static int queryPbsConfig(dmtcp::string option, dmtcp::string &pbs_config)
{
int fds[2];
const char *pbs_config_path = "pbs-config";
static const char *pbs_config_args[] = { "pbs-config", option.c_str(), NULL };
int cpid;
if( pipe(fds) == -1){
// just go away - we cannot serve this request
JTRACE("Cannot create pipe to execute pbs-config to find Torque/PBS library!");
return -1;
}
cpid = _real_fork();
if( cpid < 0 ){
JTRACE( "ERROR: cannot execute pbs-config. Will not run tm_spawn!");
return -1;
}
if( cpid == 0 ){
JTRACE ( "child process, will exec into external de-compressor");
fds[1] = dup(dup(dup(fds[1])));
close(fds[0]);
JASSERT(dup2(fds[1], STDOUT_FILENO) == STDOUT_FILENO);
close(fds[1]);
_real_execvp(pbs_config_path, (char **)pbs_config_args);
/* should not get here */
JASSERT(false)("ERROR: Failed to exec pbs-config. tm_spawn will fail with TM_BADINIT")(strerror(errno));
exit(0);
}
/* parent process */
JTRACE ( "created child process for pbs-config")(cpid);
int status;
if( waitpid(cpid,&status,0) < 0 ){
return -1;
}
if( !( WIFEXITED(status) && WEXITSTATUS(status) == 0 ) ){
return -1;
}
// set descriptor as non-blocking
// JTRACE ( "Set pipe fds[0] as non-blocking");
int flags = fcntl(fds[0], F_GETFL);
fcntl(fds[0], F_SETFL, flags | O_NONBLOCK);
//JTRACE ( "Read pbs-config output from pipe");
pbs_config = "";
char buf[256];
int count = 0;
while( (count = read(fds[0], buf, 255)) > 0 ){
buf[count] = '\0';
pbs_config += dmtcp::string() + buf;
}
JTRACE ( "pbs-config output:")(pbs_config);
return 0;
}
extern "C" pid_t fork()
{
if (isPerformingCkptRestart()) {
return _real_syscall(SYS_fork);
}
/* Acquire the wrapperExeution lock to prevent checkpoint to happen while
* processing this system call.
*/
WRAPPER_EXECUTION_GET_EXCL_LOCK();
dmtcp::DmtcpWorker::processEvent(DMTCP_EVENT_PRE_FORK, NULL);
/* Little bit cheating here: child_time should be same for both parent and
* child, thus we compute it before forking the child. */
child_time = time(NULL);
long host = dmtcp::UniquePid::ThisProcess().hostid();
dmtcp::UniquePid parent = dmtcp::UniquePid::ThisProcess();
dmtcp::string child_name = jalib::Filesystem::GetProgramName() + "_(forked)";
coordinatorAPI.createNewConnectionBeforeFork(child_name);
dmtcp::Util::setVirtualPidEnvVar(coordinatorAPI.virtualPid(), getpid());
//Enable the pthread_atfork child call
pthread_atfork_enabled = true;
pid_t childPid = _real_fork();
if (childPid == -1) {
} else if (childPid == 0) { /* child process */
/* NOTE: Any work that needs to be done for the newly created child
* should be put into pthread_atfork_child() function. That function is
* hooked to the libc:fork() and will be called right after the new
* process is created and before the fork() returns.
*
* pthread_atfork_child is registered by calling pthread_atfork() from
* within the DmtcpWorker constructor to make sure that this is the first
* registered handle.
