Re: [PATCH v3 06/12] fs/xfs: Check if the inode supports DAX under lock

From: Dave Chinner <david@fromorbit.com>
To: ira.weiny@intel.com
Cc: linux-kernel@vger.kernel.org,
	Alexander Viro <viro@zeniv.linux.org.uk>,
	"Darrick J. Wong" <darrick.wong@oracle.com>,
	Dan Williams <dan.j.williams@intel.com>,
	Christoph Hellwig <hch@lst.de>,
	"Theodore Y. Ts'o" <tytso@mit.edu>, Jan Kara <jack@suse.cz>,
	linux-ext4@vger.kernel.org, linux-xfs@vger.kernel.org,
	linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH v3 06/12] fs/xfs: Check if the inode supports DAX under lock
Date: Tue, 11 Feb 2020 17:16:39 +1100	[thread overview]
Message-ID: <20200211061639.GH10776@dread.disaster.area> (raw)
In-Reply-To: <20200208193445.27421-7-ira.weiny@intel.com>

On Sat, Feb 08, 2020 at 11:34:39AM -0800, ira.weiny@intel.com wrote:
> From: Ira Weiny <ira.weiny@intel.com>
> 
> One of the checks for an inode supporting DAX is if the inode is
> reflinked.  During a non-DAX to DAX state change we could race with
> the file being reflinked and end up with a reflinked file being in DAX
> state.
> 
> Prevent this race by checking for DAX support under the MMAP_LOCK.

The on disk inode flags are protected by the XFS_ILOCK, not the
MMAP_LOCK. i.e. the MMAPLOCK provides data access serialisation, not
metadata modification serialisation.

> 
> Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> ---
>  fs/xfs/xfs_ioctl.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
> index da1eb2bdb386..4ff402fd6636 100644
> --- a/fs/xfs/xfs_ioctl.c
> +++ b/fs/xfs/xfs_ioctl.c
> @@ -1194,10 +1194,6 @@ xfs_ioctl_setattr_dax_invalidate(
>  
>  	*join_flags = 0;
>  
> -	if ((fa->fsx_xflags & FS_XFLAG_DAX) == FS_XFLAG_DAX &&
> -	    !xfs_inode_supports_dax(ip))
> -		return -EINVAL;
> -
>  	/* If the DAX state is not changing, we have nothing to do here. */
>  	if ((fa->fsx_xflags & FS_XFLAG_DAX) &&
>  	    (ip->i_d.di_flags2 & XFS_DIFLAG2_DAX))
> @@ -1211,6 +1207,13 @@ xfs_ioctl_setattr_dax_invalidate(
>  
>  	/* lock, flush and invalidate mapping in preparation for flag change */
>  	xfs_ilock(ip, XFS_MMAPLOCK_EXCL | XFS_IOLOCK_EXCL);
> +
> +	if ((fa->fsx_xflags & FS_XFLAG_DAX) == FS_XFLAG_DAX &&
> +	    !xfs_inode_supports_dax(ip)) {
> +		error = -EINVAL;
> +		goto out_unlock;
> +	}

Yes, you might be able to get away with reflink vs dax flag
serialisation on the inode flag modification, but it is not correct and
leaves a landmine for future inode flag modifications that are done
without holding either the MMAP or IOLOCK.

e.g. concurrent calls to xfs_ioctl_setattr() setting/clearing flags
other than the on disk DAX flag are all serialised by the ILOCK_EXCL
and will no be serialised against this DAX check. Hence if there are
other flags that we add in future that affect the result of
xfs_inode_supports_dax(), this code will not be correctly
serialised.

This raciness in checking the DAX flags is the reason that
xfs_ioctl_setattr_xflags() redoes all the reflink vs dax checks once
it's called under the XFS_ILOCK_EXCL during the actual change
transaction....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com