/* Update block user and reboot as needed block_state_mutex needs to
* be locked before coming in. */
static void _sync_agent(bg_action_t *bg_action_ptr, bg_record_t *bg_record)
{
struct job_record *job_ptr = bg_action_ptr->job_ptr;
debug3("Queue sync of job %u in BG block %s ending at %ld",
job_ptr->job_id, bg_action_ptr->bg_block_id,
job_ptr->end_time);
last_bg_update = time(NULL);
ba_sync_job_to_block(bg_record, job_ptr);
set_select_jobinfo(job_ptr->select_jobinfo->data,
SELECT_JOBDATA_BLOCK_PTR,
bg_record);
num_unused_cpus -= job_ptr->total_cpus;
if (!block_ptr_exist_in_list(bg_lists->job_running, bg_record))
list_push(bg_lists->job_running, bg_record);
if (!block_ptr_exist_in_list(bg_lists->booted, bg_record))
list_push(bg_lists->booted, bg_record);
if (bg_record->state == BG_BLOCK_INITED) {
int sync_user_rc;
job_ptr->job_state &= (~JOB_CONFIGURING);
last_job_update = time(NULL);
/* Just in case reset the boot flags */
bg_record->boot_state = 0;
bg_record->boot_count = 0;
sync_user_rc = bridge_block_sync_users(bg_record);
if (sync_user_rc == SLURM_ERROR) {
slurm_mutex_unlock(&block_state_mutex);
(void) slurm_fail_job(job_ptr->job_id, JOB_BOOT_FAIL);
slurm_mutex_lock(&block_state_mutex);
}
_destroy_bg_action(bg_action_ptr);
} else {
if (bg_record->state != BG_BLOCK_BOOTING) {
error("Block %s isn't ready and isn't "
"being configured! Starting job again.",
bg_action_ptr->bg_block_id);
} else {
debug("Block %s is booting, job ok",
bg_action_ptr->bg_block_id);
}
/* the function _block_op calls will destroy the
bg_action_ptr */
_block_op(bg_action_ptr);
}
}
/* Perform job initiation work */
static void _start_agent(bg_action_t *bg_action_ptr)
{
int rc, set_user_rc = SLURM_SUCCESS;
bg_record_t *bg_record = NULL;
bg_record_t *found_record = NULL;
ListIterator itr;
List delete_list = NULL;
int requeue_job = 0;
uint32_t req_job_id = bg_action_ptr->job_ptr->job_id;
bool block_inited = 0;
bool delete_it = 0;
slurm_mutex_lock(&block_state_mutex);
bg_record = find_bg_record_in_list(bg_lists->main,
bg_action_ptr->bg_block_id);
if (!bg_record) {
bg_record->modifying = 0;
slurm_mutex_unlock(&block_state_mutex);
error("block %s not found in bg_lists->main",
bg_action_ptr->bg_block_id);
bg_requeue_job(req_job_id, 1, 0, JOB_BOOT_FAIL, false);
return;
}
if ((bg_record->job_running <= NO_JOB_RUNNING)
&& !find_job_in_bg_record(bg_record, req_job_id)) {
bg_record->modifying = 0;
// bg_reset_block(bg_record); should already happened
slurm_mutex_unlock(&block_state_mutex);
debug("job %u finished during the queueing job "
"(everything is ok)",
req_job_id);
return;
}
if ((bg_record->state == BG_BLOCK_TERM) || bg_record->free_cnt) {
/* It doesn't appear state of a small block
(conn_type) is held on a BGP system so
if we to reset it so, just set the reboot flag and
handle it later in that code. */
bg_action_ptr->reboot = 1;
}
delete_list = list_create(NULL);
itr = list_iterator_create(bg_lists->main);
while ((found_record = list_next(itr))) {
if (bg_record == found_record)
continue;
if (!blocks_overlap(bg_record, found_record)) {
debug2("block %s isn't part of %s",
found_record->bg_block_id,
bg_record->bg_block_id);
continue;
}
if (found_record->job_ptr
|| (found_record->job_list
&& list_count(found_record->job_list))) {
struct job_record *job_ptr = found_record->job_ptr;
if (!found_record->job_ptr)
job_ptr = find_job_in_bg_record(
found_record, NO_VAL);
error("Trying to start job %u on block %s, "
"but there is a job %u running on an overlapping "
"block %s it will not end until %ld. "
"This should never happen.",
req_job_id,
bg_record->bg_block_id,
job_ptr->job_id,
found_record->bg_block_id,
job_ptr->end_time);
requeue_job = 1;
break;
}
debug2("need to make sure %s is free, it's part of %s",
found_record->bg_block_id,
bg_record->bg_block_id);
list_push(delete_list, found_record);
}
list_iterator_destroy(itr);
if (requeue_job) {
FREE_NULL_LIST(delete_list);
bg_reset_block(bg_record, bg_action_ptr->job_ptr);
bg_record->modifying = 0;
slurm_mutex_unlock(&block_state_mutex);
bg_requeue_job(req_job_id, 0, 0, JOB_BOOT_FAIL, false);
return;
}
slurm_mutex_unlock(&block_state_mutex);
if (bg_conf->layout_mode == LAYOUT_DYNAMIC)
delete_it = 1;
free_block_list(req_job_id, delete_list, delete_it, 1);
FREE_NULL_LIST(delete_list);
while (1) {
slurm_mutex_lock(&block_state_mutex);
/* Failure will unlock block_state_mutex so no need to
unlock before return. No need to reset modifying
here if the block doesn't exist.
