old_environ = &empty_environ;
    315   if (!options.environ.empty())
    316     new_environ = AlterEnvironment(old_environ, options.environ);
    317 
    318   sigset_t full_sigset;
    319   sigfillset(&full_sigset);
    320   const sigset_t orig_sigmask = SetSignalMask(full_sigset);
    321 
    322   const char* current_directory = nullptr;
    323   if (!options.current_directory.empty()) {
    324     current_directory = options.current_directory.value().c_str();
    325   }
    326 
    327   pid_t pid;
    328 #if defined(OS_LINUX)
    329   if (options.clone_flags) {
    330     // Signal handling in this function assumes the creation of a new
    331     // process, so we check that a thread is not being created by mistake
    332     // and that signal handling follows the process-creation rules.
    333     RAW_CHECK(
    334         !(options.clone_flags & (CLONE_SIGHAND | CLONE_THREAD | CLONE_VM)));
    335 
    336     // We specify a null ptid and ctid.
    337     RAW_CHECK(
    338         !(options.clone_flags &
    339           (CLONE_CHILD_CLEARTID | CLONE_CHILD_SETTID | CLONE_PARENT_SETTID)));
    340 
    341     // Since we use waitpid, we do not support custom termination signals in the
    342     // clone flags.
    343     RAW_CHECK((options.clone_flags & 0xff) == 0);
    344 
    345     pid = ForkWithFlags(options.clone_flags | SIGCHLD, nullptr, nullptr);
    346   } else
    347 #endif
    348   {
    349     pid = fork();
    350   }
    351 
    352   // Always restore the original signal mask in the parent.
    353   if (pid != 0) {
    354     SetSignalMask(orig_sigmask);
    355   }
    356 
    357   if (pid < 0) {
    358     DPLOG(ERROR) << "fork";
    359     return Process();
    360   } else if (pid == 0) {
    361     // Child process
    362 
    363     // DANGER: no calls to malloc or locks are allowed from now on:
    364     // http://crbug.com/36678
    365 
    366     // DANGER: fork() rule: in the child, if you don't end up doing exec*(),
    367     // you call _exit() instead of exit(). This is because _exit() does not
    368     // call any previously-registered (in the parent) exit handlers, which
    369     // might do things like block waiting for threads that don't even exist
    370     // in the child.
    371 
    372     // If a child process uses the readline library, the process block forever.
    373     // In BSD like OSes including OS X it is safe to assign /dev/null as stdin.
    374     // See http://crbug.com/56596.
    375     base::ScopedFD null_fd(HANDLE_EINTR(open("/dev/null", O_RDONLY)));
    376     if (!null_fd.is_valid()) {
    377       RAW_LOG(ERROR, "Failed to open /dev/null");
    378       _exit(127);
    379     }
    380 
    381     int new_fd = HANDLE_EINTR(dup2(null_fd.get(), STDIN_FILENO));
    382     if (new_fd != STDIN_FILENO) {
    383       RAW_LOG(ERROR, "Failed to dup /dev/null for stdin");
    384       _exit(127);
    385     }
    386 
    387     if (options.new_process_group) {
    388       // Instead of inheriting the process group ID of the parent, the child
    389       // starts off a new process group with pgid equal to its process ID.
    390       if (setpgid(0, 0) < 0) {
    391         RAW_LOG(ERROR, "setpgid failed");
    392         _exit(127);
    393       }
    394     }
    395 
    396     if (options.maximize_rlimits) {
    397       // Some resource limits need to be maximal in this child.
    398       for (size_t i = 0; i < options.maximize_rlimits->size(); ++i) {
    399         const int resource = (*options.maximize_rlimits)[i];
    400         struct rlimit limit;
    401         if (getrlimit(resource, &limit) < 0) {
    402           RAW_LOG(WARNING, "getrlimit failed");
    403         } else if (limit.rlim_cur < limit.rlim_max) {
    404           limit.rlim_cur = limit.rlim_max;
    405           if (setrlimit(resource, &limit) < 0) {
    406             RAW_LOG(WARNING, "setrlimit failed");
    407           }
    408         }
    409       }
    410     }
    411 
    412 #if defined(OS_MACOSX)
    413     RestoreDefaultExceptionHandler();
    414 #endif  // defined(OS_MACOSX)
    415 
    416     ResetChildSignalHandlersToDefaults();
    417     SetSignalMask(orig_sigmask);
    418 
    419 #if 0
    420     // When debugging it can be helpful to check that we really aren't making
    421     // any hidden calls to malloc.
