Both the original logic and modified logic failed to lock the job
data structure prior to decrementing "prolog_running" counter.

in the forward message logic.

Previously under some conditions that boot completion was ignored
and the job kept pending.

Bug 2364

Commit fa331e30 fixes this.  The logic was bad to begin with...

uint32_t new_cpus = detail_ptr->num_tasks
	/ detail_ptr->cpus_per_task;

The / should had been * this whole time.  This was the reason we found
this in the first place.

bug 2369

If scancel is operating on large number of jobs and RPC responses from
    slurmctld daemon are slow then introduce a delay in sending the cancel job
    requests from scancel in order to reduce load on slurmctld.
bug 2256

If a job launch is delayed, the test was failing due to bad parsing.
These lines were being interpretted as a counter folloed by node
names of "queued" and "has":
  srun: job 1332712 queued and waiting for resources
  srun: job 1332712 has been allocated resources

bug 2366

Backfill scheduling properly synchronized with Cray Node Health Check.
    Prior logic could result in highest priority job getting improperly
    postponed.
bug 2350

This corrects logic from commit e5a61746
that could result in use of NULL pointer

It was previously triggered by executing "scontrol reconfig" on a
  front-end system while there was a job in completing state.

Properly account for memory, CPUs and GRES when slurmctld is reconfigured
    while there is a suspended job. Previous logic would add the CPUs, but not
    memory or GPUs. This would result in underflow/overflow errors in select
    cons_res plugin.
bug 2353

The counter is really intended to reflect the count of running or
  suspended jobs rather than running jobs alone. Previous logic
  would report an underflow for the "job_cnt_run" variable if
  1. job submitted
  2. job suspended
  3. scontrol reconfig
  4. job cancelled

Log the length of bitmaps in addition to the bits set.
Also increase the string length used for logging.

Previous logic would prevent allocation of sockets to a job unless the
entire socket was available. If there were any specialized cores, the
socket was treated as being not available and unusable. For example,
if a node had 2 sockets, then a job requesting 2 specialized cores
would reserve one core on each of the two sockets and render the job
not runnable.

There was logic in sinfo's print state function that determined
if the state was MIXED. This logic was duplicated logic from the
_query_server() function in sinfo.c and has been removed. Also
note the logic was already gone from the "short state" print
function (I noticed the discrepeancy in the print functions,
but discovered they both printed the correct state information).

Fix backfill scheduling bug which could postpone the scheduling of jobs due
    to avoidance of nodes in COMPLETING state.
bug 2350

No need to look up the Reason string for a job, we just set the value.

Bug 2255

Bug 2343

Fix for configuration of "AuthType=munge" and "AuthInfo=socket=..." with
    alternate munge socket path.
bug 2348

If a node is out of memory, then the malloc performed by slurmstepd
  periodically may fail, killing the slurmstepd and orphaning it's
  processes.
bug 2341

Backfill scheduling fix: If a job can't be started due to a "group" resource
    limit, rather than reserve resources for it when the next job ends, don't
    reserve any resources for it. The problem with the original logic is that
    if a lot of resources are reserved for such pending jobs, then jobs futher
    down the queue may defered when they really can and should be started. An
    ideal solution would track all of the TRES resources through time as jobs
    start and end, but we don't have that logic in the backfill scheduler and
    don't want that extra overhead in the backfill scheduler.
bugs 2326 and 2282