[dm-devel] [PATCH 4/8] block: support delayed holder registration

[Christoph Hellwig]

On Tue, Jul 27, 2021 at 12:06:04PM -0400, Mike Snitzer wrote:
> This header starts to shine some light on what is motivating this
> series by touching on "all kinds of bad side effects" being fixed.
> Any chance you could elaborate what you've noticed/found/hit?

The proble mis that it leaves the queue in a weird half state.  The
normal states for a gendisk are:

 1) allocated		(after *alloc_disk)
 2) registered		(after add_disk*)
 3) unregistered	(after del_gendisk)

the delayed queue registration adds a weird half state where it is
sort of registered, except for in sysfs and the elevator.  I have
some pretty big changes between how the disk and queue interact
that tripped over it, but even right now code has to be very careful
in the takedown path to deal with the half-initialized disks.