Re: use of the system work queue in ublk

[Ziyang Zhang <ZiyangZhang@linux.alibaba.com>]

On 2022/7/18 15:35, Ming Lei wrote:
> Hi Christoph,
> 
> On Sun, Jul 17, 2022 at 11:38:25PM -0700, Christoph Hellwig wrote:
>> Hi Ming,
>>
>> it seems like ublk uses schedule_work to stop the device, which
>> includes a del_gendisk.  I'm a little fearful this will gets us into
>> lockdep chains of death once syzbot or Tetsu notice it.
> 
> ublksrv has two built-in tests(generic/001, generic/002) for covering
> heavy io with device removal and killing ubq_daemon, not see lockdep
> warning when running the two tests with lockdep enabled.
> 
> Could you or Tetsu provide a bit more info about the warning?
> 
>>
>> That being said, I don't reall understand the design of
>> ublk_daemon_monitor_work, which is only used to kick off other
>> work to start with.
> 
> If the ubq daemon becomes dead, ublk_daemon_monitor_work will be
> scheduled for handling the error: abort pending io requests, and
> start to delete disk.
> 
> It has to be triggered when del_gendisk() is in-progress for making
> forward progress.
> 
> 
> Thanks,
> Ming

Hi Ming,

Just to make sure I understand usage of ublk_daemon_monitor_work correctly.

1) For a dying ubq daemon, ublk_daemon_monitor_work schedule stop work first.

2) The stop work calls del_gendisk() and it is blocked because there are
   pending blk-mq requests(maybe handling in ublksrv target).

3) Meanwhile, the monitor work aborts all pending blk-mq IOs
   (with UBLK_IO_FLAG_ACTIVE unset) by blk_mq_end_request(req, BLK_STS_IOERR).

4) After all pending blk-mq reqs are aborted,
   del_gendisk() in stop work returns and /dev/ublkbX is removed.
   No more blk-mq reqs.

5) In stop work, cancel all queued(with UBLK_IO_FLAG_ACTIVE set) ublk IOs
   by io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0) and ublksrv won't
   issue sqes again.


Hope I am correct and all of these works looks good to me.
Besides, the tests(generic/001, generic/002) run successfully for me.

Regards,
Ziyang Zhang