extern int stepd_cleanup(slurm_msg_t *msg, stepd_step_rec_t *job,
slurm_addr_t *cli, slurm_addr_t *self,
int rc, bool only_mem)
{
if (!only_mem) {
if (job->batch)
batch_finish(job, rc); /* sends batch complete message */
/* signal the message thread to shutdown, and wait for it */
eio_signal_shutdown(job->msg_handle);
pthread_join(job->msgid, NULL);
}
mpi_fini(); /* Remove stale PMI2 sockets */
#ifdef MEMORY_LEAK_DEBUG
acct_gather_conf_destroy();
(void) core_spec_g_fini();
_step_cleanup(job, msg, rc);
fini_setproctitle();
xfree(cli);
xfree(self);
xfree(conf->block_map);
xfree(conf->block_map_inv);
xfree(conf->hostname);
xfree(conf->job_acct_gather_freq);
xfree(conf->job_acct_gather_type);
xfree(conf->logfile);
xfree(conf->node_name);
xfree(conf->node_topo_addr);
xfree(conf->node_topo_pattern);
xfree(conf->spooldir);
xfree(conf->task_epilog);
xfree(conf->task_prolog);
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;
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;
}