Simplify process exit-code handling on Posix, take 2.

runtime/bin/process_android.cc

Index: runtime/bin/process_android.cc
diff --git a/runtime/bin/process_android.cc b/runtime/bin/process_android.cc
index fd97e7ae66705e5d57e732a736ec2af3ea5e76fb..fbb7649f38a66fd708e0b179f0510f65e11b5bc4 100644
--- a/runtime/bin/process_android.cc
+++ b/runtime/bin/process_android.cc
@@ -10,7 +10,6 @@
 #include <errno.h>  // NOLINT
 #include <fcntl.h>  // NOLINT
 #include <poll.h>  // NOLINT
-#include <signal.h>  // NOLINT
 #include <stdio.h>  // NOLINT
 #include <stdlib.h>  // NOLINT
 #include <string.h>  // NOLINT
@@ -111,8 +110,8 @@ ProcessInfo* ProcessInfoList::active_processes_ = NULL;
 dart::Mutex* ProcessInfoList::mutex_ = new dart::Mutex();
 
 
-// The exit code handler sets up a separate thread which is signalled
-// on SIGCHLD. That separate thread can then get the exit code from
+// The exit code handler sets up a separate thread which waits for child
+// processes to die. That separate thread can then get the exit code from
 // processes that have exited and communicate it to Dart through the
 // event loop.
 class ExitCodeHandler {
@@ -127,144 +126,97 @@ class ExitCodeHandler {
       return true;
     }
 
-    // Allocate a pipe that the signal handler can write a byte to and
-    // that the exit handler thread can poll.
-    int result = TEMP_FAILURE_RETRY(pipe(sig_chld_fds_));
-    if (result < 0) {
-      return false;
-    }
-    FDUtils::SetCloseOnExec(sig_chld_fds_[0]);
-    FDUtils::SetCloseOnExec(sig_chld_fds_[1]);
+    terminate_ = false;
 
-    // Start thread that polls the pipe and handles process exits when
-    // data is received on the pipe.
-    result = dart::Thread::Start(ExitCodeHandlerEntry, sig_chld_fds_[0]);
+    // Start thread that handles process exits when waitpid returns.
+    int result = dart::Thread::Start(ExitCodeHandlerEntry, 0);
     if (result != 0) {
       FATAL1("Failed to start exit code handler worker thread %d", result);
     }
 
-    // Mark write end non-blocking.
-    FDUtils::SetNonBlocking(sig_chld_fds_[1]);
-
     // Thread started and the ExitCodeHandler is initialized.
     initialized_ = true;
     return true;
   }
 
-  // Get the write end of the pipe.
-  static int WakeUpFd() {
-    return sig_chld_fds_[1];
-  }
-
   static void TerminateExitCodeThread() {
     MutexLocker locker(mutex_);
     if (!initialized_) {
       return;
     }
 
-    uint8_t data = kThreadTerminateByte;
-    ssize_t result =
-        TEMP_FAILURE_RETRY(write(ExitCodeHandler::WakeUpFd(), &data, 1));
-    if (result < 1) {
-      perror("Failed to write to wake-up fd to terminate exit code thread");
-    }
+    thread_terminate_monitor_->Enter();
 
-    {
-      MonitorLocker terminate_locker(thread_terminate_monitor_);
-      while (!thread_terminated_) {
-        terminate_locker.Wait();
-      }
+    terminate_ = true;
+    // Fork to wake up waitpid.
+    if (TEMP_FAILURE_RETRY(fork()) == 0) {
+      exit(0);
     }
-  }
-
-  static void ExitCodeThreadTerminated() {
-    MonitorLocker locker(thread_terminate_monitor_);
-    thread_terminated_ = true;
-    locker.Notify();
-  }
-
- private:
-  static const uint8_t kThreadTerminateByte = 1;
 