*/
if (!_make_sure_block_still_exists(bg_action_ptr, bg_record)) {
error("Problem with deallocating blocks to run job %u "
"on block %s", req_job_id,
bg_action_ptr->bg_block_id);
return;
}
/* If another thread is freeing this block we need to
wait until it is done or we will get into a state
where this job will be killed.
*/
if (!bg_record->free_cnt)
break;
debug("Waiting for block %s to free for job %u. "
"%d thread(s) trying to free it",
bg_record->bg_block_id, req_job_id,
bg_record->free_cnt);
slurm_mutex_unlock(&block_state_mutex);
sleep(1);
}
/* This was set in the start_job function to close the above
window where a job could be mistakenly requeued if another
thread is trying to free this block as we are trying to run
on it, which is fine since we will reboot it later.
*/
bg_record->modifying = 0;
if ((bg_record->job_running <= NO_JOB_RUNNING)
&& !find_job_in_bg_record(bg_record, req_job_id)) {
// bg_reset_block(bg_record); should already happened
slurm_mutex_unlock(&block_state_mutex);
debug("job %u already finished before boot",
req_job_id);
return;
}
if (bg_record->job_list
&& (bg_action_ptr->job_ptr->total_cpus != bg_record->cpu_cnt)
&& (list_count(bg_record->job_list) != 1)) {
/* We don't allow modification of a block or reboot of
a block if we are running multiple jobs on the
block.
*/
debug2("no reboot");
goto no_reboot;
}
rc = 0;
#ifdef HAVE_BGL
if (bg_action_ptr->blrtsimage
&& xstrcasecmp(bg_action_ptr->blrtsimage, bg_record->blrtsimage)) {
debug3("changing BlrtsImage from %s to %s",
bg_record->blrtsimage, bg_action_ptr->blrtsimage);
xfree(bg_record->blrtsimage);
bg_record->blrtsimage = xstrdup(bg_action_ptr->blrtsimage);
rc = 1;
}
#elif defined HAVE_BGP
if ((bg_action_ptr->conn_type[0] >= SELECT_SMALL)
&& (bg_action_ptr->conn_type[0] != bg_record->conn_type[0])) {
if (bg_conf->slurm_debug_level >= LOG_LEVEL_DEBUG3) {
char *req_conn_type =
conn_type_string_full(bg_action_ptr->conn_type);
char *conn_type =
conn_type_string_full(bg_record->conn_type);
debug3("changing small block mode from %s to %s",
conn_type, req_conn_type);
xfree(req_conn_type);
xfree(conn_type);
}
rc = 1;
# ifndef HAVE_BG_FILES
/* since we don't check state on an emulated system we
* have to change it here
*/
bg_record->conn_type[0] = bg_action_ptr->conn_type[0];
# endif
}
#endif
#ifdef HAVE_BG_L_P
if (bg_action_ptr->linuximage
&& xstrcasecmp(bg_action_ptr->linuximage, bg_record->linuximage)) {
# ifdef HAVE_BGL
debug3("changing LinuxImage from %s to %s",
bg_record->linuximage, bg_action_ptr->linuximage);
# else
debug3("changing CnloadImage from %s to %s",
bg_record->linuximage, bg_action_ptr->linuximage);
# endif
xfree(bg_record->linuximage);
bg_record->linuximage = xstrdup(bg_action_ptr->linuximage);
rc = 1;
}
if (bg_action_ptr->ramdiskimage
&& xstrcasecmp(bg_action_ptr->ramdiskimage,
bg_record->ramdiskimage)) {
# ifdef HAVE_BGL
debug3("changing RamDiskImage from %s to %s",
bg_record->ramdiskimage, bg_action_ptr->ramdiskimage);
# else
debug3("changing IoloadImage from %s to %s",
bg_record->ramdiskimage, bg_action_ptr->ramdiskimage);
# endif
xfree(bg_record->ramdiskimage);
bg_record->ramdiskimage = xstrdup(bg_action_ptr->ramdiskimage);
rc = 1;
}
#endif
if (bg_action_ptr->mloaderimage
&& xstrcasecmp(bg_action_ptr->mloaderimage,
bg_record->mloaderimage)) {
debug3("changing MloaderImage from %s to %s",
