Re: [PATCH -V7 3/9] vfs: Add name to file handle conversion support

["Aneesh Kumar K. V" <aneesh.kumar@linux.vnet.ibm.com>]

On Wed, 12 May 2010 15:49:49 -0600, Andreas Dilger <andreas.dilger@oracle.com> wrote:
> On 2010-05-12, at 09:50, Aneesh Kumar K.V wrote:
> > +static long do_sys_name_to_handle(struct path *path,
> > +			struct file_handle __user *ufh)
> > +{
> > +	if (handle_size <= f_handle.handle_size) {
> > +		/* get the uuid */
> > +		retval = sb->s_op->get_fsid(sb, &this_fs_id);
> > +		if (!retval) {
> > +			/*
> > +			 * Now verify whether we get the same vfsmount
> > +			 * if we lookup with uuid. In case we end up having
> > +			 * same uuid for the multiple file systems. When doing
> > +			 * uuid based lookup we would return the first one.So
> > +			 * with name_to_handle if we don't find the same
> > +			 * vfsmount with lookup return EOPNOTSUPP
> > +			 */
> > +			mnt = fs_get_vfsmount(current, &this_fs_id);
> > +			if (mnt != path->mnt) {
> > +				retval = -EOPNOTSUPP;
> > +				mntput(mnt);
> > +				goto err_free_out;
> > +			}
> 
> I don't see that this does anything for us except add overhead.  This
> is no protection against mounting a second filesystem with the same
> UUID after the handle is returned, since there is no expiration for
> file handles.

The new vfsmount gets added to the tail of the mount list. So if we
already have the handle then we can be sure that on open_by_handle we
get the same vfsmount. Of course if we umount the original file system
and mount it back we could get it wrong. That is also true with the
below scheme, ie if we umount both the file system mount the second one
which causes it to be added to the uuid hash table followed by the first
one which will not get added to uuid hash table because of same uuid.
In that case an open_by_handle will map the fsid to the wrong file system.

> 
> At best I think we could start by changing the list-based UUID lookup with a hash-based one, and when adding a duplicate UUID at mount time start by printing out an error message to the console in case of duplicated UUIDs, and maybe at some point in the future this might cause the mount to fail (though I don't think we can make that decision lightly or quickly).
> 
> That moves the overhead to mount time instead of for each name_to_handle() call (which would be brutal for a system with many filesystems mounted).
> 

-aneesh