/*
* assign_front_end - assign a front end node for starting a job
* job_ptr IN - job to assign a front end node (tests access control lists)
* RET pointer to the front end node to use or NULL if none found
*/
extern front_end_record_t *assign_front_end(struct job_record *job_ptr)
{
#ifdef HAVE_FRONT_END
front_end_record_t *front_end_ptr, *best_front_end = NULL;
uint32_t state_flags;
int i;
if (!job_ptr->batch_host && (job_ptr->batch_flag == 0) &&
(front_end_ptr = find_front_end_record(job_ptr->alloc_node))) {
/* Use submit host for interactive job */
if (!IS_NODE_DOWN(front_end_ptr) &&
!IS_NODE_DRAIN(front_end_ptr) &&
!IS_NODE_NO_RESPOND(front_end_ptr) &&
_front_end_access(front_end_ptr, job_ptr)) {
best_front_end = front_end_ptr;
} else {
info("%s: front-end node %s not available for job %u",
__func__, job_ptr->alloc_node, job_ptr->job_id);
return NULL;
}
} else {
for (i = 0, front_end_ptr = front_end_nodes;
i < front_end_node_cnt; i++, front_end_ptr++) {
if (job_ptr->batch_host) { /* Find specific front-end */
if (xstrcmp(job_ptr->batch_host,
front_end_ptr->name))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
break;
} else { /* Find a usable front-end node */
if (IS_NODE_DOWN(front_end_ptr) ||
IS_NODE_DRAIN(front_end_ptr) ||
IS_NODE_NO_RESPOND(front_end_ptr))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
continue;
}
if ((best_front_end == NULL) ||
(front_end_ptr->job_cnt_run <
best_front_end->job_cnt_run))
best_front_end = front_end_ptr;
}
}
if (best_front_end) {
state_flags = best_front_end->node_state & NODE_STATE_FLAGS;
best_front_end->node_state = NODE_STATE_ALLOCATED | state_flags;
best_front_end->job_cnt_run++;
return best_front_end;
} else if (job_ptr->batch_host) { /* Find specific front-end node */
error("assign_front_end: front end node %s not found",
job_ptr->batch_host);
} else { /* Find some usable front-end node */
error("assign_front_end: no available front end nodes found");
}
#endif
return NULL;
}
static char * _get_node_state(struct node_record *node_ptr)
{
static bool got_select_type = false;
static bool node_allocations;
if (!got_select_type) {
char * select_type = slurm_get_select_type();
if (select_type &&
(strcasecmp(select_type, "select/linear") == 0))
node_allocations = true;
else
node_allocations = false;
xfree(select_type);
got_select_type = true;
}
if (IS_NODE_DRAIN(node_ptr) || IS_NODE_FAIL(node_ptr))
return "Draining";
if (IS_NODE_COMPLETING(node_ptr))
return "Busy";
if (IS_NODE_DOWN(node_ptr))
return "Down";
if (IS_NODE_ALLOCATED(node_ptr)) {
if (node_allocations)
return "Busy";
else
return "Running";
}
if (IS_NODE_IDLE(node_ptr))
return "Idle";
return "Unknown";
}
/*
* assign_front_end - assign a front end node for starting a job
* RET pointer to the front end node to use or NULL if none available
*/
extern front_end_record_t *assign_front_end(void)
{
#ifdef HAVE_FRONT_END
static int last_assigned = -1;
front_end_record_t *front_end_ptr;
uint16_t state_flags;
int i;
for (i = 0; i < front_end_node_cnt; i++) {
last_assigned = (last_assigned + 1) % front_end_node_cnt;
front_end_ptr = front_end_nodes + last_assigned;
if (IS_NODE_DOWN(front_end_ptr) ||
IS_NODE_DRAIN(front_end_ptr) ||
IS_NODE_NO_RESPOND(front_end_ptr))
continue;
state_flags = front_end_nodes[last_assigned].node_state &
NODE_STATE_FLAGS;
front_end_nodes[last_assigned].node_state =
NODE_STATE_ALLOCATED | state_flags;
front_end_nodes[last_assigned].job_cnt_run++;
return front_end_ptr;
}
fatal("assign_front_end: no available front end nodes found");
#endif
return NULL;
}
/* Determine if specific slurm node is already in DOWN or DRAIN state */
extern int node_already_down(char *node_name)
{
struct node_record *node_ptr = find_node_record(node_name);
if (node_ptr) {
if (IS_NODE_DRAIN(node_ptr))
return 2;
else if (IS_NODE_DOWN(node_ptr))
return 1;
else
return 0;
}
return 0;
}
extern void
scontrol_print_completing_job(job_info_t *job_ptr,
node_info_msg_t *node_info_msg)
{
int i, c_offset = 0;
node_info_t *node_info;
hostlist_t comp_nodes, down_nodes;
char *node_buf;
comp_nodes = hostlist_create(NULL);
down_nodes = hostlist_create(NULL);
if (job_ptr->cluster && federation_flag && !local_flag)
c_offset = get_cluster_node_offset(job_ptr->cluster,
node_info_msg);
for (i = 0; job_ptr->node_inx[i] != -1; i+=2) {
int j = job_ptr->node_inx[i];
for (; j <= job_ptr->node_inx[i+1]; j++) {
int node_inx = j + c_offset;
if (node_inx >= node_info_msg->record_count)
break;
node_info = &(node_info_msg->node_array[node_inx]);
if (IS_NODE_COMPLETING(node_info))
hostlist_push_host(comp_nodes, node_info->name);
else if (IS_NODE_DOWN(node_info))
hostlist_push_host(down_nodes, node_info->name);
}
}
fprintf(stdout, "JobId=%u ", job_ptr->job_id);
node_buf = hostlist_ranged_string_xmalloc(comp_nodes);
if (node_buf && node_buf[0])
fprintf(stdout, "Nodes(COMPLETING)=%s ", node_buf);
xfree(node_buf);
