Re: [PATCH 1/2] VFS: Make d_materialise_unique() enforce directory uniqueness

[Trond Myklebust <Trond.Myklebust@netapp.com>]

On Thu, 2006-10-19 at 21:40 -0700, Andrew Morton wrote:
> On Thu, 19 Oct 2006 17:03:58 -0400
> Trond Myklebust <Trond.Myklebust@netapp.com> wrote:
> 
> > +static struct dentry *__d_unalias(struct dentry *dentry, struct dentry *alias)
> > +{
> > +	struct mutex *m1 = NULL, *m2 = NULL;
> > +	struct dentry *ret;
> > +
> > +	/* If alias and dentry share a parent, then no extra locks required */
> > +	if (alias->d_parent == dentry->d_parent)
> > +		goto out_unalias;
> > +
> > +	/* Check for loops */
> > +	ret = ERR_PTR(-ELOOP);
> > +	if (d_isparent(alias, dentry))
> > +		goto out_err;
> > +
> > +	/* See lock_rename() */
> > +	ret = ERR_PTR(-EBUSY);
> > +	if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
> > +		goto out_err;
> > +	m1 = &dentry->d_sb->s_vfs_rename_mutex;
> > +	if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex))
> > +		goto out_err;
> > +	m2 = &alias->d_parent->d_inode->i_mutex;
> > +out_unalias:
> > +	d_move_locked(alias, dentry);
> > +	ret = alias;
> > +out_err:
> > +	spin_unlock(&dcache_lock);
> > +	if (m2)
> > +		mutex_unlock(m2);
> > +	if (m1)
> > +		mutex_unlock(m1);
> > +	return ret;
> > +}
> 
> The locking in there is, of course, gruesome.  There is no way in which it
> can be made reliable?

The generic lookup() code will grab the dir->i_mutex for the parent
directory before we get anywhere near __d_alias(). That pretty much
limits us to using mutex_trylock() since otherwise we break the nesting
rules for lock_rename(), and I can't see that we can release the
problematic dir->i_mutex without causing worse races.

Cheers,
  Trond