If this failed, the step would launch as root instead of the desired
user, which could be exploited.

Bug 8084.

CVE-2019-19728.

Encourage the use for restricted permissions for slurmdbd.conf since
this file will contain your MySQL username and password on most systems.

CVE-2019-19727.

Bug 7861

In a multi-cluster environment a job may submit more jobs as part of
its workflow. This fixes situations where the variable is inherited
incorrectly on sub-jobs.

Bug 7998

srun waits for the prolog to finish before launching a job step.
In _is_prolog_finished(), slurmctld checks the state reason:

	if (job_ptr) {
		is_running = (job_ptr->state_reason != WAIT_PROLOG);
	}

But if the job is updated during the job prolog, then _update_job() will
change the state_reason, and then slurmctld will tell srun that the
prolog is completed even if it isn't. If srun launches a job step before
the extern sets up x11, then the job step won't have x11 information. To
fix this, don't change state_reason in _update_job() if it equals
WAIT_PROLOG.

Bug 7525

This in turn fixes allocation requests that weren't rejected and they
should because the requested nodes didn't have a shared network.

Bug 8210

Bug 8047

Bug 7766

Have the main scheduler and backfill scheduler make the reasons of
subsequent array tasks match the first array task, since they
sometimes didn't do this completely when the array was pending.

Bug 6814

Add _at_tpn_limit() as helper to determine when a given node is over the
tasks_per_node limit and to log when then happens.

Bug 7629.

Signed-off-by: Jason Booth <jbooth@schedmd.com>

Bug 7189

Bug 8153

Bug 7987

Co-authored-by: Broderick Gardner <broderick@schedmd.com>
Signed-off-by: Broderick Gardner <broderick@schedmd.com>

When an allocation request was done with the immediate=1 argument and
SchedulerParameters included defer, Slurm was returning a misleading
ESLURM_FRAGMENTATION error. Logic now a returns a more appropriate
ESLURM_CAN_NOT_START_IMMEDIATELY error for this scenario by decoupling
defer from the too fragmented logic in job_allocate().

Note that this doesn't change behavior as immediate + defer combination
continues having defer as the king in terms of precedence order, meaning
individual submit time allocation attempts will be avoided independently
of immediate.

Bug 5175.

This effectively reverts commit 73351553. That commit's message is,

     "Improve support for overlapping advanced reservations.
      Patch from Bill Brophy, Bull."

Jobs submitted to reservations that request more resources than are on a
node will pend forever because of that commit. Reverting that commit
causes those jobs to be immediately rejected. Also, that commit doesn't
appear to "improve support for overlapping advanced reservations" in any
way.

The job is already immediately rejected if it asks for more resources
than are on a node without being submitted to a reservation, or if the
job asks for more nodes than are currently in the reservation. So, this
commit just makes behavior consistent.

Bug 5175.

Do not assume that these sock_gres_t pointers always exist:
bits_by_sock
bits_by_sock[s]

If they don't, that means there are no current iteration socket `s`
constrained GRES and so the logic shouldn't allocate the current
iteration GRES `g`.

Analogously, do not assume that bits_any_sock sock_gres_t member pointer
is always valid. If it isn't, it means there are no socket-unconstrained
GRES available to satisfy the job request, so the logic should not
allocate the current iteration GRES `g`.

Otherwise, job/node struct members holding GRES allocation information
would end up being incorrect, leading to improper allocations and also
leading to errors logged in slurmctld log at deallocation time like:

error: gres/gpu: job <X> dealloc node <Y> GRES count underflow (0 < 1)

Bug 7827

For older RPCs we should initialize db_flags with SLURMDB_JOB_FLAG_NOTSET.
(Which is treated differently than SLURMDB_JOB_FLAG_NONE, which is 0.)

Bug 8029.

Remove the TIME_FLOAT flag from the reservation to ensure _job_overlap()
does not add the current time on top of the start_time. The prior
approach was incorrect for non-TIME_FLOAT reservations and would
lead to valid reservations being rejected.

Bug 7458, 7908.

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

Bug 7169

Previously it was only after being idle. The problem was that if the
node was downed after a job ran on the node for more than SuspendTime
the node would be suspended quickly. Now it waits SuspendTime after
being idle or down (i.e. since no jobs on the node).

Bug 6774

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

CUDA_VISIBLE_DEVICES was not being set to the correct GPU indexes when
cgroups were being used. These issues were exhibited with at least the
map_gpu and mask_gpu binding options.

The issue was that usable_gres is a bitmask of GRESs in the step's
cgroup, but bit_test() was looking at bit i, which is the index of the
global gres_list (not constrained by cgroups).

Bug 7509

In 19.05 JOB_MEM_SET flag was added along with a conditional check on
this flag that changed the pn_min_memory when validating job limits.
This caused that after an upgrade, PD jobs in earlier versions didn't
have this flag and the memory was incorrectly set when their limits were
checked before starting. The patch here addresses this issue adding this
flag to jobs from an older protocol version when loading the state
files.

Bug 8011

Previously, coordinators could delete specific associations, but could
not delete users. Allow coordinators to delete users if the users are
only part of accounts that the coordinator is over.

Bug 7413.

Bug 7103.

Bug 7875

Bug 7830

Bug 7584

Previously sched_nodes was set to the estimated nodes on the last
evaluated partition that was adding a reservation, instead of the
one offering the earliest estimated start time.

Natural continuation of fdae6a05

.

Bug 7344.

Signed-off-by: Dominik Bartkiewicz <bart@schedmd.com>

Bug 8014

Bug 7749

gres_node_config_load() requires gres_list to work properly after
fully merge slurm.conf with gres.conf logic added in 4d7df8b0.

Bug 7986

Bug 7490

If not enforcing QOS, it's possible to submit a job without a qos. If
submitting such a job to multiple partitions where at least one has a
qos, slurmctld would abort in a development build. A non-development
build didn't segfault only because _find_qos_part doesn't dereference
the NULL pointer. Prevent the abort.

Bug 7171

Bug 7712

Bug 7884.

Previous logic would only call s_p_hashtbl_create() to create the hashtable
when the file acct_gather.conf could be successfully stat()'d. This lead to
a subsequent attempt to pack the non-created hashtable into a buffer which
triggered the abort.

This makes it so the hashtable is uncondtionally created no matter if the
file is missing.

Bug 7893.

gethostbyaddr() can potentially return a fully-qualified domain name,
which breaks backwards compatibility with the shortname AllocNodes
expected pre 19.05.

Bug 7653.

Fortunately the extra arguments were provided at the end, and thus ignored on
most common platforms.

Bug 7555.