After a loss of transport due to an adapter migration or crash/disconnect
from the host partner there is a tiny window where we can race adjusting
the request_limit of the adapter. The request limit is atomically
increased/decreased to track the number of inflight requests against the
allowed limit of our VIOS partner.

After a transport loss we set the request_limit to zero to reflect this
state.  However, there is a window where the adapter may attempt to queue a
command because the transport loss event hasn't been fully processed yet
and request_limit is still greater than zero.  The hypercall to send the
event will fail and the error path will increment the request_limit as a
result.  If the adapter processes the transport event prior to this
increment the request_limit becomes out of sync with the adapter state and
can result in SCSI commands being submitted on the now reset connection
prior to an SRP Login resulting in a protocol violation.

Fix this race by protecting request_limit with the host lock when changing
the value via atomic_set() to indicate no transport.

Link: https://lore.kernel.org/r/20201025001355.4527-1-tyreld@linux.ibm.comSigned-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>