Re: Deadlock during multipath failover

[Mike Anderson <andmike@linux.vnet.ibm.com>]

Hannes Reinecke <hare@suse.de> wrote:
> Hi Christof,
>
> Christof Schmitt wrote:
>> During failover tests on a current distribution kernel, we found this
>> problem. From reading the code, the upstream kernel has the same
>> problem:
>>
>> During multipath failover tests with SCSI on System z, the kernel
>> deadlocks in this situation:
>>
>>>  STACK:
>>>  0 blk_add_timer+206 [0x2981ea]
>>>  1 blk_rq_timed_out+132 [0x2982a8]
>>>  2 blk_abort_request+114 [0x29833e]
>>>  3 blk_abort_queue+92 [0x2983a8]
>>>  4 deactivate_path+74 [0x3e00009625a]
>>>  5 run_workqueue+236 [0x149e04]
>>>  6 worker_thread+294 [0x149fce]
>>>  7 kthread+110 [0x14f436]
>>>  8 kernel_thread_starter+6 [0x10941a]
>>
>> blk_abort_queue takes the queue_lock with spinlock_irqsave and walks
>> the timer_list with list_for_each_entry_safe. Since a path to a SCSI
>> device just failed, the rport state is FC_PORTSTATE_BLOCKED. This
>> rport state triggers blk_add_timer that calls list_add_tail to move
>> the request to the end of timer_list. Thus, the
>> list_for_each_entry_safe never reaches the end of the timer_list, it
>> continously moves the requests to the end of the list.
>>
> Hmm. That would be fixes by using list_splice() here:
>
> diff --git a/block/blk-timeout.c b/block/blk-timeout.c
> index a095353..67bcc3f 100644
> --- a/block/blk-timeout.c
> +++ b/block/blk-timeout.c
> @@ -209,12 +209,15 @@ void blk_abort_queue(struct request_queue *q)
> {
>        unsigned long flags;
>        struct request *rq, *tmp;
> +       LIST_HEAD(list);
>
>        spin_lock_irqsave(q->queue_lock, flags);
>
>        elv_abort_queue(q);
>
> -       list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list)
> +       list_splice_init(&q->timeout_list, &list);
> +
> +       list_for_each_entry_safe(rq, tmp, &list, timeout_list)
>                blk_abort_request(rq);
>
>        spin_unlock_irqrestore(q->queue_lock, flags);
>
>> The rport state FC_PORTSTATE_BLOCKED would end, when the function
>> fc_timeout_deleted_rport would run to remove the rport. But this
>> function was schedules from queue_delayed_work. The timer already
>> expired, but the timer function does not run, because the timer
>> interrupt is disabled from the spinlock_irqsave call.
>>
> .. but this shouldn't happen anymore when using splice, as
> the timer will be called _after_ the irqrestore above.

If this patch does not address the deadlock another option to look into
would be to run some testing without blk_abort_request (just using
elv_abort_queue) and not try to abort in flight IOs at this time. 

We observed reduced IO delays during storage failover testing (target
responsive but timing out IOs) with this code, but I do not have good
breakdown data on the number of IOs handled by elv_abort_queue vs
blk_abort_request vs IO delay (It is also config dependent).

-andmike
--
Michael Anderson
andmike@linux.vnet.ibm.com