static int
_slurmd_init(void)
{
struct rlimit rlim;
slurm_ctl_conf_t *cf;
struct stat stat_buf;
uint32_t cpu_cnt;
/*
* Process commandline arguments first, since one option may be
* an alternate location for the slurm config file.
*/
_process_cmdline(*conf->argc, *conf->argv);
/*
* Build nodes table like in slurmctld
* This is required by the topology stack
* Node tables setup must preceed _read_config() so that the
* proper hostname is set.
*/
slurm_conf_init(conf->conffile);
init_node_conf();
/* slurm_select_init() must be called before
* build_all_nodeline_info() to be called with proper argument. */
if (slurm_select_init(1) != SLURM_SUCCESS )
return SLURM_FAILURE;
build_all_nodeline_info(true);
build_all_frontend_info(true);
/*
* Read global slurm config file, override necessary values from
* defaults and command line.
*/
_read_config();
cpu_cnt = MAX(conf->conf_cpus, conf->block_map_size);
if ((gres_plugin_init() != SLURM_SUCCESS) ||
(gres_plugin_node_config_load(cpu_cnt) != SLURM_SUCCESS))
return SLURM_FAILURE;
if (slurm_topo_init() != SLURM_SUCCESS)
return SLURM_FAILURE;
/*
* Get and set slurmd topology information
* Build node hash table first to speed up the topo build
*/
rehash_node();
slurm_topo_build_config();
_set_topo_info();
/*
* Check for cpu frequency set capabilities on this node
*/
cpu_freq_init(conf);
_print_conf();
if (slurm_proctrack_init() != SLURM_SUCCESS)
return SLURM_FAILURE;
if (slurmd_task_init() != SLURM_SUCCESS)
return SLURM_FAILURE;
if (slurm_auth_init(NULL) != SLURM_SUCCESS)
return SLURM_FAILURE;
if (spank_slurmd_init() < 0)
return SLURM_FAILURE;
if (getrlimit(RLIMIT_CPU, &rlim) == 0) {
rlim.rlim_cur = rlim.rlim_max;
setrlimit(RLIMIT_CPU, &rlim);
if (rlim.rlim_max != RLIM_INFINITY) {
error("Slurmd process CPU time limit is %d seconds",
(int) rlim.rlim_max);
}
}
if (getrlimit(RLIMIT_NOFILE, &rlim) == 0) {
rlim.rlim_cur = rlim.rlim_max;
setrlimit(RLIMIT_NOFILE, &rlim);
}
#ifndef NDEBUG
if (getrlimit(RLIMIT_CORE, &rlim) == 0) {
rlim.rlim_cur = rlim.rlim_max;
setrlimit(RLIMIT_CORE, &rlim);
}
#endif /* !NDEBUG */
/*
* Create a context for verifying slurm job credentials
*/
if (!(conf->vctx = slurm_cred_verifier_ctx_create(conf->pubkey)))
return SLURM_FAILURE;
if (!strcmp(conf->select_type, "select/serial")) {
/* Only cache credential for 5 seconds with select/serial
* for shorter cache searches and higher throughput */
slurm_cred_ctx_set(conf->vctx, SLURM_CRED_OPT_EXPIRY_WINDOW, 5);
}
/*
* Create slurmd spool directory if necessary.
*/
if (_set_slurmd_spooldir() < 0) {
error("Unable to initialize slurmd spooldir");
return SLURM_FAILURE;
}
if (conf->cleanstart) {
/*
* Need to kill any running slurmd's here
*/
_kill_old_slurmd();
stepd_cleanup_sockets(conf->spooldir, conf->node_name);
_stepd_cleanup_batch_dirs(conf->spooldir, conf->node_name);
}
if (conf->daemonize) {
bool success = false;
if (conf->logfile && (conf->logfile[0] == '/')) {
char *slash_ptr, *work_dir;
work_dir = xstrdup(conf->logfile);
slash_ptr = strrchr(work_dir, '/');
if (slash_ptr == work_dir)
work_dir[1] = '\0';
else
slash_ptr[0] = '\0';
if ((access(work_dir, W_OK) != 0) ||
(chdir(work_dir) < 0)) {
error("Unable to chdir to %s", work_dir);
} else
success = true;
xfree(work_dir);
}
if (!success) {
if ((access(conf->spooldir, W_OK) != 0) ||
(chdir(conf->spooldir) < 0)) {
error("Unable to chdir to %s", conf->spooldir);
} else
success = true;
}
if (!success) {
if ((access("/var/tmp", W_OK) != 0) ||
(chdir("/var/tmp") < 0)) {
error("chdir(/var/tmp): %m");
return SLURM_FAILURE;
} else
info("chdir to /var/tmp");
}
}
/*
* Cache the group access list
*/
cf = slurm_conf_lock();
if (cf->group_info & GROUP_CACHE)
init_gids_cache(1);
else
init_gids_cache(0);
slurm_conf_unlock();
if ((devnull = open_cloexec("/dev/null", O_RDWR)) < 0) {
error("Unable to open /dev/null: %m");
return SLURM_FAILURE;
}
/* make sure we have slurmstepd installed */
if (stat(conf->stepd_loc, &stat_buf))
fatal("Unable to find slurmstepd file at %s", conf->stepd_loc);
if (!S_ISREG(stat_buf.st_mode))
fatal("slurmstepd not a file at %s", conf->stepd_loc);
return SLURM_SUCCESS;
}
/*
* route_p_split_hostlist - logic to split an input hostlist into
* a set of hostlists to forward to.
