[SCSI] libiscsi: Fix recovery slowdown regression

We could be failing/stopping a connection due to libiscsi starting
recovery/cleanup, but the xmit path or scsi eh thread path
could be dropping the connection at the same time.

As a result the session->state gets set to failed instead of in
recovery. We end up not blocking the session
and so the replacement timeout never gets started and we only end up
failing the IO when scsi_softirq_done sees that the
cmd has been running for (cmd->allowed + 1) * rq->timeout secs.

We used to fail the IO right away so users are seeing a long
delay when using dm-multipath. This problem was added in
2.6.28.

Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: stable@kernel.org
Signed-off-by: James Bottomley <James.Bottomley@suse.de>