static int _attempt_backfill(void)
{
DEF_TIMERS;
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
time_t now, sched_start, later_start, start_res, resv_end;
node_space_map_t *node_space;
struct timeval bf_time1, bf_time2;
int sched_timeout = 2, yield_sleep = 1;
int rc = 0;
int job_test_count = 0;
uint32_t *uid = NULL, nuser = 0;
uint16_t *njobs = NULL;
bool already_counted;
uint32_t reject_array_job_id = 0;
#ifdef HAVE_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
START_TIMER;
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
END_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: ALPS inventory completed, %s", TIME_STR);
/* The Basil inventory can take a long time to complete. Process
* pending RPCs before starting the backfill scheduling logic */
_yield_locks(1);
#endif
START_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: beginning");
sched_start = now = time(NULL);
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true);
if (list_count(job_queue) == 0) {
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
gettimeofday(&bf_time1, NULL);
slurmctld_diag_stats.bf_queue_len = list_count(job_queue);
slurmctld_diag_stats.bf_queue_len_sum += slurmctld_diag_stats.
bf_queue_len;
slurmctld_diag_stats.bf_last_depth = 0;
slurmctld_diag_stats.bf_last_depth_try = 0;
slurmctld_diag_stats.bf_when_last_cycle = now;
bf_last_yields = 0;
slurmctld_diag_stats.bf_active = 1;
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt + 3));
node_space[0].begin_time = sched_start;
node_space[0].end_time = sched_start + backfill_window;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
if (max_backfill_job_per_user) {
uid = xmalloc(BF_MAX_USERS * sizeof(uint32_t));
njobs = xmalloc(BF_MAX_USERS * sizeof(uint16_t));
}
while ((job_queue_rec = (job_queue_rec_t *)
list_pop_bottom(job_queue, sort_job_queue2))) {
job_ptr = job_queue_rec->job_ptr;
orig_time_limit = job_ptr->time_limit;
if ((time(NULL) - sched_start) >= sched_timeout) {
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %d jobs, %s",
job_test_count, TIME_STR);
}
if (_yield_locks(yield_sleep) && !backfill_continue) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing %d "
"jobs", job_test_count);
}
rc = 1;
break;
}
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
job_test_count = 0;
START_TIMER;
}
part_ptr = job_queue_rec->part_ptr;
job_test_count++;
xfree(job_queue_rec);
if (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
if (!avail_front_end(job_ptr))
continue; /* No available frontend for this job */
if (job_ptr->array_task_id != (uint16_t) NO_VAL) {
if (reject_array_job_id == job_ptr->array_job_id)
continue; /* already rejected array element */
/* assume reject whole array for now, clear if OK */
reject_array_job_id = job_ptr->array_job_id;
}
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill test for job %u", job_ptr->job_id);
slurmctld_diag_stats.bf_last_depth++;
already_counted = false;
if (max_backfill_job_per_user) {
for (j = 0; j < nuser; j++) {
if (job_ptr->user_id == uid[j]) {
njobs[j]++;
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: user %u: "
"#jobs %u",
uid[j], njobs[j]);
break;
}
}
if (j == nuser) { /* user not found */
if (nuser < BF_MAX_USERS) {
uid[j] = job_ptr->user_id;
njobs[j] = 1;
nuser++;
} else {
error("backfill: too many users in "
"queue. Consider increasing "
"BF_MAX_USERS");
}
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug2("backfill: found new user %u. "
"Total #users now %u",
job_ptr->user_id, nuser);
} else {
if (njobs[j] > max_backfill_job_per_user) {
/* skip job */
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: have already "
"checked %u jobs for "
"user %u; skipping "
"job %u",
max_backfill_job_per_user,
job_ptr->user_id,
job_ptr->job_id);
continue;
}
}
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL))
continue;
if ((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)
continue;
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS))
continue;
/* Determine minimum and maximum node counts */
min_nodes = MAX(job_ptr->details->min_nodes,
part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partition's max_nodes */
continue;
}
/* Determine job's expected completion time */
if (job_ptr->time_limit == NO_VAL) {
if (part_ptr->max_time == INFINITE)
time_limit = 365 * 24 * 60; /* one year */
else
time_limit = part_ptr->max_time;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_ptr->max_time);
}
comp_time_limit = time_limit;
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE) &&
slurm_get_preempt_mode())
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
if ((time(NULL) - sched_start) >= sched_timeout) {
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks 2"
"after testing %d jobs, %s",
job_test_count, TIME_STR);
}
if (_yield_locks(yield_sleep) && !backfill_continue) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing %d "
"jobs", job_test_count);
}
rc = 1;
break;
}
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
job_test_count = 1;
START_TIMER;
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap,
&exc_core_bitmap);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
resv_end = find_resv_end(start_res);
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
if ((resv_end++) &&
((later_start == 0) || (resv_end < later_start))) {
later_start = resv_end;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap,
job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
/* Job can not start until too far in the future */
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = sched_start + backfill_window;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
if (!already_counted) {
slurmctld_diag_stats.bf_last_depth_try++;
already_counted = true;
}
j = _try_sched(job_ptr, &avail_bitmap, min_nodes, max_nodes,
req_nodes, exc_core_bitmap);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) {
uint32_t save_time_limit = job_ptr->time_limit;
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)) {
if (orig_time_limit == NO_VAL)
job_ptr->time_limit = comp_time_limit;
else
job_ptr->time_limit = orig_time_limit;
job_ptr->end_time = job_ptr->start_time +
(job_ptr->time_limit * 60);
} else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
job_ptr->time_limit = comp_time_limit;
job_ptr->end_time = job_ptr->start_time +
(comp_time_limit * 60);
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
} else {
job_ptr->time_limit = orig_time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
/* Planned to start job, but something bad
* happended. */
job_ptr->start_time = 0;
break;
} else {
/* Started this job, move to next one */
reject_array_job_id = 0;
/* Update the database if job time limit
* changed and move to next job */
if (save_time_limit != job_ptr->time_limit)
jobacct_storage_g_job_start(acct_db_conn,
job_ptr);
continue;
}
} else
job_ptr->time_limit = orig_time_limit;
if (later_start && (job_ptr->start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
/* Already have too many jobs to deal with */
break;
}
end_reserve = job_ptr->start_time + (time_limit * 60);
if (_test_resv_overlap(node_space, avail_bitmap,
job_ptr->start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
reject_array_job_id = 0;
bit_not(avail_bitmap);
_add_reservation(job_ptr->start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
}
xfree(uid);
xfree(njobs);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
gettimeofday(&bf_time2, NULL);
_do_diag_stats(&bf_time1, &bf_time2, yield_sleep);
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed testing %d jobs, %s",
job_test_count, TIME_STR);
}
return rc;
}
static char * _will_run_test2(uint32_t jobid, time_t start_time,
char *node_list,
uint32_t *preemptee, int preemptee_cnt,
int *err_code, char **err_msg)
{
struct job_record *job_ptr = NULL, *pre_ptr;
struct part_record *part_ptr;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
time_t start_res;
uint32_t min_nodes, max_nodes, req_nodes;
List preemptee_candidates = NULL, preempted_jobs = NULL;
time_t orig_start_time;
char *reply_msg = NULL;
int i, rc;
bool resv_overlap = false;
xassert(node_list);
debug2("wiki2: will_run2 job_id=%u start_time=%u node_list=%s",
jobid, (uint32_t)start_time, node_list);
job_ptr = find_job_record(jobid);
if (job_ptr == NULL) {
*err_code = -700;
*err_msg = "No such job";
error("wiki: Failed to find job %u", jobid);
return NULL;
}
if ((job_ptr->details == NULL) || (!IS_JOB_PENDING(job_ptr))) {
*err_code = -700;
