Re: [PATCH V3 4/7] ublk_drv: requeue rqs with recovery feature enabled

From: Ming Lei <ming.lei@redhat.com>
To: Ziyang Zhang <ZiyangZhang@linux.alibaba.com>
Cc: axboe@kernel.dk, xiaoguang.wang@linux.alibaba.com,
	linux-block@vger.kernel.org, linux-kernel@vger.kernel.org,
	joseph.qi@linux.alibaba.com
Subject: Re: [PATCH V3 4/7] ublk_drv: requeue rqs with recovery feature enabled
Date: Tue, 20 Sep 2022 12:41:48 +0800	[thread overview]
Message-ID: <YylEjEply6y+bs0B@T590> (raw)
In-Reply-To: <0642eab9-6124-ba42-1585-82eab1ff9e87@linux.alibaba.com>

On Tue, Sep 20, 2022 at 11:34:32AM +0800, Ziyang Zhang wrote:
> On 2022/9/20 11:18, Ming Lei wrote:
> > On Tue, Sep 20, 2022 at 11:04:30AM +0800, Ziyang Zhang wrote:
> >> On 2022/9/20 10:39, Ming Lei wrote:
> >>> On Tue, Sep 20, 2022 at 09:31:54AM +0800, Ziyang Zhang wrote:
> >>>> On 2022/9/19 20:39, Ming Lei wrote:
> >>>>> On Mon, Sep 19, 2022 at 05:12:21PM +0800, Ziyang Zhang wrote:
> >>>>>> On 2022/9/19 11:55, Ming Lei wrote:
> >>>>>>> On Tue, Sep 13, 2022 at 12:17:04PM +0800, ZiyangZhang wrote:
> >>>>>>>> With recovery feature enabled, in ublk_queue_rq or task work
> >>>>>>>> (in exit_task_work or fallback wq), we requeue rqs instead of
> >>>>>>>> ending(aborting) them. Besides, No matter recovery feature is enabled
> >>>>>>>> or disabled, we schedule monitor_work immediately.
> >>>>>>>>
> >>>>>>>> Signed-off-by: ZiyangZhang <ZiyangZhang@linux.alibaba.com>
> >>>>>>>> ---
> >>>>>>>>  drivers/block/ublk_drv.c | 34 ++++++++++++++++++++++++++++++++--
> >>>>>>>>  1 file changed, 32 insertions(+), 2 deletions(-)
> >>>>>>>>
> >>>>>>>> diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
> >>>>>>>> index 23337bd7c105..b067f33a1913 100644
> >>>>>>>> --- a/drivers/block/ublk_drv.c
> >>>>>>>> +++ b/drivers/block/ublk_drv.c
> >>>>>>>> @@ -682,6 +682,21 @@ static void ubq_complete_io_cmd(struct ublk_io *io, int res)
> >>>>>>>>  
> >>>>>>>>  #define UBLK_REQUEUE_DELAY_MS	3
> >>>>>>>>  
> >>>>>>>> +static inline void __ublk_abort_rq_in_task_work(struct ublk_queue *ubq,
> >>>>>>>> +		struct request *rq)
> >>>>>>>> +{
> >>>>>>>> +	pr_devel("%s: %s q_id %d tag %d io_flags %x.\n", __func__,
> >>>>>>>> +			(ublk_queue_can_use_recovery(ubq)) ? "requeue" : "abort",
> >>>>>>>> +			ubq->q_id, rq->tag, ubq->ios[rq->tag].flags);
> >>>>>>>> +	/* We cannot process this rq so just requeue it. */
> >>>>>>>> +	if (ublk_queue_can_use_recovery(ubq)) {
> >>>>>>>> +		blk_mq_requeue_request(rq, false);
> >>>>>>>> +		blk_mq_delay_kick_requeue_list(rq->q, UBLK_REQUEUE_DELAY_MS);
> >>>>>>>
> >>>>>>> Here you needn't to kick requeue list since we know it can't make
> >>>>>>> progress. And you can do that once before deleting gendisk
> >>>>>>> or the queue is recovered.
> >>>>>>
> >>>>>> No, kicking rq here is necessary.
> >>>>>>
> >>>>>> Consider USER_RECOVERY is enabled and everything goes well.
> >>>>>> User sends STOP_DEV, and we have kicked requeue list in
> >>>>>> ublk_stop_dev() and are going to call del_gendisk().
> >>>>>> However, a crash happens now. Then rqs may be still requeued
> >>>>>> by ublk_queue_rq() because ublk_queue_rq() sees a dying
> >>>>>> ubq_daemon. So del_gendisk() will hang because there are
> >>>>>> rqs leaving in requeue list and no one kicks them.
> >>>>>
> >>>>> Why can't you kick requeue list before calling del_gendisk().
> >>>>
> >>>> Yes, we can kick requeue list once before calling del_gendisk().
> >>>> But a crash may happen just after kicking but before del_gendisk().
> >>>> So some rqs may be requeued at this moment. But we have already
> >>>> kicked the requeue list! Then del_gendisk() will hang, right?
> >>>
> >>> ->force_abort is set before kicking in ublk_unquiesce_dev(), so
> >>> all new requests are failed immediately instead of being requeued,
> >>> right?
> >>>
> >>
> >> ->force_abort is not heplful here because there may be fallback wq running
> >> which can requeue rqs after kicking requeue list.
> > 
> > After ublk_wait_tagset_rqs_idle() returns, there can't be any
> > pending requests in fallback wq or task work, can there
> Please consider this case: a crash happens while ublk_stop_dev() is
> calling. In such case I cannot schedule quiesce_work or call
> ublk_wait_tagset_rqs_idle(). This is because quiesce_work has to
> accquire ub_mutex to quiesce request queue.

The issue can be addressed in the following way more reliably &
cleanly & consistently, then you needn't to switch between the
two modes.

ublk_stop_dev()

	if (ublk_can_use_recovery(ub)) {
		if (ub->dev_info.state == UBLK_S_DEV_LIVE)
				__ublk_quiesce_dev(ub);		//lockless version
		ublk_unquiesce_dev();
	}

Meantime not necessary to disable recovery feature in ublk_unquiesce_dev
any more.

thanks,
Ming