node_buf = hostlist_ranged_string_xmalloc(down_nodes);
if (node_buf && node_buf[0])
fprintf(stdout, "Nodes(DOWN)=%s ", node_buf);
xfree(node_buf);
fprintf(stdout, "\n");
hostlist_destroy(comp_nodes);
hostlist_destroy(down_nodes);
}
/*
* avail_front_end - test if any front end nodes are available for starting job
*/
extern bool avail_front_end(void)
{
#ifdef HAVE_FRONT_END
front_end_record_t *front_end_ptr;
int i;
for (i = 0, front_end_ptr = front_end_nodes;
i < front_end_node_cnt; i++, front_end_ptr++) {
if (IS_NODE_DOWN(front_end_ptr) ||
IS_NODE_DRAIN(front_end_ptr) ||
IS_NODE_NO_RESPOND(front_end_ptr))
continue;
return true;
}
return false;
#else
return true;
#endif
}
/*
* assign_front_end - assign a front end node for starting a job
* job_ptr IN - job to assign a front end node (tests access control lists)
* RET pointer to the front end node to use or NULL if none found
*/
extern front_end_record_t *assign_front_end(struct job_record *job_ptr)
{
#ifdef HAVE_FRONT_END
front_end_record_t *front_end_ptr, *best_front_end = NULL;
uint32_t state_flags;
int i;
for (i = 0, front_end_ptr = front_end_nodes; i < front_end_node_cnt;
i++, front_end_ptr++) {
if (job_ptr->batch_host) { /* Find specific front-end node */
if (strcmp(job_ptr->batch_host, front_end_ptr->name))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
break;
} else { /* Find some usable front-end node */
if (IS_NODE_DOWN(front_end_ptr) ||
IS_NODE_DRAIN(front_end_ptr) ||
IS_NODE_NO_RESPOND(front_end_ptr))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
continue;
}
if ((best_front_end == NULL) ||
(front_end_ptr->job_cnt_run < best_front_end->job_cnt_run))
best_front_end = front_end_ptr;
}
if (best_front_end) {
state_flags = best_front_end->node_state & NODE_STATE_FLAGS;
best_front_end->node_state = NODE_STATE_ALLOCATED | state_flags;
best_front_end->job_cnt_run++;
return best_front_end;
} else if (job_ptr->batch_host) { /* Find specific front-end node */
error("assign_front_end: front end node %s not found",
job_ptr->batch_host);
} else { /* Find some usable front-end node */
error("assign_front_end: no available front end nodes found");
}
#endif
return NULL;
}
extern void
scontrol_print_completing_job(job_info_t *job_ptr,
node_info_msg_t *node_info_msg)
{
int i;
node_info_t *node_info;
hostlist_t all_nodes, comp_nodes, down_nodes;
char *node_buf;
all_nodes = hostlist_create(job_ptr->nodes);
comp_nodes = hostlist_create("");
down_nodes = hostlist_create("");
for (i=0; i<node_info_msg->record_count; i++) {
node_info = &(node_info_msg->node_array[i]);
if (IS_NODE_COMPLETING(node_info) &&
(_in_node_bit_list(i, job_ptr->node_inx)))
hostlist_push_host(comp_nodes, node_info->name);
else if (IS_NODE_DOWN(node_info) &&
(hostlist_find(all_nodes, node_info->name) != -1))
hostlist_push_host(down_nodes, node_info->name);
}
fprintf(stdout, "JobId=%u ", job_ptr->job_id);
node_buf = hostlist_ranged_string_xmalloc(comp_nodes);
if (node_buf && node_buf[0])
fprintf(stdout, "Nodes(COMPLETING)=%s ", node_buf);
xfree(node_buf);
node_buf = hostlist_ranged_string_xmalloc(down_nodes);
if (node_buf && node_buf[0])
fprintf(stdout, "Nodes(DOWN)=%s ", node_buf);
xfree(node_buf);
fprintf(stdout, "\n");
hostlist_destroy(all_nodes);
hostlist_destroy(comp_nodes);
hostlist_destroy(down_nodes);
}
/*
* assign_front_end - assign a front end node for starting a job
* job_ptr IN - job to assign a front end node (tests access control lists)
* RET pointer to the front end node to use or NULL if none found
*/
extern front_end_record_t *assign_front_end(struct job_record *job_ptr)
{
#ifdef HAVE_FRONT_END
static int last_assigned = -1;
front_end_record_t *front_end_ptr;
uint16_t state_flags;
int i;
for (i = 0; i < front_end_node_cnt; i++) {
last_assigned = (last_assigned + 1) % front_end_node_cnt;
front_end_ptr = front_end_nodes + last_assigned;
if (job_ptr->batch_host) { /* Find specific front-end node */
if (strcmp(job_ptr->batch_host, front_end_ptr->name))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
break;
} else { /* Find some usable front-end node */
if (IS_NODE_DOWN(front_end_ptr) ||
IS_NODE_DRAIN(front_end_ptr) ||
IS_NODE_NO_RESPOND(front_end_ptr))
continue;
if (!_front_end_access(front_end_ptr, job_ptr))
continue;
}
state_flags = front_end_nodes[last_assigned].node_state &
NODE_STATE_FLAGS;
front_end_nodes[last_assigned].node_state =
NODE_STATE_ALLOCATED | state_flags;
front_end_nodes[last_assigned].job_cnt_run++;
return front_end_ptr;
}
if (job_ptr->batch_host) { /* Find specific front-end node */
error("assign_front_end: front end node %s not found",
job_ptr->batch_host);
} else { /* Find some usable front-end node */
error("assign_front_end: no available front end nodes found");
}
#endif
return NULL;
}
/**
* basil_inventory - Periodic node-state query via ALPS XML-RPC.
* This should be run immediately before each scheduling cycle.