*/
dmtcp::UniquePid child = dmtcp::UniquePid(host, getpid(), child_time);
JTRACE("fork() done [CHILD]") (child) (parent);
dmtcp::initializeMtcpEngine();
} else if (childPid > 0) { /* Parent Process */
dmtcp::UniquePid child = dmtcp::UniquePid(host, childPid, child_time);
dmtcp::ProcessInfo::instance().insertChild(childPid, child);
JTRACE("fork()ed [PARENT] done") (child);;
}
pthread_atfork_enabled = false;
if (childPid != 0) {
dmtcp::Util::setVirtualPidEnvVar(getpid(), getppid());
coordinatorAPI.closeConnection();
WRAPPER_EXECUTION_RELEASE_EXCL_LOCK();
}
return childPid;
}
pid_t fork()
{
WRAPPER_HEADER(pid_t, fork, _real_fork);
if (SYNC_IS_REPLAY) {
WRAPPER_REPLAY_TYPED(pid_t, fork);
} else if (SYNC_IS_RECORD) {
dmtcp::ThreadInfo::setOptionalEvent();
retval = _real_fork();
if (retval != 0) {
dmtcp::ThreadInfo::unsetOptionalEvent();
WRAPPER_LOG_WRITE_ENTRY(fork);
}
}
return retval;
}
extern "C" pid_t fork()
{
pid_t retval = 0;
pid_t virtualPid = getPidFromEnvVar();
dmtcp::VirtualPidTable::instance().writeVirtualTidToFileForPtrace(virtualPid);
pid_t realPid = _real_fork();
if (realPid > 0) { /* Parent Process */
retval = virtualPid;
dmtcp::VirtualPidTable::instance().updateMapping(virtualPid, realPid);
} else {
retval = realPid;
dmtcp::VirtualPidTable::instance().readVirtualTidFromFileForPtrace();
}
return retval;
}
// Copied from mtcp/mtcp_restart.c.
// Let's keep this code close to MTCP code to avoid maintenance problems.
// MTCP code in: mtcp/mtcp_restart.c:open_ckpt_to_read()
// A previous version tried to replace this with popen, causing a regression:
// (no call to pclose, and possibility of using a wrong fd).
// Returns fd; sets ext_decomp_pid, if checkpoint was compressed.
static int open_ckpt_to_read(const char *filename)
{
int fd;
int fds[2];
char fc;
const char *decomp_path;
const char **decomp_args;
const char *gzip_path = "gzip";
static const char * gzip_args[] = { "gzip", "-d", "-", NULL };
#ifdef HBICT_DELTACOMP
const char *hbict_path = "hbict";
static const char *hbict_args[] = { "hbict", "-r", NULL };
#endif
pid_t cpid;
fc = first_char(filename);
fd = open(filename, O_RDONLY);
JASSERT(fd>=0)(filename).Text("Failed to open file.");
if(fc == DMTCP_MAGIC_FIRST) /* no compression */
return fd;
#ifdef HBICT_DELTACOMP
else if (fc == GZIP_FIRST || fc == HBICT_FIRST){ /* External compression */
#else
else if(fc == GZIP_FIRST){ /* gzip */
#endif
if( fc == GZIP_FIRST ){
decomp_path = gzip_path;
decomp_args = gzip_args;
}
#ifdef HBICT_DELTACOMP
else{
decomp_path = hbict_path;
decomp_args = hbict_args;
}
#endif
JASSERT(pipe(fds) != -1)(filename).Text("Cannot create pipe to execute gunzip to decompress checkpoint file!");
cpid = _real_fork();
JASSERT(cpid != -1).Text("ERROR: Cannot fork to execute gunzip to decompress checkpoint file!");
if(cpid > 0) /* parent process */
{
JTRACE ( "created child process to uncompress checkpoint file")(cpid);
ext_decomp_pid = cpid;
close(fd);
close(fds[1]);
return fds[0];
}
else /* child process */
{
JTRACE ( "child process, will exec into external de-compressor");
fd = dup(dup(dup(fd)));
fds[1] = dup(fds[1]);
close(fds[0]);
JASSERT(fd != -1);
JASSERT(dup2(fd, STDIN_FILENO) == STDIN_FILENO);
close(fd);
JASSERT(dup2(fds[1], STDOUT_FILENO) == STDOUT_FILENO);
close(fds[1]);
_real_execvp(decomp_path, (char **)decomp_args);
JASSERT(decomp_path!=NULL)(decomp_path).Text("Failed to launch gzip.");
/* should not get here */
JASSERT(false)("ERROR: Decompression failed! No restoration will be performed! Cancelling now!");
abort();
}
} else /* invalid magic number */
JASSERT(false).Text("ERROR: Invalid magic number in this checkpoint file!");
// NOT_REACHED
return -1;
}
// See comments above for open_ckpt_to_read()
int dmtcp::CkptSerializer::openDmtcpCheckpointFile(const dmtcp::string& path){
// Function also sets dmtcp::ext_decomp_pid::ConnectionToFds
int fd = open_ckpt_to_read( path.c_str() );
// The rest of this function is for compatibility with original definition.
