Re: [PATCH 2/6] fs: FS_IOC_GETUUID

[Amir Goldstein <amir73il@gmail.com>]

On Tue, Feb 6, 2024 at 12:49 AM Kent Overstreet
<kent.overstreet@linux.dev> wrote:
>
> On Tue, Feb 06, 2024 at 09:17:58AM +1100, Dave Chinner wrote:
> > On Mon, Feb 05, 2024 at 03:05:13PM -0500, Kent Overstreet wrote:
> > > Add a new generic ioctls for querying the filesystem UUID.
> > >
> > > These are lifted versions of the ext4 ioctls, with one change: we're not
> > > using a flexible array member, because UUIDs will never be more than 16
> > > bytes.
> > >
> > > This patch adds a generic implementation of FS_IOC_GETFSUUID, which
> > > reads from super_block->s_uuid; FS_IOC_SETFSUUID is left for individual
> > > filesystems to implement.
> > >

It's fine to have a generic implementation, but the filesystem should
have the option to opt-in for a specific implementation.

There are several examples, even with xfs and btrfs where ->s_uuid
does not contain the filesystem's UUID or there is more than one
uuid and ->s_uuid is not the correct one to expose to the user.

A model like ioctl_[gs]etflags() looks much more appropriate
and could be useful for network filesystems/FUSE as well.

> > > Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
> > > Cc: Christian Brauner <brauner@kernel.org>
> > > Cc: Jan Kara <jack@suse.cz>
> > > Cc: Dave Chinner <dchinner@redhat.com>
> > > Cc: "Darrick J. Wong" <djwong@kernel.org>
> > > Cc: Theodore Ts'o <tytso@mit.edu>
> > > Cc: linux-fsdevel@vger.kernel.or
> > > Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
> > > ---
> > >  fs/ioctl.c              | 16 ++++++++++++++++
> > >  include/uapi/linux/fs.h | 16 ++++++++++++++++
> > >  2 files changed, 32 insertions(+)
> > >
> > > diff --git a/fs/ioctl.c b/fs/ioctl.c
> > > index 76cf22ac97d7..858801060408 100644
> > > --- a/fs/ioctl.c
> > > +++ b/fs/ioctl.c
> > > @@ -763,6 +763,19 @@ static int ioctl_fssetxattr(struct file *file, void __user *argp)
> > >     return err;
> > >  }
> > >
> > > +static int ioctl_getfsuuid(struct file *file, void __user *argp)
> > > +{
> > > +   struct super_block *sb = file_inode(file)->i_sb;
> > > +
> > > +   if (WARN_ON(sb->s_uuid_len > sizeof(sb->s_uuid)))
> > > +           sb->s_uuid_len = sizeof(sb->s_uuid);
> >
> > A "get"/read only ioctl should not be change superblock fields -
> > this is not the place for enforcing superblock filed constraints.
> > Make a helper function super_set_uuid(sb, uuid, uuid_len) for the
> > filesystems to call that does all the validity checking and then
> > sets the superblock fields appropriately.
>
> *nod* good thought...
>
> > > +struct fsuuid2 {
> > > +   __u32       fsu_len;
> > > +   __u32       fsu_flags;
> > > +   __u8        fsu_uuid[16];
> > > +};
> >
> > Nobody in userspace will care that this is "version 2" of the ext4
> > ioctl. I'd just name it "fs_uuid" as though the ext4 version didn't
> > ever exist.
>
> I considered that - but I decided I wanted the explicit versioning,
> because too often we live with unfixed mistakes because versioning is
> ugly, or something?
>
> Doing a new revision of an API should be a normal, frequent thing, and I
> want to start making it a convention.
>
> >
> > > +
> > >  /* extent-same (dedupe) ioctls; these MUST match the btrfs ioctl definitions */
> > >  #define FILE_DEDUPE_RANGE_SAME             0
> > >  #define FILE_DEDUPE_RANGE_DIFFERS  1
> > > @@ -215,6 +229,8 @@ struct fsxattr {
> > >  #define FS_IOC_FSSETXATTR          _IOW('X', 32, struct fsxattr)
> > >  #define FS_IOC_GETFSLABEL          _IOR(0x94, 49, char[FSLABEL_MAX])
> > >  #define FS_IOC_SETFSLABEL          _IOW(0x94, 50, char[FSLABEL_MAX])
> > > +#define FS_IOC_GETFSUUID           _IOR(0x94, 51, struct fsuuid2)
> > > +#define FS_IOC_SETFSUUID           _IOW(0x94, 52, struct fsuuid2)
> >
> > 0x94 is the btrfs ioctl space, not the VFS space - why did you
> > choose that? That said, what is the VFS ioctl space identifier? 'v',
> > perhaps?
>
> "Promoting ioctls from fs to vfs without revising and renaming
> considered harmful"... this is a mess that could have been avoided if we
> weren't taking the lazy route.
>
> And 'v' doesn't look like it to me, I really have no idea what to use
> here. Does anyone?
>

All the other hoisted FS_IOC_* use the original fs ioctl namespace they
came from. Although it is not an actual hoist, I'd use:

struct fsuuid128 {
       __u32       fsu_len;
       __u32       fsu_flags;
       __u8        fsu_uuid[16];
};

#define FS_IOC_GETFSUUID              _IOR('f', 45, struct fsuuid128)
#define FS_IOC_SETFSUUID              _IOW('f', 46, struct fsuuid128)

Technically, could also overload EXT4_IOC_[GS]ETFSUUID numbers
because of the different type:

#define FS_IOC_GETFSUUID              _IOR('f', 44, struct fsuuid128)
#define FS_IOC_SETFSUUID              _IOW('f', 44, struct fsuuid128)

and then ext4 can follow up with this patch, because as far as I can tell,
the ext4 implementation is already compatible with the new ioctls.

Thanks,
Amir.

--- a/fs/ext4/ioctl.c
+++ b/fs/ext4/ioctl.c
@@ -1613,8 +1613,10 @@ static long __ext4_ioctl(struct file *filp,
unsigned int cmd, unsigned long arg)
                return ext4_ioctl_setlabel(filp,
                                           (const void __user *)arg);

+       case FS_IOC_GETFSUUID:
         case EXT4_IOC_GETFSUUID:
                 return ext4_ioctl_getuuid(EXT4_SB(sb), (void __user *)arg);
+       case FS_IOC_SETFSUUID:
         case EXT4_IOC_SETFSUUID:
                 return ext4_ioctl_setuuid(filp, (const void __user *)arg);