* Returns non-SLURM_SUCCESS if
* - INVENTORY method failed (error)
* - no nodes are available (no point in scheduling)
* - orphaned ALPS reservation exists (wait until ALPS resynchronizes)
*/
extern int basil_inventory(void)
{
enum basil_version version = get_basil_version();
struct basil_inventory *inv;
struct basil_node *node;
struct basil_rsvn *rsvn;
int slurm_alps_mismatch = 0;
int rc = SLURM_SUCCESS;
inv = get_full_inventory(version);
if (inv == NULL) {
error("BASIL %s INVENTORY failed", bv_names_long[version]);
return SLURM_ERROR;
}
debug("BASIL %s INVENTORY: %d/%d batch nodes available",
bv_names_long[version], inv->batch_avail, inv->batch_total);
if (!inv->f->node_head || !inv->batch_avail || !inv->batch_total)
rc = ESLURM_REQUESTED_NODE_CONFIG_UNAVAILABLE;
for (node = inv->f->node_head; node; node = node->next) {
struct node_record *node_ptr;
char *reason = NULL;
node_ptr = _find_node_by_basil_id(node->node_id);
if (node_ptr == NULL) {
error("nid%05u (%s node in state %s) not in slurm.conf",
node->node_id, nam_noderole[node->role],
nam_nodestate[node->state]);
continue;
}
if (node_is_allocated(node) && !IS_NODE_ALLOCATED(node_ptr)) {
/*
* ALPS still hangs on to the node while SLURM considers
* it already unallocated. Possible causes are partition
* cleanup taking too long (can be 10sec ... minutes),
* and orphaned ALPS reservations (caught below).
*
* The converse case (SLURM hanging on to the node while
* ALPS has already freed it) happens frequently during
* job completion: select_g_job_fini() is called before
* make_node_comp(). Rely on SLURM logic for this case.
*/
slurm_alps_mismatch++;
}
if (node->state == BNS_DOWN) {
reason = "ALPS marked it DOWN";
} else if (node->state == BNS_UNAVAIL) {
reason = "node is UNAVAILABLE";
} else if (node->state == BNS_ROUTE) {
reason = "node does ROUTING";
} else if (node->state == BNS_SUSPECT) {
reason = "entered SUSPECT mode";
} else if (node->state == BNS_ADMINDOWN) {
reason = "node is ADMINDOWN";
} else if (node->state != BNS_UP) {
reason = "state not UP";
} else if (node->role != BNR_BATCH) {
reason = "mode not BATCH";
} else if (node->arch != BNA_XT) {
reason = "arch not XT/XE";
}
if (reason) {
if (!IS_NODE_DOWN(node_ptr)) {
xfree(node_ptr->reason);
debug("MARKING %s DOWN (%s)",
node_ptr->name, reason);
/* set_node_down also kills any running jobs */
set_node_down(node_ptr->name, reason);
}
} else if (IS_NODE_DOWN(node_ptr)) {
xfree(node_ptr->reason);
/* Reset state, make_node_idle figures out the rest */
node_ptr->node_state &= NODE_STATE_FLAGS;
node_ptr->node_state |= NODE_STATE_UNKNOWN;
make_node_idle(node_ptr, NULL);
}
}
if (slurm_alps_mismatch)
debug("ALPS: %d node(s) still held", slurm_alps_mismatch);
/*
* Check that each ALPS reservation corresponds to a SLURM job.
* Purge orphaned reservations, which may result from stale or
* messed up system state, or are indicative of ALPS problems
* (stuck in pending cancel calls).
*/
for (rsvn = inv->f->rsvn_head; rsvn; rsvn = rsvn->next) {
ListIterator job_iter = list_iterator_create(job_list);
struct job_record *job_ptr;
uint32_t resv_id;
if (job_iter == NULL)
fatal("list_iterator_create: malloc failure");
while ((job_ptr = (struct job_record *)list_next(job_iter))) {
if (_get_select_jobinfo(job_ptr->select_jobinfo->data,
SELECT_JOBDATA_RESV_ID,
&resv_id) == SLURM_SUCCESS
&& resv_id == rsvn->rsvn_id)
break;
}
list_iterator_destroy(job_iter);
if (job_ptr == NULL) {
error("orphaned ALPS reservation %u, trying to remove",
rsvn->rsvn_id);
basil_safe_release(rsvn->rsvn_id, inv);
slurm_alps_mismatch = true;
}
}
free_inv(inv);
if (slurm_alps_mismatch)
/* ALPS will take some time, do not schedule now. */
return ESLURM_REQUESTED_NODE_CONFIG_UNAVAILABLE;
return rc;
}
/* Perform any power change work to nodes */
static void _do_power_work(time_t now)
{
static time_t last_log = 0, last_work_scan = 0;
int i, wake_cnt = 0, sleep_cnt = 0, susp_total = 0;
time_t delta_t;
uint32_t susp_state;
bitstr_t *wake_node_bitmap = NULL, *sleep_node_bitmap = NULL;
struct node_record *node_ptr;
bool run_suspend = false;
/* Set limit on counts of nodes to have state changed */
delta_t = now - last_work_scan;
if (delta_t >= 60) {
suspend_cnt_f = 0.0;
resume_cnt_f = 0.0;
} else {
float rate = (60 - delta_t) / 60.0;
suspend_cnt_f *= rate;
resume_cnt_f *= rate;
}
suspend_cnt = (suspend_cnt_f + 0.5);
resume_cnt = (resume_cnt_f + 0.5);
if (now > (last_suspend + suspend_timeout)) {
/* ready to start another round of node suspends */
run_suspend = true;
if (last_suspend) {
bit_nclear(suspend_node_bitmap, 0,
(node_record_count - 1));
bit_nclear(resume_node_bitmap, 0,
(node_record_count - 1));
last_suspend = (time_t) 0;
}
}
last_work_scan = now;
/* Build bitmaps identifying each node which should change state */
for (i = 0, node_ptr = node_record_table_ptr;
i < node_record_count; i++, node_ptr++) {
