Bug 4483

Signed-off-by: Boris Karasev <karasev.b@gmail.com>

Bug 4541

The slurm and pmix RPMs both attempt to install a libpmi. While Slurm's
version should be preferred, the best way to handle this is to have both
pmix and slurm packages split this off into a separate RPM that can
conflict with each other, giving the admin a way to resolve the conflict
manually.

Bug 4511.

The cluster name could have been free'd before it would be strdup'ed.
Happens when a federated persisent connection is established and the
cluster needs to sync up with the remote cluster.

Bug 4503

Remove the origin job when a remote sibling job was cancelled while the
origin was down.

e.g.
job originated on cluster1
job runs on cluster2
cluster1 goes down
job is cancelled on cluster2
cluster1 comes backup. The origin job should be removed if the sibling
job was cancelled.

Replace some references to "pack job" with "heterogenous job".
bug 4546

The user_name is only populated if LaunchParameters=send_gids is set.
Without this field set, the pam_setup call will get NULL instead of the
user_name, leading to batch job launch failures. (The step launch code
already handles this same issue separately.)

Bug 4412.

Previously written to "admin_comment" in commit 540d0e5c
bug 4529

On a job [pack]allocation RPC request, if the allocation succeed but the
send response message back to the client failed (i.e. srun was killed before
it could receive the response), then modify the job_record pointer so that
the job_state is set to FAILED, the exit_code as if the job got a SIGTERM
signal and the state_reason to FAIL_LAUNCH. Then users when querying
the job with sacct can discern that something bad happend for this scenario,
instead of STATE being showed as COMPLETED and the ExitCode as 0:0.

Bug 4513.

Use FREE_NULL_BUFFER instead, otherwise we could attempt to
free_buffer this a second time if we jump to the rwfail label.

Bug 4484.

When printing fields in sacct with user specified units (--units), the
nnodes field showed an incorrect string. This commit reverts a65fa572
and avoids the unit conversion, which does not make sense outside the
context of a Blue Gene systems (deprecated) anyways.

Bug 4490.

Slurm may generate empty manifest files depending on configuration
and library availability. Disable the new empty manifest check to
allow builds to proceed with rpm 4.13+ / Fedora 25+.

Bug 4453.

(Fixing tim@schedmd.com's mistake on prior commit.)

Bug 4467.

When set, do not flush the Lustre cache through the alpsc_flush_lustre()
call, and do not drop caches through /proc/sys/vm/drop_caches either.

This avoids a potential source of SIGBUG errors for other jobs sharing
the node.

Bug 4309.

before printing anything for a connection.

field.

Bug 4529

fails. The description of the failure will be in the job's "Reason" field.

Bug 4529

buffer error.

Bug 4529

Bug 4222.

on startup. Just use the checkpointed job_id_sequence. get_next_job_id()
will not use a jobid if it's in use in the system.

Bug 4538

node_features/knl_generic - If plugin can not fully load then do not spawn
    a background pthread (which will fail with invalid memory reference).

Add "Force=1" to knl_generic.conf to override (for testing)
bug 4487

This patch fix commit b31bb7 for big-endian machines inverting the
__builtin_clzll and __builtin_cltll calls depending on the architecture.
This caused to get the incorrect first and last bit set on the bitmap in
bit_ffs bit_fls.

bug 4494

completely instead of right after the parent job finishes.

Bug 4516

bug 3582

when a job requests no tasks and more memory than MaxMemPer{CPU|NODE}.

e.g.
sbatch --wrap="sleep 10"

Bug 4515

This will give expected results. Found while working on Bug 4515.

Bug 4478 comment 25.

And print an appropriate fatal error message rather than relying
upon random errno value.

Bug 4523

Bug 4478.

Based off of Ryan Cox's original contribs/pam_slurm_adopt/README.

Bug 3567.

In the federation case, the origin job is completed in the database when
a sibling job starts the job. The complete message is then sent again to
the database when the job is completed on the sibling cluster but
it is updated with the sibling job's exit code.

The jobcomp plugin didn't handle the multiple updates to the record.
This change allows the existing record to be updated.

Bug 4493