Re: iSCSI Abort Task and WRITE PENDING

[Konstantin Shelekhin <k.shelekhin@yadro.com>]

On Wed, Oct 13, 2021 at 09:58:32PM +0300, Konstantin Shelekhin wrote:
> On Wed, Oct 13, 2021 at 01:30:32PM -0500, Mike Christie wrote:
> > On 10/13/21 1:24 PM, Mike Christie wrote:
> > > On 10/13/21 1:08 PM, Konstantin Shelekhin wrote:
> > >> On Wed, Oct 13, 2021 at 12:51:32PM -0500, Mike Christie wrote:
> > >>> On 10/13/21 8:21 AM, Konstantin Shelekhin wrote:
> > >>>> Hi,
> > >>>>
> > >>>> I really need the collective wisdom.
> > >>>>
> > >>>> Not long ago we've uncovered the problem with iSCSI and ABORT TASK
> > >>>> handling. Currently it's not possible to abort a WRITE_10 command in
> > >>>> TRANSPORT_WRITE_PENDING state, because ABORT TASK  will hang itself in
> > >>>> the process:
> > >>>>
> > >>>>   # dmesg | tail -2
> > >>>>   [   83.563505] ABORT_TASK: Found referenced iSCSI task_tag: 3372979269
> > >>>>   [   84.593545] Unable to recover from DataOut timeout while in ERL=0, closing iSCSI connection for I_T Nexus <nexus>
> > >>>>
> > >>>>   # ps aux | awk '$8 ~/D/'
> > >>>>   root        32  0.0  0.0      0     0 ?        D    15:19   0:00 [kworker/0:1+events]
> > >>>>   root      1187  0.0  0.0      0     0 ?        D    15:20   0:00 [iscsi_ttx]
> > >>>>
> > >>>>   # cat /proc/32/stack
> > >>>>   [<0>] target_put_cmd_and_wait+0x68/0xa0
> > >>>>   [<0>] core_tmr_abort_task.cold+0x16b/0x192
> > >>>>   [<0>] target_tmr_work+0x9e/0xe0
> > >>>>   [<0>] process_one_work+0x1d4/0x370
> > >>>>   [<0>] worker_thread+0x48/0x3d0
> > >>>>   [<0>] kthread+0x122/0x140
> > >>>>   [<0>] ret_from_fork+0x22/0x30
> > >>>>
> > >>>>   # cat /proc/1187/stack
> > >>>>   [<0>] __transport_wait_for_tasks+0xaf/0x100
> > >>>>   [<0>] transport_generic_free_cmd+0xe9/0x180
> > >>>>   [<0>] iscsit_free_cmd+0x50/0xb0
> > >>>>   [<0>] iscsit_close_connection+0x47d/0x8c0
> > >>>>   [<0>] iscsit_take_action_for_connection_exit+0x6f/0xf0
> > >>>>   [<0>] iscsi_target_tx_thread+0x184/0x200
> > >>>>   [<0>] kthread+0x122/0x140
> > >>>>   [<0>] ret_from_fork+0x22/0x30
> > >>>>
> > >>>> What happens:
> > >>>>
> > >>>>   1. Initiator sends WRITE_10 CDB
> > >>>>   2. Target parses the CDB and sends R2T
> > >>>>   3. Target starts the Data-Out timer
> > >>>>   4. Initiator sends ABORT TASK; no new data from Initiator after this
> > >>>>   5. Target starts aborting WRITE_10, gets into core_tmr_abort_task()
> > >>>>      and starts waiting for the request completion
> > >>>>   6. Nothing happens
> > >>>>   7. The Data-Out timers expires, connection teardown starts and gets
> > >>>>      stuck waiting for ABORT TASK that waits for WRITE_10
> > >>>>
> > >>>> The ABORT TASK processing looks roughly like this:
> > >>>>
> > >>>>   iscsi_rx_opcode
> > >>>>     iscsi_handle_task_mgt_cmd
> > >>>>       iscsi_tmr_abort_task
> > >>>>       transport_generic_handle_tmr
> > >>>>         if (tmr_cmd->transport_state & CMD_T_ABORTED)
> > >>>>           target_handle_abort
> > >>>>         else
> > >>>>           target_tmr_work
> > >>>>             if (tmr_cmd->transport_state & CMD_T_ABORTED)
> > >>>>               target_handle_abort
> > >>>>             else
> > >>>>               core_tmr_abort_task
> > >>>>                 ret = __target_check_io_state
> > >>>>                   if (write_cmd->transport_state & CMD_T_STOP)
> > >>>>                     return -1
> > >>>>                   write_cmd->transport_state |= CMD_T_ABORTED
> > >>>>                   return 0
> > >>>>                 if (!ret)
> > >>>>                   list_move_tail(&write_cmd->state_list, &aborted)
> > >>>>                   target_put_cmd_and_wait(&write_cmd)
> > >>>>
> > >>>> As I see it, the main problem is that the abort path can't initiate the
> > >>>> command termination, it simply waits for the request to handle this on
> > >>>> the execution path like in target_execute_cmd():
> > >>>>
> > >>>>   target_execute_cmd
> > >>>>     target_cmd_interrupted
> > >>>>       INIT_WORK(&cmd->work, target_abort_work)
> > >>>>
> > >>>> However, in this case the request is not going to be executed because
> > >>>> Initiator will not send the Data-Out buffer.
> > >>>>
> > >>>> I have a couple of ideas on how to fix this, but they all look kinda
> > >>>> ugly. The one that currently works around this for me:
> > >>>>
> > >>>>   core_tmr_abort_task():
