Re: [RFC PATCH v2] hostfs: handle idmapped mounts

[Glenn Washburn]

On Thu, 2 Mar 2023 09:39:28 +0100
Christian Brauner <brauner@xxxxxxxxxx> wrote:

> On Tue, Feb 28, 2023 at 07:50:02PM -0600, Glenn Washburn wrote:
> > Let hostfs handle idmapped mounts. This allows to have the same
> > hostfs mount appear in multiple locations with different id
> > mappings.
> > 
> > root@(none):/media# id
> > uid=0(root) gid=0(root) groups=0(root)
> > root@(none):/media# mkdir mnt idmapped
> > root@(none):/media# mount -thostfs -o/home/user hostfs mnt
> > 
> > root@(none):/media# touch mnt/aaa
> > root@(none):/media# mount-idmapped --map-mount u:`id -u user`:0:1
> > --map-mount g:`id -g user`:0:1 /media/mnt /media/idmapped
> > root@(none):/media# ls -l mnt/aaa idmapped/aaa -rw-r--r-- 1 root
> > root 0 Jan 28 01:23 idmapped/aaa -rw-r--r-- 1 user user 0 Jan 28
> > 01:23 mnt/aaa
> > 
> > root@(none):/media# touch idmapped/bbb
> > root@(none):/media# ls -l mnt/bbb idmapped/bbb
> > -rw-r--r-- 1 root root 0 Jan 28 01:26 idmapped/bbb
> > -rw-r--r-- 1 user user 0 Jan 28 01:26 mnt/bbb
> > 
> > Signed-off-by: Glenn Washburn <development@xxxxxxxxxxxxxxx>
> > ---
> > Changes from v1:
> >  * Rebase on to tip. The above commands work and have the results
> > expected. The __vfsuid_val(make_vfsuid(...)) seems ugly to get the
> > uid_t, but it seemed like the best one I've come across. Is there a
> > better way?
> 
> Sure, I can help you with that. ;)

Thank you!

> > 
> > Glenn
> > ---
> >  fs/hostfs/hostfs_kern.c | 13 +++++++------
> >  1 file changed, 7 insertions(+), 6 deletions(-)
> > 
> > diff --git a/fs/hostfs/hostfs_kern.c b/fs/hostfs/hostfs_kern.c
> > index c18bb50c31b6..9459da99a0db 100644
> > --- a/fs/hostfs/hostfs_kern.c
> > +++ b/fs/hostfs/hostfs_kern.c
> > @@ -786,7 +786,7 @@ static int hostfs_permission(struct mnt_idmap
> > *idmap, err = access_file(name, r, w, x);
> >  	__putname(name);
> >  	if (!err)
> > -		err = generic_permission(&nop_mnt_idmap, ino,
> > desired);
> > +		err = generic_permission(idmap, ino, desired);
> >  	return err;
> >  }
> >  
> > @@ -794,13 +794,14 @@ static int hostfs_setattr(struct mnt_idmap
> > *idmap, struct dentry *dentry, struct iattr *attr)
> >  {
> >  	struct inode *inode = d_inode(dentry);
> > +	struct user_namespace *fs_userns = i_user_ns(inode);
> 
> Fyi, since hostfs can't be mounted in a user namespace
> fs_userns == &init_user_ns
> so it doesn't really matter what you use.

What would you suggest as preferable?

> >  	struct hostfs_iattr attrs;
> >  	char *name;
> >  	int err;
> >  
> >  	int fd = HOSTFS_I(inode)->fd;
> >  
> > -	err = setattr_prepare(&nop_mnt_idmap, dentry, attr);
> > +	err = setattr_prepare(idmap, dentry, attr);
> >  	if (err)
> >  		return err;
> >  
> > @@ -814,11 +815,11 @@ static int hostfs_setattr(struct mnt_idmap
> > *idmap, }
> >  	if (attr->ia_valid & ATTR_UID) {
> >  		attrs.ia_valid |= HOSTFS_ATTR_UID;
> > -		attrs.ia_uid = from_kuid(&init_user_ns,
> > attr->ia_uid);
> > +		attrs.ia_uid = __vfsuid_val(make_vfsuid(idmap,
> > fs_userns, attr->ia_uid)); }
> >  	if (attr->ia_valid & ATTR_GID) {
> >  		attrs.ia_valid |= HOSTFS_ATTR_GID;
> > -		attrs.ia_gid = from_kgid(&init_user_ns,
> > attr->ia_gid);
> > +		attrs.ia_gid = __vfsgid_val(make_vfsgid(idmap,
> > fs_userns, attr->ia_gid));
> 
> Heh, if you look include/linux/fs.h:
> 
>         /*
>          * The two anonymous unions wrap structures with the same
> member. *
>          * Filesystems raising FS_ALLOW_IDMAP need to use
> ia_vfs{g,u}id which
>          * are a dedicated type requiring the filesystem to use the
> dedicated
>          * helpers. Other filesystem can continue to use ia_{g,u}id
> until they
>          * have been ported.
>          *
>          * They always contain the same value. In other words
> FS_ALLOW_IDMAP
>          * pass down the same value on idmapped mounts as they would
> on regular
>          * mounts.
>          */
>         union {
>                 kuid_t          ia_uid;
>                 vfsuid_t        ia_vfsuid;
>         };
>         union {
>                 kgid_t          ia_gid;
>                 vfsgid_t        ia_vfsgid;
>         };
> 
> this just is:
> 
> attrs.ia_uid = from_vfsuid(idmap, fs_userns, attr->ia_vfsuid));
> attrs.ia_gid = from_vfsgid(idmap, fs_userns, attr->ia_vfsgid));

Its easy to miss from this patch because of lack of context, but attrs
is a struct hostfs_iattr, not struct iattr. And attrs.ia_uid is of type
uid_t, not kuid_t. So the above fails to compile. This is why I needed
to wrap make_vfsuid() in __vfsuid_val() (to get the uid_t).

I had decided against using from_vfsuid() because then I thought I'd
need to use from_kuid() to get the uid_t. And from_kuid() takes the
namespace (again), which seemed uglier.

Based on this, what do you suggest?

Glenn