/* Convert a SLURM hostlist expression into the equivalent node index
* value expression.
*/
static char *_get_nids(opt_t *opt_local)
{
hostlist_t hl;
char *nids = NULL;
int node_cnt;
if (!opt_local->nodelist)
return NULL;
hl = hostlist_create(opt_local->nodelist);
if (!hl) {
error("Invalid hostlist: %s", opt_local->nodelist);
return NULL;
}
//info("input hostlist: %s", nodelist);
hostlist_uniq(hl);
/* aprun needs the hostlist to be the exact size requested.
So if it doesn't set it.
*/
node_cnt = hostlist_count(hl);
if (opt_local->nodes_set_opt && (node_cnt != opt_local->min_nodes)) {
error("You requested %d nodes and %d hosts. These numbers "
"must be the same, so setting number of nodes to %d",
opt_local->min_nodes, node_cnt, node_cnt);
}
opt_local->min_nodes = node_cnt;
opt_local->nodes_set = 1;
nids = cray_nodelist2nids(hl, NULL);
hostlist_destroy(hl);
//info("output node IDs: %s", nids);
return nids;
}
static int _run_nhc(nhc_info_t *nhc_info)
{
#ifdef HAVE_NATIVE_CRAY
int argc = 11, status = 1, wait_rc, i = 0;
char *argv[argc];
pid_t cpid;
char *jobid_char = NULL, *apid_char = NULL, *nodelist_nids = NULL,
*exit_char = NULL;
DEF_TIMERS;
START_TIMER;
apid_char = xstrdup_printf("%"PRIu64"", nhc_info->apid);
exit_char = xstrdup_printf("%u", nhc_info->exit_code);
jobid_char = xstrdup_printf("%u", nhc_info->jobid);
nodelist_nids = cray_nodelist2nids(NULL, nhc_info->nodelist);
argv[i++] = "/opt/cray/nodehealth/default/bin/xtcleanup_after";
argv[i++] = "-a";
argv[i++] = apid_char;
argv[i++] = "-e";
argv[i++] = exit_char;
argv[i++] = "-r";
argv[i++] = jobid_char;
argv[i++] = "-m";
argv[i++] = nhc_info->step ? "application" : "reservation";
argv[i++] = nodelist_nids;
argv[i++] = NULL;
if (debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("Calling NHC for jobid %u and apid %"PRIu64" "
"on nodes %s(%s) exit code %u",
nhc_info->jobid, nhc_info->apid,
nhc_info->nodelist, nodelist_nids,
nhc_info->exit_code);
}
if (!nhc_info->nodelist || !nodelist_nids) {
/* already done */
goto fini;
}
if ((cpid = fork()) < 0) {
error("_run_nhc fork error: %m");
goto fini;
}
if (cpid == 0) {
#ifdef SETPGRP_TWO_ARGS
setpgrp(0, 0);
#else
setpgrp();
#endif
execvp(argv[0], argv);
exit(127);
}
while (1) {
wait_rc = waitpid(cpid, &status, 0);
if (wait_rc < 0) {
if (errno == EINTR)
continue;
error("_run_nhc waitpid error: %m");
break;
} else if (wait_rc > 0) {
killpg(cpid, SIGKILL); /* kill children too */
break;
}
}
END_TIMER;
if (status != 0) {
error("_run_nhc jobid %u and apid %"PRIu64" exit "
"status %u:%u took: %s",
nhc_info->jobid, nhc_info->apid, WEXITSTATUS(status),
WTERMSIG(status), TIME_STR);
} else if (debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("_run_nhc jobid %u and apid %"PRIu64" completed took: %s",
nhc_info->jobid, nhc_info->apid, TIME_STR);
fini:
xfree(apid_char);
xfree(exit_char);
xfree(jobid_char);
xfree(nodelist_nids);
return status;
#else
if (debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("simluating calling NHC for jobid %u "
"and apid %"PRIu64" on nodes %s",
nhc_info->jobid, nhc_info->apid, nhc_info->nodelist);
/* simulate sleeping */
sleep(2);
return 0;
#endif
}
static int _run_nhc(uint64_t id, char *nodelist, bool step)
{
#ifdef HAVE_NATIVE_CRAY
int argc = 5, status = 1, wait_rc;
char *argv[argc];
pid_t cpid;
DEF_TIMERS;
START_TIMER;
argv[0] = "/opt/cray/nodehealth/default/bin/xtcleanup_after";
if (step)
argv[1] = "-a";
else
argv[1] = "-r";
argv[2] = xstrdup_printf("%"PRIu64"", id);
argv[3] = cray_nodelist2nids(NULL, nodelist);
argv[4] = NULL;
if (debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("Calling NHC for id %"PRIu64" on nodes %s(%s)",
id, nodelist, argv[3]);
if ((cpid = fork()) < 0) {
error("_run_nhc fork error: %m");
goto fini;
}
if (cpid == 0) {
#ifdef SETPGRP_TWO_ARGS
setpgrp(0, 0);
#else
setpgrp();
#endif
execvp(argv[0], argv);
exit(127);
}
while (1) {
wait_rc = waitpid(cpid, &status, 0);
if (wait_rc < 0) {
if (errno == EINTR)
continue;
error("_run_nhc waitpid error: %m");
break;
} else if (wait_rc > 0) {
killpg(cpid, SIGKILL); /* kill children too */
break;
}
}
END_TIMER;
if (status != 0) {
error("_run_nhc %s %"PRIu64" exit status %u:%u took: %s",
step ? "step" : "job", step ? id : id,
WEXITSTATUS(status), WTERMSIG(status), TIME_STR);
} else if (debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("_run_nhc %s %"PRIu64" completed took: %s",
step ? "step" : "job", step ? id : id, TIME_STR);
fini:
xfree(argv[2]);
xfree(argv[3]);
return status;
#else
if (debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("simluating calling NHC for id %"PRIu64" on nodes %s",
id, nodelist);
/* simulate sleeping */
sleep(2);
return 0;
#endif
}