*err_msg = "WillRun not applicable to non-pending job";
error("wiki: WillRun on non-pending job %u", jobid);
return NULL;
}
part_ptr = job_ptr->part_ptr;
if (part_ptr == NULL) {
*err_code = -700;
*err_msg = "Job lacks a partition";
error("wiki: Job %u lacks a partition", jobid);
return NULL;
}
if (node_name2bitmap(node_list, false, &avail_bitmap) != 0) {
*err_code = -700;
*err_msg = "Invalid available nodes value";
error("wiki: Attempt to set invalid available node "
"list for job %u, %s", jobid, node_list);
return NULL;
}
/* Enforce reservation: access control, time and nodes */
start_res = start_time;
rc = job_test_resv(job_ptr, &start_res, true, &resv_bitmap,
&exc_core_bitmap, &resv_overlap);
if (rc != SLURM_SUCCESS) {
*err_code = -730;
*err_msg = "Job denied access to reservation";
error("wiki: reservation access denied for job %u", jobid);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
bit_and(avail_bitmap, resv_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
/* Only consider nodes that are not DOWN or DRAINED */
bit_and(avail_bitmap, avail_node_bitmap);
/* Consider only nodes in this job's partition */
if (part_ptr->node_bitmap)
bit_and(avail_bitmap, part_ptr->node_bitmap);
else {
*err_code = -730;
*err_msg = "Job's partition has no nodes";
error("wiki: no nodes in partition %s for job %u",
part_ptr->name, jobid);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (job_req_node_filter(job_ptr, avail_bitmap) != SLURM_SUCCESS) {
/* Job probably has invalid feature list */
*err_code = -730;
*err_msg = "Job's required features not available "
"on selected nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap, job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
if ((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) {
*err_code = -730;
*err_msg = "Job's required nodes not available";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
min_nodes = MAX(job_ptr->details->min_nodes, part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partitions max_nodes */
*err_code = -730;
*err_msg = "Job's min_nodes > max_nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (preemptee_cnt) {
preemptee_candidates = list_create(NULL);
for (i=0; i<preemptee_cnt; i++) {
if ((pre_ptr = find_job_record(preemptee[i])))
list_append(preemptee_candidates, pre_ptr);
}
}
orig_start_time = job_ptr->start_time;
rc = select_g_job_test(job_ptr, avail_bitmap, min_nodes, max_nodes,
req_nodes, SELECT_MODE_WILL_RUN,
preemptee_candidates, &preempted_jobs,
exc_core_bitmap);
FREE_NULL_LIST(preemptee_candidates);
if (rc == SLURM_SUCCESS) {
char *hostlist, *sep, tmp_str[128];
uint32_t pre_cnt = 0, proc_cnt = 0;
#ifdef HAVE_BG
select_g_select_jobinfo_get(job_ptr->select_jobinfo,
SELECT_JOBDATA_NODE_CNT, &proc_cnt);
#else
proc_cnt = job_ptr->total_cpus;
#endif
snprintf(tmp_str, sizeof(tmp_str),
"STARTINFO=%u TASKS=%u STARTTIME=%u NODES=",
job_ptr->job_id, proc_cnt,
(uint32_t) job_ptr->start_time);
xstrcat(reply_msg, tmp_str);
hostlist = bitmap2node_name(avail_bitmap);
xstrcat(reply_msg, hostlist);
xfree(hostlist);
if (preempted_jobs) {
while ((pre_ptr = list_pop(preempted_jobs))) {
if (pre_cnt++)
sep = ",";
else
sep = " PREEMPT=";
snprintf(tmp_str, sizeof(tmp_str), "%s%u",
sep, pre_ptr->job_id);
xstrcat(reply_msg, tmp_str);
}
FREE_NULL_LIST(preempted_jobs);
}
} else {
xstrcat(reply_msg, "Jobs not runable on selected nodes");
error("wiki: jobs not runnable on nodes");
}
/* Restore pending job's expected start time */
job_ptr->start_time = orig_start_time;
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return reply_msg;
}
/*
* Attempt to start a job
* jobid (IN) - job id
* task_cnt (IN) - total count of tasks to start
* hostlist (IN) - SLURM hostlist expression with no repeated hostnames
* tasklist (IN/OUT) - comma separated list of hosts with tasks to be started,
* list hostname once per task to start
* comment_ptr (IN) - new comment field for the job or NULL for no change
* err_code (OUT) - Moab error code
* err_msg (OUT) - Moab error message
*/
static int _start_job(uint32_t jobid, int task_cnt, char *hostlist,
char *tasklist, char *comment_ptr,
int *err_code, char **err_msg)
{
int rc = 0, old_task_cnt = 1;
struct job_record *job_ptr;
/* Write lock on job info, read lock on node info */
slurmctld_lock_t job_write_lock = {
NO_LOCK, WRITE_LOCK, READ_LOCK, NO_LOCK };
char *new_node_list = NULL;
static char tmp_msg[128];
bitstr_t *new_bitmap = (bitstr_t *) NULL;
bitstr_t *save_req_bitmap = (bitstr_t *) NULL;
bitoff_t i, bsize;
int ll; /* layout info index */
char *node_name, *node_idx, *node_cur, *save_req_nodes = NULL;
size_t node_name_len;
static uint32_t cr_test = 0, cr_enabled = 0;
if (cr_test == 0) {
select_g_get_info_from_plugin(SELECT_CR_PLUGIN, NULL,
&cr_enabled);
cr_test = 1;
}
lock_slurmctld(job_write_lock);
job_ptr = find_job_record(jobid);
if (job_ptr == NULL) {
*err_code = -700;
*err_msg = "No such job";
error("wiki: Failed to find job %u", jobid);
rc = -1;
goto fini;
}
if ((job_ptr->details == NULL) || (!IS_JOB_PENDING(job_ptr))) {
*err_code = -700;
*err_msg = "Job not pending, can't start";
error("wiki: Attempt to start job %u in state %s",
jobid, job_state_string(job_ptr->job_state));
rc = -1;
goto fini;
}
if (comment_ptr) {
char *reserved = strstr(comment_ptr, "RESERVED:");
if (reserved) {
reserved += 9;
job_ptr->details->reserved_resources =
strtol(reserved, NULL, 10);
}
xfree(job_ptr->comment);
job_ptr->comment = xstrdup(comment_ptr);
}
if (task_cnt) {
new_node_list = xstrdup(hostlist);
if (node_name2bitmap(new_node_list, false, &new_bitmap) != 0) {
*err_code = -700;
*err_msg = "Invalid TASKLIST";
error("wiki: Attempt to set invalid node list for "
"job %u, %s",
jobid, hostlist);
xfree(new_node_list);
rc = -1;
goto fini;
}
if (!bit_super_set(new_bitmap, avail_node_bitmap)) {
/* Selected node is UP and not responding
* or it just went DOWN */
*err_code = -700;
*err_msg = "TASKLIST includes non-responsive node";
error("wiki: Attempt to use non-responsive nodes for "
"job %u, %s",
jobid, hostlist);
xfree(new_node_list);
FREE_NULL_BITMAP(new_bitmap);
rc = -1;
goto fini;
}
/* User excluded node list incompatible with Wiki
* Exclude all nodes not explicitly requested */
FREE_NULL_BITMAP(job_ptr->details->exc_node_bitmap);
job_ptr->details->exc_node_bitmap = bit_copy(new_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
/* Build layout information from tasklist (assuming that Moab
* sends a non-bracketed list of nodes, repeated as many times
* as cpus should be used per node); at this point, node names
* are comma-separated. This is _not_ a fast algorithm as it
* performs many string compares. */
xfree(job_ptr->details->req_node_layout);
if (task_cnt && cr_enabled) {
uint16_t cpus_per_task = MAX(1, job_ptr->details->cpus_per_task);
job_ptr->details->req_node_layout = (uint16_t *)
xmalloc(bit_set_count(new_bitmap) * sizeof(uint16_t));
bsize = bit_size(new_bitmap);
for (i = 0, ll = -1; i < bsize; i++) {
if (!bit_test(new_bitmap, i))
continue;
ll++;
node_name = node_record_table_ptr[i].name;
node_name_len = strlen(node_name);
if (node_name_len == 0)
continue;
node_cur = tasklist;
while (*node_cur) {
if ((node_idx = strstr(node_cur, node_name))) {
if ((node_idx[node_name_len] == ',') ||
(node_idx[node_name_len] == '\0')) {
job_ptr->details->
req_node_layout[ll] +=
cpus_per_task;
}
node_cur = strchr(node_idx, ',');
if (node_cur)
continue;
}
break;
}
}
}
/* save and update job state to start now */
save_req_nodes = job_ptr->details->req_nodes;
job_ptr->details->req_nodes = new_node_list;
save_req_bitmap = job_ptr->details->req_node_bitmap;
job_ptr->details->req_node_bitmap = new_bitmap;
old_task_cnt = job_ptr->details->min_cpus;
job_ptr->details->min_cpus = MAX(task_cnt, old_task_cnt);
job_ptr->priority = 100000000;
fini: unlock_slurmctld(job_write_lock);
if (rc)
return rc;
/* No errors so far */
(void) schedule(INFINITE); /* provides own locking */
/* Check to insure the job was actually started */
lock_slurmctld(job_write_lock);
if (job_ptr->job_id != jobid)
job_ptr = find_job_record(jobid);
if (job_ptr && (job_ptr->job_id == jobid) &&
(!IS_JOB_RUNNING(job_ptr))) {
uint16_t wait_reason = 0;
char *wait_string;
if (IS_JOB_FAILED(job_ptr))
wait_string = "Invalid request, job aborted";
else {
wait_reason = job_ptr->state_reason;
if (wait_reason == WAIT_HELD) {
/* some job is completing, slurmctld did
* not even try to schedule this job */
wait_reason = WAIT_RESOURCES;
}
wait_string = job_reason_string(wait_reason);