bg_record->mloaderimage, bg_action_ptr->mloaderimage);
xfree(bg_record->mloaderimage);
bg_record->mloaderimage = xstrdup(bg_action_ptr->mloaderimage);
rc = 1;
}
if (rc || bg_action_ptr->reboot) {
bg_record->modifying = 1;
/* Increment free_cnt to make sure we don't loose this
* block since bg_free_block will unlock block_state_mutex.
*/
bg_record->free_cnt++;
bg_free_block(bg_record, 1, 1);
bg_record->free_cnt--;
#if defined HAVE_BG_FILES && defined HAVE_BG_L_P
#ifdef HAVE_BGL
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_BlrtsImg,
bg_record->blrtsimage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_BlrtsImg): %s",
bg_err_str(rc));
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_LinuxImg,
bg_record->linuximage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_LinuxImg): %s",
bg_err_str(rc));
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_RamdiskImg,
bg_record->ramdiskimage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_RamdiskImg): %s",
bg_err_str(rc));
#elif defined HAVE_BGP
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_CnloadImg,
bg_record->linuximage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_CnloadImg): %s",
bg_err_str(rc));
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_IoloadImg,
bg_record->ramdiskimage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_IoloadImg): %s",
bg_err_str(rc));
if (bg_action_ptr->conn_type[0] > SELECT_SMALL) {
char *conn_type = NULL;
switch(bg_action_ptr->conn_type[0]) {
case SELECT_HTC_S:
conn_type = "s";
break;
case SELECT_HTC_D:
conn_type = "d";
break;
case SELECT_HTC_V:
conn_type = "v";
break;
case SELECT_HTC_L:
conn_type = "l";
break;
default:
break;
}
/* the option has to be set before the pool can be
set */
if ((rc = bridge_block_modify(
bg_record->bg_block_id,
RM_MODIFY_Options,
conn_type)) != SLURM_SUCCESS)
error("bridge_set_data(RM_MODIFY_Options): %s",
bg_err_str(rc));
}
#endif
if ((rc = bridge_block_modify(bg_record->bg_block_id,
RM_MODIFY_MloaderImg,
bg_record->mloaderimage))
!= SLURM_SUCCESS)
error("bridge_block_modify(RM_MODIFY_MloaderImg): %s",
bg_err_str(rc));
#endif
bg_record->modifying = 0;
}
no_reboot:
if (bg_record->state == BG_BLOCK_FREE) {
if ((rc = bridge_block_boot(bg_record)) != SLURM_SUCCESS) {
char reason[200];
bg_record->boot_state = 0;
bg_record->boot_count = 0;
if (rc == BG_ERROR_INVALID_STATE)
snprintf(reason, sizeof(reason),
"Block %s is in an incompatible "
"state. This usually means "
"hardware is allocated "
"by another block (maybe outside "
"of SLURM).",
bg_record->bg_block_id);
else
snprintf(reason, sizeof(reason),
"Couldn't boot block %s: %s",
bg_record->bg_block_id,
bg_err_str(rc));
slurm_mutex_unlock(&block_state_mutex);
requeue_and_error(bg_record, reason);
return;
}
} else if (bg_record->state == BG_BLOCK_BOOTING) {
#ifdef HAVE_BG_FILES
bg_record->boot_state = 1;
#else
if (!block_ptr_exist_in_list(bg_lists->booted, bg_record))
list_push(bg_lists->booted, bg_record);
bg_record->state = BG_BLOCK_INITED;
last_bg_update = time(NULL);
#endif
}
if ((bg_record->job_running <= NO_JOB_RUNNING)
&& !find_job_in_bg_record(bg_record, req_job_id)) {
slurm_mutex_unlock(&block_state_mutex);
debug("job %u finished during the start of the boot "
"(everything is ok)",
req_job_id);
return;
}
/* Don't reset boot_count, it will be reset when state
changes, and needs to outlast a job allocation.
