sr.c kobject refcounting got buggered [Re: Ooops unmounting a defect DVD]

[Al Viro <viro@parcelfarce.linux.theplanet.co.uk>]

On Fri, Jan 28, 2005 at 06:42:44PM +0100, Oliver Neukum wrote:
> I got this oops unmounting by "eject" a defect DVD on a genuine
> SCSI drive.

Looks like failing IO + close afterwards - umount is irrelevant here.
And oops itself looks like cdrom_release((void *)0x18, whatever),
called from sr_block_release().  Which is
static int sr_block_release(struct inode *inode, struct file *file)
{
        int ret;
        struct scsi_cd *cd = scsi_cd(inode->i_bdev->bd_disk);
        ret = cdrom_release(&cd->cdi, file);
and since cdi is at offset 0x18 on i386, we have
	scsi_cd(inode->i_bdev->bd_disk) == NULL
IOW,
	inode->i_bdev->bd_disk->private_data == NULL
at the time of sr_block_release().  Which would be a problem, indeed.
AFAICS, the only place that could cause that crap is
static void sr_kref_release(struct kref *kref)
{
        struct scsi_cd *cd = container_of(kref, struct scsi_cd, kref);
        struct gendisk *disk = cd->disk;

        spin_lock(&sr_index_lock);
        clear_bit(disk->first_minor, sr_index_bits);
        spin_unlock(&sr_index_lock);

        unregister_cdrom(&cd->cdi);

        disk->private_data = NULL;

        put_disk(disk);

        kfree(cd);
}

so we have scsi_cd refcount reaching zero (and scsi_cd being freed) before
the final close of /dev/sr<whatever>...