job_ptr->state_reason = WAIT_HELD;
xfree(job_ptr->state_desc);
}
*err_code = -910 - wait_reason;
snprintf(tmp_msg, sizeof(tmp_msg),
"Could not start job %u(%s): %s",
jobid, new_node_list, wait_string);
*err_msg = tmp_msg;
error("wiki: %s", tmp_msg);
/* restore some of job state */
job_ptr->priority = 0;
job_ptr->details->min_cpus = old_task_cnt;
rc = -1;
}
if (job_ptr && (job_ptr->job_id == jobid) && job_ptr->details) {
/* Restore required node list in case job requeued */
xfree(job_ptr->details->req_nodes);
job_ptr->details->req_nodes = save_req_nodes;
FREE_NULL_BITMAP(job_ptr->details->req_node_bitmap);
job_ptr->details->req_node_bitmap = save_req_bitmap;
FREE_NULL_BITMAP(job_ptr->details->exc_node_bitmap);
xfree(job_ptr->details->req_node_layout);
} else {
error("wiki: start_job(%u) job missing", jobid);
xfree(save_req_nodes);
FREE_NULL_BITMAP(save_req_bitmap);
}
unlock_slurmctld(job_write_lock);
schedule_node_save(); /* provides own locking */
schedule_job_save(); /* provides own locking */
return rc;
}
int
main(int argc, char *argv[])
{
note("Testing static decl");
{
bitstr_t bit_decl(bs, 65);
/*bitstr_t *bsp = bs;*/
bit_set(bs,9);
bit_set(bs,14);
TEST(bit_test(bs,9), "bit 9 set");
TEST(!bit_test(bs,12), "bit 12 not set");
TEST(bit_test(bs,14), "bit 14 set" );
/*bit_free(bsp);*/ /* triggers TEST in bit_free - OK */
}
note("Testing basic vixie functions");
{
bitstr_t *bs = bit_alloc(16), *bs2;
/*bit_set(bs, 42);*/ /* triggers TEST in bit_set - OK */
bit_set(bs,9);
bit_set(bs,14);
TEST(bit_test(bs,9), "bit 9 set");
TEST(!bit_test(bs,12), "bit 12 not set" );
TEST(bit_test(bs,14), "bit 14 set");
bs2 = bit_copy(bs);
bit_fill_gaps(bs2);
TEST(bit_ffs(bs2) == 9, "first bit set = 9 ");
TEST(bit_fls(bs2) == 14, "last bit set = 14");
TEST(bit_set_count(bs2) == 6, "bitstring");
TEST(bit_test(bs2,12), "bitstring");
TEST(bit_super_set(bs,bs2) == 1, "bitstring");
TEST(bit_super_set(bs2,bs) == 0, "bitstring");
bit_clear(bs,14);
TEST(!bit_test(bs,14), "bitstring");
bit_nclear(bs,9,14);
TEST(!bit_test(bs,9), "bitstring");
TEST(!bit_test(bs,12), "bitstring");
TEST(!bit_test(bs,14), "bitstring");
bit_nset(bs,9,14);
TEST(bit_test(bs,9), "bitstring");
TEST(bit_test(bs,12), "bitstring");
TEST(bit_test(bs,14), "bitstring");
TEST(bit_ffs(bs) == 9, "ffs");
TEST(bit_ffc(bs) == 0, "ffc");
bit_nset(bs,0,8);
TEST(bit_ffc(bs) == 15, "ffc");
bit_free(bs);
/*bit_set(bs,9); */ /* triggers TEST in bit_set - OK */
}
note("Testing and/or/not");
{
bitstr_t *bs1 = bit_alloc(128);
bitstr_t *bs2 = bit_alloc(128);
bit_set(bs1, 100);
bit_set(bs1, 104);
bit_set(bs2, 100);
bit_and(bs1, bs2);
TEST(bit_test(bs1, 100), "and");
TEST(!bit_test(bs1, 104), "and");
bit_set(bs2, 110);
bit_set(bs2, 111);
bit_set(bs2, 112);
bit_or(bs1, bs2);
TEST(bit_test(bs1, 100), "or");
TEST(bit_test(bs1, 110), "or");
TEST(bit_test(bs1, 111), "or");
TEST(bit_test(bs1, 112), "or");
bit_not(bs1);
TEST(!bit_test(bs1, 100), "not");
TEST(bit_test(bs1, 12), "not");
bit_free(bs1);
bit_free(bs2);
}
note("testing bit selection");
{
bitstr_t *bs1 = bit_alloc(128), *bs2;
bit_set(bs1, 21);
bit_set(bs1, 100);
bit_fill_gaps(bs1);
bs2 = bit_pick_cnt(bs1,20);
if (bs2) {
TEST(bit_set_count(bs2) == 20, "pick");
TEST(bit_ffs(bs2) == 21, "pick");
TEST(bit_fls(bs2) == 40, "pick");
bit_free(bs2);
}
else
TEST(0, "alloc fail");
bit_free(bs1);
}
note("Testing realloc");
{
bitstr_t *bs = bit_alloc(1);
TEST(bit_ffs(bs) == -1, "bitstring");
bit_set(bs,0);
/*bit_set(bs, 1000);*/ /* triggers TEST in bit_set - OK */
bs = bit_realloc(bs,1048576);
bit_set(bs,1000);
bit_set(bs,1048575);
TEST(bit_test(bs, 0), "bitstring");
TEST(bit_test(bs, 1000), "bitstring");
TEST(bit_test(bs, 1048575), "bitstring");
TEST(bit_set_count(bs) == 3, "bitstring");
bit_clear(bs,0);
bit_clear(bs,1000);
TEST(bit_set_count(bs) == 1, "bitstring");
TEST(bit_ffs(bs) == 1048575, "bitstring");
bit_free(bs);
}
note("Testing bit_fmt");
{
char tmpstr[1024];
bitstr_t *bs = bit_alloc(1024);
TEST(!strcmp(bit_fmt(tmpstr,sizeof(tmpstr),bs), ""), "bitstring");
bit_set(bs,42);
TEST(!strcmp(bit_fmt(tmpstr,sizeof(tmpstr),bs), "42"), "bitstring");
bit_set(bs,102);
TEST(!strcmp(bit_fmt(tmpstr,sizeof(tmpstr),bs), "42,102"), "bitstring");
bit_nset(bs,9,14);
TEST(!strcmp(bit_fmt(tmpstr,sizeof(tmpstr), bs),
"9-14,42,102"), "bitstring");
}
note("Testing bit_nffc/bit_nffs");
{
bitstr_t *bs = bit_alloc(1024);
bit_set(bs, 2);
bit_set(bs, 6);
bit_set(bs, 7);
bit_nset(bs,12,1018);
TEST(bit_nffc(bs, 2) == 0, "bitstring");
TEST(bit_nffc(bs, 3) == 3, "bitstring");
TEST(bit_nffc(bs, 4) == 8, "bitstring");
TEST(bit_nffc(bs, 5) == 1019, "bitstring");
TEST(bit_nffc(bs, 6) == -1, "bitstring");
TEST(bit_nffs(bs, 1) == 2, "bitstring");
TEST(bit_nffs(bs, 2) == 6, "bitstring");
TEST(bit_nffs(bs, 100) == 12, "bitstring");
TEST(bit_nffs(bs, 1023) == -1, "bitstring");
bit_free(bs);
}
note("Testing bit_unfmt");
{
bitstr_t *bs = bit_alloc(1024);
bitstr_t *bs2 = bit_alloc(1024);
char tmpstr[4096];
bit_set(bs,1);
bit_set(bs,3);
bit_set(bs,30);
bit_nset(bs,42,64);
bit_nset(bs,97,1000);
bit_fmt(tmpstr, sizeof(tmpstr), bs);
TEST(bit_unfmt(bs2, tmpstr) != -1, "bitstring");
TEST(bit_equal(bs, bs2), "bitstring");
}
totals();
return failed;
}
static char * _will_run_test(uint32_t jobid, time_t start_time,
char *node_list, int *err_code, char **err_msg)
{
struct job_record *job_ptr = NULL;
struct part_record *part_ptr;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
char *hostlist, *reply_msg = NULL;
uint32_t min_nodes, max_nodes, req_nodes;
int rc;
time_t start_res, orig_start_time;
List preemptee_candidates;
debug2("wiki2: will_run job_id=%u start_time=%u node_list=%s",
jobid, (uint32_t)start_time, node_list);
job_ptr = find_job_record(jobid);
if (job_ptr == NULL) {
*err_code = -700;
*err_msg = "No such job";
error("wiki: Failed to find job %u", jobid);
return NULL;
}
if ((job_ptr->details == NULL) || (!IS_JOB_PENDING(job_ptr))) {
*err_code = -700;
*err_msg = "WillRun not applicable to non-pending job";
error("wiki: WillRun on non-pending job %u", jobid);
return NULL;
}
part_ptr = job_ptr->part_ptr;
if (part_ptr == NULL) {
*err_code = -700;
*err_msg = "Job lacks a partition";
error("wiki: Job %u lacks a partition", jobid);
return NULL;
}
if ((node_list == NULL) || (node_list[0] == '\0')) {
/* assume all nodes available to job for testing */
avail_bitmap = bit_copy(avail_node_bitmap);
} else if (node_name2bitmap(node_list, false, &avail_bitmap) != 0) {
*err_code = -700;
*err_msg = "Invalid available nodes value";
error("wiki: Attempt to set invalid available node "
"list for job %u, %s", jobid, node_list);
return NULL;
}
/* Enforce reservation: access control, time and nodes */
start_res = start_time;
rc = job_test_resv(job_ptr, &start_res, true, &resv_bitmap,
&exc_core_bitmap);
if (rc != SLURM_SUCCESS) {
*err_code = -730;
*err_msg = "Job denied access to reservation";
error("wiki: reservation access denied for job %u", jobid);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
start_time = MAX(start_time, start_res);
bit_and(avail_bitmap, resv_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
/* Only consider nodes that are not DOWN or DRAINED */
bit_and(avail_bitmap, avail_node_bitmap);
/* Consider only nodes in this job's partition */
if (part_ptr->node_bitmap)
bit_and(avail_bitmap, part_ptr->node_bitmap);
else {
*err_code = -730;
*err_msg = "Job's partition has no nodes";
error("wiki: no nodes in partition %s for job %u",
part_ptr->name, jobid);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
if (job_req_node_filter(job_ptr, avail_bitmap) != SLURM_SUCCESS) {
/* Job probably has invalid feature list */
*err_code = -730;
*err_msg = "Job's required features not available "
"on selected nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap, job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
if ((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) {
*err_code = -730;
*err_msg = "Job's required nodes not available";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
min_nodes = MAX(job_ptr->details->min_nodes, part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partitions max_nodes */
*err_code = -730;
*err_msg = "Job's min_nodes > max_nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
preemptee_candidates = slurm_find_preemptable_jobs(job_ptr);
orig_start_time = job_ptr->start_time;
rc = select_g_job_test(job_ptr, avail_bitmap,
min_nodes, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL, exc_core_bitmap);
if (preemptee_candidates)
