Re: [PATCH v2 09/13] scsi: target: iscsit: Fix isert disconnect handling during login

[Sagi Grimberg <sagi@grimberg.me>]

On 1/18/23 04:14, Mike Christie wrote:
> On 1/13/23 08:08, Dmitry Bogdanov wrote:
>> On Wed, Jan 11, 2023 at 09:08:28PM -0600, Mike Christie wrote:
>>>
>>> If we get a disconnect event while logging in we can end up in a state
>>> where will never be able to relogin. This happens when:
>>>
>>> 1. login thread has put us into TARG_CONN_STATE_IN_LOGIN
>>> 2. isert then does
>>>
>>> isert_disconnected_handler -> iscsit_cause_connection_reinstatement
>>>
>>> which sets the conn connection_reinstatement flag. Nothing else happens
>>> because we are only in IN_LOGIN. The tx/rx threads are not running yet
>>> so we can't start recovery from those contexts at this time.
>>>
>>> 3. The login thread finishes processing the login pdu and thinks login is
>>> done. It sets us into TARG_CONN_STATE_LOGGED_IN/TARG_SESS_STATE_LOGGED_IN.
>>> This starts the rx/tx threads.
>>>
>>> 4. The initiator thought it disconnected the connection at 2, and has
>>> since sent a new connect which is now handled. This leads us to eventually
>>> run:
>>>
>>> iscsi_check_for_session_reinstatement-> iscsit_stop_session ->
>>> iscsit_cause_connection_reinstatement
>>>
>>> iscsit_stop_session sees the old conn and does
>>> iscsit_cause_connection_reinstatement which sees connection_reinstatement
>>> is set so it just returns instead of trying to kill the tx/rx threads
>>> which would have caused recovery to start.
>>>
>>> 5. iscsit_stop_session then waits on session_wait_comp which will never
>>> happen since we didn't kill the tx/rx threads.
>>>
>>> This has the iscsit login code check if a fabric driver ran
>>> iscsit_cause_connection_reinstatement during the login process similar
>>> to how we check for the sk state for tcp based iscsit. This will prevent
>>> us from getting into 3 and creating a ghost session.
>>>
>>> Signed-off-by: Mike Christie <michael.christie@oracle.com>
>>> ---
>>>   drivers/target/iscsi/iscsi_target_nego.c | 10 ++++++++++
>>>   1 file changed, 10 insertions(+)
>>>
>>> diff --git a/drivers/target/iscsi/iscsi_target_nego.c b/drivers/target/iscsi/iscsi_target_nego.c
>>> index ff49c8f3fe24..2dd81752d4c9 100644
>>> --- a/drivers/target/iscsi/iscsi_target_nego.c
>>> +++ b/drivers/target/iscsi/iscsi_target_nego.c
>>> @@ -350,6 +350,16 @@ static int iscsi_target_do_tx_login_io(struct iscsit_conn *conn, struct iscsi_lo
>>>                                              ISCSI_LOGIN_STATUS_NO_RESOURCES);
>>>                          return -1;
>>>                  }
>>> +
>>> +               /*
>>> +                * isert doesn't know the iscsit state and uses
>>> +                * iscsit_cause_connection_reinstatement as a generic error
>>> +                * notification system. It may call it before we are in FFP.
>>> +                * Handle this now in case it signaled a failure before the
>>> +                * rx/tx threads were up and could start recovery.
>>> +                */
>>> +               if (atomic_read(&conn->connection_reinstatement))
>>> +                       goto err;
>>
>> Why only for login->login_complete case? In other case the session will
>> not hang? Will it be droppped on login timeout or something else?
> 
> It will not hang. If you hit an error with isert before we think we can go into
> FFP then the login timeout currently cleans up the conn and session.
> 
>>
>> May be the root cause is point 2 itself - calling iscsit_cause_connection_reinstatement
>> in not ISER_CONN_FULL_FEATURE state where there are no TX_RX threads?
>> I mean that was a misuse of iscsit_cause_connection_reinstatement?
> 
> Let me drop this patch for now. After writing the response above about normally it just
> times out, and thinking about your question here I think to really fix this we want to
> fully integrate isert login into iscsit.
> 
> The root problem is that isert login is not integrated into iscsit, so there is really
> no error handling. iscsit_cause_connection_reinstatement was supposed to do it, but it
> doesn't do anything.
> 
> So we need to separate the LOGIN_FLAGS_CLOSED tests and setting from the iscsi_target_sk*
> code. Add a helper to set that bit and do some state checks, and make the checks generic
> in the login code (not tied to having a socket). Convert iscsit and isert to use the new
> helper. Then handle the LOGIN_FLAGS_READ_ACTIVE/LOGIN_FLAGS_WRITE_ACTIVE and login_work
> stuff. Then review cxgb?
> 
> I'll do that later after fixing the command cleanup stuff in this patchsdet and the
> other patchsets I have outstanding.

Perhaps the fix is as simple as checking the iscsi state to be also 
full-featured and if not, do a simple error handling within
iser_disconnected_handler?