susp_state = IS_NODE_POWER_SAVE(node_ptr);
if (susp_state)
susp_total++;
/* Resume nodes as appropriate */
if (susp_state &&
((resume_rate == 0) || (resume_cnt < resume_rate)) &&
(bit_test(suspend_node_bitmap, i) == 0) &&
(IS_NODE_ALLOCATED(node_ptr) ||
(node_ptr->last_idle > (now - idle_time)))) {
if (wake_node_bitmap == NULL) {
wake_node_bitmap =
bit_alloc(node_record_count);
}
wake_cnt++;
resume_cnt++;
resume_cnt_f++;
node_ptr->node_state &= (~NODE_STATE_POWER_SAVE);
node_ptr->node_state |= NODE_STATE_POWER_UP;
node_ptr->node_state |= NODE_STATE_NO_RESPOND;
bit_clear(power_node_bitmap, i);
bit_clear(avail_node_bitmap, i);
node_ptr->last_response = now + resume_timeout;
bit_set(wake_node_bitmap, i);
bit_set(resume_node_bitmap, i);
}
/* Suspend nodes as appropriate */
if (run_suspend &&
(susp_state == 0) &&
((suspend_rate == 0) || (suspend_cnt < suspend_rate)) &&
(IS_NODE_IDLE(node_ptr) || IS_NODE_DOWN(node_ptr)) &&
(node_ptr->sus_job_cnt == 0) &&
(!IS_NODE_COMPLETING(node_ptr)) &&
(!IS_NODE_POWER_UP(node_ptr)) &&
(node_ptr->last_idle != 0) &&
(node_ptr->last_idle < (now - idle_time)) &&
((exc_node_bitmap == NULL) ||
(bit_test(exc_node_bitmap, i) == 0))) {
if (sleep_node_bitmap == NULL) {
sleep_node_bitmap =
bit_alloc(node_record_count);
}
sleep_cnt++;
suspend_cnt++;
suspend_cnt_f++;
node_ptr->node_state |= NODE_STATE_POWER_SAVE;
node_ptr->node_state &= (~NODE_STATE_NO_RESPOND);
if (!IS_NODE_DOWN(node_ptr) &&
!IS_NODE_DRAIN(node_ptr))
bit_set(avail_node_bitmap, i);
bit_set(power_node_bitmap, i);
bit_set(sleep_node_bitmap, i);
bit_set(suspend_node_bitmap, i);
last_suspend = now;
}
}
if (((now - last_log) > 600) && (susp_total > 0)) {
info("Power save mode: %d nodes", susp_total);
last_log = now;
}
if (sleep_node_bitmap) {
char *nodes;
nodes = bitmap2node_name(sleep_node_bitmap);
if (nodes)
_do_suspend(nodes);
else
error("power_save: bitmap2nodename");
xfree(nodes);
FREE_NULL_BITMAP(sleep_node_bitmap);
/* last_node_update could be changed already by another thread!
last_node_update = now; */
}
if (wake_node_bitmap) {
char *nodes;
nodes = bitmap2node_name(wake_node_bitmap);
if (nodes)
_do_resume(nodes);
else
error("power_save: bitmap2nodename");
xfree(nodes);
FREE_NULL_BITMAP(wake_node_bitmap);
/* last_node_update could be changed already by another thread!
last_node_update = now; */
}
}
/**
* basil_inventory - Periodic node-state query via ALPS XML-RPC.
* This should be run immediately before each scheduling cycle.
* Returns non-SLURM_SUCCESS if
* - INVENTORY method failed (error)
* - no nodes are available (no point in scheduling)
* - orphaned ALPS reservation exists (wait until ALPS resynchronizes)
*/
extern int basil_inventory(void)
{
enum basil_version version = get_basil_version();
struct basil_inventory *inv;
struct basil_node *node;
struct basil_rsvn *rsvn;
int slurm_alps_mismatch = 0;
int rc = SLURM_SUCCESS;
int rel_rc;
time_t now = time(NULL);
static time_t slurm_alps_mismatch_time = (time_t) 0;
static bool logged_sync_timeout = false;
static time_t last_inv_run = 0;
if ((now - last_inv_run) < inv_interval)
return SLURM_SUCCESS;
last_inv_run = now;
inv = get_full_inventory(version);
if (inv == NULL) {
error("BASIL %s INVENTORY failed", bv_names_long[version]);
return SLURM_ERROR;
}
debug("BASIL %s INVENTORY: %d/%d batch nodes available",
bv_names_long[version], inv->batch_avail, inv->batch_total);
/* Avoid checking for inv->batch_avail here since if we are
gang scheduling returning an error for a full system is
probably the wrong thing to do. (the schedule() function
in the slurmctld will never run ;)).
*/
if (!inv->f->node_head || !inv->batch_total)
rc = ESLURM_REQUESTED_NODE_CONFIG_UNAVAILABLE;
for (node = inv->f->node_head; node; node = node->next) {
int node_inx;
struct node_record *node_ptr;
char *reason = NULL;
/* This will ignore interactive nodes when iterating through
* the apbasil inventory. If we don't do this, SLURM is
* unable to resolve the ID to a nidXXX name since it's not in
* the slurm.conf file. (Chris North)
*/
if (node->role == BNR_INTER)
continue;
node_ptr = _find_node_by_basil_id(node->node_id);
if (node_ptr == NULL) {
error("nid%05u (%s node in state %s) not in slurm.conf",
node->node_id, nam_noderole[node->role],
nam_nodestate[node->state]);
continue;
}
node_inx = node_ptr - node_record_table_ptr;
if (node_is_allocated(node) && !IS_NODE_ALLOCATED(node_ptr)) {
/*
* ALPS still hangs on to the node while SLURM considers
* it already unallocated. Possible causes are partition
* cleanup taking too long (can be 10sec ... minutes),
* and orphaned ALPS reservations (caught below).
*
* The converse case (SLURM hanging on to the node while
* ALPS has already freed it) happens frequently during
* job completion: select_g_job_fini() is called before
* make_node_comp(). Rely on SLURM logic for this case.
