int
main (int argc, char *argv[])
{
slurm_addr_t *cli;
slurm_addr_t *self;
slurm_msg_t *msg;
slurmd_job_t *job;
int ngids;
gid_t *gids;
int rc = 0;
if (process_cmdline (argc, argv) < 0)
fatal ("Error in slurmstepd command line");
xsignal_block(slurmstepd_blocked_signals);
conf = xmalloc(sizeof(*conf));
conf->argv = &argv;
conf->argc = &argc;
init_setproctitle(argc, argv);
if (slurm_select_init(1) != SLURM_SUCCESS )
fatal( "failed to initialize node selection plugin" );
/* Receive job parameters from the slurmd */
_init_from_slurmd(STDIN_FILENO, argv, &cli, &self, &msg,
&ngids, &gids);
/* Fancy way of closing stdin that keeps STDIN_FILENO from being
* allocated to any random file. The slurmd already opened /dev/null
* on STDERR_FILENO for us. */
dup2(STDERR_FILENO, STDIN_FILENO);
/* Create the slurmd_job_t, mostly from info in a
launch_tasks_request_msg_t or a batch_job_launch_msg_t */
if(!(job = _step_setup(cli, self, msg))) {
_send_fail_to_slurmd(STDOUT_FILENO);
rc = SLURM_FAILURE;
goto ending;
}
job->ngids = ngids;
job->gids = gids;
/* fork handlers cause mutexes on some global data structures
to be re-initialized after the fork. */
list_install_fork_handlers();
slurm_conf_install_fork_handlers();
/* sets job->msg_handle and job->msgid */
if (msg_thr_create(job) == SLURM_ERROR) {
_send_fail_to_slurmd(STDOUT_FILENO);
rc = SLURM_FAILURE;
goto ending;
}
_send_ok_to_slurmd(STDOUT_FILENO);
/* Fancy way of closing stdout that keeps STDOUT_FILENO from being
* allocated to any random file. The slurmd already opened /dev/null
* on STDERR_FILENO for us. */
dup2(STDERR_FILENO, STDOUT_FILENO);
/* This does most of the stdio setup, then launches all the tasks,
and blocks until the step is complete */
rc = job_manager(job);
/* signal the message thread to shutdown, and wait for it */
eio_signal_shutdown(job->msg_handle);
pthread_join(job->msgid, NULL);
if (job->batch)
batch_finish(job, rc); /* sends batch complete message */
ending:
#ifdef MEMORY_LEAK_DEBUG
_step_cleanup(job, msg, rc);
xfree(cli);
xfree(self);
xfree(conf->hostname);
xfree(conf->block_map);
xfree(conf->block_map_inv);
xfree(conf->spooldir);
xfree(conf->node_name);
xfree(conf->node_topo_addr);
xfree(conf->node_topo_pattern);
xfree(conf->logfile);
xfree(conf);
#endif
info("done with job");
return rc;
}
int
main (int argc, char **argv)
{
slurm_addr_t *cli;
slurm_addr_t *self;
slurm_msg_t *msg;
stepd_step_rec_t *job;
int ngids;
gid_t *gids;
int rc = 0;
char *launch_params;
if (_process_cmdline (argc, argv) < 0)
fatal ("Error in slurmstepd command line");
xsignal_block(slurmstepd_blocked_signals);
conf = xmalloc(sizeof(*conf));
conf->argv = &argv;
conf->argc = &argc;
init_setproctitle(argc, argv);
if (slurm_select_init(1) != SLURM_SUCCESS )
fatal( "failed to initialize node selection plugin" );
if (slurm_auth_init(NULL) != SLURM_SUCCESS)
fatal( "failed to initialize authentication plugin" );
/* Receive job parameters from the slurmd */
_init_from_slurmd(STDIN_FILENO, argv, &cli, &self, &msg,
&ngids, &gids);
/* Create the stepd_step_rec_t, mostly from info in a
* launch_tasks_request_msg_t or a batch_job_launch_msg_t */
if (!(job = _step_setup(cli, self, msg))) {
_send_fail_to_slurmd(STDOUT_FILENO);
rc = SLURM_FAILURE;
goto ending;
}
job->ngids = ngids;
job->gids = gids;
/* fork handlers cause mutexes on some global data structures
* to be re-initialized after the fork. */
list_install_fork_handlers();
slurm_conf_install_fork_handlers();
/* sets job->msg_handle and job->msgid */
if (msg_thr_create(job) == SLURM_ERROR) {
_send_fail_to_slurmd(STDOUT_FILENO);
rc = SLURM_FAILURE;
goto ending;
}
_send_ok_to_slurmd(STDOUT_FILENO);
_got_ack_from_slurmd(STDIN_FILENO);
/* Fancy way of closing stdin that keeps STDIN_FILENO from being
* allocated to any random file. The slurmd already opened /dev/null
* on STDERR_FILENO for us. */
dup2(STDERR_FILENO, STDIN_FILENO);
/* Fancy way of closing stdout that keeps STDOUT_FILENO from being
* allocated to any random file. The slurmd already opened /dev/null
* on STDERR_FILENO for us. */
dup2(STDERR_FILENO, STDOUT_FILENO);
/* slurmstepd is the only daemon that should survive upgrade. If it
* had been swapped out before upgrade happened it could easily lead
* to SIGBUS at any time after upgrade. Avoid that by locking it
* in-memory. */
launch_params = slurm_get_launch_params();
if (launch_params && strstr(launch_params, "slurmstepd_memlock")) {
#ifdef _POSIX_MEMLOCK
int flags = MCL_CURRENT;
if (strstr(launch_params, "slurmstepd_memlock_all"))
flags |= MCL_FUTURE;
if (mlockall(flags) < 0)
info("failed to mlock() slurmstepd pages: %m");
else
debug("slurmstepd locked in memory");
#else
info("mlockall() system call does not appear to be available");
#endif
}
xfree(launch_params);
/* This does most of the stdio setup, then launches all the tasks,
* and blocks until the step is complete */
rc = job_manager(job);
return stepd_cleanup(msg, job, cli, self, rc, 0);
ending:
return stepd_cleanup(msg, job, cli, self, rc, 1);
}