list_destroy(preemptee_candidates);
if (rc == SLURM_SUCCESS) {
char tmp_str[128];
*err_code = 0;
uint32_t proc_cnt = 0;
xstrcat(reply_msg, "STARTINFO=");
#ifdef HAVE_BG
select_g_select_jobinfo_get(job_ptr->select_jobinfo,
SELECT_JOBDATA_NODE_CNT,
&proc_cnt);
#else
proc_cnt = job_ptr->total_cpus;
#endif
snprintf(tmp_str, sizeof(tmp_str), "%u:%u@%u,",
jobid, proc_cnt, (uint32_t) job_ptr->start_time);
xstrcat(reply_msg, tmp_str);
hostlist = bitmap2node_name(avail_bitmap);
xstrcat(reply_msg, hostlist);
xfree(hostlist);
} else {
xstrcat(reply_msg, "Jobs not runable on selected nodes");
error("wiki: jobs not runnable on nodes");
}
/* Restore pending job's expected start time */
job_ptr->start_time = orig_start_time;
FREE_NULL_BITMAP(avail_bitmap);
return reply_msg;
}
/*
* 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;
}
extern int create_full_system_block(List bg_found_block_list)
{
int rc = SLURM_SUCCESS;
ListIterator itr;
bg_record_t *bg_record = NULL;
char *name = NULL;
List records = NULL;
uint16_t geo[SYSTEM_DIMENSIONS];
int i;
select_ba_request_t blockreq;
List results = NULL;
struct part_record *part_ptr = NULL;
bitstr_t *bitmap = bit_alloc(node_record_count);
static int *dims = NULL;
bool larger = 0;
char start_char[SYSTEM_DIMENSIONS+1];
char geo_char[SYSTEM_DIMENSIONS+1];
if (!dims) {
dims = select_g_ba_get_dims();
memset(start_char, 0, sizeof(start_char));
memset(geo_char, 0, sizeof(geo_char));
}
/* Locks are already in place to protect part_list here */
itr = list_iterator_create(part_list);
while ((part_ptr = list_next(itr))) {
/* we only want to use mps that are in
* partitions
*/
if (!part_ptr->node_bitmap) {
debug4("Partition %s doesn't have any nodes in it.",
part_ptr->name);
continue;
}
bit_or(bitmap, part_ptr->node_bitmap);
}
list_iterator_destroy(itr);
bit_not(bitmap);
if (bit_ffs(bitmap) != -1) {
error("We don't have the entire system covered by partitions, "
"can't create full system block");
FREE_NULL_BITMAP(bitmap);
return SLURM_ERROR;
}
FREE_NULL_BITMAP(bitmap);
/* Here we are adding a block that in for the entire machine
just in case it isn't in the bluegene.conf file.
*/
slurm_mutex_lock(&block_state_mutex);
for (i=0; i<SYSTEM_DIMENSIONS; i++) {
geo[i] = dims[i] - 1;
if (geo[i] > 0)
larger = 1;
geo_char[i] = alpha_num[geo[i]];
start_char[i] = alpha_num[0];
}
i = (10+strlen(bg_conf->slurm_node_prefix));
name = xmalloc(i);
if (!larger)
snprintf(name, i, "%s%s",
bg_conf->slurm_node_prefix, start_char);
else
snprintf(name, i, "%s[%sx%s]",
bg_conf->slurm_node_prefix, start_char, geo_char);
if (bg_found_block_list) {
itr = list_iterator_create(bg_found_block_list);
while ((bg_record = (bg_record_t *) list_next(itr)) != NULL) {
if (!strcmp(name, bg_record->mp_str)) {
xfree(name);
list_iterator_destroy(itr);
/* don't create total already there */
goto no_total;
}
}
list_iterator_destroy(itr);
} else {
error("create_full_system_block: no bg_found_block_list 2");
}
if (bg_lists->main) {
itr = list_iterator_create(bg_lists->main);
while ((bg_record = (bg_record_t *) list_next(itr))
!= NULL) {
if (!strcmp(name, bg_record->mp_str)) {
xfree(name);
list_iterator_destroy(itr);
/* don't create total already there */
goto no_total;
}
}
list_iterator_destroy(itr);
} else {
xfree(name);
error("create_overlapped_blocks: no bg_lists->main 3");
rc = SLURM_ERROR;
goto no_total;
}
records = list_create(destroy_bg_record);
memset(&blockreq, 0, sizeof(select_ba_request_t));
blockreq.save_name = name;
for (i=0; i<SYSTEM_DIMENSIONS; i++)
blockreq.conn_type[i] = SELECT_TORUS;
add_bg_record(records, NULL, &blockreq, 0 , 0);
xfree(name);
bg_record = (bg_record_t *) list_pop(records);
if (!bg_record) {
error("Nothing was returned from full system create");
rc = SLURM_ERROR;
goto no_total;
}
reset_ba_system(false);
for(i=0; i<SYSTEM_DIMENSIONS; i++) {
geo_char[i] = alpha_num[bg_record->geo[i]];
start_char[i] = alpha_num[bg_record->start[i]];
}
debug2("adding %s %s %s", bg_record->mp_str, start_char, geo_char);
if (bg_record->ba_mp_list)
list_flush(bg_record->ba_mp_list);
else
bg_record->ba_mp_list = list_create(destroy_ba_mp);
#ifdef HAVE_BGQ
results = list_create(destroy_ba_mp);
#else
results = list_create(NULL);
#endif
name = set_bg_block(results,
bg_record->start,
bg_record->geo,
bg_record->conn_type);
if (!name) {
error("I was unable to make the full system block.");
list_destroy(results);
list_iterator_destroy(itr);
slurm_mutex_unlock(&block_state_mutex);
return SLURM_ERROR;
}
xfree(name);
if (bg_record->ba_mp_list)
list_destroy(bg_record->ba_mp_list);
#ifdef HAVE_BGQ
bg_record->ba_mp_list = results;
results = NULL;
#else
bg_record->ba_mp_list = list_create(destroy_ba_mp);
copy_node_path(results, &bg_record->ba_mp_list);
list_destroy(results);
#endif
if ((rc = bridge_block_create(bg_record)) == SLURM_ERROR) {
error("create_full_system_block: "
"unable to configure block in api");
destroy_bg_record(bg_record);
goto no_total;
}
print_bg_record(bg_record);
list_append(bg_lists->main, bg_record);
no_total:
if (records)
list_destroy(records);
slurm_mutex_unlock(&block_state_mutex);
return rc;
}
/* cr_job_test - does most of the real work for select_p_job_test(), which
* includes contiguous selection, load-leveling and max_share logic
*
* PROCEDURE:
*
* Step 1: compare nodes in "avail" bitmap with current node state data
* to find available nodes that match the job request
*
* Step 2: check resources in "avail" bitmap with allocated resources from
* higher priority partitions (busy resources are UNavailable)
*
* Step 3: select resource usage on remaining resources in "avail" bitmap
* for this job, with the placement influenced by existing
* allocations
*/
extern int cr_job_test(struct job_record *job_ptr, bitstr_t *bitmap, int mode,
uint16_t cr_type, enum node_cr_state job_node_req,
uint32_t cr_node_cnt,
struct part_res_record *cr_part_ptr,
struct node_use_record *node_usage)
{
static int gang_mode = -1;
int error_code = SLURM_SUCCESS;
bitstr_t *orig_map, *avail_cores, *free_cores;
bitstr_t *tmpcore = NULL;
bool test_only;
uint32_t c, i, j, k, n, csize, save_mem = 0;
job_resources_t *job_res;
struct job_details *details_ptr;
struct part_res_record *p_ptr, *jp_ptr;
uint16_t *cpu_count;
if (gang_mode == -1) {
if (slurm_get_preempt_mode() & PREEMPT_MODE_GANG)
gang_mode = 1;
else
gang_mode = 0;
}
details_ptr = job_ptr->details;
free_job_resources(&job_ptr->job_resrcs);
if (mode == SELECT_MODE_TEST_ONLY)
test_only = true;
else /* SELECT_MODE_RUN_NOW || SELECT_MODE_WILL_RUN */
test_only = false;
/* check node_state and update the node bitmap as necessary */
if (!test_only) {
error_code = _verify_node_state(cr_part_ptr, job_ptr,
bitmap, cr_type, node_usage,
job_node_req);
if (error_code != SLURM_SUCCESS)
return error_code;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: evaluating job %u on %u nodes",
job_ptr->job_id, bit_set_count(bitmap));
}
orig_map = bit_copy(bitmap);
avail_cores = _make_core_bitmap(bitmap);
/* test to make sure that this job can succeed with all avail_cores
* if 'no' then return FAIL
* if 'yes' then we will seek the optimal placement for this job
* within avail_cores
*/
free_cores = bit_copy(avail_cores);
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt, free_cores,
node_usage, cr_type, test_only);
if (cpu_count == NULL) {
/* job cannot fit */
FREE_NULL_BITMAP(orig_map);
FREE_NULL_BITMAP(free_cores);
FREE_NULL_BITMAP(avail_cores);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 0 fail: "
"insufficient resources");
}
return SLURM_ERROR;
} else if (test_only) {
FREE_NULL_BITMAP(orig_map);
FREE_NULL_BITMAP(free_cores);
FREE_NULL_BITMAP(avail_cores);
xfree(cpu_count);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE)
info("select/serial: cr_job_test: test 0 pass: "******"test_only");
return SLURM_SUCCESS;
}
if (cr_type == CR_MEMORY) {
/* CR_MEMORY does not care about existing CPU allocations,
* so we can jump right to job allocation from here */
goto alloc_job;
}
xfree(cpu_count);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 0 pass - "
"job fits on given resources");
}
/* now that we know that this job can run with the given resources,
* let's factor in the existing allocations and seek the optimal set
* of resources for this job. Here is the procedure:
*
* Step 1: Seek idle CPUs across all partitions. If successful then
* place job and exit. If not successful, then continue. Two
* related items to note:
* 1. Jobs that don't share CPUs finish with step 1.