> > >>>>
> > >>>>     [...]
> > >>>>
> > >>>>     spin_lock_irqsave(&se_cmd->t_state_lock, flags);
> > >>>>     write_pending = se_cmd->t_state == TRANSPORT_WRITE_PENDING;
> > >>>>     spin_unlock_irqrestore(&se_cmd->t_state_lock, flags);
> > >>>>     
> > >>>>     if (write_pending && se_cmd->se_tfo->abort_write_pending)
> > >>>>             se_cmd->se_tfo->abort_write_pending(se_cmd);
> > >>>>     
> > >>>>     target_put_cmd_and_wait(se_cmd);
> > >>>>
> > >>>>     [...]
> > >>>>
> > >>>> The new method abort_write_pending() is defined only for iSCSI and calls
> > >>>> target_handle_abort(). However, this opens up another can of worms
> > >>>> because this code heavily races with R2T sending and requires a couple
> > >>>> of checks to "work most of the time". Not ideal, by far.
> > >>>>
> > >>>> I can make this one better by introducing R2T list draining that ensures
> > >>>> the proper order during cleanup, but maybe there is a much easier way
> > >>>> that I'm not seeing here.
> > >>>
> > >>> Ccing Maurizio to make sure I don't add his original bug back.
> > >>>
> > >>> If I understand you, I think I added this bug in:
> > >>>
> > >>> commit f36199355c64a39fe82cfddc7623d827c7e050da
> > >>> Author: Mike Christie <michael.christie@oracle.com>
> > >>> Date:   Fri Nov 13 19:46:18 2020 -0600
> > >>>
> > >>>     scsi: target: iscsi: Fix cmd abort fabric stop race
> > >>>
> > >>> With that patch if the abort or a lun reset has got to lio core then we
> > >>> are going to be stuck waiting for the data which won't come because we
> > >>> killed the iscsi threads.
> > >>>
> > >>> Can go back to always having the iscsi target clean up the cmd, but if
> > >>> LIO has started to abort the cmd we take an extra ref so we don't free
> > >>> the cmd from under each other.
> > >>>
> > >>> This patch is completely untested:
> > >>>
> > >>>
> > >>> diff --git a/drivers/target/iscsi/iscsi_target.c b/drivers/target/iscsi/iscsi_target.c
> > >>> index 2c54c5d8412d..d221e9be7468 100644
> > >>> --- a/drivers/target/iscsi/iscsi_target.c
> > >>> +++ b/drivers/target/iscsi/iscsi_target.c
> > >>> @@ -4090,12 +4090,13 @@ static void 	(struct iscsi_conn *conn)
> > >>>  			spin_lock_irq(&se_cmd->t_state_lock);
> > >>>  			if (se_cmd->transport_state & CMD_T_ABORTED) {
> > >>>  				/*
> > >>> -				 * LIO's abort path owns the cleanup for this,
> > >>> -				 * so put it back on the list and let
> > >>> -				 * aborted_task handle it.
> > >>> +				 * The LIO TMR handler owns the cmd but if
> > >>> +				 * we were waiting for data from the initiator
> > >>> +				 * then we need to internally cleanup to be
> > >>> +				 * able to complete it. Get an extra ref so
> > >>> +				 * we don't free the cmd from under LIO core.
> > >>>  				 */
> > >>> -				list_move_tail(&cmd->i_conn_node,
> > >>> -					       &conn->conn_cmd_list);
> > >>> +				target_get_sess_cmd(se_cmd, false);
> > >>>  			} else {
> > >>>  				se_cmd->transport_state |= CMD_T_FABRIC_STOP;
> > >>>  			}
> > >>
> > >> The bug was there before. I had to backport this patch in order to
> > >> introduce my fix. I can revert it and check what is different, but it's
> > >> there in some form.
> > >>
> > > 
> > > Don't waste your time. It's because iscsit_free_cmd's call to
> > > transport_generic_free_cmd has wait_for_tasks=true.
> > > 
> > > We then do transport_generic_free_cmd -> target_wait_free_cmd ->
> > > __transport_wait_for_tasks like you posted above.
> > 
> > That's wrong, it's the transport_generic_free_cmd wait:
> > 
> >         if (aborted) {
> >                 pr_debug("Detected CMD_T_ABORTED for ITT: %llu\n", cmd->tag);
> >                 wait_for_completion(&compl);
> >                 ret = 1;
> 
> I'm not sure I'm following you. The wait on master comes from this:
> 
>  core_tmr_abort_task():
>   [...]
>   target_put_cmd_and_wait(se_cmd);
>   [...]
> 
> IMO it's kinda legit wait. And we can't just drop the references and
> call it a day, because a request has to go through the
> target_handle_abort() because it (at least) does some TAS magic. And
> currently there is no code that calls target_handle_abort() for
> WRITE_PENDING requests.

Ah, sorry, I should've mentioned it I guess. The
transport_generic_free_cmd() bit is connection teardown waiting on ABORT TASK
that waits for WRITE_10 task. That's part is completely fine IMO.