select()/FD_ISSET() does not work for file descriptor larger than 1023.

avoid add/remove node resource of job if the node is lost by resize

 I found another case that idle node can not be allocated. It can be
reproduced as follows:

1. run a job with -k option:

    [root@mn0 ~]# srun -w cn[18-28] -k sleep 1000
    srun: error: Node failure on cn28
    srun: error: Node failure on cn28
    srun: error: cn28: task 10: Killed
    ^Csrun: interrupt (one more within 1 sec to abort)
    srun: tasks 0-9: running
    srun: task 10: exited abnormally
    ^Csrun: sending Ctrl-C to job 106120.0
    srun: Job step aborted: Waiting up to 2 seconds for job step to
finish.

2. set a node down and then set it idle:

    [root@mn0 ~]# scontrol update nodename=cn28 state=down reason="hjcao
test"
    [root@mn0 ~]# scontrol update nodename=cn28 state=idle

3. restart slurmctld

    [root@mn0 ~]# service slurm restart
    stopping slurmctld:                                        [  OK  ]
    slurmctld is stopped
    starting slurmctld:                                        [  OK  ]

4. cancel the job

then, the node set down will be left unavailable:

    [root@mn0 ~]# sinfo -n cn[18-28]
    PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
    work*        up   infinite     11   idle cn[18-28]

    [root@mn0 ~]# srun -w cn[18-28] hostname
    srun: job 106122 queued and waiting for resources

    [root@mn0 slurm]# grep cn28 slurmctld.log
    [2013-03-18T15:28:02+08:00] debug3: cons_res: _vns: node cn28 in
exclusive use
    [2013-03-18T15:29:02+08:00] debug3: cons_res: _vns: node cn28 in
exclusive use

I made an attempt to fix this by the attached patch. Please review it.

I don't believe save_time_limit was redundant.  At least in this case:

if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)){
    if (orig_time_limit == NO_VAL)
        orig_time_limit = comp_time_limit;
    job_ptr->time_limit = orig_time_limit;
[...]

So later, when updating the db,

    if (save_time_limit != job_ptr->time_limit)
        jobacct_storage_g_job_start(acct_db_conn,
                        job_ptr);
will cause the db to be updated, while,

        if (orig_time_limit != job_ptr->time_limit)
        jobacct_storage_g_job_start(acct_db_conn,
                        job_ptr);

will not because job_ptr->time_limit now equals orig_time_limit.

Without this change, if the job's time limit is modified down
toward --time-min by the backfill scheduler, update the job's
time limit in the database.

If step requests more CPUs than possible in specified node count of job
allocation then return ESLURM_TOO_MANY_REQUESTED_CPUS rather than
ESLURM_NODES_BUSY and retrying.

I found a bug in cons_res/select_p_select_nodeinfo_set_all.

If a node is part of two (or more) partitions the code will only count the number of cores/cpus in the partition that has the most running jobs on that node.

Patch attached to fix the problem.

I also added an new function to bitstring to count the number of bits in an range (bit_set_count_range) and made a minor improvement of (bit_set_count) while reviewing the range version.

Best regards,
Magnus

This permits default reservation names to be more easily managed

Without this change, when the sbatch --export option is used, many
Slurm environment variables are not set unless explicitly exported.

success

This problem would effect systems in which specific GRES are associated
with specific CPUs.
One possible result is the CPUs identified as usable could be inappropriate
and job would be held when trying to layout out the tasks on CPUs (all
done as part of the job allocation process).
The other problem is that if multiple GRES are linked to specific CPUs,
there was a CPU bitmap OR which should have been an AND, resulting in
some CPUs being identified as usable, but not available to all GRES.