Re: [RFC] open_by_handle() vs. EA inodes

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

On Fri, Jun 29, 2018 at 07:38:30PM +0000, Trond Myklebust wrote:
> On Fri, 2018-06-29 at 19:19 +0100, Al Viro wrote:
> > 	On ea_inode-enabled ext4 open_by_handle() (as well as knfsd,
> > etc.)
> > can get to EA inodes as long as it knows their inumbers - just pass
> > it
> > an fhandle with zeroed version bytes and the right inumber in it.
> > 
> > 	AFAICS, it's Not Nice(tm), especially since you can write to
> > those,
> > whether they are shared or not.
> > 
> > 	Should we make ext4_nfs_get_inode() check for EXT4_EA_INODE_FL
> > and fail if it's set?
> 
> handle_to_path() requires CAP_DAC_READ_SEARCH capabilities. Isn't that
> sufficiently restrictive for open_by_handle()?
> 
> Concerning knfsd, people can in theory enable subtree checking to
> enforce checking whether or not you are in an exported subtree. In
> practice that breaks rename, so people are strongly encouraged to
> disable subtree checking, and only to export complete filesystems.

Umm...  Do we ever want those accessed via fhandles, capabilities or
no capabilities?  IOW, is there any reason for ext4 ->fh_to_dentry()
to give access to such inodes?  Those are implementation internals,
same as e.g. journal inode...