*/
slurm_alps_mismatch++;
}
if (node->state == BNS_DOWN) {
reason = "ALPS marked it DOWN";
} else if (node->state == BNS_UNAVAIL) {
reason = "node is UNAVAILABLE";
} else if (node->state == BNS_ROUTE) {
reason = "node does ROUTING";
} else if (node->state == BNS_SUSPECT) {
reason = "entered SUSPECT mode";
} else if (node->state == BNS_ADMINDOWN) {
reason = "node is ADMINDOWN";
} else if (node->state != BNS_UP) {
reason = "state not UP";
} else if (node->role != BNR_BATCH) {
reason = "mode not BATCH";
} else if (node->arch != BNA_XT) {
reason = "arch not XT/XE";
}
/* Base state entirely derives from ALPS */
if (reason) {
if (node_ptr->down_time == 0)
node_ptr->down_time = now;
if (IS_NODE_DOWN(node_ptr)) {
/* node still down */
} else if ((slurmctld_conf.slurmd_timeout == 0) ||
((now - node_ptr->down_time) <
slurmctld_conf.slurmd_timeout)) {
node_ptr->node_state |= NODE_STATE_NO_RESPOND;
bit_clear(avail_node_bitmap, node_inx);
} else {
xfree(node_ptr->reason);
info("MARKING %s DOWN (%s)",
node_ptr->name, reason);
/* set_node_down also kills any running jobs */
set_node_down_ptr(node_ptr, reason);
}
} else if (IS_NODE_DOWN(node_ptr)) {
xfree(node_ptr->reason);
node_ptr->down_time = 0;
info("MARKING %s UP", node_ptr->name);
/* Reset state, make_node_idle figures out the rest */
node_ptr->node_state &= NODE_STATE_FLAGS;
node_ptr->node_state &= (~NODE_STATE_NO_RESPOND);
node_ptr->node_state |= NODE_STATE_UNKNOWN;
make_node_idle(node_ptr, NULL);
if (!IS_NODE_DRAIN(node_ptr) &&
!IS_NODE_FAIL(node_ptr)) {
xfree(node_ptr->reason);
node_ptr->reason_time = 0;
node_ptr->reason_uid = NO_VAL;
clusteracct_storage_g_node_up(
acct_db_conn, node_ptr, now);
}
} else if (IS_NODE_NO_RESPOND(node_ptr)) {
node_ptr->node_state &= (~NODE_STATE_NO_RESPOND);
if (!IS_NODE_DRAIN(node_ptr) &&
!IS_NODE_FAIL(node_ptr)) {
bit_set(avail_node_bitmap, node_inx);
}
}
}
if (slurm_alps_mismatch)
debug("ALPS: %d node(s) still held", slurm_alps_mismatch);
/*
* Check that each ALPS reservation corresponds to a SLURM job.
* Purge orphaned reservations, which may result from stale or
* messed up system state, or are indicative of ALPS problems
* (stuck in pending cancel calls).
*/
for (rsvn = inv->f->rsvn_head; rsvn; rsvn = rsvn->next) {
ListIterator job_iter = list_iterator_create(job_list);
struct job_record *job_ptr;
uint32_t resv_id;
while ((job_ptr = (struct job_record *)list_next(job_iter))) {
if (_get_select_jobinfo(job_ptr->select_jobinfo->data,
SELECT_JOBDATA_RESV_ID,
&resv_id) == SLURM_SUCCESS
&& resv_id == rsvn->rsvn_id)
break;
}
list_iterator_destroy(job_iter);
/*
* Changed to ignore reservations for "UNKNOWN" batch
* ids (e.g. the interactive region) (Chris North)
*/
if ((job_ptr == NULL) && (xstrcmp(rsvn->batch_id, "UNKNOWN"))) {
error("orphaned ALPS reservation %u, trying to remove",
rsvn->rsvn_id);
rel_rc = basil_safe_release(rsvn->rsvn_id, inv);
if (rel_rc) {
error("ALPS reservation %u removal FAILED: %s",
rsvn->rsvn_id, basil_strerror(rel_rc));
} else {
debug("ALPS reservation %u removed",
rsvn->rsvn_id);
}
slurm_alps_mismatch = true;
}
}
free_inv(inv);
if (slurm_alps_mismatch) {
/* If SLURM and ALPS state are not in synchronization,
* do not schedule any more jobs until waiting at least
* SyncTimeout seconds. */
if (slurm_alps_mismatch_time == 0) {
slurm_alps_mismatch_time = now;
} else if (cray_conf->sync_timeout == 0) {
/* Wait indefinitely */
} else if (difftime(now, slurm_alps_mismatch_time) <
cray_conf->sync_timeout) {
return ESLURM_REQUESTED_NODE_CONFIG_UNAVAILABLE;
} else if (!logged_sync_timeout) {
error("Could not synchronize SLURM with ALPS for %u "
"seconds, proceeding with job scheduling",
cray_conf->sync_timeout);
logged_sync_timeout = true;
}
} else {
slurm_alps_mismatch_time = 0;
logged_sync_timeout = false;
}
return rc;
}
/*
* ping_nodes - check that all nodes and daemons are alive,
* get nodes in UNKNOWN state to register
*/
void ping_nodes (void)
{
static bool restart_flag = true; /* system just restarted */
static int offset = 0; /* mutex via node table write lock on entry */
static int max_reg_threads = 0; /* max node registration threads
* this can include DOWN nodes, so
* limit the number to avoid huge
* communication delays */
int i;
time_t now, still_live_time, node_dead_time;
static time_t last_ping_time = (time_t) 0;
hostlist_t down_hostlist = NULL;
char *host_str = NULL;
agent_arg_t *ping_agent_args = NULL;
agent_arg_t *reg_agent_args = NULL;
#ifdef HAVE_FRONT_END
front_end_record_t *front_end_ptr = NULL;
#else
struct node_record *node_ptr = NULL;
#endif
now = time (NULL);
ping_agent_args = xmalloc (sizeof (agent_arg_t));
ping_agent_args->msg_type = REQUEST_PING;
ping_agent_args->retry = 0;
ping_agent_args->hostlist = hostlist_create("");
reg_agent_args = xmalloc (sizeof (agent_arg_t));
reg_agent_args->msg_type = REQUEST_NODE_REGISTRATION_STATUS;
reg_agent_args->retry = 0;
reg_agent_args->hostlist = hostlist_create("");
/*
* If there are a large number of down nodes, the node ping
* can take a long time to complete:
* ping_time = down_nodes * agent_timeout / agent_parallelism
* ping_time = down_nodes * 10_seconds / 10
* ping_time = down_nodes (seconds)
* Because of this, we extend the SlurmdTimeout by the
* time needed to complete a ping of all nodes.
