diff mbox

[v3,3/6] ext4: Online defrag not supported with DAX

Message ID 1455680059-20126-4-git-send-email-ross.zwisler@linux.intel.com (mailing list archive)
State Accepted
Commit 73f34a5e2ced
Headers show

Commit Message

Ross Zwisler Feb. 17, 2016, 3:34 a.m. UTC
Online defrag operations for ext4 are hard coded to use the page cache.
See ext4_ioctl() -> ext4_move_extents() -> move_extent_per_page()

When combined with DAX I/O, which circumvents the page cache, this can
result in data corruption.  This was observed with xfstests ext4/307 and
ext4/308.

Fix this by only allowing online defrag for non-DAX files.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
---
 fs/ext4/ioctl.c | 5 +++++
 1 file changed, 5 insertions(+)

Comments

Jan Kara Feb. 17, 2016, 9:34 p.m. UTC | #1
On Tue 16-02-16 20:34:16, Ross Zwisler wrote:
> Online defrag operations for ext4 are hard coded to use the page cache.
> See ext4_ioctl() -> ext4_move_extents() -> move_extent_per_page()
> 
> When combined with DAX I/O, which circumvents the page cache, this can
> result in data corruption.  This was observed with xfstests ext4/307 and
> ext4/308.
> 
> Fix this by only allowing online defrag for non-DAX files.
> 
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>

We need to handle this eventually but for now we are fine. You can add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
>  fs/ext4/ioctl.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/fs/ext4/ioctl.c b/fs/ext4/ioctl.c
> index 0f6c369..e32c86f 100644
> --- a/fs/ext4/ioctl.c
> +++ b/fs/ext4/ioctl.c
> @@ -583,6 +583,11 @@ group_extend_out:
>  				 "Online defrag not supported with bigalloc");
>  			err = -EOPNOTSUPP;
>  			goto mext_out;
> +		} else if (IS_DAX(inode)) {
> +			ext4_msg(sb, KERN_ERR,
> +				 "Online defrag not supported with DAX");
> +			err = -EOPNOTSUPP;
> +			goto mext_out;
>  		}
>  
>  		err = mnt_want_write_file(filp);
> -- 
> 2.5.0
> 
>
Ross Zwisler Feb. 17, 2016, 9:50 p.m. UTC | #2
On Tue, Feb 16, 2016 at 08:34:16PM -0700, Ross Zwisler wrote:
> Online defrag operations for ext4 are hard coded to use the page cache.
> See ext4_ioctl() -> ext4_move_extents() -> move_extent_per_page()
> 
> When combined with DAX I/O, which circumvents the page cache, this can
> result in data corruption.  This was observed with xfstests ext4/307 and
> ext4/308.
> 
> Fix this by only allowing online defrag for non-DAX files.

Jan,

Thinking about this a bit more, it's probably the case that the data
corruption I was observing was due to us skipping the writeback of the dirty
page cache pages because S_DAX was set.

I do think we have a problem with defrag because it is doing the extent
swapping using the page cache, and we won't flush the dirty pages due to
S_DAX being set.

This patch is the quick and easy answer, and is perhaps appropriate for v4.5.

Looking forward, though, what do you think the correct solution is?  Making an
extent swapper that doesn't use the page cache (as I believe XFS has? see
xfs_swap_extents()), or maybe just unsetting S_DAX while we do the defrag and
being careful to block out page faults and I/O?  Or is it acceptable to just
say that DAX and defrag are mutually exclusive for ext4?

