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.
--
Dave Chinner
david(a)fromorbit.com