terminate slurmd with a fatal error and don't crash slurmctld.

address after the call to realloc.

we can avoid dereferencing a null pointer

* Increased the size of the buffer used to store libstate from
  128KB to 1MB

* Decreased size of SNI adapter name storage from 16 to 5 bytes.
  We should never see a name with any more than 5 bytes, e.g. (sniX)

* Increased realloc block size from 4KB to 128K.  Fixed memory
  corruption problems on larger clusters

* Added lid->adapter_name caching for required for POE

* Fixed bug where next_adapter/next_window were always reset to zero
  after a call to fed_unpack_nodeinfo()

* Reformatted the output fed_print_nodeinfo()

* Increased lock section in fed_unpack_nodeinfo() to address
  race condition bug

* Fixed potential infinite loop in setup_table_entry.  Now we only
  iterate once through all windows and return error if none are
  found

* Added lock section to _copy_libstate and fed_free_libstate

complete RPC including current switch state info.

arguments.

to slurmd, there is no body to the request message.

    (srun -A would generate a segmentation fault, Chris Holmes, HP).

    env var (Brian O'Sullivan, pathscale)

    external scheduler control of RoolOnly partition (Chris Holmes, HP).

switch_p_job_postfini().

While not ideal, it makes for one less thing to worry about updating
when a new release and thus should be better in the long run.

happening correctly.