-  // GetProcessExitCodes is called on a separate thread when a SIGCHLD
-  // signal is received to retrieve the exit codes and post them to
-  // dart.
-  static void GetProcessExitCodes() {
-    pid_t pid = 0;
-    int status = 0;
-    while ((pid = TEMP_FAILURE_RETRY(waitpid(-1, &status, WNOHANG))) > 0) {
-      int exit_code = 0;
-      int negative = 0;
-      if (WIFEXITED(status)) {
-        exit_code = WEXITSTATUS(status);
-      }
-      if (WIFSIGNALED(status)) {
-        exit_code = WTERMSIG(status);
-        negative = 1;
-      }
-      intptr_t exit_code_fd = ProcessInfoList::LookupProcessExitFd(pid);
-      if (exit_code_fd != 0) {
-        int message[2] = { exit_code, negative };
-        ssize_t result =
-            FDUtils::WriteToBlocking(exit_code_fd, &message, sizeof(message));
-        // If the process has been closed, the read end of the exit
-        // pipe has been closed. It is therefore not a problem that
-        // write fails with a broken pipe error. Other errors should
-        // not happen.
-        if (result != -1 && result != sizeof(message)) {
-          FATAL("Failed to write entire process exit message");
-        } else if (result == -1 && errno != EPIPE) {
-          FATAL1("Failed to write exit code: %d", errno);
-        }
-        ProcessInfoList::RemoveProcess(pid);
-      }
+    while (!initialized_) {

[Mads Ager (google), 2013/09/09 16:35:11]

Shouldn't you be looping while initialized_ is true?

You are waiting for the thread to terminate. When the thread terminates it sets
initialized_ to false. So you have to loop while initialized_ is true. When it
becomes false there is no thread running and you have successfully waited for
termination?

[Anders Johnsen, 2013/09/09 17:55:36]

You are absolutely right, I'll rename to better explain the usage.

+      thread_terminate_monitor_->Wait(dart::Monitor::kNoTimeout);
     }
+    thread_terminate_monitor_->Exit();
   }
 
-
+ private:
   // Entry point for the separate exit code handler thread started by
   // the ExitCodeHandler.
   static void ExitCodeHandlerEntry(uword param) {
-    struct pollfd pollfds;
-    pollfds.fd = param;
-    pollfds.events = POLLIN;
-    while (true) {
-      int result = TEMP_FAILURE_RETRY(poll(&pollfds, 1, -1));
-      if (result == -1) {
-        ASSERT(EAGAIN == EWOULDBLOCK);
-        if (errno != EWOULDBLOCK) {
-          perror("ExitCodeHandler poll failed");
+    pid_t pid = 0;
+    int status = 0;
+    while (!terminate_) {
+      if ((pid = TEMP_FAILURE_RETRY(waitpid(-1, &status, 0))) > 0) {
+        int exit_code = 0;
+        int negative = 0;
+        if (WIFEXITED(status)) {
+          exit_code = WEXITSTATUS(status);
+        }
+        if (WIFSIGNALED(status)) {
+          exit_code = WTERMSIG(status);
+          negative = 1;
         }
-      } else {
-        // Read the byte from the wake-up fd.
-        ASSERT(result = 1);
-        intptr_t data = 0;
-        ssize_t read_bytes = FDUtils::ReadFromBlocking(pollfds.fd, &data, 1);
-        if (read_bytes < 1) {
-          perror("Failed to read from wake-up fd in exit-code handler");
+        intptr_t exit_code_fd = ProcessInfoList::LookupProcessExitFd(pid);
+        if (exit_code_fd != 0) {
+          int message[2] = { exit_code, negative };
+          ssize_t result =
+              FDUtils::WriteToBlocking(exit_code_fd, &message, sizeof(message));
+          // If the process has been closed, the read end of the exit
+          // pipe has been closed. It is therefore not a problem that
+          // write fails with a broken pipe error. Other errors should
+          // not happen.
+          if (result != -1 && result != sizeof(message)) {
+            FATAL("Failed to write entire process exit message");
+          } else if (result == -1 && errno != EPIPE) {
+            FATAL1("Failed to write exit code: %d", errno);
+          }
+          ProcessInfoList::RemoveProcess(pid);
         }
-        if (data == ExitCodeHandler::kThreadTerminateByte) {
-          ExitCodeThreadTerminated();
+      } else if (errno == ECHILD) {
+        if (mutex_->TryLock()) {
+          initialized_ = false;
+          terminate_ = true;

[Mads Ager (google), 2013/09/09 16:35:11]

I guess setting terminate_ here doesn't really do anything because of the return
below. You could remove the return or remove the assignment to terminate_?

Actually, I think you have to remove the return here. You are terminating the
thread and therefore you should be signalling the thread_terminate_monitor_,
right?

[Anders Johnsen, 2013/09/09 17:55:36]

As it's able to take the lock, it know that it's _not_ in the process of
terminating, but that we simply don't have any child processes. Now that it have
the lock, it can do a quick shutdown, by resetting the settings and returning
(stopping the thread).

