Fix shutdown race condition that could cause the plugin to deadlock.
Improve validation of burst buffer options.
Hold job with an burst_buffer specification later discovered to be bad.

intended for the head of the tree.

Handle out-of-band (i.e. outside of Slurm) persistent burst buffer
  creation/deletion with respect to their reservations, as used
  for limit enforcement.

Avoid creating a duplicate burst buffer record due to a race condition
between the "create_persistent" thread and the "load_state" thread.

Improve precisions of poll timeout logic

Previously we reported an underflow of the node's comp_job_cnt
  (actually the message reported an underflow of the run_job_cnt,
  but was really checking the value of comp_job_cnt, in any case,
  the count is cleared when the node goes down as we are not
  counting on getting any responses from the down node).

Major refactoring of burst buffer option parsing for interactive
  job support

The --bbf option reads the job's burst buffer options from a file
rather than an inline option

meaning meaning the limit is removed.

wouldn't get the correct protocol version to launch a step.

wait for job TRES limit check before starting burst buffer allocation

Trigger the job scheduler as soon as the last persistent buffer
operation for a pending job completes. At that point, the job
buffer can be allocated.

This reverts commit 7bdf6917.

I believe c4545110 should fix this.  I have verified this doesn't
happen with just this commit.

Make the "pre_run" logic run asynchronously (in separate thread)
  and launch the application as soon as the function completes.
  The function can not be started until after the allocation has
  taken place and we know the nodes to be used.

Remove vestigial function bb_limit_test(), made redundant by TRES.
Restore logic to set buffer access/type info
Correct read config logic

Conflicts:
	src/common/slurm_protocol_api.c
	src/slurmctld/proc_req.c
	src/slurmd/slurmd/req.c

This change adds the AuthInfo configuration parameter to the
g_slurm_auth_get_uid() and g_slurm_auth_get_gid() functions.

Then continuing on - src/common/slurm_auth.h shows that the following
functions are intended to take the auth_info argument:
extern void *   g_slurm_auth_create( void *hosts, int timeout, char
*auth_info );
 extern int      g_slurm_auth_verify( void *cred, void *hosts, int
timeout, char *auth_info );
extern uid_t    g_slurm_auth_get_uid( void *cred, char *auth_info );
extern gid_t    g_slurm_auth_get_gid( void *cred, char *auth_info );

g_slurm_auth_create and g_slurm_auth_verify seems to be OK now - but
g_slurm_auth_get_uid and g_slurm_auth_get_gid are not (most cases will
work anyway since the munge auth plugin will only use auth_info if the
cred has not yet been verified - and in many instances it has - still
I would assume passing the information to be the safe thing to do).
I've attached two blind patches that just replaces null with
slurm_get_auth_info() for these two functions.

This is an refinement of commit 47f96cae
with improved logging and a similar fix to another place with
similar logic
bug 1880

Fix gang scheduling/preemption issue that could cancel job at startup.
I have not been able to reproduce the reported problem, but this
should prevent the reported problem.
bug 1880