Use a workqueue for processing RDMA completions instead of creating one
kernel thread per RDMA channel. This change improves performance if the
number of RDMA channels is large by reducing the number of context
switches between kernel threads while processing RDMA completions. An
additional change is that srpt_set_ch_state() no longer wakes up the
RDMA completion context but that srpt_cm_rtu_recv() wakes up the RDMA
completion context instead.
See also https://github.com/bvanassche/scst/issues/33 .
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@9178 d57e44dd-8a1f-0410-8b47-8ef2f437770f
