Re: use of the system work queue in ublk

[Ming Lei <ming.lei@redhat.com>]

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