bool
NamedPipeWriter::write_data(void* buffer, int len)
{
assert(m_initialized);
// if we're writing to a pipe that has multiple writers,
// we need to make sure our messages are no larger than
// PIPE_BUF to guarantee atomic writes
//
assert(len <= PIPE_BUF);
// if we have a watchdog, we don't go right into a blocking
// write. instead, we select with both the real pipe and the
// watchdog pipe, which will close if our peer shuts down or
// crashes
//
if (m_watchdog != NULL) {
int watchdog_pipe = m_watchdog->get_file_descriptor();
Selector selector;
selector.add_fd( m_pipe, Selector::IO_WRITE );
selector.add_fd( watchdog_pipe, Selector::IO_READ );
selector.execute();
if ( selector.failed() || selector.signalled() ) {
dprintf(D_ALWAYS,
"select error: %s (%d)\n",
strerror(selector.select_errno()),
selector.select_errno());
return false;
}
if ( selector.fd_ready( watchdog_pipe, Selector::IO_READ ) ) {
dprintf(D_ALWAYS,
"error writing to named pipe: "
"watchdog pipe has closed\n");
return false;
}
}
// do the write
//
int bytes = write(m_pipe, buffer, len);
if (bytes != len) {
if (bytes == -1) {
dprintf(D_ALWAYS,
"write error: %s (%d)\n",
strerror(errno),
errno);
}
else {
dprintf(D_ALWAYS,
"error: wrote %d of %d bytes\n",
bytes,
len);
}
return false;
}
return true;
}
bool
VanillaProc::JobReaper(int pid, int status)
{
dprintf(D_FULLDEBUG,"Inside VanillaProc::JobReaper()\n");
//
// Run all the reapers first, since some of them change the exit status.
//
if( m_pid_ns_status_filename.length() > 0 ) {
status = pidNameSpaceReaper( status );
}
bool jobExited = OsProc::JobReaper( pid, status );
if( pid != JobPid ) { return jobExited; }
#if defined(LINUX)
// On newer kernels if memory.use_hierarchy==1, then we cannot disable
// the OOM killer. Hence, we have to be ready for a SIGKILL to be delivered
// by the kernel at the same time we get the notification. Hence, if we
// see an exit signal, we must also check the event file descriptor.
//
// outOfMemoryEvent() is aware of checkpointing and will mention that
// the OOM event happened during a checkpoint.
int efd = -1;
if( (m_oom_efd >= 0) && daemonCore->Get_Pipe_FD(m_oom_efd, &efd) && (efd != -1) ) {
Selector selector;
selector.add_fd(efd, Selector::IO_READ);
selector.set_timeout(0);
selector.execute();
if( !selector.failed() && !selector.timed_out() && selector.has_ready() && selector.fd_ready(efd, Selector::IO_READ) ) {
outOfMemoryEvent( m_oom_efd );
}
}
#endif
//
// We have three cases to consider:
// * if we're checkpointing; or
// * if we see a special checkpoint exit code; or
// * there's no special case to consider.
//
bool wantsFileTransferOnCheckpointExit = false;
JobAd->LookupBool( ATTR_WANT_FT_ON_CHECKPOINT, wantsFileTransferOnCheckpointExit );
int checkpointExitCode = 0;
JobAd->LookupInteger( ATTR_CHECKPOINT_EXIT_CODE, checkpointExitCode );
int checkpointExitSignal = 0;
JobAd->LookupInteger( ATTR_CHECKPOINT_EXIT_SIGNAL, checkpointExitSignal );
bool checkpointExitBySignal = 0;
JobAd->LookupBool( ATTR_CHECKPOINT_EXIT_BY_SIGNAL, checkpointExitBySignal );
int successfulCheckpointStatus = 0;
if( checkpointExitBySignal ) {
successfulCheckpointStatus = checkpointExitSignal;
} else if( checkpointExitCode != 0 ) {
successfulCheckpointStatus = checkpointExitCode << 8;
#if defined( WINDOWS )
successfulCheckpointStatus = checkpointExitCode;
#endif
}
if( isCheckpointing ) {
dprintf( D_FULLDEBUG, "Inside VanillaProc::JobReaper() during a checkpoint\n" );
if( exit_status == successfulCheckpointStatus ) {
if( isSoftKilling ) {
notifySuccessfulEvictionCheckpoint();
return true;
}
restartCheckpointedJob();
isCheckpointing = false;
return false;
} else {
// The job exited without taking a checkpoint. If we don't do
// anything, it will be reported as if the error code or signal
// had happened naturally (and the job will usually exit the
// queue). This could confuse the users.
//
// Instead, we'll put the job on hold, figuring that if the job
// requested that we (periodically) send it a signal, and we
// did, that it's not our fault that the job failed. This has
// the convenient side-effect of not overwriting the job's
// previous checkpoint(s), if any (since file transfer doesn't
// occur when the job goes on hold).
killFamilyIfWarranted();
recordFinalUsage();
std::string holdMessage;
formatstr( holdMessage, "Job did not exit as promised when sent its checkpoint signal. "
"Promised exit was %s %u, actual exit status was %s %u.",
checkpointExitBySignal ? "on signal" : "with exit code",
checkpointExitBySignal ? checkpointExitSignal : checkpointExitCode,
WIFSIGNALED( exit_status ) ? "on signal" : "with exit code",
WIFSIGNALED( exit_status ) ? WTERMSIG( exit_status ) : WEXITSTATUS( exit_status ) );
Starter->jic->holdJob( holdMessage.c_str(), CONDOR_HOLD_CODE_FailedToCheckpoint, exit_status );
Starter->Hold();
return true;
}
} else if( wantsFileTransferOnCheckpointExit && exit_status == successfulCheckpointStatus ) {
dprintf( D_FULLDEBUG, "Inside VanillaProc::JobReaper() and the job self-checkpointed.\n" );
if( isSoftKilling ) {
notifySuccessfulEvictionCheckpoint();
return true;
} else {
restartCheckpointedJob();
return false;
}
} else {
// If the parent job process died, clean up all of the job's processes.
killFamilyIfWarranted();
// Record final usage stats for this process family, since
// once the reaper returns, the family is no longer
// registered with DaemonCore and we'll never be able to
// get this information again.
recordFinalUsage();
return jobExited;
}
}
