[PATCH] xfs: break layouts before swapping extents

[PATCH] xfs: break layouts before swapping extents

[Daejun Park via B4 Relay]

From: Daejun Park <daejun7.park@samsung.com>

XFS_IOC_SWAPEXT does not break pNFS layouts. A client can keep writing
to blocks that the swap moved to the temporary file. xfs_fsr frees
those blocks when it closes that file. The writes then land in whatever
reuses them.

Lock with xfs_ilock2_io_mmap(), as the exchange-range ioctls do.

Fixes: 781355c6e5ae ("xfs: recall pNFS layouts on conflicting access")
Cc: stable@vger.kernel.org
Signed-off-by: Daejun Park <daejun7.park@samsung.com>
---
xfs_fsr uses SWAPEXT on kernels without XFS_IOC_COMMIT_RANGE, on
filesystems without exchange-range, and in all xfsprogs before 6.12.

When both files are DAX, xfs_ilock2_io_mmap() also breaks DAX layouts.
With two DAX files whose pages had been faulted in, SWAPEXT hit the
WARN_ON_ONCE in truncate_folio_batch_exceptionals() in 3 of 3 boots
before this patch and in 0 of 3 after it. A DAX file swapped with a
non-DAX file still hits it.

Tested with a pynfs client holding an RW block layout while the MDS ran
"xfs_io -c swapext". Before, the client's writes overwrote other files'
data and a directory block. After, the swap recalls the layout first.
The fstests fsr and swapext tests pass with and without the patch.
---
 fs/xfs/xfs_bmap_util.c | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c
index 865aec3ae5a..595b7ea3ac9 100644
--- a/fs/xfs/xfs_bmap_util.c
+++ b/fs/xfs/xfs_bmap_util.c
@@ -1554,9 +1554,9 @@ xfs_swap_extents(
 	 * page cache safely. Once we have done this we can take the ilocks and
 	 * do the rest of the checks.
 	 */
-	lock_two_nondirectories(VFS_I(ip), VFS_I(tip));
-	filemap_invalidate_lock_two(VFS_I(ip)->i_mapping,
-				    VFS_I(tip)->i_mapping);
+	error = xfs_ilock2_io_mmap(ip, tip);
+	if (error)
+		return error;
 
 	/* Verify that both files have the same format */
 	if ((VFS_I(ip)->i_mode & S_IFMT) != (VFS_I(tip)->i_mode & S_IFMT)) {
@@ -1762,9 +1762,7 @@ xfs_swap_extents(
 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 	xfs_iunlock(tip, XFS_ILOCK_EXCL);
 out_unlock:
-	filemap_invalidate_unlock_two(VFS_I(ip)->i_mapping,
-				      VFS_I(tip)->i_mapping);
-	unlock_two_nondirectories(VFS_I(ip), VFS_I(tip));
+	xfs_iunlock2_io_mmap(ip, tip);
 	return error;
 
 out_trans_cancel:

---
base-commit: 1e17fc56469d62d1c1ac45c137a2456d14456f1d
change-id: 20261109-xfs-swapext-break-layout-89fc24f5be8f

Best regards,
-- 
Daejun Park <daejun7.park@samsung.com>

Re: [PATCH] xfs: break layouts before swapping extents

[Christoph Hellwig]

Looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>