* 2. The remaining steps assume sharing or preemption.
*
* Step 2: Remove resources that are in use by higher-priority
* partitions, and test that job can still succeed. If not
* then exit.
*
* Step 3: Seek idle nodes among the partitions with the same
* priority as the job's partition. If successful then
* goto Step 6. If not then continue:
*
* Step 4: Seek placement within the job's partition. Search
* row-by-row. If no placement if found, then exit. If a row
* is found, then continue:
*
* Step 5: Place job and exit. FIXME! Here is where we need a
* placement algorithm that recognizes existing job
* boundaries and tries to "overlap jobs" as efficiently
* as possible.
*
* Step 6: Place job and exit. FIXME! here is we use a placement
* algorithm similar to Step 5 on jobs from lower-priority
* partitions.
*/
/*** Step 1 ***/
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
/* remove all existing allocations from free_cores */
tmpcore = bit_copy(free_cores);
for (p_ptr = cr_part_ptr; p_ptr; p_ptr = p_ptr->next) {
if (!p_ptr->row)
continue;
for (i = 0; i < p_ptr->num_rows; i++) {
if (!p_ptr->row[i].row_bitmap)
continue;
bit_copybits(tmpcore, p_ptr->row[i].row_bitmap);
bit_not(tmpcore); /* set bits now "free" resources */
bit_and(free_cores, tmpcore);
}
}
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt, free_cores,
node_usage, cr_type, test_only);
if (cpu_count) {
/* job fits! We're done. */
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 1 pass - "
"idle resources found");
}
goto alloc_job;
}
if ((gang_mode == 0) && (job_node_req == NODE_CR_ONE_ROW)) {
/* This job CANNOT share CPUs regardless of priority,
* so we fail here. Note that Shared=EXCLUSIVE was already
* addressed in _verify_node_state() and job preemption
* removes jobs from simulated resource allocation map
* before this point. */
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 1 fail - "
"no idle resources available");
}
goto alloc_job;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 1 fail - "
"not enough idle resources");
}
/*** Step 2 ***/
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
for (jp_ptr = cr_part_ptr; jp_ptr; jp_ptr = jp_ptr->next) {
if (jp_ptr->part_ptr == job_ptr->part_ptr)
break;
}
if (!jp_ptr) {
fatal("select/serial: could not find partition for job %u",
job_ptr->job_id);
return SLURM_ERROR; /* Fix CLANG false positive */
}
/* remove existing allocations (jobs) from higher-priority partitions
* from avail_cores */
for (p_ptr = cr_part_ptr; p_ptr; p_ptr = p_ptr->next) {
if ((p_ptr->part_ptr->priority <= jp_ptr->part_ptr->priority) &&
(p_ptr->part_ptr->preempt_mode != PREEMPT_MODE_OFF))
continue;
if (!p_ptr->row)
continue;
for (i = 0; i < p_ptr->num_rows; i++) {
if (!p_ptr->row[i].row_bitmap)
continue;
bit_copybits(tmpcore, p_ptr->row[i].row_bitmap);
bit_not(tmpcore); /* set bits now "free" resources */
bit_and(free_cores, tmpcore);
}
}
/* make these changes permanent */
bit_copybits(avail_cores, free_cores);
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt, free_cores,
node_usage, cr_type, test_only);
if (!cpu_count) {
/* job needs resources that are currently in use by
* higher-priority jobs, so fail for now */
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 2 fail - "
"resources busy with higher priority jobs");
}
goto alloc_job;
}
xfree(cpu_count);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 2 pass - "
"available resources for this priority");
}
/*** Step 3 ***/
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
/* remove existing allocations (jobs) from same-priority partitions
* from avail_cores */
for (p_ptr = cr_part_ptr; p_ptr; p_ptr = p_ptr->next) {
if (p_ptr->part_ptr->priority != jp_ptr->part_ptr->priority)
continue;
if (!p_ptr->row)
continue;
for (i = 0; i < p_ptr->num_rows; i++) {
if (!p_ptr->row[i].row_bitmap)
continue;
bit_copybits(tmpcore, p_ptr->row[i].row_bitmap);
bit_not(tmpcore); /* set bits now "free" resources */
bit_and(free_cores, tmpcore);
}
}
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt, free_cores,
node_usage, cr_type, test_only);
if (cpu_count) {
/* jobs from low-priority partitions are the only thing left
* in our way. for now we'll ignore them, but FIXME: we need
* a good placement algorithm here that optimizes "job overlap"
* between this job (in these idle nodes) and the low-priority
* jobs */
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 3 pass - "
"found resources");
}
goto alloc_job;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 3 fail - "
"not enough idle resources in same priority");
}
/*** Step 4 ***/
/* try to fit the job into an existing row
*
* tmpcore = worker core_bitmap
* free_cores = core_bitmap to be built
* avail_cores = static core_bitmap of all available cores
*/
if (!jp_ptr || !jp_ptr->row) {
/* there's no existing jobs in this partition, so place
* the job in avail_cores. FIXME: still need a good
* placement algorithm here that optimizes "job overlap"
* between this job (in these idle nodes) and existing
* jobs in the other partitions with <= priority to
* this partition */
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt,
free_cores, node_usage, cr_type,
test_only);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 4 pass - "
"first row found");
}
goto alloc_job;
}
cr_sort_part_rows(jp_ptr);
c = jp_ptr->num_rows;
if (job_node_req != NODE_CR_AVAILABLE)
c = 1;
for (i = 0; i < c; i++) {
if (!jp_ptr->row[i].row_bitmap)
break;
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
bit_copybits(tmpcore, jp_ptr->row[i].row_bitmap);
bit_not(tmpcore);
bit_and(free_cores, tmpcore);
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt,
free_cores, node_usage, cr_type,
test_only);
if (cpu_count) {
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: "
"test 4 pass - row %i", i);
}
break;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: "
"test 4 fail - row %i", i);
}
}
if ((i < c) && !jp_ptr->row[i].row_bitmap) {
/* we've found an empty row, so use it */
bit_copybits(bitmap, orig_map);
bit_copybits(free_cores, avail_cores);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: "
"test 4 trying empty row %i",i);
}
