When a delete event is queued for an rport, set state to NONE so that no
other processing is done on the rport as it is being removed.
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
After lport_destroy, the local port should not be used again. Transition
to state NONE, any incoming frames or link up should not transition out
of this state since we are deleting exchange table and cleaning up the
local port. Also, mark link as down.
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
We want to generate the rport queue event (from the logoff)
before flushing the queue otherwise the event may still be
in the queue when we logoff.
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Rogue ports are currently not tracked on any list. The only reference
to them is through any outstanding exchanges pending on the rogue ports.
If the module is removed while a retry is set on a rogue port
(say a Plogi retry for instance), this retry is not cancelled because there
is no reference to the rogue port in the discovery rports list. Thus the
local port can clean itself up, delete the exchange pool, and then the
rogue port timeout can fire and try to start up another exchange.
This patch tracks the rogue ports in a new list disc->rogue_rports. Creating
a new list instead of using the disc->rports list keeps remote port code
change to a minimum.
1) Whenever a rogue port is created, it is immediately added to the
disc->rogue_rports list.
2) When the rogues port goes to ready, it is removed from the rogue list
and the real remote port is added to the disc->rports list
3) The removal of the rogue from the disc->rogue_rports list is done in
the context of the fc_rport_work() workQ thread in discovery callback.
4) Real rports are removed from the disc->rports list like before. Lookup
is done only in the real rports list. This avoids making large changes
to the remote port code.
5) In fc_disc_stop_rports, the rogues list is traversed in addition to the
real list to stop the rogue ports and issue logoffs on them. This way, rogue
ports get cleaned up when the local port goes away.
6) rogue remote ports are not removed from the list right away, but
removed late in fc_rport_work() context, multiple threads can find the same
remote port in the list and call rport_logoff(). Rport_logoff() only
continues with the logoff if port is not in NONE state, thus preventing
multiple logoffs and multiple list deletions.
7) Since the rport is removed from the disc list at a later stage
(in the disc callback), incoming frames can find the rport even if
rport_logoff() has been called on the rport. When rport_logoff() is called,
the rport state is set to NONE, and we are trying to cancel all exchanges
and retries on that port. While in this state, if an incoming
Plogi/Prli/Logo or other frames match the rport, we should not reply
because the rport is in the NONE state. Just drop the frame, since the
rport will be deleted soon in the disc callback (fc_rport_work)
8) In fc_disc_single(), remove rport lookup and call to fc_disc_del_target.
fc_disc_single() is called from recv_rscn_req() where rport lookup
and rport_logoff is already done.
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
For instance, if there is a Plogi pending (remote port is in Plogi state),
and the state changes to say NONE (because the port is being logged off),
then when the Plogi resp times out, do not start a retry.
This patch partially reverts an earlier patch (libfc: check for err when
recv and state is incorrect), by moving the state check back to before
checking for error. However, if the state does not match, then there is
an additional check to see if its an error ptr or a real frame before
jumping to err or out respectively.
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
gpn_ft_resp processing currently does not hold the discovery lock.
disc_done() thus gets called from gpn_ft_resp or from gpn_ft_parse
without the lock held. This then sets disc->pending to zero or calls
gpn_ft_req() without disc_lock held.
- Hold disc mutex during gpn_ft resp processing
- In disc_done, release the disc mutex while calling lport callback
Signed-off-by: Abhijeet Joglekar <abjoglek@cisco.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
(version 2)
Fixed a bug in calculating ddp map range when search for free entries:
it was going beyond the end by one, thus corrupting gl_skb[0].
Signed-off-by: Karen Xie <kxie@chelsio.com>
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
When doing a lot (128) of large writes (256K) we can hit the cxgb3_snd_win
check pretty easily. The driver's xmit thread then takes 100% of the cpu.
The driver should not be returning -EAGAIN for this problem. It should
be returing -ENOBUFS, then when the window is opened again it should
queue the xmit thread (it already wakes the xmit thread).