*/
if ((slurmctld_conf.slurmd_timeout == 0) ||
(last_ping_time == (time_t) 0)) {
node_dead_time = (time_t) 0;
} else {
node_dead_time = last_ping_time -
slurmctld_conf.slurmd_timeout;
}
still_live_time = now - (slurmctld_conf.slurmd_timeout / 3);
last_ping_time = now;
if (max_reg_threads == 0) {
max_reg_threads = MAX(slurm_get_tree_width(), 1);
}
offset += max_reg_threads;
if ((offset > node_record_count) &&
(offset >= (max_reg_threads * MAX_REG_FREQUENCY)))
offset = 0;
#ifdef HAVE_FRONT_END
for (i = 0, front_end_ptr = front_end_nodes;
i < front_end_node_cnt; i++, front_end_ptr++) {
if ((slurmctld_conf.slurmd_timeout == 0) &&
(!restart_flag) &&
(!IS_NODE_UNKNOWN(front_end_ptr)) &&
(!IS_NODE_NO_RESPOND(front_end_ptr)))
continue;
if ((front_end_ptr->last_response != (time_t) 0) &&
(front_end_ptr->last_response <= node_dead_time) &&
(!IS_NODE_DOWN(front_end_ptr))) {
if (down_hostlist)
(void) hostlist_push_host(down_hostlist,
front_end_ptr->name);
else {
down_hostlist =
hostlist_create(front_end_ptr->name);
if (down_hostlist == NULL)
fatal("hostlist_create: malloc error");
}
set_front_end_down(front_end_ptr, "Not responding");
front_end_ptr->not_responding = false;
continue;
}
if (restart_flag) {
front_end_ptr->last_response =
slurmctld_conf.last_update;
}
/* Request a node registration if its state is UNKNOWN or
* on a periodic basis (about every MAX_REG_FREQUENCY ping,
* this mechanism avoids an additional (per node) timer or
* counter and gets updated configuration information
* once in a while). We limit these requests since they
* can generate a flood of incoming RPCs. */
if (IS_NODE_UNKNOWN(front_end_ptr) || restart_flag ||
((i >= offset) && (i < (offset + max_reg_threads)))) {
hostlist_push(reg_agent_args->hostlist,
front_end_ptr->name);
reg_agent_args->node_count++;
continue;
}
if ((!IS_NODE_NO_RESPOND(front_end_ptr)) &&
(front_end_ptr->last_response >= still_live_time))
continue;
/* Do not keep pinging down nodes since this can induce
* huge delays in hierarchical communication fail-over */
if (IS_NODE_NO_RESPOND(front_end_ptr) &&
IS_NODE_DOWN(front_end_ptr))
continue;
hostlist_push(ping_agent_args->hostlist, front_end_ptr->name);
ping_agent_args->node_count++;
}
#else
for (i=0, node_ptr=node_record_table_ptr;
i<node_record_count; i++, node_ptr++) {
if (IS_NODE_FUTURE(node_ptr) || IS_NODE_POWER_SAVE(node_ptr))
continue;
if ((slurmctld_conf.slurmd_timeout == 0) &&
(!restart_flag) &&
(!IS_NODE_UNKNOWN(node_ptr)) &&
(!IS_NODE_NO_RESPOND(node_ptr)))
continue;
if ((node_ptr->last_response != (time_t) 0) &&
(node_ptr->last_response <= node_dead_time) &&
(!IS_NODE_DOWN(node_ptr))) {
if (down_hostlist)
(void) hostlist_push_host(down_hostlist,
node_ptr->name);
else {
down_hostlist =
hostlist_create(node_ptr->name);
if (down_hostlist == NULL)
fatal("hostlist_create: malloc error");
}
set_node_down_ptr(node_ptr, "Not responding");
node_ptr->not_responding = false; /* logged below */
continue;
}
if (restart_flag)
node_ptr->last_response = slurmctld_conf.last_update;
/* Request a node registration if its state is UNKNOWN or
* on a periodic basis (about every MAX_REG_FREQUENCY ping,
* this mechanism avoids an additional (per node) timer or
* counter and gets updated configuration information
* once in a while). We limit these requests since they
* can generate a flood of incoming RPCs. */
if (IS_NODE_UNKNOWN(node_ptr) || restart_flag ||
((i >= offset) && (i < (offset + max_reg_threads)))) {
hostlist_push(reg_agent_args->hostlist,
node_ptr->name);
reg_agent_args->node_count++;
continue;
}
if ((!IS_NODE_NO_RESPOND(node_ptr)) &&
(node_ptr->last_response >= still_live_time))
continue;
/* Do not keep pinging down nodes since this can induce
* huge delays in hierarchical communication fail-over */
if (IS_NODE_NO_RESPOND(node_ptr) && IS_NODE_DOWN(node_ptr))
continue;
hostlist_push(ping_agent_args->hostlist, node_ptr->name);
ping_agent_args->node_count++;
}
#endif
restart_flag = false;
if (ping_agent_args->node_count == 0) {
hostlist_destroy(ping_agent_args->hostlist);
xfree (ping_agent_args);
} else {
hostlist_uniq(ping_agent_args->hostlist);
host_str = hostlist_ranged_string_xmalloc(
ping_agent_args->hostlist);
debug("Spawning ping agent for %s", host_str);
xfree(host_str);
ping_begin();
agent_queue_request(ping_agent_args);
}
if (reg_agent_args->node_count == 0) {
hostlist_destroy(reg_agent_args->hostlist);
xfree (reg_agent_args);
} else {
hostlist_uniq(reg_agent_args->hostlist);
host_str = hostlist_ranged_string_xmalloc(
reg_agent_args->hostlist);
debug("Spawning registration agent for %s %d hosts",
host_str, reg_agent_args->node_count);
xfree(host_str);
ping_begin();
agent_queue_request(reg_agent_args);
}
if (down_hostlist) {
hostlist_uniq(down_hostlist);
host_str = hostlist_ranged_string_xmalloc(down_hostlist);
error("Nodes %s not responding, setting DOWN", host_str);
xfree(host_str);
hostlist_destroy(down_hostlist);
}
}
/**
* basil_geometry - Check node attributes, resolve (X,Y,Z) coordinates.
