static int _sync_block_lists(List full_list, List incomp_list)
{
ListIterator itr;
bg_record_t *new_record = NULL;
int count = 0;
itr = list_iterator_create(full_list);
while ((new_record = list_next(itr))) {
/* Make sure we aren't adding any block that doesn't
have a block_id. If the record has an original
then we don't need to add either, (since it is
already in the list).
*/
if (!new_record->bg_block_id || new_record->original)
continue;
list_remove(itr);
if (bg_conf->slurm_debug_flags & DEBUG_FLAG_BG_PICK)
info("sync: adding %s %p",
new_record->bg_block_id, new_record);
list_append(incomp_list, new_record);
last_bg_update = time(NULL);
count++;
}
list_iterator_destroy(itr);
if (count)
sort_bg_record_inc_size(incomp_list);
return count;
}
/*
* create_defined_blocks - create the static blocks that will be used
* for scheduling, all partitions must be able to be created and booted
* at once.
* IN - int overlapped, 1 if partitions are to be overlapped, 0 if they are
* static.
* RET - success of fitting all configurations
*/
extern int create_defined_blocks(bg_layout_t overlapped,
List bg_found_block_list)
{
int rc = SLURM_SUCCESS;
ListIterator itr;
bg_record_t *bg_record = NULL;
int i;
uint16_t geo[SYSTEM_DIMENSIONS];
char temp[256];
struct part_record *part_ptr = NULL;
bitstr_t *usable_mp_bitmap = bit_alloc(node_record_count);
/* 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(usable_mp_bitmap, part_ptr->node_bitmap);
}
list_iterator_destroy(itr);
if (bit_ffs(usable_mp_bitmap) == -1) {
fatal("We don't have any nodes in any partitions. "
"Can't create blocks. "
"Please check your slurm.conf.");
}
slurm_mutex_lock(&block_state_mutex);
reset_ba_system(false);
ba_set_removable_mps(usable_mp_bitmap, 1);
if (bg_lists->main) {
itr = list_iterator_create(bg_lists->main);
while ((bg_record = list_next(itr))) {
if (bg_record->mp_count > 0
&& !bg_record->full_block
&& bg_record->cpu_cnt >= bg_conf->cpus_per_mp) {
char *name = NULL;
char start_char[SYSTEM_DIMENSIONS+1];
char geo_char[SYSTEM_DIMENSIONS+1];
if (overlapped == LAYOUT_OVERLAP) {
reset_ba_system(false);
ba_set_removable_mps(usable_mp_bitmap,
1);
}
/* we want the mps that aren't
* in this record to mark them as used
*/
if (ba_set_removable_mps(
bg_record->mp_bitmap, 1)
!= SLURM_SUCCESS)
fatal("It doesn't seem we have a "
"bitmap for %s",
bg_record->bg_block_id);
for (i=0; i<SYSTEM_DIMENSIONS; i++) {
geo[i] = bg_record->geo[i];
start_char[i] = alpha_num[
bg_record->start[i]];
geo_char[i] = alpha_num[geo[i]];
}
start_char[i] = '\0';
geo_char[i] = '\0';
debug2("adding %s %s %s",
bg_record->mp_str,
start_char, geo_char);
if (bg_record->ba_mp_list
&& list_count(bg_record->ba_mp_list)) {
if ((rc = check_and_set_mp_list(
bg_record->ba_mp_list))
!= SLURM_SUCCESS) {
debug2("something happened in "
"the load of %s"
"Did you use smap to "
"make the "
"bluegene.conf file?",
bg_record->bg_block_id);
break;
}
} else {
#ifdef HAVE_BGQ
List results =
list_create(destroy_ba_mp);
#else
List results = list_create(NULL);
#endif
name = set_bg_block(
results,
bg_record->start,
geo,
bg_record->conn_type);
ba_reset_all_removed_mps();
if (!name) {
error("I was unable to "
"make the "
"requested block.");
list_destroy(results);
rc = SLURM_ERROR;
break;
}
snprintf(temp, sizeof(temp), "%s%s",
bg_conf->slurm_node_prefix,
name);
xfree(name);
if (strcmp(temp, bg_record->mp_str)) {
fatal("given list of %s "
"but allocated %s, "
"your order might be "
"wrong in bluegene.conf",
bg_record->mp_str,
temp);
}
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 (!block_exist_in_list(
bg_found_block_list, bg_record)) {
if (bg_record->full_block) {
/* if this is defined we need
to remove it since we are
going to try to create it
later on overlap systems
