Re: [RFC PATCH v2 16/18] ceph: add support to readdir for encrypted filenames

[Jeff Layton <jlayton@kernel.org>]

On Mon, 2020-09-07 at 22:34 -0700, Eric Biggers wrote:
> On Fri, Sep 04, 2020 at 12:05:35PM -0400, Jeff Layton wrote:
> > diff --git a/fs/ceph/crypto.h b/fs/ceph/crypto.h
> > index 1b11e9af165e..88c672ccdcf8 100644
> > --- a/fs/ceph/crypto.h
> > +++ b/fs/ceph/crypto.h
> > @@ -6,6 +6,8 @@
> >  #ifndef _CEPH_CRYPTO_H
> >  #define _CEPH_CRYPTO_H
> >  
> > +#include <linux/fscrypt.h>
> > +
> >  #ifdef CONFIG_FS_ENCRYPTION
> >  
> >  #define	CEPH_XATTR_NAME_ENCRYPTION_CONTEXT	"encryption.ctx"
> > @@ -16,6 +18,29 @@ int ceph_fscrypt_set_ops(struct super_block *sb);
> >  int ceph_fscrypt_prepare_context(struct inode *dir, struct inode *inode,
> >  				 struct ceph_acl_sec_ctx *as);
> >  
> > +static inline int ceph_fname_alloc_buffer(struct inode *parent, struct fscrypt_str *fname)
> > +{
> > +	if (!IS_ENCRYPTED(parent))
> > +		return 0;
> > +	return fscrypt_fname_alloc_buffer(NAME_MAX, fname);
> > +}
> > +
> > +static inline void ceph_fname_free_buffer(struct inode *parent, struct fscrypt_str *fname)
> > +{
> > +	if (IS_ENCRYPTED(parent))
> > +		fscrypt_fname_free_buffer(fname);
> > +}
> > +
> > +static inline int ceph_get_encryption_info(struct inode *inode)
> > +{
> > +	if (!IS_ENCRYPTED(inode))
> > +		return 0;
> > +	return fscrypt_get_encryption_info(inode);
> > +}
> > +
> > +int ceph_fname_to_usr(struct inode *parent, char *name, u32 len,
> > +			struct fscrypt_str *tname, struct fscrypt_str *oname);
> > +
> >  #else /* CONFIG_FS_ENCRYPTION */
> >  
> >  #define DUMMY_ENCRYPTION_ENABLED(fsc) (0)
> > @@ -31,6 +56,28 @@ static inline int ceph_fscrypt_prepare_context(struct inode *dir, struct inode *
> >  	return 0;
> >  }
> >  
> > +static inline int ceph_fname_alloc_buffer(struct inode *parent, struct fscrypt_str *fname)
> > +{
> > +	return 0;
> > +}
> > +
> > +static inline void ceph_fname_free_buffer(struct inode *parent, struct fscrypt_str *fname)
> > +{
> > +}
> > +
> > +static inline int ceph_get_encryption_info(struct inode *inode)
> > +{
> > +	return 0;
> > +}
> 
> This makes it so that readdir will succeed on encrypted directories when
> !CONFIG_FS_ENCRYPTION.  The other filesystems instead return an error code,
> which seems much better.  Can you check what the other filesystems handle
> readdir?
>

Maybe. I'm not sure it's better.

It would be nice to be able to allow such clients to be able to clean
out an encrypted tree (given appropriate permissions, of course).

A network filesystem like this is a much different case than a local
one. We may have a swath of varying client kernel versions and
configurations that are operating on the same filesystem.

Where we draw that line is still being determined though.

> > +static bool fscrypt_key_status_change(struct dentry *dentry)
> > +{
> > +	struct inode *dir;
> > +	bool encrypted_name, have_key;
> > +
> > +	lockdep_assert_held(&dentry->d_lock);
> > +
> > +	dir = d_inode(dentry->d_parent);
> > +	if (!IS_ENCRYPTED(dir))
> > +		return false;
> > +
> > +	encrypted_name = dentry->d_flags & DCACHE_ENCRYPTED_NAME;
> > +	have_key = fscrypt_has_encryption_key(dir);
> > +
> > +	if (encrypted_name == have_key)
> > +		ceph_dir_clear_complete(dir);
> > +
> > +	dout("%s encrypted_name=%d have_key=%d\n", __func__, encrypted_name, have_key);
> > +	return encrypted_name == have_key;
> > +}
> > +
> 
> Only the no-key => key case needs to be handled, not key => no-key.
> Also, the caller already has 'dir', so there's no need to use ->d_parent.
> 
> What's wrong with just:
> 
>                 di = ceph_dentry(dentry);
>                 if (d_unhashed(dentry) ||
>                     d_really_is_negative(dentry) ||
>                     di->lease_shared_gen != shared_gen ||
> +                   ((dentry->d_flags & DCACHE_ENCRYPTED_NAME) &&
> +                    fscrypt_has_encryption_key(dir)))  {
>                         spin_unlock(&dentry->d_lock);
>                         dput(dentry);
>                         err = -EAGAIN;
>                         goto out;
>                 }

Ok, I didn't realize that I didn't need to worry about key removal. Your
proposed scheme is simpler.

> >  /*
> >   * When possible, we try to satisfy a readdir by peeking at the
> >   * dcache.  We make this work by carefully ordering dentries on
> > @@ -238,11 +261,11 @@ static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
> >  			goto out;
> >  		}
> >  
> > -		spin_lock(&dentry->d_lock);
> 
> Why delete this spin_lock()?
> 

Yikes -- good catch! Fixed.

> > +			if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode)) {
> > +				spin_lock(&dn->d_lock);
> > +				dn->d_flags |= DCACHE_ENCRYPTED_NAME;
> > +				spin_unlock(&dn->d_lock);
> > +			}
> 
> This is racy because fscrypt_has_encryption_key() could have been false when the
> dentry was created, then true here.
> 
> Take a look at how __fscrypt_prepare_lookup() solves this problem.

Blech.

I guess I need to have ceph_fname_to_usr tell whether the name is a
nokey name, but that info is not currently returned
by fscrypt_fname_disk_to_usr. I guess it will need to be...

-- 
Jeff Layton <jlayton@kernel.org>