Re: [PATCH v6 3/3] xfs: reject removal of realtime flag when datadev doesn't support DAX

[Dave Chinner <david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org>]

On Tue, Feb 20, 2018 at 03:01:09PM -0800, Darrick J. Wong wrote:
> On Sun, Feb 18, 2018 at 11:23:17AM +1100, Dave Chinner wrote:
> > On Fri, Feb 16, 2018 at 09:22:47AM -0800, Darrick J. Wong wrote:
> > > On Fri, Feb 16, 2018 at 10:04:26AM -0700, Dave Jiang wrote:
> > > > In a situation where the rt_dev is DAX and data_dev is not DAX, if the user
> > > > requests to remove the realtime flag via ioctl we can no longer support DAX
> > > > for that file. Dynamic changing of S_DAX on the inode is not supported due
> > > > to various complications in the existing implementation. Therefore until we
> > > > address the dynamic S_DAX change issues, we must disallow realtime flag
> > > > being removed.
> > > > 
> > > > Signed-off-by: Dave Jiang <dave.jiang-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
> > > > Reviewed-by: Christoph Hellwig <hch-jcswGhMUV9g@public.gmane.org>
> > > > ---
> > > >  fs/xfs/xfs_ioctl.c |   14 ++++++++++++++
> > > >  1 file changed, 14 insertions(+)
> > > > 
> > > > diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
> > > > index 2c70a0a4f59f..edd97d527fe8 100644
> > > > --- a/fs/xfs/xfs_ioctl.c
> > > > +++ b/fs/xfs/xfs_ioctl.c
> > > > @@ -1030,6 +1030,20 @@ xfs_ioctl_setattr_xflags(
> > > >  {
> > > >  	struct xfs_mount	*mp = ip->i_mount;
> > > >  	uint64_t		di_flags2;
> > > > +	struct inode		*inode = VFS_I(ip);
> > > > +	struct super_block	*sb = inode->i_sb;
> > > > +
> > > > +	/*
> > > > +	 * In the case that the inode is realtime, and we are trying to remove
> > > > +	 * the realtime flag, and the rtdev supports DAX but the datadev does
> > > > +	 * not support DAX, we can't allow the realtime flag to be removed
> > > > +	 * since we do not support dynamic S_DAX flag removal yet.
> > > > +	 */
> > > > +	if (XFS_IS_REALTIME_INODE(ip) &&
> > > > +	    !(fa->fsx_xflags & FS_XFLAG_REALTIME) &&
> > > > +	    bdev_dax_supported(mp->m_rtdev_targp->bt_bdev, sb->s_blocksize) &&
> > > > +	    !bdev_dax_supported(mp->m_ddev_targp->bt_bdev, sb->s_blocksize))
> > > 
> > > What happens here if we have a non-rt file that we're trying to turn
> > > into an rt file and the data dev supports dax but not the rt dev?
> > > 
> > > Changing the rt flag is only supported on files with no data blocks (no
> > > extents, no delalloc blocks), so why can't we remove S_DAX from an empty
> > > file?  There aren't any memory mappings or page cache to get in the way,
> > > correct?
> > 
> > File size can be non-zero, so you can have DAX read-over-hole
> > mappings present. I simply don't think it's safe to remove/add S_DAX
> > flags via ioctls right now. If we have a DAX capable rtdev, then the
> > only way we should allow rtdev+dax to be used right now is via the
> > RT inherit bit on the dir that creates files in the rtdev right from
> > the start. i.e. we can't set/remove the RT inode flag on an inode
> > via ioctl if rtdev+dax is enabled until the whole dynamic S_DAX
> > inode flag thing is resolved.
> 
> Could we deal with the restriction that the DAX flag can't change
> (whether by user ioctl or by toggling the rt flag) unless the file size
> is zero?  That adds another way setting/clearing the realtime flag can
> fail, but at least it'd be the same EINVAL.

I thought we still mmap a zero length file and get a page fault that
returns a zeroed page? Or does that segv?

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org