*
* Checks both SDB database and ALPS inventory for consistency. The inventory
* part is identical to basil_inventory(), with the difference of being called
* before valid bitmaps exist, from select_g_node_init().
* Its dependencies are:
* - it needs reset_job_bitmaps() in order to rebuild node_bitmap fields,
* - it relies on _sync_nodes_to_jobs() to
* o kill active jobs on nodes now marked DOWN,
* o reset node state to ALLOCATED if it has been marked IDLE here (which is
* an error case, since there is no longer an ALPS reservation for the job,
* this is caught by the subsequent basil_inventory()).
*/
extern int basil_geometry(struct node_record *node_ptr_array, int node_cnt)
{
struct node_record *node_ptr, *end = node_ptr_array + node_cnt;
enum basil_version version = get_basil_version();
struct basil_inventory *inv;
/* General mySQL */
MYSQL *handle;
MYSQL_STMT *stmt = NULL;
/* Input parameters */
unsigned int node_id;
/*
* Use a left outer join here since the attributes table may not be
* populated for a given nodeid (e.g. when the node has been disabled
* on the SMW via 'xtcli disable').
* The processor table has more authoritative information, if a nodeid
* is not listed there, it does not exist.
*/
const char query[] = "SELECT x_coord, y_coord, z_coord,"
" cab_position, cab_row, cage, slot, cpu,"
" LOG2(coremask+1), availmem, "
" processor_type "
"FROM processor LEFT JOIN attributes "
"ON processor_id = nodeid "
"WHERE processor_id = ? ";
const int PARAM_COUNT = 1; /* node id */
MYSQL_BIND params[PARAM_COUNT];
int x_coord, y_coord, z_coord;
int cab, row, cage, slot, cpu;
unsigned int node_cpus, node_mem;
char proc_type[BASIL_STRING_SHORT];
MYSQL_BIND bind_cols[COLUMN_COUNT];
my_bool is_null[COLUMN_COUNT];
my_bool is_error[COLUMN_COUNT];
int is_gemini, i;
time_t now = time(NULL);
memset(params, 0, sizeof(params));
params[0].buffer_type = MYSQL_TYPE_LONG;
params[0].is_unsigned = true;
params[0].is_null = (my_bool *)0;
params[0].buffer = (char *)&node_id;
memset(bind_cols, 0, sizeof(bind_cols));
for (i = 0; i < COLUMN_COUNT; i ++) {
bind_cols[i].is_null = &is_null[i];
bind_cols[i].error = &is_error[i];
if (i == COL_TYPE) {
bind_cols[i].buffer_type = MYSQL_TYPE_STRING;
bind_cols[i].buffer_length = sizeof(proc_type);
bind_cols[i].buffer = proc_type;
} else {
bind_cols[i].buffer_type = MYSQL_TYPE_LONG;
bind_cols[i].is_unsigned = (i >= COL_CORES);
}
}
bind_cols[COL_X].buffer = (char *)&x_coord;
bind_cols[COL_Y].buffer = (char *)&y_coord;
bind_cols[COL_Z].buffer = (char *)&z_coord;
bind_cols[COL_CAB].buffer = (char *)&cab;
bind_cols[COL_ROW].buffer = (char *)&row;
bind_cols[COL_CAGE].buffer = (char *)&cage;
bind_cols[COL_SLOT].buffer = (char *)&slot;
bind_cols[COL_CPU].buffer = (char *)&cpu;
bind_cols[COL_CORES].buffer = (char *)&node_cpus;
bind_cols[COL_MEMORY].buffer = (char *)&node_mem;
inv = get_full_inventory(version);
if (inv == NULL)
fatal("failed to get initial BASIL inventory");
info("BASIL %s initial INVENTORY: %d/%d batch nodes available",
bv_names_long[version], inv->batch_avail, inv->batch_total);
handle = cray_connect_sdb();
if (handle == NULL)
fatal("can not connect to XTAdmin database on the SDB");
is_gemini = cray_is_gemini_system(handle);
if (is_gemini < 0)
fatal("can not determine Cray XT/XE system type");
stmt = prepare_stmt(handle, query, params, PARAM_COUNT,
bind_cols, COLUMN_COUNT);
if (stmt == NULL)
fatal("can not prepare statement to resolve Cray coordinates");
for (node_ptr = node_record_table_ptr; node_ptr < end; node_ptr++) {
struct basil_node *node;
char *reason = NULL;
if ((node_ptr->name == NULL) ||
(sscanf(node_ptr->name, "nid%05u", &node_id) != 1)) {
error("can not read basil_node_id from %s",
node_ptr->name);
continue;
}
if (exec_stmt(stmt, query, bind_cols, COLUMN_COUNT) < 0)
fatal("can not resolve %s coordinates", node_ptr->name);
if (fetch_stmt(stmt) == 0) {
#if _DEBUG
info("proc_type:%s cpus:%u memory:%u",
proc_type, node_cpus, node_mem);
info("row:%u cage:%u slot:%u cpu:%u xyz:%u:%u:%u",
row, cage, slot, cpu, x_coord, y_coord, z_coord);
#endif
if (strcmp(proc_type, "compute") != 0) {
/*
* Switching a compute node to be a service node
* can not happen at runtime: requires a reboot.
*/
fatal("Node '%s' is a %s node. "
"Only compute nodes can appear in slurm.conf.",
node_ptr->name, proc_type);
} else if (is_null[COL_CORES] || is_null[COL_MEMORY]) {
/*
* This can happen if a node has been disabled
* on the SMW (using 'xtcli disable <nid>'). The
* node will still be listed in the 'processor'
* table, but have no 'attributes' entry (NULL
* values for CPUs/memory). Also, the node will
* be invisible to ALPS, which is why we need to
* set it down here already.
*/
node_cpus = node_mem = 0;
reason = "node data unknown - disabled on SMW?";
} else if (is_null[COL_X] || is_null[COL_Y]
|| is_null[COL_Z]) {
/*
* Similar case to the one above, observed when
* a blade has been removed. Node will not
* likely show up in ALPS.
