IB/iser: Enhance disconnection logic for multi-pathing

The iser connection teardown flow isn't over until the underlying
Connection Manager (e.g the IB CM) delivers a disconnected or timeout
event through the RDMA-CM.  When the remote (target) side isn't
reachable, e.g when some HW e.g port/hca/switch isn't functioning or
taken down administratively, the CM timeout flow is used and the event
may be generated only after relatively long time -- on the order of
tens of seconds.

The current iser code exposes this possibly long delay to higher
layers, specifically to the iscsid daemon and iscsi kernel stack. As a
result, the iscsi stack doesn't respond well: this low-level CM delay
is added to the fail-over time under HA schemes such as the one
provided by DM multipath through the multipathd(8) service.

This patch enhances the reference counting scheme on iser's IB
connections so that the disconnect flow initiated by iscsid from user
space (ep_disconnect) doesn't wait for the CM to deliver the
disconnect/timeout event.  (The connection teardown isn't done from
iser's view point until the event is delivered)

The iser ib (rdma) connection object is destroyed when its reference
count reaches zero.  When this happens on the RDMA-CM callback
context, extra care is taken so that the RDMA-CM does the actual
destroying of the associated ID, since doing it in the callback is
prohibited.

The reference count of iser ib connection normally reaches three,
where the <ref, deref> relations are

 1. conn <init, terminate>
 2. conn <bind, stop/destroy>
 3. cma id <create, disconnect/error/timeout callbacks>

With this patch, multipath fail-over time is about 30 seconds, while
without this patch, multipath fail-over time is about 130 seconds.

Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>