cpu_count = _select_nodes(job_ptr, bitmap, cr_node_cnt,
free_cores, node_usage, cr_type,
test_only);
}
if (!cpu_count) {
/* job can't fit into any row, so exit */
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: test 4 fail - "
"busy partition");
}
goto alloc_job;
}
/*** CONSTRUCTION ZONE FOR STEPs 5 AND 6 ***
* Note that while the job may have fit into a row, it should
* still be run through a good placement algorithm here that
* optimizes "job overlap" between this job (in these idle nodes)
* and existing jobs in the other partitions with <= priority to
* this partition */
alloc_job:
/* at this point we've found a good set of
* bits to allocate to this job:
* - bitmap is the set of nodes to allocate
* - free_cores is the set of allocated cores
* - cpu_count is the number of cpus per allocated node
*
* Next steps are to cleanup the worker variables,
* create the job_resources struct,
* distribute the job on the bits, and exit
*/
FREE_NULL_BITMAP(orig_map);
FREE_NULL_BITMAP(avail_cores);
FREE_NULL_BITMAP(tmpcore);
if (!cpu_count) {
/* we were sent here to cleanup and exit */
FREE_NULL_BITMAP(free_cores);
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: exiting cr_job_test with no "
"allocation");
}
return SLURM_ERROR;
}
/* At this point we have:
* - a bitmap of selected nodes
* - a free_cores bitmap of usable cores on each selected node
* - a per-alloc-node cpu_count array
*/
if ((mode != SELECT_MODE_WILL_RUN) && (job_ptr->part_ptr == NULL))
error_code = EINVAL;
if ((error_code == SLURM_SUCCESS) && (mode == SELECT_MODE_WILL_RUN))
job_ptr->total_cpus = 1;
if ((error_code != SLURM_SUCCESS) || (mode != SELECT_MODE_RUN_NOW)) {
FREE_NULL_BITMAP(free_cores);
xfree(cpu_count);
return error_code;
}
n = bit_ffs(bitmap);
if (n < 0) {
FREE_NULL_BITMAP(free_cores);
xfree(cpu_count);
return error_code;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: distributing job %u",
job_ptr->job_id);
}
/** create the struct_job_res **/
job_res = create_job_resources();
job_res->node_bitmap = bit_copy(bitmap);
job_res->nodes = bitmap2node_name(bitmap);
job_res->nhosts = bit_set_count(bitmap);
job_res->ncpus = job_res->nhosts;
if (job_ptr->details->ntasks_per_node)
job_res->ncpus *= details_ptr->ntasks_per_node;
job_res->ncpus = MAX(job_res->ncpus,
details_ptr->min_cpus);
job_res->ncpus = MAX(job_res->ncpus,
details_ptr->pn_min_cpus);
job_res->node_req = job_node_req;
job_res->cpus = cpu_count;
job_res->cpus_used = xmalloc(job_res->nhosts *
sizeof(uint16_t));
job_res->memory_allocated = xmalloc(job_res->nhosts *
sizeof(uint32_t));
job_res->memory_used = xmalloc(job_res->nhosts *
sizeof(uint32_t));
/* store the hardware data for the selected nodes */
error_code = build_job_resources(job_res, node_record_table_ptr,
select_fast_schedule);
if (error_code != SLURM_SUCCESS) {
free_job_resources(&job_res);
FREE_NULL_BITMAP(free_cores);
return error_code;
}
c = 0;
csize = bit_size(job_res->core_bitmap);
j = cr_get_coremap_offset(n);
k = cr_get_coremap_offset(n + 1);
for (; j < k; j++, c++) {
if (!bit_test(free_cores, j))
continue;
if (c >= csize) {
error("select/serial: cr_job_test "
"core_bitmap index error on node %s",
select_node_record[n].node_ptr->name);
drain_nodes(select_node_record[n].node_ptr->name,
"Bad core count", getuid());
free_job_resources(&job_res);
FREE_NULL_BITMAP(free_cores);
return SLURM_ERROR;
}
bit_set(job_res->core_bitmap, c);
break;
}
if (select_debug_flags & DEBUG_FLAG_SELECT_TYPE) {
info("select/serial: cr_job_test: job %u ncpus %u cbits %u/%d "
"nbits %u", job_ptr->job_id,
job_res->ncpus, bit_set_count(free_cores), 1,
job_res->nhosts);
}
FREE_NULL_BITMAP(free_cores);
/* distribute the tasks and clear any unused cores */
job_ptr->job_resrcs = job_res;
error_code = cr_dist(job_ptr, cr_type);
if (error_code != SLURM_SUCCESS) {
free_job_resources(&job_ptr->job_resrcs);
return error_code;
}
/* translate job_res->cpus array into format with rep count */
job_ptr->total_cpus = build_job_resources_cpu_array(job_res);
if (!(cr_type & CR_MEMORY))
return error_code;
/* load memory allocated array */
save_mem = details_ptr->pn_min_memory;
if (save_mem & MEM_PER_CPU) {
/* memory is per-cpu */
save_mem &= (~MEM_PER_CPU);
job_res->memory_allocated[0] = job_res->cpus[0] * save_mem;
} else {
/* memory is per-node */
job_res->memory_allocated[0] = save_mem;
}
return error_code;
}
static int _attempt_backfill(void)
{
DEF_TIMERS;
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr, **bf_part_ptr = NULL;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit, part_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL, *non_cg_bitmap = NULL;
time_t now, sched_start, later_start, start_res, resv_end, window_end;
node_space_map_t *node_space;
struct timeval bf_time1, bf_time2;
int rc = 0;
int job_test_count = 0;
uint32_t *uid = NULL, nuser = 0, bf_parts = 0, *bf_part_jobs = NULL;
uint16_t *njobs = NULL;
bool already_counted;
uint32_t reject_array_job_id = 0;
struct part_record *reject_array_part = NULL;
uint32_t job_start_cnt = 0, start_time;
time_t config_update = slurmctld_conf.last_update;
time_t part_update = last_part_update;
struct timeval start_tv;
bf_last_yields = 0;
#ifdef HAVE_ALPS_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
START_TIMER;
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
END_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: ALPS inventory completed, %s", TIME_STR);
/* The Basil inventory can take a long time to complete. Process
* pending RPCs before starting the backfill scheduling logic */
_yield_locks(1000000);
#endif
START_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: beginning");
else
debug("backfill: beginning");
sched_start = now = time(NULL);
gettimeofday(&start_tv, NULL);
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true, true);
if (list_count(job_queue) == 0) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: no jobs to backfill");
else
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
gettimeofday(&bf_time1, NULL);
non_cg_bitmap = bit_copy(cg_node_bitmap);
bit_not(non_cg_bitmap);
slurmctld_diag_stats.bf_queue_len = list_count(job_queue);
slurmctld_diag_stats.bf_queue_len_sum += slurmctld_diag_stats.
