It is unavoidable that zfcp_scsi_queuecommand() has to finish requests
with DID_IMM_RETRY (like fc_remote_port_chkready()) during the time
window when zfcp detected an unavailable rport but
fc_remote_port_delete(), which is asynchronous via
zfcp_scsi_schedule_rport_block(), has not yet blocked the rport.

However, for the case when the rport becomes available again, we should
prevent unblocking the rport too early.  In contrast to other FCP LLDDs,
zfcp has to open each LUN with the FCP channel hardware before it can
send I/O to a LUN.  So if a port already has LUNs attached and we
unblock the rport just after port recovery, recoveries of LUNs behind
this port can still be pending which in turn force
zfcp_scsi_queuecommand() to unnecessarily finish requests with
DID_IMM_RETRY.

This also opens a time window with unblocked rport (until the followup
LUN reopen recovery has finished).  If a scsi_cmnd timeout occurs during
this time window fc_timed_out() cannot work as desired and such command
would indeed time out and trigger scsi_eh. This prevents a clean and
timely path failover.  This should not happen if the path issue can be
recovered on FC transport layer such as path issues involving RSCNs.

Fix this by only calling zfcp_scsi_schedule_rport_register(), to
asynchronously trigger fc_remote_port_add(), after all LUN recoveries as
children of the rport have finished and no new recoveries of equal or
higher order were triggered meanwhile.  Finished intentionally includes
any recovery result no matter if successful or failed (still unblock
rport so other successful LUNs work).  For simplicity, we check after
each finished LUN recovery if there is another LUN recovery pending on
the same port and then do nothing.  We handle the special case of a
successful recovery of a port without LUN children the same way without
changing this case's semantics.

For debugging we introduce 2 new trace records written if the rport
unblock attempt was aborted due to still unfinished or freshly triggered
recovery. The records are only written above the default trace level.

Benjamin noticed the important special case of new recovery that can be
triggered between having given up the erp_lock and before calling
zfcp_erp_action_cleanup() within zfcp_erp_strategy().  We must avoid the
following sequence:

ERP thread                 rport_work      other context
-------------------------  --------------  --------------------------------
port is unblocked, rport still blocked,
 due to pending/running ERP action,
 so ((port->status & ...UNBLOCK) != 0)
 and (port->rport == NULL)
unlock ERP
zfcp_erp_action_cleanup()
case ZFCP_ERP_ACTION_REOPEN_LUN:
zfcp_erp_try_rport_unblock()
((status & ...UNBLOCK) != 0) [OLD!]
                                           zfcp_erp_port_reopen()
                                           lock ERP
                                           zfcp_erp_port_block()
                                           port->status clear ...UNBLOCK
                                           unlock ERP
                                           zfcp_scsi_schedule_rport_block()
                                           port->rport_task = RPORT_DEL
                                           queue_work(rport_work)
                           zfcp_scsi_rport_work()
                           (port->rport_task != RPORT_ADD)
                           port->rport_task = RPORT_NONE
                           zfcp_scsi_rport_block()
                           if (!port->rport) return
zfcp_scsi_schedule_rport_register()
port->rport_task = RPORT_ADD
queue_work(rport_work)
                           zfcp_scsi_rport_work()
                           (port->rport_task == RPORT_ADD)
                           port->rport_task = RPORT_NONE
                           zfcp_scsi_rport_register()
                           (port->rport == NULL)
                           rport = fc_remote_port_add()
                           port->rport = rport;

Now the rport was erroneously unblocked while the zfcp_port is blocked.
This is another situation we want to avoid due to scsi_eh
potential. This state would at least remain until the new recovery from
the other context finished successfully, or potentially forever if it
failed.  In order to close this race, we take the erp_lock inside
zfcp_erp_try_rport_unblock() when checking the status of zfcp_port or
LUN.  With that, the possible corresponding rport state sequences would
be: (unblock[ERP thread],block[other context]) if the ERP thread gets
erp_lock first and still sees ((port->status & ...UNBLOCK) != 0),
(block[other context],NOP[ERP thread]) if the ERP thread gets erp_lock
after the other context has already cleard ...UNBLOCK from port->status.

Since checking fields of struct erp_action is unsafe because they could
have been overwritten (re-used for new recovery) meanwhile, we only
check status of zfcp_port and LUN since these are only changed under
erp_lock elsewhere. Regarding the check of the proper status flags (port
or port_forced are similar to the shown adapter recovery):

[zfcp_erp_adapter_shutdown()]
zfcp_erp_adapter_reopen()
 zfcp_erp_adapter_block()
  * clear UNBLOCK ---------------------------------------+
 zfcp_scsi_schedule_rports_block()                       |
 write_lock_irqsave(&adapter->erp_lock, flags);-------+  |
 zfcp_erp_action_enqueue()                            |  |
  zfcp_erp_setup_act()                                |  |
   * set ERP_INUSE -----------------------------------|--|--+
 write_unlock_irqrestore(&adapter->erp_lock, flags);--+  |  |
.context-switch.                                         |  |
zfcp_erp_thread()                                        |  |
 zfcp_erp_strategy()                                     |  |
  write_lock_irqsave(&adapter->erp_lock, flags);------+  |  |
  ...                                                 |  |  |
  zfcp_erp_strategy_check_target()                    |  |  |
   zfcp_erp_strategy_check_adapter()                  |  |  |
    zfcp_erp_adapter_unblock()                        |  |  |
     * set UNBLOCK -----------------------------------|--+  |
  zfcp_erp_action_dequeue()                           |     |
   * clear ERP_INUSE ---------------------------------|-----+
  ...                                                 |
  write_unlock_irqrestore(&adapter->erp_lock, flags);-+

Hence, we should check for both UNBLOCK and ERP_INUSE because they are
interleaved.  Also we need to explicitly check ERP_FAILED for the link
down case which currently does not clear the UNBLOCK flag in
zfcp_fsf_link_down_info_eval().
Signed-off-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: 8830271c ("[SCSI] zfcp: Dont fail SCSI commands when transitioning to blocked fc_rport")
Fixes: a2fa0aed ("[SCSI] zfcp: Block FC transport rports early on errors")
Fixes: 5f852be9 ("[SCSI] zfcp: Fix deadlock between zfcp ERP and SCSI")
Fixes: 338151e0 ("[SCSI] zfcp: make use of fc_remote_port_delete when target port is unavailable")
Fixes: 3859f6a2 ("[PATCH] zfcp: add rports to enable scsi_add_device to work again")
Cc: <stable@vger.kernel.org> #2.6.32+
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>