Thanks,
- Ross

> 
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> ---
>  fs/ext4/ioctl.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/fs/ext4/ioctl.c b/fs/ext4/ioctl.c
> index 0f6c369..e32c86f 100644
> --- a/fs/ext4/ioctl.c
> +++ b/fs/ext4/ioctl.c
> @@ -583,6 +583,11 @@ group_extend_out:
>  				 "Online defrag not supported with bigalloc");
>  			err = -EOPNOTSUPP;
>  			goto mext_out;
> +		} else if (IS_DAX(inode)) {
> +			ext4_msg(sb, KERN_ERR,
> +				 "Online defrag not supported with DAX");
> +			err = -EOPNOTSUPP;
> +			goto mext_out;
>  		}
>  
>  		err = mnt_want_write_file(filp);
> -- 
> 2.5.0
>
Jan Kara Feb. 17, 2016, 10:10 p.m. UTC | #3
On Wed 17-02-16 14:50:37, Ross Zwisler wrote:
> On Tue, Feb 16, 2016 at 08:34:16PM -0700, Ross Zwisler wrote:
> > Online defrag operations for ext4 are hard coded to use the page cache.
> > See ext4_ioctl() -> ext4_move_extents() -> move_extent_per_page()
> > 
> > When combined with DAX I/O, which circumvents the page cache, this can
> > result in data corruption.  This was observed with xfstests ext4/307 and
> > ext4/308.
> > 
> > Fix this by only allowing online defrag for non-DAX files.
> 
> Jan,
> 
> Thinking about this a bit more, it's probably the case that the data
> corruption I was observing was due to us skipping the writeback of the dirty
> page cache pages because S_DAX was set.
> 
> I do think we have a problem with defrag because it is doing the extent
> swapping using the page cache, and we won't flush the dirty pages due to
> S_DAX being set.
> 
> This patch is the quick and easy answer, and is perhaps appropriate for v4.5.
> 
> Looking forward, though, what do you think the correct solution is?  Making an
> extent swapper that doesn't use the page cache (as I believe XFS has? see
> xfs_swap_extents()), or maybe just unsetting S_DAX while we do the defrag and
> being careful to block out page faults and I/O?  Or is it acceptable to just
> say that DAX and defrag are mutually exclusive for ext4?

For 4.5 I'd just say just make them exclusive. Long term, we could just
avoid using page cache for copying data - grab all the necessary locks,
wait for all DIO to complete, evict anything in pagecache, copy data, swap
extents. It could even result in a simpler code.

								Honza


> > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> > ---
> >  fs/ext4/ioctl.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> > 
> > diff --git a/fs/ext4/ioctl.c b/fs/ext4/ioctl.c
> > index 0f6c369..e32c86f 100644
> > --- a/fs/ext4/ioctl.c
> > +++ b/fs/ext4/ioctl.c
> > @@ -583,6 +583,11 @@ group_extend_out:
> >  				 "Online defrag not supported with bigalloc");
> >  			err = -EOPNOTSUPP;
> >  			goto mext_out;
> > +		} else if (IS_DAX(inode)) {
> > +			ext4_msg(sb, KERN_ERR,
> > +				 "Online defrag not supported with DAX");
> > +			err = -EOPNOTSUPP;
> > +			goto mext_out;
> >  		}
> >  
> >  		err = mnt_want_write_file(filp);
> > -- 
> > 2.5.0
> >
Dave Chinner Feb. 18, 2016, 12:12 a.m. UTC | #4
On Wed, Feb 17, 2016 at 02:50:37PM -0700, Ross Zwisler wrote:
> On Tue, Feb 16, 2016 at 08:34:16PM -0700, Ross Zwisler wrote:
> > Online defrag operations for ext4 are hard coded to use the page cache.
> > See ext4_ioctl() -> ext4_move_extents() -> move_extent_per_page()
> > 
> > When combined with DAX I/O, which circumvents the page cache, this can
> > result in data corruption.  This was observed with xfstests ext4/307 and
> > ext4/308.
> > 
> > Fix this by only allowing online defrag for non-DAX files.
> 
> Jan,
> 
> Thinking about this a bit more, it's probably the case that the data
> corruption I was observing was due to us skipping the writeback of the dirty
> page cache pages because S_DAX was set.
> 
> I do think we have a problem with defrag because it is doing the extent
> swapping using the page cache, and we won't flush the dirty pages due to
> S_DAX being set.
> 
> This patch is the quick and easy answer, and is perhaps appropriate for v4.5.
> 
> Looking forward, though, what do you think the correct solution is?  Making an
> extent swapper that doesn't use the page cache (as I believe XFS has? see
> xfs_swap_extents()),

XFS does the data copy in userspace using direct IO so we don't
care about whether DAX is enabled or not on either the source or
destination inode. i.e. xfs_swap_extents() is a pure
metadata operation, swapping the entire extent tree between two
inodes if the source data has not changed while the copy was in
progress.

Cheers,

Dave.
diff mbox

Patch

diff --git a/fs/ext4/ioctl.c b/fs/ext4/ioctl.c
index 0f6c369..e32c86f 100644
--- a/fs/ext4/ioctl.c
+++ b/fs/ext4/ioctl.c
@@ -583,6 +583,11 @@  group_extend_out:
 				 "Online defrag not supported with bigalloc");
 			err = -EOPNOTSUPP;
 			goto mext_out;
+		} else if (IS_DAX(inode)) {
+			ext4_msg(sb, KERN_ERR,
+				 "Online defrag not supported with DAX");
+			err = -EOPNOTSUPP;
+			goto mext_out;
 		}
 
 		err = mnt_want_write_file(filp);