If the pointer is NULL then the size_val should be zero or you'll be
packing garbage.

Compare the release locks to the acquired locks.

More interestingly, in further commits this will allow us to annotate
functions to verify the appropriate locks are held at runtime, and
head off potential sources of corruption.

Only enabled on development builds.

Calling lock_slurmctld() while already holding locks leads to deadlock.
Intentionally cause such instances to abort() on development builds to
force these locations to the surface.

when using fed_mgr_is_active(). It should only be used outside of any
locks.

Prior security check only permitted root or the SlurmUser to update.

No way to do this through sbatch, but the API would permit it.

Bug 3631.

When not in a federation, or using --local, just go directly to the
cluster without spawning a thread.

Instead of packing an empty federation rec, just pack a NULL and return
a NULL.

This introduced a race condition between RPC messages starting
to be processed ahead of the plugin init() calls completing.

This reverts commit 6e65b5cb.

Bug 3646.

Only happens in a SPANK plugin on the setenv() call.
dyn_spank_set_job_env() in unaffected, and needs the NULL value for proper
backwards-compatibility.

Bug 3503.

This should have been part of commit f84456f3

This fixes test 7.17 to work with commit a152196f

Move the working_cluster_rec reference out of the following functions:
slurm_send_recv_controller_msg(), slurm_send_recv_controller_rc_msg(),
slurm_open_controller_conn_spec() and slurm_send_only_controller_msg()
so that communcations to multiple clusters can be performed in parallel
(using threads)

If a cluster in our federation is down, it's "control_host" value
is "" and an attempt to communicate with it will abort.

Previous content has news about version 16.05 release. Updated
for version 17.02 release.

[min|max]_cpus if not already set when a job comes in.

Change slurm_load_jobs(), slurm_load_job_user() and slurm_load_job()
to all use the same underlying function, which was formerly just
used by slurm_load_jobs(). This function will get data from all
of the clusters in a federation, but in a serial fashion for now.

The clang error isn't entirely clear, but this should fix it.

the job may still point to the qos_id of the removed qos.