On 2/12/2015 10:47 AM, Nicholas A. Bellinger wrote:
On Wed, 2015-02-11 at 10:17 +0200, Sagi Grimberg wrote:
Hey Nic,
So Our QA guys recently stepped on this bug when performing stress
login-logout from a single initiator to 10 targets each exposed over
4 portals, so overall 40 sessions (needless to say we are talking on
iser...). So there are lots of logins in parallel with lots of logouts.
It seems that the connection termination causes iscsi_tx_thread to
access the connection after it is freed or something (list corruption
probably coming from iscsit_handle_immediate_queue or
iscsit_handle_response_queue, and NULL deref coming from
iscsit_take_action_for_connection_exit).
Note, isert_wait_conn waits for session commands and QP flush which is
normally pretty fast, the conn termination is done in a work that waits
for DISCONNECTED event which might take longer (which is why we do it
outside wait_conn context to avoid blocking it).
I didn't get too far with this until now, do you have any idea on what
might have happened?
Mmm, it looks like iscsit_take_action_for_connection_exit() in TX thread
context is calling iscsi_close_connection() after hitting the following
check in iscsi_target_erl0.c:
if (conn->conn_state == TARG_CONN_STATE_IN_LOGOUT) {
spin_unlock_bh(&conn->state_lock);
iscsit_close_connection(conn);
return;
}
.. once iscsi_close_connection() has already being called earlier by
iser-target code.
Not sure I understand where iscsit_close_connection is called earlier
by iser target. The iser code usually only notifies any problems to the
iscsi layer to do it's thing.
Care to explain how iscsit_close_connection might be called twice?
AFAICT, this check is specific to traditional iscsi-target during the
iscsit_handle_logout_cmd() failure case in iscsi_target_rx_opcode().
Here's a quick patch to make this case ISCSI_TCP only.
I'll give it a shot.
Thanks,
Sagi.
--
To unsubscribe from this list: send the line "unsubscribe target-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