JASSERT(fd>=0)(path).Text("Failed to open file.");
char buf[512];
const int len = strlen(DMTCP_FILE_HEADER);
JASSERT(_real_read(fd, buf, len)==len)(path).Text("_real_read() failed");
if(strncmp(buf, DMTCP_FILE_HEADER, len)==0){
JTRACE("opened checkpoint file [uncompressed]")(path);
}else{
close_ckpt_to_read(fd);
fd = open_ckpt_to_read( path.c_str() ); /* Re-open from beginning */
}
return fd;
}
// Copied from mtcp/mtcp_restart.c.
// Let's keep this code close to MTCP code to avoid maintenance problems.
// MTCP code in: mtcp/mtcp_restart.c:open_ckpt_to_read()
// A previous version tried to replace this with popen, causing a regression:
// (no call to pclose, and possibility of using a wrong fd).
// Returns fd; sets ext_decomp_pid, if checkpoint was compressed.
static int open_ckpt_to_read(const char *filename)
{
int fd;
int fds[2];
char fc;
const char *decomp_path;
const char **decomp_args;
const char *gzip_path = "gzip";
static const char * gzip_args[] = { "gzip", "-d", "-", NULL };
#ifdef HBICT_DELTACOMP
const char *hbict_path = "hbict";
static const char *hbict_args[] = { "hbict", "-r", NULL };
#endif
pid_t cpid;
fc = first_char(filename);
fd = open(filename, O_RDONLY);
JASSERT(fd>=0)(filename).Text("Failed to open file.");
if (fc == DMTCP_MAGIC_FIRST) { /* no compression */
return fd;
}
else if (fc == GZIP_FIRST
#ifdef HBICT_DELTACOMP
|| fc == HBICT_FIRST
#endif
) {
if (fc == GZIP_FIRST) {
decomp_path = gzip_path;
decomp_args = gzip_args;
}
#ifdef HBICT_DELTACOMP
else {
decomp_path = hbict_path;
decomp_args = hbict_args;
}
#endif
JASSERT(pipe(fds) != -1) (filename)
.Text("Cannot create pipe to execute gunzip to decompress ckpt file!");
cpid = _real_fork();
JASSERT(cpid != -1)
.Text("ERROR: Cannot fork to execute gunzip to decompress ckpt file!");
if (cpid > 0) { /* parent process */
JTRACE("created child process to uncompress checkpoint file") (cpid);
ext_decomp_pid = cpid;
close(fd);
close(fds[1]);
return fds[0];
} else { /* child process */
JTRACE ( "child process, will exec into external de-compressor");
fd = dup(dup(dup(fd)));
fds[1] = dup(fds[1]);
close(fds[0]);
JASSERT(fd != -1);
JASSERT(dup2(fd, STDIN_FILENO) == STDIN_FILENO);
close(fd);
JASSERT(dup2(fds[1], STDOUT_FILENO) == STDOUT_FILENO);
close(fds[1]);
_real_execvp(decomp_path, (char **)decomp_args);
JASSERT(decomp_path!=NULL) (decomp_path)
.Text("Failed to launch gzip.");
/* should not get here */
JASSERT(false)
.Text("Decompression failed! No restoration will be performed!");
}
} else { /* invalid magic number */
JASSERT(false)
.Text("ERROR: Invalid magic number in this checkpoint file!");
}
return -1;
}