If it doesn't get the lock, we are terminating. Simply continue, as the
terminating-setting will make sure we get down to the monitor notification.

That said, you are 100% right about setting terminate. Removed and overall
renamed the booleans to reflect this.

+          mutex_->Unlock();
           return;
         }
-        // Get the exit code from all processes that have died.
-        GetProcessExitCodes();
       }
     }
+    thread_terminate_monitor_->Enter();
+    initialized_ = false;

[Mads Ager (google), 2013/09/09 16:35:11]

I don't like that we have collapsed thread_terminated_ and initialized_ into one
thing. The problem is that  initialized is now guarded by two different things:
the mutex and the thread terminate monitor. I can't convince myself that this
will always work correctly. I particular, here you are updating initialized and
you are guarding it by holding the thread_terminate_monitor. Some other thread
could be inspecting the value in EnsureInitialized which is guarded by the
mutex.

[Anders Johnsen, 2013/09/09 17:55:36]

Yes, the rename should have made this much clearer.

+    thread_terminate_monitor_->Notify();
+    thread_terminate_monitor_->Exit();
   }
 
   static dart::Mutex* mutex_;
   static bool initialized_;
-  static int sig_chld_fds_[2];
-  static bool thread_terminated_;
+  static bool terminate_;
   static dart::Monitor* thread_terminate_monitor_;
 };
 
 
 dart::Mutex* ExitCodeHandler::mutex_ = new dart::Mutex();
 bool ExitCodeHandler::initialized_ = false;
-int ExitCodeHandler::sig_chld_fds_[2] = { 0, 0 };
-bool ExitCodeHandler::thread_terminated_ = false;
+bool ExitCodeHandler::terminate_ = false;
 dart::Monitor* ExitCodeHandler::thread_terminate_monitor_ = new dart::Monitor();
 
 
@@ -276,21 +228,6 @@ static void SetChildOsErrorMessage(char** os_error_message) {
 }
 
 
-static void SigChldHandler(int process_signal, siginfo_t* siginfo, void* tmp) {
-  // Save errno so it can be restored at the end.
-  int entry_errno = errno;
-  // Signal the exit code handler where the actual processing takes
-  // place.
-  ssize_t result =
-      TEMP_FAILURE_RETRY(write(ExitCodeHandler::WakeUpFd(), "", 1));
-  if (result < 1) {
-    perror("Failed to write to wake-up fd in SIGCHLD handler");
-  }
-  // Restore errno.
-  errno = entry_errno;
-}
-
-
 static void ReportChildError(int exec_control_fd) {
   // In the case of failure in the child process write the errno and
   // the OS error message to the exec control pipe and exit.
@@ -330,14 +267,6 @@ int Process::Start(const char* path,
   int exec_control[2];  // Pipe to get the result from exec.
   int result;
 
-  bool initialized = ExitCodeHandler::EnsureInitialized();
-  if (!initialized) {
-    SetChildOsErrorMessage(os_error_message);
-    Log::PrintErr("Error initializing exit code handler: %s\n",
-                  *os_error_message);
-    return errno;
-  }
-
   result = TEMP_FAILURE_RETRY(pipe(read_in));
   if (result < 0) {
     SetChildOsErrorMessage(os_error_message);
@@ -413,13 +342,6 @@ int Process::Start(const char* path,
     program_environment[environment_length] = NULL;
   }
 
-  struct sigaction act;
-  bzero(&act, sizeof(act));
-  act.sa_sigaction = SigChldHandler;
-  act.sa_flags = SA_NOCLDSTOP | SA_SIGINFO;
-  if (sigaction(SIGCHLD, &act, 0) != 0) {
-    perror("Process start: setting signal handler failed");
-  }
   pid = TEMP_FAILURE_RETRY(fork());
   if (pid < 0) {
     SetChildOsErrorMessage(os_error_message);
@@ -563,6 +485,15 @@ int Process::Start(const char* path,
   *err = read_err[0];
   TEMP_FAILURE_RETRY(close(read_err[1]));
 
+  // Be sure to listen for exit-codes, now we have a child-process.
+  bool initialized = ExitCodeHandler::EnsureInitialized();
+  if (!initialized) {
+    SetChildOsErrorMessage(os_error_message);
+    Log::PrintErr("Error initializing exit code handler: %s\n",
+                  *os_error_message);
+    return errno;
+  }
+
   *id = pid;
   return 0;
 }