*/
/* bg_record->boot_count = 0; */
if (bg_record->state == BG_BLOCK_INITED) {
debug("block %s is already ready.", bg_record->bg_block_id);
/* Just in case reset the boot flags */
bg_record->boot_state = 0;
bg_record->boot_count = 0;
set_user_rc = bridge_block_sync_users(bg_record);
block_inited = 1;
}
slurm_mutex_unlock(&block_state_mutex);
/* This lock needs to happen after the block_state_mutex to
avoid deadlock.
*/
if (block_inited && bg_action_ptr->job_ptr) {
slurmctld_lock_t job_write_lock = {
NO_LOCK, WRITE_LOCK, NO_LOCK, NO_LOCK, NO_LOCK };
lock_slurmctld(job_write_lock);
bg_action_ptr->job_ptr->job_state &= (~JOB_CONFIGURING);
last_job_update = time(NULL);
unlock_slurmctld(job_write_lock);
}
if (set_user_rc == SLURM_ERROR) {
sleep(2);
/* wait for the slurmd to begin
the batch script, slurm_fail_job()
is a no-op if issued prior
to the script initiation do clean up just
in case the fail job isn't ran */
(void) slurm_fail_job(req_job_id, JOB_BOOT_FAIL);
}
}
extern int bg_status_update_block_state(bg_record_t *bg_record,
uint16_t state,
List kill_job_list)
{
bool skipped_dealloc = false;
kill_job_struct_t *freeit = NULL;
int updated = 0;
uint16_t real_state = bg_record->state & (~BG_BLOCK_ERROR_FLAG);
if (real_state == state)
return 0;
debug("state of Block %s was %s and now is %s",
bg_record->bg_block_id,
bg_block_state_string(bg_record->state),
bg_block_state_string(state));
/*
check to make sure block went
through freeing correctly
*/
if ((real_state != BG_BLOCK_TERM
&& !(bg_record->state & BG_BLOCK_ERROR_FLAG))
&& state == BG_BLOCK_FREE)
skipped_dealloc = 1;
else if ((real_state == BG_BLOCK_INITED)
&& (state == BG_BLOCK_BOOTING)) {
/* This means the user did a reboot through
mpirun but we missed the state
change */
debug("Block %s skipped rebooting, "
"but it really is.",
bg_record->bg_block_id);
} else if ((real_state == BG_BLOCK_TERM)
&& (state == BG_BLOCK_BOOTING))
/* This is a funky state IBM says
isn't a bug, but all their
documentation says this doesn't
happen, but IBM says oh yeah, you
weren't really suppose to notice
that. So we will just skip this
state and act like this didn't happen. */
goto nochange_state;
real_state = state;
if (bg_record->state & BG_BLOCK_ERROR_FLAG)
state |= BG_BLOCK_ERROR_FLAG;
bg_record->state = state;
if (real_state == BG_BLOCK_TERM || skipped_dealloc)
_block_is_deallocating(bg_record, kill_job_list);
else if (real_state == BG_BLOCK_BOOTING) {
debug("Setting bootflag for %s", bg_record->bg_block_id);
bg_record->boot_state = 1;
} else if (real_state == BG_BLOCK_FREE) {
/* Make sure block is cleaned up. If there are
* running jobs on the block this happens when they
* are cleaned off. */
if (bg_record->job_running == NO_JOB_RUNNING
&& (!bg_record->job_list
|| !list_count(bg_record->job_list)))
bg_reset_block(bg_record, NULL);
remove_from_bg_list(bg_lists->booted, bg_record);
} else if (real_state & BG_BLOCK_ERROR_FLAG) {
if (bg_record->boot_state)
error("Block %s in an error state while booting.",
bg_record->bg_block_id);
else
error("Block %s in an error state.",
bg_record->bg_block_id);
remove_from_bg_list(bg_lists->booted, bg_record);
trigger_block_error();
} else if (real_state == BG_BLOCK_INITED) {
if (!block_ptr_exist_in_list(bg_lists->booted, bg_record))
list_push(bg_lists->booted, bg_record);
}
updated = 1;
last_bg_update = time(NULL);
nochange_state:
/* check the boot state */
debug3("boot state for block %s is %d",
bg_record->bg_block_id, bg_record->boot_state);
if (bg_record->boot_state) {
if (bg_record->state & BG_BLOCK_ERROR_FLAG) {
/* If we get an error on boot that
* means it is a transparent L3 error
* and should be trying to fix
* itself. If this is the case we
* just hang out waiting for the state
* to go to free where we will try to
* boot again below.