bf_queue_len;
slurmctld_diag_stats.bf_last_depth = 0;
slurmctld_diag_stats.bf_last_depth_try = 0;
slurmctld_diag_stats.bf_when_last_cycle = now;
slurmctld_diag_stats.bf_active = 1;
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt * 2 + 1));
node_space[0].begin_time = sched_start;
window_end = sched_start + backfill_window;
node_space[0].end_time = window_end;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_node_space_table(node_space);
if (max_backfill_job_per_part) {
ListIterator part_iterator;
struct part_record *part_ptr;
bf_parts = list_count(part_list);
bf_part_ptr = xmalloc(sizeof(struct part_record *) * bf_parts);
bf_part_jobs = xmalloc(sizeof(int) * bf_parts);
part_iterator = list_iterator_create(part_list);
i = 0;
while ((part_ptr = (struct part_record *)
list_next(part_iterator))) {
bf_part_ptr[i++] = part_ptr;
}
list_iterator_destroy(part_iterator);
}
if (max_backfill_job_per_user) {
uid = xmalloc(BF_MAX_USERS * sizeof(uint32_t));
njobs = xmalloc(BF_MAX_USERS * sizeof(uint16_t));
}
sort_job_queue(job_queue);
while (1) {
job_queue_rec = (job_queue_rec_t *) list_pop(job_queue);
if (!job_queue_rec) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: reached end of job queue");
break;
}
if (slurmctld_config.shutdown_time)
break;
if (((defer_rpc_cnt > 0) &&
(slurmctld_config.server_thread_count >= defer_rpc_cnt)) ||
(_delta_tv(&start_tv) >= sched_timeout)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
if ((_yield_locks(yield_sleep) && !backfill_continue) ||
(slurmctld_conf.last_update != config_update) ||
(last_part_update != part_update)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing "
"%u(%d) jobs",
slurmctld_diag_stats.bf_last_depth,
job_test_count);
}
rc = 1;
xfree(job_queue_rec);
break;
}
/* cg_node_bitmap may be changed */
bit_copybits(non_cg_bitmap, cg_node_bitmap);
bit_not(non_cg_bitmap);
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
gettimeofday(&start_tv, NULL);
job_test_count = 0;
START_TIMER;
}
job_ptr = job_queue_rec->job_ptr;
/* With bf_continue configured, the original job could have
* been cancelled and purged. Validate pointer here. */
if ((job_ptr->magic != JOB_MAGIC) ||
(job_ptr->job_id != job_queue_rec->job_id)) {
xfree(job_queue_rec);
continue;
}
orig_time_limit = job_ptr->time_limit;
part_ptr = job_queue_rec->part_ptr;
job_test_count++;
slurmctld_diag_stats.bf_last_depth++;
already_counted = false;
xfree(job_queue_rec);
if (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
if (!avail_front_end(job_ptr))
continue; /* No available frontend for this job */
if (job_ptr->array_task_id != NO_VAL) {
if ((reject_array_job_id == job_ptr->array_job_id) &&
(reject_array_part == part_ptr))
continue; /* already rejected array element */
/* assume reject whole array for now, clear if OK */
reject_array_job_id = job_ptr->array_job_id;
reject_array_part = part_ptr;
}
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill test for JobID=%u Prio=%u Partition=%s",
job_ptr->job_id, job_ptr->priority,
job_ptr->part_ptr->name);
}
if (max_backfill_job_per_part) {
bool skip_job = false;
for (j = 0; j < bf_parts; j++) {
if (bf_part_ptr[j] != job_ptr->part_ptr)
continue;
if (bf_part_jobs[j]++ >=
max_backfill_job_per_part)
skip_job = true;
break;
}
if (skip_job) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: have already "
"checked %u jobs for "
"partition %s; skipping "
"job %u",
max_backfill_job_per_part,
job_ptr->part_ptr->name,
job_ptr->job_id);
continue;
}
}
if (max_backfill_job_per_user) {
for (j = 0; j < nuser; j++) {
if (job_ptr->user_id == uid[j]) {
njobs[j]++;
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: user %u: "
"#jobs %u",
uid[j], njobs[j]);
break;
}
}
if (j == nuser) { /* user not found */
static bool bf_max_user_msg = true;
if (nuser < BF_MAX_USERS) {
uid[j] = job_ptr->user_id;
njobs[j] = 1;
nuser++;
} else if (bf_max_user_msg) {
bf_max_user_msg = false;
error("backfill: too many users in "
"queue. Consider increasing "
"BF_MAX_USERS");
}
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug2("backfill: found new user %u. "
"Total #users now %u",
job_ptr->user_id, nuser);
} else {
if (njobs[j] >= max_backfill_job_per_user) {
/* skip job */
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: have already "
"checked %u jobs for "
"user %u; skipping "
"job %u",
max_backfill_job_per_user,
job_ptr->user_id,
job_ptr->job_id);
continue;
}
}
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL) ||
((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: partition %s not usable",
job_ptr->part_ptr->name);
continue;
}
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS)) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u not runable now",
job_ptr->job_id);
continue;
}
/* Determine minimum and maximum node counts */
min_nodes = MAX(job_ptr->details->min_nodes,
part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u node count too high",
job_ptr->job_id);
continue;
}
/* Determine job's expected completion time */
if (part_ptr->max_time == INFINITE)
part_time_limit = 365 * 24 * 60; /* one year */
else
part_time_limit = part_ptr->max_time;
if (job_ptr->time_limit == NO_VAL) {
time_limit = part_time_limit;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_time_limit);
}
comp_time_limit = time_limit;
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE) &&
slurm_get_preempt_mode())
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
if (slurmctld_config.shutdown_time)
break;
if (((defer_rpc_cnt > 0) &&
(slurmctld_config.server_thread_count >= defer_rpc_cnt)) ||
(_delta_tv(&start_tv) >= sched_timeout)) {
uint32_t save_job_id = job_ptr->job_id;
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
if ((_yield_locks(yield_sleep) && !backfill_continue) ||
(slurmctld_conf.last_update != config_update) ||
(last_part_update != part_update)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing "
"%u(%d) jobs",
slurmctld_diag_stats.bf_last_depth,
job_test_count);
}
rc = 1;
break;
}
/* cg_node_bitmap may be changed */
bit_copybits(non_cg_bitmap, cg_node_bitmap);
bit_not(non_cg_bitmap);
/* With bf_continue configured, the original job could
* have been scheduled or cancelled and purged.
* Revalidate job the record here. */
if ((job_ptr->magic != JOB_MAGIC) ||
(job_ptr->job_id != save_job_id))
continue;
if (!IS_JOB_PENDING(job_ptr))
continue;
if (!avail_front_end(job_ptr))
continue; /* No available frontend */
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
gettimeofday(&start_tv, NULL);
job_test_count = 1;
START_TIMER;
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap,
&exc_core_bitmap);
if (j != SLURM_SUCCESS) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u reservation defer",
job_ptr->job_id);
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
resv_end = find_resv_end(start_res);
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
bit_and(avail_bitmap, non_cg_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
if (resv_end && (++resv_end < window_end) &&
((later_start == 0) || (resv_end < later_start))) {
later_start = resv_end;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap,
job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features OR
* no change since previously tested nodes (only changes
* in other partition nodes) */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
/* Job can not start until too far in the future */
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = sched_start + backfill_window;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
if (!already_counted) {
slurmctld_diag_stats.bf_last_depth_try++;
already_counted = true;
}
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_job_test(job_ptr, avail_bitmap, start_res);
j = _try_sched(job_ptr, &avail_bitmap, min_nodes, max_nodes,
req_nodes, exc_core_bitmap);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) { /* Can start now */
uint32_t save_time_limit = job_ptr->time_limit;
uint32_t hard_limit;
bool reset_time = false;
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)) {
if (orig_time_limit == NO_VAL) {
acct_policy_alter_job(
job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
} else {
acct_policy_alter_job(
job_ptr, orig_time_limit);
job_ptr->time_limit = orig_time_limit;
}
} else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
acct_policy_alter_job(job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
reset_time = true;
} else if (orig_time_limit == NO_VAL) {
acct_policy_alter_job(job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
} else {
acct_policy_alter_job(job_ptr, orig_time_limit);
job_ptr->time_limit = orig_time_limit;
}
if (job_ptr->time_limit == INFINITE)
hard_limit = 365 * 24 * 60; /* one year */
else
hard_limit = job_ptr->time_limit;
job_ptr->end_time = job_ptr->start_time +
(hard_limit * 60);
if (reset_time) {
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: planned start of job %u"
" failed: %s", job_ptr->job_id,
slurm_strerror(rc));
}
/* Drop through and reserve these resources.
* Likely due to state changes during sleep.
* Make best-effort based upon original state */
job_ptr->time_limit = orig_time_limit;
later_start = 0;
} else {
/* Started this job, move to next one */
reject_array_job_id = 0;
reject_array_part = NULL;
/* Update the database if job time limit
* changed and move to next job */
if (save_time_limit != job_ptr->time_limit)
jobacct_storage_g_job_start(acct_db_conn,
job_ptr);
job_start_cnt++;
if (max_backfill_jobs_start &&
(job_start_cnt >= max_backfill_jobs_start)){
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: bf_max_job_start"
" limit of %d reached",
max_backfill_jobs_start);
}
break;
}
continue;
}
} else {
job_ptr->time_limit = orig_time_limit;
}
start_time = job_ptr->start_time;
end_reserve = job_ptr->start_time + (time_limit * 60);
start_time = (start_time / backfill_resolution) *
backfill_resolution;
end_reserve = (end_reserve / backfill_resolution) *
backfill_resolution;
if (later_start && (start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve,
avail_bitmap);
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: table size limit of %u reached",
max_backfill_job_cnt);
}
break;
}
if ((job_ptr->start_time > now) &&
