Bug 5708

This matches the functionality in code.

Bug 5726

Bug 5795

Continuation of 06582da8 (17.11.9)

Poll was timing out too quickly due to an incorrect conversion of
MessageTimeout. Added a multiplier so timeout reflects the correct
millisecond value.

Bug 5553

Coverity CID 188467

"set errno to indicate error" made some think that the function had
to explicitly set the errno before returning, when it is often done
by sub functions.

Bug 5376

Replace these with either SLURM_ERROR or SLURM_SUCCESS

Bug 5392

# Conflicts:
#	src/common/gres.c

# Conflicts:
#	src/common/slurm_protocol_pack.c
#	src/slurmctld/controller.c

This adds the ability for a step to explicitly clear a job's gres counter.
For example, the job requests --gpus-per-task=2 and the step requests
--gpus-per-task=0.

change some comments and log messages.
no change in functionality

failing to propagate gpus-per-task for job step

Since sacct exits the program when it finds unknown parameters, the
debug2() statements should be escalated to error() statements.
The first debug2() even had "Error: " in it.
Get rid of extra newline.

Bug 5421

Bug 5786

Bug 5165

Bug 5743.

Co-authored-by: Dominik Bartkiewicz <bart@schedmd.com>

Move reused GRES parsing logic to a shared function.

(Corrected author; was previously mis-committed as 433257c6.)

Bug 5743.

This reverts commit 433257c6.

This just tests the gpus-per-node option. Additional tests are
needed for gpus-per-task, etc. as well as the logic to support
those options for job steps.

Bug 5165

It was still possible for job priority to go down to 0, because I was
letting last_prio get set down to 1, then later on we were using
last_prio-1 as the new job priority. Since last_prio was 1, then new
job priority was 1-1=0. This happened when the job priorities were all
at 2 (which I never tested). Michael found it by running test2.26 in a
case where initial job priorities were 2:

spawn /home/marshall/slurm/18.08/voyager/bin/squeue --name=test2.26 -o JOB_ID=%A PRIO=%Q
JOB_ID=JOBID PRIO=PRIORITY
JOB_ID=10702 PRIO=2
JOB_ID=10703 PRIO=2
JOB_ID=10704 PRIO=2
Reset and test job priorities

spawn /home/marshall/slurm/18.08/voyager/bin/scontrol top 10704,10703
spawn /home/marshall/slurm/18.08/voyager/bin/squeue --name=test2.26 -o JOB_ID=%A PRIO=%Q
JOB_ID=JOBID PRIO=PRIORITY
JOB_ID=10704 PRIO=2
JOB_ID=10703 PRIO=1
JOB_ID=10702 PRIO=0
Reset and test job priorities

Fix memory leak in previous patch.

Bug 5479

Coverity 188464

The format for this field is "<inst>:<svc_id>:<comp_id>"; if inst or svc_id
are unset then they're blank, and the parsing code had assumed they would
always be blank as the systems slurmsmwd were developed on never used them.

Bug 5411.

If a job with running steps is resized, that will result in the job's
core_bitmap size changing, but not that of the step, resulting in an
error message. This adds a job resized flag and prints a more meaningful
message when appropriate.

change to bit_ffs/bit_fls rather than full scan, improves performance

check for TaskPlugin != none at start
Print all hostnames associated with launched job (to diagnose problems)

Prevents job priorities from being lowered to 0 or from underflowing.

Bug 5165

Bug 5231

Change slurmctld/controller.c's purge_thread_lock from static to
global extern, so the mutex can be used in slurmctld/job_mgr.c

slurm_cond_signal() needs to always be wrapped in the same mutex as
slurm_cond_[timed]wait(), or else there is a possibility that
slurm_cond_signal() will trigger before slurm_cond_[timed]wait()
is even listening, most likely causing a deadlock.

Bug 5703

initialize old fairshare association pointer correctly.

Bug 5744

Bug 5458

Bug 5403

Coverity CID 187779

Previous logic could abort in some cases