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

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Nov 25, 2008 at 01:08:09AM +0900, Tejun Heo wrote:

> Devices can be killed from userland via sysfs for SCSI or mdadm for
> md.  It's true that such approach is less convenient for unloading but
> if it can make usual cases easier, why not?

_What_ usual cases?

> > And the right way to deal with that is to have explicit boundaries for
> > "opened or in process of being opened"; we almost have them (probe and
> > final release), so the only point we are missing is on failure exit from
> > __blkdev_get()...
> > 
> > I really think that it's much saner than trying to change the lifetime
> > rules for gendisk, etc.
> 
> Well, I don't know.  It seems like a lot of trouble just to allow
> "rmmod something" without first killing the devices and as people are
> now so used to reference counted objects and ->release, not having it
> on cdev or gendisk is quite a PITA.  (BTW, Greg, can you please drop
> cdev->release patch for now, it's wrong as it currently stands).

gendisks *ARE* reference counted, damnit.  So are net_device and a lot
of other things.  And no, it's not true that "struct net_device exists"
implies "the low-level objects that once might have been related to it
still exist" either.

> Can you see any problem with caching ->disk_release existence on
> registration and wrap __module_get/put() around its invocation?  It
> wouldn't change behavior of any existing drivers and md can use it if
> it wants.  Doing "mdadm --stop --scan" would be enough to unload the
> module and md can do whatever forward or back reference it wants to do
> to work out the weird userland interface.

Other than general ugliness and special-casing where none is really needed?
Special-casing as "very different life cycle if special method is present"...

If anything, we need to go in opposite direction - give the drivers a way
to say "my underlying object is gone, STFU and don't bother me with that
gendisk ever again; free it when you are done with it, but from now on
any access to it would better fail.  Oh, and I might find a new device
in place of that any time now, so new open() would better get not fail
just something in VFS still has a reference to that gendisk".

Which is doable - note that we can unhash block_device, dissociate inodes
from it and let new open() DTRT.  Earlier opened files will still have
a reference to address_space of original block_device (which is why we
have file->f_mapping instead of going through ->f_dentry->d_inode->i_mapping),
so we are fine.