_test_resv_overlap(node_space, avail_bitmap,
start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve, avail_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
reject_array_job_id = 0;
reject_array_part = NULL;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve, avail_bitmap);
xfree(job_ptr->sched_nodes);
job_ptr->sched_nodes = bitmap2node_name(avail_bitmap);
bit_not(avail_bitmap);
_add_reservation(start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_node_space_table(node_space);
}
xfree(bf_part_jobs);
xfree(bf_part_ptr);
xfree(uid);
xfree(njobs);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
FREE_NULL_BITMAP(non_cg_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
gettimeofday(&bf_time2, NULL);
_do_diag_stats(&bf_time1, &bf_time2, yield_sleep);
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
return rc;
}
static int _attempt_backfill(void)
{
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
time_t now = time(NULL), sched_start, later_start, start_res;
node_space_map_t *node_space;
static int sched_timeout = 0;
int this_sched_timeout = 0, rc = 0;
sched_start = now;
if (sched_timeout == 0) {
sched_timeout = slurm_get_msg_timeout() / 2;
sched_timeout = MAX(sched_timeout, 1);
sched_timeout = MIN(sched_timeout, 10);
}
this_sched_timeout = sched_timeout;
#ifdef HAVE_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
#endif
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true);
if (list_count(job_queue) <= 1) {
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt + 3));
node_space[0].begin_time = sched_start;
node_space[0].end_time = sched_start + backfill_window;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
while ((job_queue_rec = (job_queue_rec_t *)
list_pop_bottom(job_queue, sort_job_queue2))) {
job_ptr = job_queue_rec->job_ptr;
part_ptr = job_queue_rec->part_ptr;
xfree(job_queue_rec);
if (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill test for job %u", job_ptr->job_id);
if ((job_ptr->state_reason == WAIT_ASSOC_JOB_LIMIT) ||
(job_ptr->state_reason == WAIT_ASSOC_RESOURCE_LIMIT) ||
(job_ptr->state_reason == WAIT_ASSOC_TIME_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_JOB_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_RESOURCE_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_TIME_LIMIT) ||
!acct_policy_job_runnable(job_ptr)) {
debug2("backfill: job %u is not allowed to run now. "
"Skipping it. State=%s. Reason=%s. Priority=%u",
job_ptr->job_id,
job_state_string(job_ptr->job_state),
job_reason_string(job_ptr->state_reason),
job_ptr->priority);
continue;
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL))
continue;
if ((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)
continue;
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS))
continue;
/* Determine minimum and maximum node counts */
min_nodes = MAX(job_ptr->details->min_nodes,
part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partition's max_nodes */
continue;
}
/* Determine job's expected completion time */
if (job_ptr->time_limit == NO_VAL) {
if (part_ptr->max_time == INFINITE)
time_limit = 365 * 24 * 60; /* one year */
else
time_limit = part_ptr->max_time;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_ptr->max_time);
}
comp_time_limit = time_limit;
orig_time_limit = job_ptr->time_limit;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
FREE_NULL_BITMAP(avail_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap,
job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
job_ptr->time_limit = orig_time_limit;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
if ((time(NULL) - sched_start) >= this_sched_timeout) {
debug("backfill: loop taking too long, yielding locks");
if (_yield_locks()) {
debug("backfill: system state changed, "
"breaking out");
rc = 1;
break;
} else {
this_sched_timeout += sched_timeout;
}
}
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
j = _try_sched(job_ptr, &avail_bitmap,
min_nodes, max_nodes, req_nodes);
debug2("backfill: finished _try_sched for job %u.",
job_ptr->job_id);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) {
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
job_ptr->time_limit = orig_time_limit;
else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
job_ptr->time_limit = comp_time_limit;
job_ptr->end_time = job_ptr->start_time +
(comp_time_limit * 60);
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
} else {
job_ptr->time_limit = orig_time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
/* Planned to start job, but something bad
* happended. */
job_ptr->start_time = 0;
break;
} else {
/* Started this job, move to next one */
continue;
}
} else
job_ptr->time_limit = orig_time_limit;
if (later_start && (job_ptr->start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
/* Already have too many jobs to deal with */
break;
}
end_reserve = job_ptr->start_time + (time_limit * 60);
if (_test_resv_overlap(node_space, avail_bitmap,
job_ptr->start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
bit_not(avail_bitmap);
_add_reservation(job_ptr->start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
return rc;
}
static void _validate_switches(void)
{
slurm_conf_switches_t *ptr, **ptr_array;
int depth, i, j;
struct switch_record *switch_ptr, *prior_ptr;
hostlist_t hl, invalid_hl = NULL;
char *child, *buf;
bool have_root = false;
bitstr_t *multi_homed_bitmap = NULL; /* nodes on >1 leaf switch */
bitstr_t *switches_bitmap = NULL; /* nodes on any leaf switch */
bitstr_t *tmp_bitmap = NULL;
_free_switch_record_table();
switch_record_cnt = _read_topo_file(&ptr_array);
if (switch_record_cnt == 0) {
error("No switches configured");
s_p_hashtbl_destroy(conf_hashtbl);
return;
}
switch_record_table = xmalloc(sizeof(struct switch_record) *
switch_record_cnt);
multi_homed_bitmap = bit_alloc(node_record_count);
switch_ptr = switch_record_table;
for (i=0; i<switch_record_cnt; i++, switch_ptr++) {
ptr = ptr_array[i];
switch_ptr->name = xstrdup(ptr->switch_name);
/* See if switch name has already been defined. */
prior_ptr = switch_record_table;
for (j=0; j<i; j++, prior_ptr++) {
if (strcmp(switch_ptr->name, prior_ptr->name) == 0) {
fatal("Switch (%s) has already been defined",
prior_ptr->name);
}
}
switch_ptr->link_speed = ptr->link_speed;
if (ptr->nodes) {
switch_ptr->level = 0; /* leaf switch */
switch_ptr->nodes = xstrdup(ptr->nodes);
if (_node_name2bitmap(ptr->nodes,
&switch_ptr->node_bitmap,
&invalid_hl)) {
fatal("Invalid node name (%s) in switch "
"config (%s)",
ptr->nodes, ptr->switch_name);
}
if (switches_bitmap) {
tmp_bitmap = bit_copy(switch_ptr->node_bitmap);
bit_and(tmp_bitmap, switches_bitmap);
bit_or(multi_homed_bitmap, tmp_bitmap);
FREE_NULL_BITMAP(tmp_bitmap);
bit_or(switches_bitmap,
switch_ptr->node_bitmap);
} else {
switches_bitmap = bit_copy(switch_ptr->
node_bitmap);
}
} else if (ptr->switches) {
switch_ptr->level = -1; /* determine later */
switch_ptr->switches = xstrdup(ptr->switches);
} else {
fatal("Switch configuration (%s) lacks children",
ptr->switch_name);
}
}
for (depth=1; ; depth++) {
bool resolved = true;
switch_ptr = switch_record_table;
for (i=0; i<switch_record_cnt; i++, switch_ptr++) {
if (switch_ptr->level != -1)
continue;
hl = hostlist_create(switch_ptr->switches);
if (!hl) {
fatal("Invalid switches: %s",
switch_ptr->switches);
}
while ((child = hostlist_pop(hl))) {
j = _get_switch_inx(child);
if ((j < 0) || (j == i)) {
fatal("Switch configuration %s has "
"invalid child (%s)",
switch_ptr->name, child);
}
if (switch_record_table[j].level == -1) {
/* Children not resolved */
resolved = false;
switch_ptr->level = -1;
FREE_NULL_BITMAP(switch_ptr->
node_bitmap);
free(child);
break;
}
if (switch_ptr->level == -1) {
switch_ptr->level = 1 +
switch_record_table[j].level;
switch_ptr->node_bitmap =
bit_copy(switch_record_table[j].
node_bitmap);
} else {
switch_ptr->level =
MAX(switch_ptr->level,
(switch_record_table[j].
level + 1));
bit_or(switch_ptr->node_bitmap,
switch_record_table[j].
node_bitmap);
}
free(child);
}
hostlist_destroy(hl);
}
if (resolved)
break;
if (depth > 20) /* Prevent infinite loop */
fatal("Switch configuration is not a tree");
}
switch_levels = 0;
switch_ptr = switch_record_table;
for (i=0; i<switch_record_cnt; i++, switch_ptr++) {
switch_levels = MAX(switch_levels, switch_ptr->level);
if (switch_ptr->node_bitmap == NULL)
error("switch %s has no nodes", switch_ptr->name);
}
if (switches_bitmap) {
bit_not(switches_bitmap);
i = bit_set_count(switches_bitmap);
if (i > 0) {
child = bitmap2node_name(switches_bitmap);
error("WARNING: switches lack access to %d nodes: %s",
i, child);
xfree(child);
}
FREE_NULL_BITMAP(switches_bitmap);
} else
fatal("switches contain no nodes");
if (invalid_hl) {
buf = hostlist_ranged_string_xmalloc(invalid_hl);
error("WARNING: Invalid hostnames in switch configuration: %s",
buf);
xfree(buf);
hostlist_destroy(invalid_hl);
}
/* Report nodes on multiple leaf switches,
* possibly due to bad configuration file */
i = bit_set_count(multi_homed_bitmap);
if (i > 0) {
child = bitmap2node_name(multi_homed_bitmap);
error("WARNING: Multiple leaf switches contain nodes: %s",
child);
xfree(child);
}
FREE_NULL_BITMAP(multi_homed_bitmap);
/* Create array of indexes of children of each switch,
* and see if any switch can reach all nodes */
for (i = 0; i < switch_record_cnt; i++) {
if (switch_record_table[i].level != 0) {
_find_child_switches (i);
}
if (node_record_count ==
bit_set_count(switch_record_table[i].node_bitmap)) {
have_root = true;
}
}
if (!have_root) {
info("TOPOLOGY: warning -- no switch can reach all nodes"
" through its descendants."
"Do not use route/topology");
}
s_p_hashtbl_destroy(conf_hashtbl);
_log_switches();
}
/*
* Test all combinations of x and y
*/
static void multi_test(bit_t x, bit_t y) {
test_or(x, y);
test_or(bit_not(x), y);
test_or(x, bit_not(y));
test_or(bit_not(x), bit_not(y));
test_or(y, x);
test_or(y, bit_not(x));
test_or(bit_not(y), x);
test_or(bit_not(y), bit_not(x));
test_xor(x, y);
test_xor(bit_not(x), y);
test_xor(x, bit_not(y));
test_xor(bit_not(x), bit_not(y));
test_xor(y, x);
test_xor(bit_not(y), x);
test_xor(y, bit_not(x));
test_xor(bit_not(y), bit_not(x));
}