this doesn't matter, but
since we don't clear the
table on static mode we
can't do it here or it just
won't work since other
wires will be or are
already set
*/
list_remove(itr);
continue;
}
if ((rc = bridge_block_create(bg_record))
!= SLURM_SUCCESS)
break;
print_bg_record(bg_record);
}
}
list_iterator_destroy(itr);
if (rc != SLURM_SUCCESS)
goto end_it;
} else {
error("create_defined_blocks: no bg_lists->main 2");
rc = SLURM_ERROR;
goto end_it;
}
slurm_mutex_unlock(&block_state_mutex);
create_full_system_block(bg_found_block_list);
slurm_mutex_lock(&block_state_mutex);
sort_bg_record_inc_size(bg_lists->main);
end_it:
ba_reset_all_removed_mps();
FREE_NULL_BITMAP(usable_mp_bitmap);
slurm_mutex_unlock(&block_state_mutex);
#ifdef _PRINT_BLOCKS_AND_EXIT
if (bg_lists->main) {
itr = list_iterator_create(bg_lists->main);
debug("\n\n");
while ((found_record = (bg_record_t *) list_next(itr))
!= NULL) {
print_bg_record(found_record);
}
list_iterator_destroy(itr);
} else {
error("create_defined_blocks: no bg_lists->main 5");
}
exit(0);
#endif /* _PRINT_BLOCKS_AND_EXIT */
//exit(0);
return rc;
}
/*
* This could potentially lock the node lock in the slurmctld with
* slurm_drain_node, or slurm_fail_job so if slurmctld_locked is called we
* will call the functions without locking the locks again.
*/
extern int down_nodecard(char *mp_name, bitoff_t io_start,
bool slurmctld_locked)
{
List requests = NULL;
List delete_list = NULL;
ListIterator itr = NULL;
bg_record_t *bg_record = NULL, *found_record = NULL, tmp_record;
bg_record_t *smallest_bg_record = NULL;
struct node_record *node_ptr = NULL;
int mp_bit = 0;
static int io_cnt = NO_VAL;
static int create_size = NO_VAL;
static select_ba_request_t blockreq;
int rc = SLURM_SUCCESS;
char *reason = "select_bluegene: nodecard down";
xassert(mp_name);
if (io_cnt == NO_VAL) {
io_cnt = 1;
/* Translate 1 nodecard count to ionode count */
if ((io_cnt *= bg_conf->io_ratio))
io_cnt--;
/* make sure we create something that is able to be
created */
if (bg_conf->smallest_block < bg_conf->nodecard_cnode_cnt)
create_size = bg_conf->nodecard_cnode_cnt;
else
create_size = bg_conf->smallest_block;
}
node_ptr = find_node_record(mp_name);
if (!node_ptr) {
error ("down_sub_node_blocks: invalid node specified '%s'",
mp_name);
return EINVAL;
}
/* this is here for sanity check to make sure we don't core on
these bits when we set them below. */
if (io_start >= bg_conf->ionodes_per_mp
|| (io_start+io_cnt) >= bg_conf->ionodes_per_mp) {
debug("io %d-%d not configured on this "
"system, only %d ionodes per midplane",
io_start, io_start+io_cnt, bg_conf->ionodes_per_mp);
return EINVAL;
}
mp_bit = (node_ptr - node_record_table_ptr);
memset(&blockreq, 0, sizeof(select_ba_request_t));
blockreq.conn_type[0] = SELECT_SMALL;
blockreq.save_name = mp_name;
debug3("here setting node %d of %d and ionodes %d-%d of %d",
mp_bit, node_record_count, io_start,
io_start+io_cnt, bg_conf->ionodes_per_mp);
memset(&tmp_record, 0, sizeof(bg_record_t));
tmp_record.mp_count = 1;
tmp_record.cnode_cnt = bg_conf->nodecard_cnode_cnt;
tmp_record.mp_bitmap = bit_alloc(node_record_count);
bit_set(tmp_record.mp_bitmap, mp_bit);
tmp_record.ionode_bitmap = bit_alloc(bg_conf->ionodes_per_mp);
bit_nset(tmp_record.ionode_bitmap, io_start, io_start+io_cnt);
slurm_mutex_lock(&block_state_mutex);
itr = list_iterator_create(bg_lists->main);
while ((bg_record = list_next(itr))) {
if (!bit_test(bg_record->mp_bitmap, mp_bit))
continue;
if (!blocks_overlap(bg_record, &tmp_record))
continue;
if (bg_record->job_running > NO_JOB_RUNNING) {
if (slurmctld_locked)
job_fail(bg_record->job_running);
else
slurm_fail_job(bg_record->job_running);
}
/* If Running Dynamic mode and the block is
smaller than the create size just continue on.