*/
x_coord = y_coord = z_coord = 0;
reason = "unknown coordinates - hardware failure?";
} else if (node_cpus < node_ptr->config_ptr->cpus) {
/*
* FIXME: Might reconsider this policy.
*
* FastSchedule is ignored here, it requires the
* slurm.conf to be consistent with hardware.
*
* Assumption is that CPU/Memory do not change
* at runtime (Cray has no hot-swappable parts).
*
* Hence checking it in basil_inventory() would
* mean a lot of runtime overhead.
*/
fatal("slurm.conf: node %s has only Procs=%d",
node_ptr->name, node_cpus);
} else if (node_mem < node_ptr->config_ptr->real_memory) {
fatal("slurm.conf: node %s has RealMemory=%d",
node_ptr->name, node_mem);
}
} else if (is_gemini) {
fatal("Non-existing Gemini node '%s' in slurm.conf",
node_ptr->name);
} else {
fatal("Non-existing SeaStar node '%s' in slurm.conf",
node_ptr->name);
}
if (!is_gemini) {
/*
* SeaStar: each node has unique coordinates
*/
if (node_ptr->arch == NULL)
node_ptr->arch = xstrdup("XT");
} else {
/*
* Gemini: each 2 nodes share the same network
* interface (i.e., nodes 0/1 and 2/3 each have
* the same coordinates).
*/
if (node_ptr->arch == NULL)
node_ptr->arch = xstrdup("XE");
}
xfree(node_ptr->node_hostname);
xfree(node_ptr->comm_name);
/*
* Convention: since we are using SLURM in frontend-mode,
* we use Node{Addr,HostName} as follows.
*
* NodeAddr: <X><Y><Z> coordinates in base-36 encoding
*
* NodeHostName: c#-#c#s#n# using the NID convention
* <cabinet>-<row><chassis><slot><node>
* - each cabinet can accommodate 3 chassis (c1..c3)
* - each chassis has 8 slots (s0..s7)
* - each slot contains 2 or 4 nodes (n0..n3)
* o either 2 service nodes (n0/n3)
* o or 4 compute nodes (n0..n3)
* o or 2 gemini chips (g0/g1 serving n0..n3)
*
* Example: c0-0c1s0n1
* - c0- = cabinet 0
* - 0 = row 0
* - c1 = chassis 1
* - s0 = slot 0
* - n1 = node 1
*/
node_ptr->node_hostname = xstrdup_printf("c%u-%uc%us%un%u", cab,
row, cage, slot, cpu);
node_ptr->comm_name = xstrdup_printf("%c%c%c",
_enc_coord(x_coord),
_enc_coord(y_coord),
_enc_coord(z_coord));
dim_size[0] = MAX(dim_size[0], (x_coord - 1));
dim_size[1] = MAX(dim_size[1], (y_coord - 1));
dim_size[2] = MAX(dim_size[2], (z_coord - 1));
#if _DEBUG
info("%s %s %s cpus=%u, mem=%u reason=%s", node_ptr->name,
node_ptr->node_hostname, node_ptr->comm_name,
node_cpus, node_mem, reason);
#endif
/*
* Check the current state reported by ALPS inventory, unless it
* is already evident that the node has some other problem.
*/
if (reason == NULL) {
for (node = inv->f->node_head; node; node = node->next)
if (node->node_id == node_id)
break;
if (node == NULL) {
reason = "not visible to ALPS - check hardware";
} else if (node->state == BNS_DOWN) {
reason = "ALPS marked it DOWN";
} else if (node->state == BNS_UNAVAIL) {
reason = "node is UNAVAILABLE";
} else if (node->state == BNS_ROUTE) {
reason = "node does ROUTING";
} else if (node->state == BNS_SUSPECT) {
reason = "entered SUSPECT mode";
} else if (node->state == BNS_ADMINDOWN) {
reason = "node is ADMINDOWN";
} else if (node->state != BNS_UP) {
reason = "state not UP";
} else if (node->role != BNR_BATCH) {
reason = "mode not BATCH";
} else if (node->arch != BNA_XT) {
reason = "arch not XT/XE";
}
}
/* Base state entirely derives from ALPS
* NOTE: The node bitmaps are not defined when this code is
* initially executed. */
node_ptr->node_state &= NODE_STATE_FLAGS;
if (reason) {
if (node_ptr->down_time == 0)
node_ptr->down_time = now;
if (IS_NODE_DOWN(node_ptr)) {
/* node still down */
debug("Initial DOWN node %s - %s",
node_ptr->name, node_ptr->reason);
} else if (slurmctld_conf.slurmd_timeout &&
((now - node_ptr->down_time) <
slurmctld_conf.slurmd_timeout)) {
node_ptr->node_state |= NODE_STATE_NO_RESPOND;
} else {
info("Initial DOWN node %s - %s",
node_ptr->name, reason);
node_ptr->reason = xstrdup(reason);
/* Node state flags preserved above */
node_ptr->node_state |= NODE_STATE_DOWN;
clusteracct_storage_g_node_down(acct_db_conn,
node_ptr,
now, NULL,
slurm_get_slurm_user_id());
}
} else {
bool node_up_flag = IS_NODE_DOWN(node_ptr) &&
!IS_NODE_DRAIN(node_ptr) &&
!IS_NODE_FAIL(node_ptr);
node_ptr->down_time = 0;
if (node_is_allocated(node))
node_ptr->node_state |= NODE_STATE_ALLOCATED;
else
node_ptr->node_state |= NODE_STATE_IDLE;
node_ptr->node_state &= (~NODE_STATE_NO_RESPOND);
xfree(node_ptr->reason);
if (node_up_flag) {
info("ALPS returned node %s to service",
node_ptr->name);
clusteracct_storage_g_node_up(acct_db_conn,
node_ptr, now);
}
}
free_stmt_result(stmt);
}
if (stmt_close(stmt))
error("error closing statement: %s", mysql_stmt_error(stmt));
cray_close_sdb(handle);
free_inv(inv);
return SLURM_SUCCESS;
}