[PATCH V2 09/13] scsi: force unfreezing queue into atomic mode

[Christoph Hellwig]

On Tue, Jan 25, 2022 at 04:54:43PM +0800, Ming Lei wrote:
> > So what is the advantage of trying to remove the freeze from where
> > it belongs (common unregister code) while keeping it where it is a bandaid
> > (driver specific unregister code)?
> 
> freeze in common unregister code is actually not good, because it provide
> nothing for bio based driver, so we can't move blk-cgroup shutdown into
> del_gendisk. Also we can't move elevator shutdown to del_gendisk for
> similar reason.
> 
> Secondly freeze is pretty slow in percpu mode, so why slow down removing every
> disk just for scsi's bandaid?

I'd frame this differently:

 - del_gendisk is the right place to freeze the queue, as that is where the
   gendisk is unregistered and all fs I/O needs to stop.  If we don't get
   all aspects right we need to fix.  As mentioned I'm already looking into
   keeping a reference for the bio life time for bio based drivers.
 - SCSI is the only driver that ever submits passthrough I/O without the
   gendisk.  So doing an additional freeze during request_queue teardown
   is only needed for SCSI and we can eventually remove that for everyone
   else.  And most of your series already does really good work towards that
   goal!

> 
> 
> Thanks,
> Ming
---end quoted text---