Re: [PATCH 1/2] md: make devices disappear when they are no longer needed.

[Tejun Heo <tj@kernel.org>]

(cc'ing Jens)

Neil Brown wrote:
> On Monday November 24, tj@kernel.org wrote:
>> (cc'ing Greg)
>>
>> NeilBrown wrote:
>>> Currently md devices, once created, never disappear until the module
>>> is unloaded.  This is essentially because the gendisk holds a
>>> reference to the mddev, and the mddev holds a reference to the
>>> gendisk, this a circular reference.
>>>
>>> If we drop the reference from mddev to gendisk, then we need to ensure
>>> that the mddev is destroyed when the gendisk is destroyed.  However it
>>> is not possible to hook into the gendisk destruction process to enable
>>> this.
>>>
>>> So we drop the reference from the gendisk to the mddev and destroy the
>>> gendisk when the mddev gets destroyed.  However this has a
>>> complication.
>>> Between the call
>>>    __blkdev_get->get_gendisk->kobj_lookup->md_probe
>>> and the call
>>>    __blkdev_get->md_open
>>>
>>> there is no obvious way to hold a reference on the mddev any more, so
>>> unless something is done, it will disappear and gendisk will be
>>> destroyed prematurely.
>>>
>>> Also, once we decide to destroy the mddev, there will be an unlockable
>>> moment before the gendisk is unlinked (blk_unregister_region) during
>>> which a new reference to the gendisk can be created.  We need to
>>> ensure that this reference can not be used.  i.e. the ->open must
>>> fail.
>> Ah... I'm not really sure I'm following all of this correctly but would
>> it be possible to just add ->release to genhd and do regular reference
>> counting rather than this complex dancing?  ->release was recently added
>> to cdev so it'll be nicely parallel.
> 
> Maybe...
> 
> If genhd.c:disk_release called e.g.
> 	disk->fops->final_put(disk)
> 
> then I could possibly link in to that to destroy the md state when the
> gendisk finally disappears.
> 
> When I want to kill the gendisk I would call blk_unregister_region
> directly (not through del_gendisk) to allow it to disappear.
> If md_probe then gets called before the final_put, I'd need to
> call blk_register_region again to re-install it.
> 
> I think that would work.
> 
> Would 'block_device_operations' be the right place for this
> 'final_put' or 'final_release' ??

I suppose so.  Maybe just void (*release)(struct gendisk *) but Jens is
the maintainer.  Jens, what do you think?

Thanks.

-- 
tejun