*/
if ((bg_conf->layout_mode == LAYOUT_DYNAMIC)
&& (bg_record->cnode_cnt < create_size)) {
if (!delete_list)
delete_list = list_create(NULL);
list_append(delete_list, bg_record);
continue;
}
/* keep track of the smallest size that is at least
the size of create_size. */
if (!smallest_bg_record ||
(smallest_bg_record->cnode_cnt > bg_record->cnode_cnt))
smallest_bg_record = bg_record;
}
list_iterator_destroy(itr);
slurm_mutex_unlock(&block_state_mutex);
if (bg_conf->layout_mode != LAYOUT_DYNAMIC) {
debug3("running non-dynamic mode");
/* This should never happen, but just in case... */
if (delete_list)
list_destroy(delete_list);
/* If we found a block that is smaller or equal to a
midplane we will just mark it in an error state as
opposed to draining the node.
*/
if (smallest_bg_record
&& (smallest_bg_record->cnode_cnt < bg_conf->mp_cnode_cnt)){
if (smallest_bg_record->state & BG_BLOCK_ERROR_FLAG) {
rc = SLURM_NO_CHANGE_IN_DATA;
goto cleanup;
}
rc = put_block_in_error_state(
smallest_bg_record, reason);
goto cleanup;
}
debug("No block under 1 midplane available for this nodecard. "
"Draining the whole node.");
if (!node_already_down(mp_name)) {
if (slurmctld_locked)
drain_nodes(mp_name, reason,
slurm_get_slurm_user_id());
else
slurm_drain_nodes(mp_name, reason,
slurm_get_slurm_user_id());
}
rc = SLURM_SUCCESS;
goto cleanup;
}
/* below is only for Dynamic mode */
if (delete_list) {
int cnt_set = 0;
bitstr_t *iobitmap = bit_alloc(bg_conf->ionodes_per_mp);
/* don't lock here since it is handled inside
the put_block_in_error_state
*/
itr = list_iterator_create(delete_list);
while ((bg_record = list_next(itr))) {
debug2("combining smaller than nodecard "
"dynamic block %s",
bg_record->bg_block_id);
while (bg_record->job_running > NO_JOB_RUNNING)
sleep(1);
bit_or(iobitmap, bg_record->ionode_bitmap);
cnt_set++;
}
list_iterator_destroy(itr);
list_destroy(delete_list);
if (!cnt_set) {
FREE_NULL_BITMAP(iobitmap);
rc = SLURM_ERROR;
goto cleanup;
}
/* set the start to be the same as the start of the
ionode_bitmap. If no ionodes set (not a small
block) set io_start = 0. */
if ((io_start = bit_ffs(iobitmap)) == -1) {
io_start = 0;
if (create_size > bg_conf->nodecard_cnode_cnt)
blockreq.small128 = 4;
else
blockreq.small32 = 16;
} else if (create_size <= bg_conf->nodecard_cnode_cnt)
blockreq.small32 = 1;
else
/* this should never happen */
blockreq.small128 = 1;
FREE_NULL_BITMAP(iobitmap);
} else if (smallest_bg_record) {
debug2("smallest dynamic block is %s",
smallest_bg_record->bg_block_id);
if (smallest_bg_record->state & BG_BLOCK_ERROR_FLAG) {
rc = SLURM_NO_CHANGE_IN_DATA;
goto cleanup;
}
while (smallest_bg_record->job_running > NO_JOB_RUNNING)
sleep(1);
if (smallest_bg_record->cnode_cnt == create_size) {
rc = put_block_in_error_state(
smallest_bg_record, reason);
goto cleanup;
}
if (create_size > smallest_bg_record->cnode_cnt) {
/* we should never get here. This means we
* have a create_size that is bigger than a
* block that is already made.
