This also clears up a potential race around ping_thread_cnt
as it was protected by ping_mutex in one location and shutdown_mutex
in another.

- Do not look it up again in the backup controller.
- In the backup controller, stop comparing it at all and instead
  use the backup_idx value to decide if we're outselves.

This is not a valid response here - backup and primary
must always be running the same version, so do not attempt to
handle this here.

There's no point in pinging controllers with a lower priority
than yourself - they'll already be pinging you. As we did nothing
with that data, don't bother to collect it, especially as lower
priority controllers being unavailable will delay the next pass
through this loop.

Reference backup_inx directly after startup, and exit much
earlier if this host is not a valid controller. Return a
non-zero exit code in this situation as well.

Collapse into a single function so we can appropriately warn
if a mix of options are in use.

This also avoids a confusing-looking xmalloc with the count padded by two,
which was being used to build out space for ControlMachine if SlurmctldHost
was not defined. This would have also masked off a series of off-by-one
errors, and has lead to attempts to connect to 0.0.0.0 instead of a segfault.
(Some code was intentionally using this over-provisioning as a way to
treat this as a NULL-terminated list, but this was then technically
incorrect in cases where the old-style BackupController was set since the
NULL would happen at the third position in the array, which is an invalid
memory access.)

And document why these are handled the way they are here.

This results in an out-of-bounds access (if control_machine was not
being intentionally over-alloced to avoid it), the wrong address,
and other subtle problems.

C's order of operations meant this was resolving as:
i = (_backup_index() != -1);
which is either 0 or 1.

Through sheer luck, this still results in the correct answer for the primary
(_backup_index() is -1, and then i = (-1 != -1) is still 0 which is correct),
and first backup controller (_backup_index() is 1, and then i = (1 != -1) is
still 1 which is also correct), but any further backups controllers will end
up with the address of the first backup.

Use the already-established slurmctld_primary bool instead.

The second NULL check is a duplicate of the first.

The first check is also unnecessary - this field existing is managed by
the control_cnt variable. (At one point in development control_cnt did not
exist, and control_machine was a 0-terminated array instead.)

Otherwise cross-architecture failover will break in confusing ways.

Bug 5256.

plugin.

Bug 5583

Caught by verify_lock() annotation in job_submit_plugin_modify().

Bug 5578.

Caught by verify_lock() annotations in validate_job_create_req()
and job_allocate() respectively.

Bug 5578.

Brian approved

Brian approved

Brian approved

This is only to quite coverity.  I don't think this is a real problem.

Brian approved.

Fix coverity 187747

(i.e. gres=gpu/tesla) it would get a count of 0.

Bug 5570

Note that pipe2() is Linux-specific, but this whole cgroup plugin is
already Linux-specific in design, and the eventfd() call itself is
Linux-specific as well.

Bug 5570.

Didn't realize centos 6 was so far behind.

Bug5503

"IN" tres_usage is reads
"OUT" tres_usage is writes

page faults are only "ins"

In 18.08 the jobacct_info stats were changed to tres_usage arrays which
are init'ed to INFINITE64 but < 18.08 it was init'ed to 0's.

Bug 5554

In 18.08, usage values are stored as bytes but in <18.08 values were
stored in KB, MB, etc.

Also handle cputime adjustments. Related commit: efc161ef.