    422     void *malloc_thunk =
    423         reinterpret_cast<void*>(reinterpret_cast<intptr_t>(malloc) & ~4095);
    424     mprotect(malloc_thunk, 4096, PROT_READ | PROT_WRITE | PROT_EXEC);
    425     memset(reinterpret_cast<void*>(malloc), 0xff, 8);
    426 #endif  // 0
    427 
    428 #if defined(OS_CHROMEOS)
    429     if (options.ctrl_terminal_fd >= 0) {
    430       // Set process' controlling terminal.
    431       if (HANDLE_EINTR(setsid()) != -1) {
    432         if (HANDLE_EINTR(
    433                 ioctl(options.ctrl_terminal_fd, TIOCSCTTY, NULL)) == -1) {
    434           RAW_LOG(WARNING, "ioctl(TIOCSCTTY), ctrl terminal not set");
    435         }
    436       } else {
    437         RAW_LOG(WARNING, "setsid failed, ctrl terminal not set");
    438       }
    439     }
    440 #endif  // defined(OS_CHROMEOS)
    441 
    442     if (options.fds_to_remap) {
    443       // Cannot use STL iterators here, since debug iterators use locks.
    444       for (size_t i = 0; i < options.fds_to_remap->size(); ++i) {
    445         const FileHandleMappingVector::value_type& value =
    446             (*options.fds_to_remap)[i];
    447         fd_shuffle1.push_back(InjectionArc(value.first, value.second, false));
    448         fd_shuffle2.push_back(InjectionArc(value.first, value.second, false));
    449       }
    450     }
    451 
    452     if (!options.environ.empty() || options.clear_environ)
    453       SetEnvironment(new_environ.get());
    454 
    455     // fd_shuffle1 is mutated by this call because it cannot malloc.
    456     if (!ShuffleFileDescriptors(&fd_shuffle1))
    457       _exit(127);
    458 
    459     CloseSuperfluousFds(fd_shuffle2);
    460 
    461     // Set NO_NEW_PRIVS by default. Since NO_NEW_PRIVS only exists in kernel
    462     // 3.5+, do not check the return value of prctl here.
    463 #if defined(OS_LINUX)
    464 #ifndef PR_SET_NO_NEW_PRIVS
    465 #define PR_SET_NO_NEW_PRIVS 38
    466 #endif
    467     if (!options.allow_new_privs) {
    468       if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) && errno != EINVAL) {
    469         // Only log if the error is not EINVAL (i.e. not supported).
    470         RAW_LOG(FATAL, "prctl(PR_SET_NO_NEW_PRIVS) failed");
    471       }
    472     }
    473 
    474     if (options.kill_on_parent_death) {
    475       if (prctl(PR_SET_PDEATHSIG, SIGKILL) != 0) {
    476         RAW_LOG(ERROR, "prctl(PR_SET_PDEATHSIG) failed");
    477         _exit(127);
    478       }
    479     }
    480 #endif
    481 
    482     if (current_directory != nullptr) {
    483       RAW_CHECK(chdir(current_directory) == 0);
    484     }
    485 
    486     if (options.pre_exec_delegate != nullptr) {
    487       options.pre_exec_delegate->RunAsyncSafe();
    488     }
    489 
    490     execvp(argv_cstr[0], argv_cstr.get());
    491 
    492     RAW_LOG(ERROR, "LaunchProcess: failed to execvp:");
    493     RAW_LOG(ERROR, argv_cstr[0]);
    494     _exit(127);
    495   } else {
    496     // Parent process
    497     if (options.wait) {
    498       // While this isn't strictly disk IO, waiting for another process to
    499       // finish is the sort of thing ThreadRestrictions is trying to prevent.
    500       base::ThreadRestrictions::AssertIOAllowed();
    501       pid_t ret = HANDLE_EINTR(waitpid(pid, 0, 0));
    502       DPCHECK(ret > 0);
    503     }
    504   }
    505 
    506   return Process(pid);
    507 }
    508 
    509 void RaiseProcessToHighPriority() {
    510   // On POSIX, we don't actually do anything here.  We could try to nice() or
    511   // setpriority() or sched_getscheduler, but these all require extra rights.
    512 }
    513 
    514 // Return value used by GetAppOutputInternal to encapsulate the various exit
    515 // scenarios from the function.