*/
return updated;
}
switch (real_state) {
case BG_BLOCK_BOOTING:
if (bg_record->job_ptr
&& !IS_JOB_CONFIGURING(bg_record->job_ptr)) {
debug3("Setting job %u on block %s "
"to configuring",
bg_record->job_ptr->job_id,
bg_record->bg_block_id);
bg_record->job_ptr->job_state |=
JOB_CONFIGURING;
last_job_update = time(NULL);
} else if (bg_record->job_list
&& list_count(bg_record->job_list)) {
struct job_record *job_ptr;
ListIterator job_itr =
list_iterator_create(
bg_record->job_list);
while ((job_ptr = list_next(job_itr))) {
if (job_ptr->magic != JOB_MAGIC) {
error("bg_status_update_"
"block_state: 1 "
"bad magic found when "
"looking at block %s",
bg_record->bg_block_id);
list_delete_item(job_itr);
continue;
}
job_ptr->job_state |= JOB_CONFIGURING;
}
list_iterator_destroy(job_itr);
last_job_update = time(NULL);
}
break;
case BG_BLOCK_FREE:
if (bg_record->boot_count < RETRY_BOOT_COUNT) {
bridge_block_boot(bg_record);
if (bg_record->magic == BLOCK_MAGIC) {
debug("boot count for block %s is %d",
bg_record->bg_block_id,
bg_record->boot_count);
bg_record->boot_count++;
}
} else {
char *reason = (char *)
"status_check: Boot fails ";
error("Couldn't boot Block %s",
bg_record->bg_block_id);
/* We can't push on the kill_job_list
here since we have to put this
block in an error and that means
the killing has to take place
before the erroring of the block.
*/
slurm_mutex_unlock(&block_state_mutex);
unlock_slurmctld(job_read_lock);
requeue_and_error(bg_record, reason);
lock_slurmctld(job_read_lock);
slurm_mutex_lock(&block_state_mutex);
bg_record->boot_state = 0;
bg_record->boot_count = 0;
remove_from_bg_list(bg_lists->booted,
bg_record);
}
break;
case BG_BLOCK_INITED:
debug("block %s is ready.",
bg_record->bg_block_id);
if (bg_record->job_ptr
&& IS_JOB_CONFIGURING(bg_record->job_ptr)) {
bg_record->job_ptr->job_state &=
(~JOB_CONFIGURING);
last_job_update = time(NULL);
} else if (bg_record->job_list
&& list_count(bg_record->job_list)) {
struct job_record *job_ptr;
ListIterator job_itr =
list_iterator_create(
bg_record->job_list);
while ((job_ptr = list_next(job_itr))) {
if (job_ptr->magic != JOB_MAGIC) {
error("bg_status_update_"
"block_state: 2 "
"bad magic found when "
"looking at block %s",
bg_record->bg_block_id);
list_delete_item(job_itr);
continue;
}
job_ptr->job_state &=
(~JOB_CONFIGURING);
}
list_iterator_destroy(job_itr);
last_job_update = time(NULL);
}
bg_record->boot_state = 0;
bg_record->boot_count = 0;
if (kill_job_list &&
bridge_block_sync_users(bg_record)
== SLURM_ERROR) {
freeit = (kill_job_struct_t *)
xmalloc(sizeof(kill_job_struct_t));
freeit->jobid = bg_record->job_running;
list_push(kill_job_list, freeit);
}
break;
case BG_BLOCK_TERM:
debug2("Block %s is in a deallocating state "
"during a boot. Doing nothing until "
"free state.",
bg_record->bg_block_id);
break;
case BG_BLOCK_REBOOTING:
debug2("Block %s is rebooting.",
bg_record->bg_block_id);
break;
default:
debug("Hey the state of block "
"%s is %d(%s) doing nothing.",
bg_record->bg_block_id,
real_state,
bg_block_state_string(bg_record->state));
break;
}
}
return updated;
}
