On Fri 04-08-17 17:07:11, Mimi Zohar wrote:
> On Thu, 2017-08-03 at 12:56 +0200, Jan Kara wrote:
> > On Wed 02-08-17 13:11:52, Mimi Zohar wrote:
> > > On Wed, 2017-08-02 at 10:01 +0200, Jan Kara wrote:
> > > > On Tue 01-08-17 16:24:30, Mimi Zohar wrote:
> > > > > From: Christoph Hellwig <hch@xxxxxx>
> > > > >
> > > > > Add a new ->integrity_read file operation to read data for integrity
> > > > > hash collection. This is defined to be equivalent to ->read_iter,
> > > > > except that it will be called with the i_rwsem held exclusively.
> > > > >
> > > > > Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> > > > > Cc: Matthew Garrett <matthew.garrett@xxxxxxxxxx>
> > > > > Cc: Jan Kara <jack@xxxxxxxx>
> > > > > Cc: "Theodore Ts'o" <tytso@xxxxxxx>
> > > > > Cc: Andreas Dilger <adilger.kernel@xxxxxxxxx>
> > > > > Cc: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > > > Cc: Chao Yu <yuchao0@xxxxxxxxxx>
> > > > > Cc: Steven Whitehouse <swhiteho@xxxxxxxxxx>
> > > > > Cc: Bob Peterson <rpeterso@xxxxxxxxxx>
> > > > > Cc: David Woodhouse <dwmw2@xxxxxxxxxxxxx>
> > > > > Cc: Dave Kleikamp <shaggy@xxxxxxxxxx>
> > > > > Cc: Ryusuke Konishi <konishi.ryusuke@xxxxxxxxxxxxx>
> > > > > Cc: Mark Fasheh <mfasheh@xxxxxxxxxxx>
> > > > > Cc: Joel Becker <jlbec@xxxxxxxxxxxx>
> > > > > Cc: Richard Weinberger <richard@xxxxxx>
> > > > > Cc: "Darrick J. Wong" <darrick.wong@xxxxxxxxxx>
> > > > > Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> > > > > Cc: Chris Mason <clm@xxxxxx>
> > > > > Signed-off-by: Mimi Zohar <zohar@xxxxxxxxxxxxxxxxxx>
> > > >
> > > > ...
> > > >
> > > > > +static ssize_t ext4_file_integrity_read_iter(struct kiocb *iocb,
> > > > > + struct iov_iter *to)
> > > > > +{
> > > > > + struct inode *inode = file_inode(iocb->ki_filp);
> > > > > + int o_direct = iocb->ki_flags & IOCB_DIRECT;
> > > > > +
> > > > > + lockdep_assert_held(&inode->i_rwsem);
> > > > > + if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
> > > > > + return -EIO;
> > > > > +
> > > > > + if (!iov_iter_count(to))
> > > > > + return 0; /* skip atime */
> > > > > +
> > > > > +#ifdef CONFIG_FS_DAX
> > > > > + if (IS_DAX(inode))
> > > > > + return dax_iomap_rw(iocb, to, &ext4_iomap_ops);
> > > > > +#endif
> > > > > + if (o_direct)
> > > > > + return -EINVAL;
> > > > > + return generic_file_read_iter(iocb, to);
> > > > > +}
> > > >
> > > > I have noticed this o_direct check - why is it only in ext4 and shouldn't
> > > > rather higher layers make sure IOCB_DIRECT iocbs cannot reach
> > > > .integrity_read() methods?
> > >
> > > This failure happens when opening a file with O_DIRECT on a block
> > > device that does not support dax (eg. loop). xfs makes it to here too,
> > > but the call to generic_file_read_iter() fails properly with -EINVAL.
> > > (Only tested on those filesystems included that support dax (eg. ext2,
> > > ext4, and xfs).)
> >
> > Well, yes, ext2 and ext4 will silently fall back to buffered read (as do
> > pretty much all filesystems except for XFS). However I fail to see why IMA
> > should care (which is probably due to my lack of knowledge about IMA).
> > Is O_DIRECT somehow excepted from IMA? But then why it is not for DAX?
>
> Thank you for the explanation! (I was confused about the relationship
> between O_DIRECT and DAX.) You're correct. IMA does not support
> O_DIRECT in the buffered case for two reasons, locking and updating
> the file hash, which are described in commit f9b2a735bddd "ima: audit
> log files opened with O_DIRECT flag". After reverting this commit,
> the O_DIRECT check is needed before calling generic_file_read_iter().
Thanks for the pointer. This cleaned up the issue for me.
> Most likely the same would need to be done for other filesystems that
> support O_DIRECT. Probably a generic_integrity_file_read_iter()
> should be defined.
Yeah, then please define some common helper that takes care of refusing
direct IO - IMO you should check this even before calling into
->integrity_read helper.
> For DAX, unlike do_blockdev_direct_IO() which takes the lock, reading
> the file with O_DIRECT is fine, as dax_iomap_rw() only checks that the
> lock has been taken. Assuming the file system is mounted with
> i_version, the file hash is updated properly.
Yes, for DAX direct IO is basically no different but frankly I would just
refuse O_DIRECT on DAX inodes as well just for the consistency sake.
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
