[PATCH v2 00/25] fs: use type-safe uid representation for filesystem capabilities

[Christian Brauner]

On Wed, Feb 21, 2024 at 03:24:31PM -0600, Seth Forshee (DigitalOcean) wrote:
> This series converts filesystem capabilities from passing around raw
> xattr data to using a kernel-internal representation with type safe
> uids, similar to the conversion done previously for posix ACLs.
> Currently fscaps representations in the kernel have two different
> instances of unclear or confused types:
> 
> - fscaps are generally passed around in the raw xattr form, with the
>   rootid sometimes containing the user uid value and at other times
>   containing the filesystem value.
> - The existing kernel-internal representation of fscaps,
>   cpu_vfs_cap_data, uses the kuid_t type, but the value stored is
>   actually a vfsuid.
> 
> This series eliminates this confusion by converting the xattr data to
> the kernel representation near the userspace and filesystem boundaries,
> using the kernel representation within the vfs and commoncap code. The
> internal representation is renamed to vfs_caps to reflect this broader
> use, and the rootid is changed to a vfsuid_t to correctly identify the
> type of uid which it contains.
> 
> New vfs interfaces are added to allow for getting and setting fscaps
> using the kernel representation. This requires the addition of new inode
> operations to allow overlayfs to handle fscaps properly; all other
> filesystems fall back to a generic implementation. The top-level vfs
> xattr interfaces will now reject fscaps xattrs, though the lower-level
> interfaces continue to accept them for reading and writing the raw xattr
> data.
> 
> Based on previous feedback, new security hooks are added for fscaps
> operations. These are really only needed for EVM, and the selinux and
> smack implementations just peform the same operations that the
> equivalent xattr hooks would have done. Note too that this has not yet
> been updated based on the changes to make EVM into an LSM.
> 
> The remainder of the changes are preparatory work, addition of helpers
> for converting between the xattr and kernel fscaps representation, and
> various updates to use the kernel representation and new interfaces.

I still think that the generic_{get,set,remove}_fscaps() helpers falling
back to plain *vfs_*xattr() calls is a hackish. So ideally I'd like to
see this killed in a follow-up series and make all fses that support
them use the inode operation.

> 
> I have tested this code with xfstests, ltp, libcap2, and libcap-ng with
> no regressions found.

+1