    516 enum GetAppOutputInternalResult {
    517   EXECUTE_FAILURE,
    518   EXECUTE_SUCCESS,
    519   GOT_MAX_OUTPUT,
    520 };
    521 
    522 // Executes the application specified by |argv| and wait for it to exit. Stores
    523 // the output (stdout) in |output|. If |do_search_path| is set, it searches the
    524 // path for the application; in that case, |envp| must be null, and it will use
    525 // the current environment. If |do_search_path| is false, |argv[0]| should fully
    526 // specify the path of the application, and |envp| will be used as the
    527 // environment. If |include_stderr| is true, includes stderr otherwise redirects
    528 // it to /dev/null.
    529 // If we successfully start the application and get all requested output, we
    530 // return GOT_MAX_OUTPUT, or if there is a problem starting or exiting
    531 // the application we return RUN_FAILURE. Otherwise we return EXECUTE_SUCCESS.
    532 // The GOT_MAX_OUTPUT return value exists so a caller that asks for limited
    533 // output can treat this as a success, despite having an exit code of SIG_PIPE
    534 // due to us closing the output pipe.
    535 // In the case of EXECUTE_SUCCESS, the application exit code will be returned
    536 // in |*exit_code|, which should be checked to determine if the application
    537 // ran successfully.
    538 static GetAppOutputInternalResult GetAppOutputInternal(
    539     const std::vector<std::string>& argv,
    540     char* const envp[],
    541     bool include_stderr,
    542     std::string* output,
    543     size_t max_output,
    544     bool do_search_path,
    545     int* exit_code) {
    546   // Doing a blocking wait for another command to finish counts as IO.
    547   base::ThreadRestrictions::AssertIOAllowed();
    548   // exit_code must be supplied so calling function can determine success.
    549   DCHECK(exit_code);
    550   *exit_code = EXIT_FAILURE;
    551 
    552   int pipe_fd[2];
    553   pid_t pid;
    554   InjectiveMultimap fd_shuffle1, fd_shuffle2;
    555   scoped_ptr<char*[]> argv_cstr(new char*[argv.size() + 1]);
    556 
    557   fd_shuffle1.reserve(3);
    558   fd_shuffle2.reserve(3);
    559 
    560   // Either |do_search_path| should be false or |envp| should be null, but not
    561   // both.
    562   DCHECK(!do_search_path ^ !envp);
    563 
    564   if (pipe(pipe_fd) < 0)
    565     return EXECUTE_FAILURE;
    566 
    567   switch (pid = fork()) {
    568     case -1:  // error
    569       close(pipe_fd[0]);
    570       close(pipe_fd[1]);
    571       return EXECUTE_FAILURE;
    572     case 0:  // child
    573       {
    574         // DANGER: no calls to malloc or locks are allowed from now on:
    575         // http://crbug.com/36678
    576 
    577 #if defined(OS_MACOSX)
    578         RestoreDefaultExceptionHandler();
    579 #endif
    580 
    581         // Obscure fork() rule: in the child, if you don't end up doing exec*(),
    582         // you call _exit() instead of exit(). This is because _exit() does not
    583         // call any previously-registered (in the parent) exit handlers, which
    584         // might do things like block waiting for threads that don't even exist
    585         // in the child.
    586         int dev_null = open("/dev/null", O_WRONLY);
    587         if (dev_null < 0)
    588           _exit(127);
    589 
    590         fd_shuffle1.push_back(InjectionArc(pipe_fd[1], STDOUT_FILENO, true));
    591         fd_shuffle1.push_back(InjectionArc(
    592             include_stderr ? pipe_fd[1] : dev_null,
    593             STDERR_FILENO, true));
    594         fd_shuffle1.push_back(InjectionArc(dev_null, STDIN_FILENO, true));
    595         // Adding another element here? Remeber to increase the argument to
    596         // reserve(), above.
    597 
    598         for (size_t i = 0; i < fd_shuffle1.size(); ++i)
    599           fd_shuffle2.push_back(fd_shuffle1[i]);
    600 
    601         if (!ShuffleFileDescriptors(&fd_shuffle1))
    602           _exit(127);
    603 
    604         CloseSuperfluousFds(fd_shuffle2);
    605 
    606         for (size_t i = 0; i < argv.size(); i++)
    607           argv_cstr[i] = const_cast<char*>(argv[i].c_str());
    608         argv_cstr[argv.size()] = NULL;
    609         if (do_search_path)
    610           execvp(argv_cstr[0], argv_cstr.get());
    611         else
    612           execve(argv_cstr[0], argv_cstr.get(), envp);
    613         _exit(127);
    614       }
    615     default:  // parent
    616       {
    617         // Close our writing end of pipe now. Otherwise later read would not
    618         // be able to detect end of child's output (in theory we could still
    619         // write to the pipe).