*/
rc = put_block_in_error_state(
smallest_bg_record, reason);
goto cleanup;
}
debug3("node count is %d", smallest_bg_record->cnode_cnt);
switch(smallest_bg_record->cnode_cnt) {
#ifndef HAVE_BGL
case 64:
blockreq.small32 = 2;
break;
case 256:
blockreq.small32 = 8;
break;
#endif
case 128:
blockreq.small32 = 4;
break;
case 512:
default:
blockreq.small32 = 16;
break;
}
if (create_size != bg_conf->nodecard_cnode_cnt) {
blockreq.small128 = blockreq.small32 / 4;
blockreq.small32 = 0;
io_start = 0;
} else if ((io_start =
bit_ffs(smallest_bg_record->ionode_bitmap)) == -1)
/* set the start to be the same as the start of the
ionode_bitmap. If no ionodes set (not a small
block) set io_start = 0. */
io_start = 0;
} else {
switch(create_size) {
#ifndef HAVE_BGL
case 64:
blockreq.small64 = 8;
break;
case 256:
blockreq.small256 = 2;
#endif
case 32:
blockreq.small32 = 16;
break;
case 128:
blockreq.small128 = 4;
break;
case 512:
if (!node_already_down(mp_name)) {
char *reason = "select_bluegene: nodecard down";
if (slurmctld_locked)
drain_nodes(mp_name, reason,
slurm_get_slurm_user_id());
else
slurm_drain_nodes(
mp_name, reason,
slurm_get_slurm_user_id());
}
rc = SLURM_SUCCESS;
goto cleanup;
break;
default:
error("Unknown create size of %d", create_size);
break;
}
/* since we don't have a block in this midplane
we need to start at the beginning. */
io_start = 0;
/* we also need a bg_block to pretend to be the
smallest block that takes up the entire midplane. */
}
/* Here we need to add blocks that take up nodecards on this
midplane. Since Slurm only keeps track of midplanes
natively this is the only want to handle this case.
*/
requests = list_create(destroy_bg_record);
add_bg_record(requests, NULL, &blockreq, 1, io_start);
slurm_mutex_lock(&block_state_mutex);
delete_list = list_create(NULL);
while ((bg_record = list_pop(requests))) {
itr = list_iterator_create(bg_lists->main);
while ((found_record = list_next(itr))) {
if (!blocks_overlap(bg_record, found_record))
continue;
list_push(delete_list, found_record);
list_remove(itr);
}
list_iterator_destroy(itr);
/* we need to add this record since it doesn't exist */
if (bridge_block_create(bg_record) == SLURM_ERROR) {
destroy_bg_record(bg_record);
error("down_sub_node_blocks: "
"unable to configure block in api");
continue;
}
debug("adding block %s to fill in small blocks "
"around bad nodecards",
bg_record->bg_block_id);
print_bg_record(bg_record);
list_append(bg_lists->main, bg_record);
if (bit_overlap(bg_record->ionode_bitmap,
tmp_record.ionode_bitmap)) {
/* here we know the error block doesn't exist
so just set the state here */
slurm_mutex_unlock(&block_state_mutex);
rc = put_block_in_error_state(bg_record, reason);
slurm_mutex_lock(&block_state_mutex);
}
}
list_destroy(requests);
if (delete_list) {
slurm_mutex_unlock(&block_state_mutex);
free_block_list(NO_VAL, delete_list, 0, 0);
list_destroy(delete_list);
}
slurm_mutex_lock(&block_state_mutex);
sort_bg_record_inc_size(bg_lists->main);
slurm_mutex_unlock(&block_state_mutex);
last_bg_update = time(NULL);
cleanup:
FREE_NULL_BITMAP(tmp_record.mp_bitmap);
FREE_NULL_BITMAP(tmp_record.ionode_bitmap);
return rc;
}
