void
createNewConnToCoord(CoordinatorMode mode)
{
int sockfd = -1;
if (mode & COORD_JOIN) {
sockfd = createNewSocketToCoordinator(mode);
JASSERT(sockfd != -1) (JASSERT_ERRNO)
.Text("Coordinator not found, but --join was specified. Exiting.");
} else if (mode & COORD_NEW) {
startNewCoordinator(mode);
sockfd = createNewSocketToCoordinator(mode);
JASSERT(sockfd != -1) (JASSERT_ERRNO)
.Text("Error connecting to newly started coordinator.");
} else if (mode & COORD_ANY) {
sockfd = createNewSocketToCoordinator(mode);
if (sockfd == -1) {
JTRACE("Coordinator not found, trying to start a new one.");
startNewCoordinator(mode);
sockfd = createNewSocketToCoordinator(mode);
JASSERT(sockfd != -1) (JASSERT_ERRNO)
.Text("Error connecting to newly started coordinator.");
}
} else {
JASSERT(false).Text("Not Reached");
}
Util::changeFd(sockfd, PROTECTED_COORD_FD);
JASSERT(Util::isValidFd(coordinatorSocket));
}
void dmtcp::DmtcpCoordinatorAPI::startCoordinatorIfNeeded(int modes,
int isRestart)
{
const static int CS_OK = DMTCP_FAIL_RC+1;
const static int CS_NO = DMTCP_FAIL_RC+2;
int coordinatorStatus = -1;
if (modes & COORD_BATCH) {
startNewCoordinator ( modes, isRestart );
return;
}
//fork a child process to probe the coordinator
if (fork()==0) {
//fork so if we hit an error parent won't die
dup2(2,1); //copy stderr to stdout
dup2(open("/dev/null",O_RDWR), 2); //close stderr
int result[DMTCPMESSAGE_NUM_PARAMS];
DmtcpCoordinatorAPI coordinatorAPI;
{
if ( coordinatorAPI.tryConnectToCoordinator() == false ) {
JTRACE("Coordinator not found. Will try to start a new one.");
_real_exit(DMTCP_FAIL_RC);
}
}
coordinatorAPI.sendUserCommand('s',result);
coordinatorAPI._coordinatorSocket.close();
// result[0] == numPeers of coord; bool result[1] == computation is running
if(result[0]==0 || result[1] ^ isRestart){
if(result[0] != 0) {
int num_processes = result[0];
JTRACE("Joining existing computation.") (num_processes);
}
_real_exit(CS_OK);
}else{
JTRACE("Existing computation not in a running state," \
" perhaps checkpoint in progress?");
_real_exit(CS_NO);
}
}
errno = 0;
// FIXME: wait() could return -1 if a signal happened before child exits
JASSERT(::wait(&coordinatorStatus)>0)(JASSERT_ERRNO);
JASSERT(WIFEXITED(coordinatorStatus));
//is coordinator running?
if (WEXITSTATUS(coordinatorStatus) != CS_OK) {
//is coordinator in funny state?
if(WEXITSTATUS(coordinatorStatus) == CS_NO){
JASSERT (false) (isRestart)
.Text ("Coordinator in a funny state. Peers exist, not restarting," \
"\n but not in a running state. Often this means you are" \
" connecting to\n a stale coordinator from a previous" \
" computation.\n Try killing the other coordinator," \
" or using a different port for the new comp.");
}else if (WEXITSTATUS(coordinatorStatus) == DMTCP_FAIL_RC) {
JTRACE("Coordinator not found. Starting a new one.");
}else{
JTRACE("Bad result found for coordinator. Will try start a new one.");
}
startNewCoordinator ( modes, isRestart );
}else{
if (modes & COORD_FORCE_NEW) {
JTRACE("Forcing new coordinator. --new-coordinator flag given.");
startNewCoordinator ( modes, isRestart );
return;
}
JASSERT( modes & COORD_JOIN )
.Text("Coordinator already running, but '--new' flag was given.");
}
}
