Whe PrologFlags is not set, the job prolog runs just before the first
step of the job.

Previously, when PrologFlags was not set, the job was not killed on a
prolog failure. Although the step that triggered the prolog failure
would not run, the job's batch script would continue to run and execute
any remaining steps without re-running the prolog.

Bug 10353

slurmd relied on ESLURM_BATCH_ONLY being returned to know if it was a
salloc/srun job and sent a message to the job's stdout about prolog
failure. However, in _job_requeue_op(), a few conditions for returning
ESLURM_DISABLED happen before the condition for returning ESLURM_BATCH_ONLY
so this isn't reliable. Instead this commit always sends the message to
the job's stdout about a prolog failure (if there was a prolog failure)
even to batch jobs.

Bug 10353

When a prolog fails, slurmd sends a request to slurmctld to requeue the
job. If the requeue doesn't work (the job used --no-requeue or
JobRequeue=0) then slurmd sends a request to slurmctld to complete the
batch job. This can be wrong for a couple of reasons:

(1) The job isn't necessarily a batch job. It could have been started by
salloc or srun.

(2) If a multi-node job had a prolog fail on a node that wasn't the batch
host and requeuing was disabled, when slurmd sends the complete batch
job request to slurmctld, slurmctld ignores that request because it didn't
originate from the batch host. See _slurm_rpc_complete_batch_script().
Therefore, the job isn't killed and runs to completion, even though the
prolog failed and the node was drained.

This commit sends a kill job request instead of a complete batch job
request, ensuring that the job is killed.

Bug 10353

Bug 10353

beforehand.

Bug 10675

Ensure the credential is created with the same protocol as the agent,
which should be the oldest version from job's nodes

Bug 10884

Bug 10561

By sending in the local cluster list and changing the query to only
give the default wckey first and then the newest one will always be the
one we just added making it not needed to have FOR UPDATE specified.

This also makes it so we only query non-deleted wckeys.

Bug 10561

By sending in the local cluster list and changing the query to only
give the default assoc first and then the newest one will always be the
one we just added making it not needed to have FOR UPDATE specified.

This also makes it so we only query non-deleted assocs.

Bug 10561

Bug 10561

Bug 10561

EnforcePartLimits=any and partition based associations.

Bug 10229

This will be used in the following commit.

Bug 10229

Regression typo added merging up on f8f6dd96.

Bug 8811

Bug 10917

Bug 10891.

This corrects a condition where the bad line is "scrontab" and one where
an invalid option won't be detected if its not the first one in the list.

Bug 10891

Extra refactor removing get_node_cnt_in_part.

The following procs have been replaced by the new ones:
- get_idle_node_in_part
- get_partition_nodes
- available_nodes_hostnames
- available_nodes

Bug 8811

Signed-off-by: Marcin Stolarek <cinek@schedmd.com>

Bug 8811

Signed-off-by: Marcin Stolarek <cinek@schedmd.com>

Was converting unsigned to signed int.

Bug 10568

Use get_partition_param instead of available_nodes_hostnames to get all
nodes in partition regardless of state.

Bug 8811.

Get rid of strtok() and the extra xstrdup() that was covering up the
issue from the previous commit and use xstrncasecmp() instead.

Bug 10895

strtok() is meant to run until it returns NULL. This is because it actually
modifies the first argument in the meantime. Since we only called it once
while parsing SchedulerParameters in the main scheduler, it basically
truncated everything after bf_hetjob_prio. The solution is to simply use
xstrncasecmp() instead.

The backfill scheduler is unaffected due to 5b833ff3.

Bug 10895

Update slurm.spec as well.

Bug 10876

When using containers (such as linux containers or docker), slurmds won't
start because when they call xcgroup_ns_is_available to check that they are
in a cgroup, it looks for the presence of the release_agent file. However,
the release_agent file is unique in that it is only present in the root
cgroup. Since containers are namespace-bound to a container-specific child
subdirectory in each control group, it fails to find release_agent and the
slurmd will fail to start.

The fix relies on checking a 'tasks' file instead of 'release_agent'.

Bug 10468

Introduced in 30af622a, the patch added a new case in reverse_tree_info()
which filled in parent_rank with -1 when width > nodes. This lead to callers
like _send_slurmstepd_init() to attempt to find the slurm_addr_t of this
node's parent slurmd in the step host list with parent_rank = -1 when not
applicable.

This resulted in misleading errors and messages being logged (even with
malformed node names) and incorrect reverse-tree info sent to the stepd.

Bug 10838.

Bug 10798

Bug 10798

On broken glibc versions, the zeroes in the association state file will
be saved as "nan" in packlongdouble(). Detect if this has happened in
unpacklongdouble() and convert back to zero.

https://bugzilla.redhat.com/show_bug.cgi?id=1925204

Bug 10824.

full_queue needs to be reset after the call to _schedule().
Regression from commit 2427f291.

Bug 10818.

Bug 10542.

To specify a file to write the json testsuite results to.

Bug 10798

sched was removed as an option for TaskPluginParam

Bug 10780

Signed-off-by: Tim Wickberg <tim@schedmd.com>

Bug 10673