Found using test38.17

Otherwise poll() cannot monitor these ports properly,
leading to potential network traffic problems.

Bug 4467.

until the prolog and extern step are fully ran/launched.

Only matters if running with PrologFlags=[contain|alloc].

patch 2 of 2

Bug 4458

Patch 1 of 2

Bug 4458

Due to the way Cray builds Slurm, the prefix and bindir paths include the
Slurm version (/opt/slurm/<version>). This means every time we update to a
new Slurm version we must update the Slurm ansible playbook. It also means
that the slurm_playbook.yaml file must be built with Slurm to be used
(it can't simply be copied directly).

The attached patch updates the playbook to determine the version of Slurm
to use from the module file, and hardcodes the sysconfdir setting we give
in our Slurm installation guide. If a customer uses different paths, they
can update the playbook to meet their needs.

Bug 4360.

when trying to signal a step that is still running a prolog.

Bug 4446

Bug 4446

Bug 4446

Bug 4131

Simplify the step prefix process
and move it as early as possible in the step.

Since NO_VAL = SLURM_BATCH_SCRIPT, the else statement would only compare
the job_id and not the step_id, thus when a batch step was removed all
the steps from that job would be removed too. Then when attempting to
remove the extern step in the next iteration, it was already removed
and we were incorrectly erroring out.

Bug 4458.

Missed some debug statements

Bug 4455

This makes it so if the user doesn't request any GRES they are denied
all the GRES with devices.  Before this would only happen for the GRES
type the user requested.

Bug 4455

Fix to purge old jobs using burst buffer if slurmctld daemon restarted
    after the job's burst buffer work was already completed.

Bug 4455

Bug 4378.

Fix memory leak of MailDomain configuration string when slurmctld daemon
   is reconfigured.
bug 4272 (comment 35)

Bug 4384

Bug 4450

the libslurmfull.

In my testing with the cray job_submit plugin enabled it caused a
delay of the last step most likely because the gres wasn't ready quick
enough.

This reverts commit f9977ee5.

Since it's not needed in 17.11

slurm_load_job() prior to 17.11 returns the error code in errno and not
in rc.  With the addition of 47175901 if a job is removed from memory
before sbatch checks for the job again, sbatch could get in an loop
checking for the job. This only happens if you have a very small
MinJobAge (<10) -- which is not recommended.

Bug 4445

Bug 4444

Update slurm.spec and slurm.spec-legacy as well

Bug 4323