[PATCH v9 14/14] xfs: test for realtime and reflink under the pNFS locks

[Chris Mason <mason@kernel.org>]

A pNFS client can write to blocks that its file shares with another
file, or that are no part of its file, if a LAYOUTGET races with a clone
or with FS_IOC_FSSETXATTR.

xfs_fs_map_blocks() hands out no layout for a reflinked or a realtime
inode, since the client writes in place and only knows the data device.
But it tests for both before it takes the IOLOCK, and both flags are set
under that lock:

    LAYOUTGET                           clone
    =========                           =====
                                        takes the IOLOCK
                                        breaks the layouts
    xfs_fs_map_blocks()
      sees no reflink flag
      waits for the IOLOCK
                                        sets the reflink flag
                                        shares the blocks
                                        drops the IOLOCK
      maps the shared blocks
    /* the client writes in place */

If FS_IOC_FSSETXATTR sets the realtime flag of an empty file instead,
the blocks come from the realtime device, and nfsd points the client at
the same address on the data device.

The fix used here is to test for both once the locks are held.

Fixes: e54b5bf9d7ae ("xfs: recognize the reflink feature bit")
Fixes: 527851124d10 ("xfs: implement pNFS export operations")
Signed-off-by: Chris Mason <mason@kernel.org>
---

Notes:
    v8->v9:
    - New patch.

 fs/xfs/xfs_pnfs.c | 32 +++++++++++++++++---------------
 1 file changed, 17 insertions(+), 15 deletions(-)

diff --git a/fs/xfs/xfs_pnfs.c b/fs/xfs/xfs_pnfs.c
index 065875c8b7da..945f7a3e11ca 100644
--- a/fs/xfs/xfs_pnfs.c
+++ b/fs/xfs/xfs_pnfs.c
@@ -147,21 +147,6 @@ xfs_fs_map_blocks(
 	if (xfs_is_shutdown(mp))
 		return -EIO;
 
-	/*
-	 * We can't export inodes residing on the realtime device.  The realtime
-	 * device doesn't have a UUID to identify it, so the client has no way
-	 * to find it.
-	 */
-	if (XFS_IS_REALTIME_INODE(ip))
-		return -ENXIO;
-
-	/*
-	 * The pNFS block layout spec actually supports reflink like
-	 * functionality, but the Linux pNFS server doesn't implement it yet.
-	 */
-	if (xfs_is_reflink_inode(ip))
-		return -ENXIO;
-
 	/*
 	 * Lock out any other I/O before we flush and invalidate the pagecache,
 	 * and then hand out a layout to the remote system.  This is very
@@ -174,6 +159,23 @@ xfs_fs_map_blocks(
 	 */
 	xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);
 
+	/*
+	 * We can't export inodes residing on the realtime device.  The realtime
+	 * device doesn't have a UUID to identify it, so the client has no way
+	 * to find it.  The flag is set under the IOLOCK, so test only now.
+	 */
+	error = -ENXIO;
+	if (XFS_IS_REALTIME_INODE(ip))
+		goto out_unlock;
+
+	/*
+	 * The pNFS block layout spec actually supports reflink like
+	 * functionality, but the Linux pNFS server doesn't implement it yet.
+	 * This flag is set under the IOLOCK as well.
+	 */
+	if (xfs_is_reflink_inode(ip))
+		goto out_unlock;
+
 	error = -EINVAL;
 	limit = mp->m_super->s_maxbytes;
 	if (!write)
-- 
2.55.0