*
* IN: hl - hostlist_t - list of every node to send message to
* will be empty on return;
* OUT: sp_hl - hostlist_t** - the array of hostlists that will be malloced
* OUT: count - int* - the count of created hostlists
* RET: SLURM_SUCCESS - int
*
* Note: created hostlist will have to be freed independently using
* hostlist_destroy by the caller.
* Note: the hostlist_t array will have to be xfree.
*/
extern int route_p_split_hostlist(hostlist_t hl,
hostlist_t** sp_hl,
int* count)
{
int i, j, k, hl_ndx, msg_count, sw_count, lst_count;
char *buf;
bitstr_t *nodes_bitmap = NULL; /* nodes in message list */
bitstr_t *fwd_bitmap = NULL; /* nodes in forward list */
msg_count = hostlist_count(hl);
if (switch_record_cnt == 0) {
/* configs have not already been processed */
slurm_conf_init(NULL);
if (init_node_conf()) {
fatal("ROUTE: Failed to init slurm config");
}
if (build_all_nodeline_info(false)) {
fatal("ROUTE: Failed to build node config");
}
rehash_node();
if (slurm_topo_build_config() != SLURM_SUCCESS) {
fatal("ROUTE: Failed to build topology config");
}
}
*sp_hl = (hostlist_t*) xmalloc(switch_record_cnt * sizeof(hostlist_t));
/* create bitmap of nodes to send message too */
if (hostlist2bitmap (hl, false, &nodes_bitmap) != SLURM_SUCCESS) {
buf = hostlist_ranged_string_xmalloc(hl);
fatal("ROUTE: Failed to make bitmap from hostlist=%s.", buf);
}
/* Find lowest level switch containing all the nodes in the list */
j = 0;
for (i = 0; i <= switch_levels; i++) {
for (j=0; j<switch_record_cnt; j++) {
if (switch_record_table[j].level == i) {
if (bit_super_set(nodes_bitmap,
switch_record_table[j].
node_bitmap)) {
/* All nodes in message list are in
* this switch */
break;
}
}
}
if (j < switch_record_cnt) {
/* Got here via break after bit_super_set */
break; // 'j' is our switch
} /* else, no switches at this level reach all nodes */
}
if (i > switch_levels) {
/* This can only happen if trying to schedule multiple physical
* clusters as a single logical cluster under the control of a
* single slurmctld daemon, and sending something like a
* node_registation request to all nodes.
* Revert to default behavior*/
if (debug_flags & DEBUG_FLAG_ROUTE) {
buf = hostlist_ranged_string_xmalloc(hl);
debug("ROUTE: didn't find switch containing nodes=%s",
buf);
xfree(buf);
}
FREE_NULL_BITMAP(nodes_bitmap);
xfree(*sp_hl);
return route_split_hostlist_treewidth(hl, sp_hl, count);
}
if (switch_record_table[j].level == 0) {
/* This is a leaf switch. Construct list based on TreeWidth */
FREE_NULL_BITMAP(nodes_bitmap);
xfree(*sp_hl);
return route_split_hostlist_treewidth(hl, sp_hl, count);
}
/* loop through children, construction a hostlist for each child switch
* with nodes in the message list */
hl_ndx = 0;
lst_count = 0;
for (i=0; i < switch_record_table[j].num_switches; i++) {
k = switch_record_table[j].switch_index[i];
fwd_bitmap = bit_copy(switch_record_table[k].node_bitmap);
bit_and(fwd_bitmap, nodes_bitmap);
sw_count = bit_set_count(fwd_bitmap);
if (sw_count == 0) {
continue; /* no nodes on this switch in message list */
}
(*sp_hl)[hl_ndx] = bitmap2hostlist(fwd_bitmap);
/* Now remove nodes from this switch from message list */
bit_not(fwd_bitmap);
bit_and(nodes_bitmap, fwd_bitmap);
FREE_NULL_BITMAP(fwd_bitmap);
if (debug_flags & DEBUG_FLAG_ROUTE) {
buf = hostlist_ranged_string_xmalloc((*sp_hl)[hl_ndx]);
debug("ROUTE: ... sublist[%d] switch=%s :: %s",
i, switch_record_table[i].name, buf);
xfree(buf);
}
hl_ndx++;
lst_count += sw_count;
if (lst_count == msg_count)
break; /* all nodes in message are in a child list */
}
FREE_NULL_BITMAP(nodes_bitmap);
*count = hl_ndx;
return SLURM_SUCCESS;
}