    620         close(pipe_fd[1]);
    621 
    622         output->clear();
    623         char buffer[256];
    624         size_t output_buf_left = max_output;
    625         ssize_t bytes_read = 1;  // A lie to properly handle |max_output == 0|
    626                                  // case in the logic below.
    627 
    628         while (output_buf_left > 0) {
    629           bytes_read = HANDLE_EINTR(read(pipe_fd[0], buffer,
    630                                     std::min(output_buf_left, sizeof(buffer))));
    631           if (bytes_read <= 0)
    632             break;
    633           output->append(buffer, bytes_read);
    634           output_buf_left -= static_cast<size_t>(bytes_read);
    635         }
    636         close(pipe_fd[0]);
    637 
    638         // Always wait for exit code (even if we know we'll declare
    639         // GOT_MAX_OUTPUT).
    640         Process process(pid);
    641         bool success = process.WaitForExit(exit_code);
    642 
    643         // If we stopped because we read as much as we wanted, we return
    644         // GOT_MAX_OUTPUT (because the child may exit due to |SIGPIPE|).
    645         if (!output_buf_left && bytes_read > 0)
    646           return GOT_MAX_OUTPUT;
    647         else if (success)
    648           return EXECUTE_SUCCESS;
    649         return EXECUTE_FAILURE;
    650       }
    651   }
    652 }
    653 
    654 bool GetAppOutput(const CommandLine& cl, std::string* output) {
    655   return GetAppOutput(cl.argv(), output);
    656 }
    657 
    658 bool GetAppOutput(const std::vector<std::string>& argv, std::string* output) {
    659   // Run |execve()| with the current environment and store "unlimited" data.
    660   int exit_code;
    661   GetAppOutputInternalResult result = GetAppOutputInternal(
    662       argv, NULL, false, output, std::numeric_limits<std::size_t>::max(), true,
    663       &exit_code);
    664   return result == EXECUTE_SUCCESS && exit_code == EXIT_SUCCESS;
    665 }
    666 
    667 bool GetAppOutputAndError(const CommandLine& cl, std::string* output) {
    668   // Run |execve()| with the current environment and store "unlimited" data.
    669   int exit_code;
    670   GetAppOutputInternalResult result = GetAppOutputInternal(
    671       cl.argv(), NULL, true, output, std::numeric_limits<std::size_t>::max(),
    672       true, &exit_code);
    673   return result == EXECUTE_SUCCESS && exit_code == EXIT_SUCCESS;
    674 }
    675 
    676 // TODO(viettrungluu): Conceivably, we should have a timeout as well, so we
    677 // don't hang if what we're calling hangs.
    678 bool GetAppOutputRestricted(const CommandLine& cl,
    679                             std::string* output, size_t max_output) {
    680   // Run |execve()| with the empty environment.
    681   char* const empty_environ = NULL;
    682   int exit_code;
    683   GetAppOutputInternalResult result = GetAppOutputInternal(
    684       cl.argv(), &empty_environ, false, output, max_output, false, &exit_code);
    685   return result == GOT_MAX_OUTPUT || (result == EXECUTE_SUCCESS &&
    686                                       exit_code == EXIT_SUCCESS);
    687 }
    688 
    689 bool GetAppOutputWithExitCode(const CommandLine& cl,
    690                               std::string* output,
    691                               int* exit_code) {
    692   // Run |execve()| with the current environment and store "unlimited" data.
    693   GetAppOutputInternalResult result = GetAppOutputInternal(
    694       cl.argv(), NULL, false, output, std::numeric_limits<std::size_t>::max(),
    695       true, exit_code);
    696   return result == EXECUTE_SUCCESS;
    697 }
    698 
    699 #endif  // !defined(OS_NACL_NONSFI)
    700 
    701 #if defined(OS_LINUX) || defined(OS_NACL_NONSFI)
    702 namespace {
    703 
    704 bool IsRunningOnValgrind() {
    705   return RUNNING_ON_VALGRIND;
    706 }
    707 
    708 // This function runs on the stack specified on the clone call. It uses longjmp
    709 // to switch back to the original stack so the child can return from sys_clone.
