Avoid blindly overwriting job wait reason when job unable to run due to
partition configuration. As jobs finish, they can cause pending jobs to
randomly change reason to priority and back to the real reason for their
waiting.

Bug 8347.

Command line tools option is stored as signed int, which is rather
int16_t than uint32_t. Value of INFINITE is effectively -1 which is used
by sbatch/salloc as "no val" effectively preventing the
job_desc.ntasks_per_core being set as a result of --hint=multithread.

On the slurmctld side, ntasks_per_core == INFINITE16 (see
_set_multi_core_data()) is used as a "no val". In the case of
CR_ONE_TASK_PER_CORE, using --multithread wasn't overriding this and
allowing more than one task per core.

Bug 8147

Error's out pointing to possible database corruption.

Bug 7583

The test runs sacctmgr show for various entities and checks if all
headers are accepted in format= specification.

Bug 7201

Both MaxSubmitPA and MaxSubmitPU are displayed in sacctmgr show qos
header, so we should handle it correctly when specified in format=
option.

Bug 7201

Bug 8834.

No functional change.

Bug 8834.

Adjust regex in response to changes in GRES format in scontrol show job
added on commit (51a95b70)

Bug 8799

Commit 6eef5739 broke 1.88 and 38.7 on systems without strlcpy.
We don't want to add any extra/slurm dependency for MPI jobs in tests
1.88 and 38.7, and we want them to run on systems without strlcpy too.
This commit restore the original code of such tests.

Changes on test7.17 are kept as it is already depending on slurm,
and a candidate to become a unit test.

Bug 7265

Signed-off-by: Nate Rini <nate@schedmd.com>

Bug 8563

This is a partial revert of commit 74184e4d. From that commit message:

"At the start of a scheduling cycle, the job's "reason" field can
be cleared. If the scheduler fails to reach that job and set its
value to a new reason, the original reason was lost and the state
reports would report NoReason."

This isn't true anymore so we revert that change (packing the previous
state_reason) and just pack the job's state_reason.

Bug 8848

# Conflicts:
#	src/plugins/select/cons_tres/dist_tasks.c

Revert the changes from 6dcb598c and break appropriate loop.

Bug 8818.

Co-authored-by: Danny Auble <da@schedmd.com>

This was missed in commit 60c67e82

Bug 8859

(1) In _pick_step_nodes(), we get a bitmap (nodes_idle) of "idle" nodes
in the job - in other words, nodes that don't have any steps running
on them.

(2) A little farther down, we check if nodes_idle has any nodes, and
try to allocate resources from the idle nodes for the step.

(3) Then we check if there are any "available" nodes in the job (nodes
that have some resources free but aren't completely idle), and try
to allocate resources from them. We will fill up a node before
allocating from the next one.

The problem: when we check if nodes are idle[1], we check which nodes
are being used by all running steps, *including* the extern and batch
steps. This means that with PrologFlags=contain (which enables the
extern step), there will never be any "idle" nodes for a job because
the extern step runs on all the nodes. We will skip section (2) and go
straight to section (3). Steps will still run, but where the steps are
allocated is different than without the extern or batch steps.

The buggy behavior is this: We pack all the steps onto the first node,
then we pack all the steps onto the second node, and so on.

The fixed behavior is the same as without a batch step or extern steps:
we run new steps on idle nodes first, then we pack subsequent steps on
each node until it is filled (only moving on to the next node once the
current node is filled or doesn't have enough resources for the step
request).

Bug 8020.

Bug 8584

Co-authored-by: Brian Christiansen <brian@schedmd.com>

# Conflicts:
#	src/plugins/select/cons_tres/select_cons_tres.c

Bug 8370

# Conflicts:
#	src/plugins/select/cons_tres/job_test.c

Continuation of commit b00ce193, fixes declaration of opt variable in
scancel.

Bug 8370

GCC 10 defaults to -fno-common which implies we cannot declare static
global variables in .h files. The patch here fixes the building issues
for the current version removing an unnecessary block in cons_tres code.

Bug 8370

Previously if a job was ended for example with OUT_OF_MEMORY and we call
wait_for_job DONE, the function will keep pulling until
max_job_state_delay.
Now all DONE state are checked, in order.
NOT_FOUND is not a real state, but manually created in the function.

Bug 8837

Signed-off-by: Michael Hinton <hinton@schedmd.com>

Test 1.35 was getting OOM in the steps.
The patch also specifies the memory unit to avoid problems
if user "default_gbytes".

Bug 8837

Signed-off-by: Michael Hinton <hinton@schedmd.com>

We need to ensure that memset is not optimized by removing the -O at
compile time, and touching some of the memory, plus marking it volatile.

Bug 8755

to be reverted in 20.11 when network is no longer overloaded with a new
option.

Continuation of 6ee50489

Bug 7039

Signed-off-by: Danny Auble <da@schedmd.com>

Continuation of 6ee50489

Bug 7039

Signed-off-by: Danny Auble <da@schedmd.com>

Bug 8734

Bug 7265

Add XCC plugin existence in acct_gather man page and html,
and in slurm.conf

Bug 8484

e.g. if job_hash[JOB_HASH_INX()] contains jobs but not the job we are
looking to remove.

Bug 8861

The documentation of the preemption parameters tries to be as close
as possible to their documentation in slurm.conf.

Bug 8710

There is no GraceTime related to an assoc, it was always
showing 0 using the default print_routine with NULL value.

Bug 8710

Improves and fixes the documentation about Preemption and Gang,
specially for preempt/qos and Suspend on both slurm.conf and
sacctmgr man pages.

Bug 8710