On Wed, Mar 04, 2020 at 06:29:07PM +0100, Greg KH wrote:

The rule should be, "whenever you pass a pointer to a device off, the
reference count is incremented".  Somehow that is not happening here and
RCU is not going to solve the issue really, it's only going to delay the
problem from showing up until much later.
...
The struct device refcount should be all that is needed, don't use RCU
just to "delay freeing this object until some later time because someone
else might have a pointer to id".  That's ripe for disaster.

I agree that this is a better fix than trying to continue to paper
over the problem.

That being said, I also think it would be better if we could *also*
send a notification to the file system (or device mapper) when a block
device has disappeared, so we can set a flag in struct super
indicating, "this is an ex-device" so that we don't have to have
potentially hundreds of I/O errors clogging up the console and/or any
error notification ifrastructure we might want to add in the future,
as we attempt to send I/O to a device is not coming back.  This would
allow us to short-circuit things like writeback, instead of letting
everything drain via pointless io_submits sending bios that will never
go anywhere useful.

					- Ted