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
cxgb3i was setting can_queue to only 128 commands, and was
setting the can_queue and cmd_per_lun to the same value.
This sets the can_queue to 1024 commands, and sets the cmd_per_lun
to a safer default of 32.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Set target can queue limit to the number of preallocated
session tasks we have.
This along with the cxgb3i can_queue patch will fix a throughput
problem where it could only queue one LU worth of data at a time.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
If a command's scsi cmd pdu setup fails then we can just fail
the IO to the scsi layer. If a DATA_OUT for a R2T fails then
we will want to drop the session, because it means we got a
bad request from the target (iscsi protocol error).
This patch has us propogate the error upwards so libiscsi_tcp
or libiscsi can decide what the best action is to take. It
also fixes a bug where we could try to grab the session lock
while holding it, because if iscsi_tcp drops the session in the
pdu setup callout the session lock is held when setting up the
scsi cmd pdu.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The zfcp_port might have been removed, while the FC fast_io_fail timer
is still running and could trigger the terminate_rport_io callback.
Set the pointer to the zfcp_port to NULL and check accordingly
before using it.
Reviewed-by: Martin Petermann <martin@linux.vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
If this problem appears zfcp ports cannot be de-queued since it is
checked for a zero refcount. The port reference counting is wrong for
existing zfcp ports when e.g. an adapter gets on-line again. During
port scanning the reference counting for existing ports should not be
changed.
Signed-off-by: Martin Petermann <martin.petermann@de.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The current sbal counting can be wrong if a fsf request is
waiting for free sbals and at the same time qdio request queue
is shutdown and re-opened. Revering a previous patch fixes this
issue.
Signed-off-by: Martin Petermann <martin.petermann@de.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
When the abort handler cannot find a pending FSF request, the request
completion could just be running. This means we cannot return SUCCESS,
since this would lead to call to scsi_done after exiting the SCSI
error handler which is not allowed.
Reviewed-by: Martin Petermann <martin@linux.vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
A remote port remains in error state even if we receive a RSCN
stating that the connection is re-established. The port recovery
is not started due to a flag which is not reset.
The solution is to clear the flag in question before we trigger a ERP.
Signed-off-by: Swen Schillig <swen@vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Error codes specific to the control file requests are evaluated by the
actcli tool, so don't report -ENXIO for those. Generic problems are
still checked for outside the command specific handler.
Reviewed-by: Martin Petermann <martin@linux.vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The remote port remains in error state even if the connection
is re-established. A wrong precondition check was performed on
the port status leading to a cancellation of the port reopen.
Remove the pre-req check because it's not required and better
handled within the ERP.
Signed-off-by: Swen Schillig <swen@vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The ERP thread is performing a task before it is executing the
corresponding down on the semaphore. The response handler of the
just started exchange config should wait for the completion by
performing a down on this semaphore. Since this semaphore is still
positive from the ERP enqueue the handler won't wait and therefore
the exchange config will always fail leaving the adapter in error.
The problem can be solved by performing the down on the semaphore
before starting an ERP task. This is the logically correct order.
Only walk the ERP loop if there is a task to perform.
Signed-off-by: Swen Schillig <swen@vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
The nameserver port might be in state online when the adapter is
offlined. On adapter reactivation the nameserver port is not
re-opened due to the PORT_ONLINE status. This results in an
unsuccessful recovery. In forcing the nameserver port status
to offline on all adapter offline events this issue is prevented.
Waiting for the reference count to drop to zero in
zfcp_wka_port_offline is not required, so remove it.
Signed-off-by: Swen Schillig <swen@vnet.ibm.com>
Signed-off-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Chris Leech
Dan Carpenter
Abhijeet Joglekar
kxie@chelsio.com
Mike Christie
Christof Schmitt
Martin Petermann
Swen Schillig
Martin Petermann