    710 int CloneHelper(void* arg) {
    711   jmp_buf* env_ptr = reinterpret_cast<jmp_buf*>(arg);
    712   longjmp(*env_ptr, 1);
    713 
    714   // Should not be reached.
    715   RAW_CHECK(false);
    716   return 1;
    717 }
    718 
    719 // This function is noinline to ensure that stack_buf is below the stack pointer
    720 // that is saved when setjmp is called below. This is needed because when
    721 // compiled with FORTIFY_SOURCE, glibc's longjmp checks that the stack is moved
    722 // upwards. See crbug.com/442912 for more details.
    723 #if defined(ADDRESS_SANITIZER)
    724 // Disable AddressSanitizer instrumentation for this function to make sure
    725 // |stack_buf| is allocated on thread stack instead of ASan's fake stack.
    726 // Under ASan longjmp() will attempt to clean up the area between the old and
    727 // new stack pointers and print a warning that may confuse the user.
    728 __attribute__((no_sanitize_address))
    729 #endif
    730 NOINLINE pid_t CloneAndLongjmpInChild(unsigned long flags,
    731                                       pid_t* ptid,
    732                                       pid_t* ctid,
    733                                       jmp_buf* env) {
    734   // We use the libc clone wrapper instead of making the syscall
    735   // directly because making the syscall may fail to update the libc's
    736   // internal pid cache. The libc interface unfortunately requires
    737   // specifying a new stack, so we use setjmp/longjmp to emulate
    738   // fork-like behavior.
    739   char stack_buf[PTHREAD_STACK_MIN];
    740 #if defined(ARCH_CPU_X86_FAMILY) || defined(ARCH_CPU_ARM_FAMILY) || \
    741     defined(ARCH_CPU_MIPS64_FAMILY) || defined(ARCH_CPU_MIPS_FAMILY)
    742   // The stack grows downward.
    743   void* stack = stack_buf + sizeof(stack_buf);
    744 #else
    745 #error "Unsupported architecture"
    746 #endif
    747   return clone(&CloneHelper, stack, flags, env, ptid, nullptr, ctid);
    748 }
    749 
    750 }  // anonymous namespace
    751 
    752 pid_t ForkWithFlags(unsigned long flags, pid_t* ptid, pid_t* ctid) {
    753   const bool clone_tls_used = flags & CLONE_SETTLS;
    754   const bool invalid_ctid =
    755       (flags & (CLONE_CHILD_SETTID | CLONE_CHILD_CLEARTID)) && !ctid;
    756   const bool invalid_ptid = (flags & CLONE_PARENT_SETTID) && !ptid;
    757 
    758   // We do not support CLONE_VM.
    759   const bool clone_vm_used = flags & CLONE_VM;
    760 
    761   if (clone_tls_used || invalid_ctid || invalid_ptid || clone_vm_used) {
    762     RAW_LOG(FATAL, "Invalid usage of ForkWithFlags");
    763   }
    764 
    765   // Valgrind's clone implementation does not support specifiying a child_stack
    766   // without CLONE_VM, so we cannot use libc's clone wrapper when running under
    767   // Valgrind. As a result, the libc pid cache may be incorrect under Valgrind.
    768   // See crbug.com/442817 for more details.
    769   if (IsRunningOnValgrind()) {
    770     // See kernel/fork.c in Linux. There is different ordering of sys_clone
    771     // parameters depending on CONFIG_CLONE_BACKWARDS* configuration options.
    772 #if defined(ARCH_CPU_X86_64)
    773     return syscall(__NR_clone, flags, nullptr, ptid, ctid, nullptr);
    774 #elif defined(ARCH_CPU_X86) || defined(ARCH_CPU_ARM_FAMILY) || \
    775     defined(ARCH_CPU_MIPS_FAMILY) || defined(ARCH_CPU_MIPS64_FAMILY)
    776     // CONFIG_CLONE_BACKWARDS defined.
    777     return syscall(__NR_clone, flags, nullptr, ptid, nullptr, ctid);
    778 #else
    779 #error "Unsupported architecture"
    780 #endif
    781   }
    782 
    783   jmp_buf env;
    784   if (setjmp(env) == 0) {
    785     return CloneAndLongjmpInChild(flags, ptid, ctid, &env);
    786   }
    787 
    788   return 0;
    789 }
    790 #endif  // defined(OS_LINUX) || defined(OS_NACL_NONSFI)
    791 
    792 }  // namespace base
    793