diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index 4e73dd4a4d82f5..8c7cbe7f47eff0 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -2297,11 +2297,26 @@ xfs_ifree_cluster(
 		 * This buffer may not have been correctly initialised as we
 		 * didn't read it from disk. That's not important because we are
 		 * only using to mark the buffer as stale in the log, and to
-		 * attach stale cached inodes on it. That means it will never be
-		 * dispatched for IO. If it is, we want to know about it, and we
-		 * want it to fail. We can acheive this by adding a write
-		 * verifier to the buffer.
+		 * attach stale cached inodes on it.
+		 *
+		 * For the inode that triggered the cluster freeing, this
+		 * attachment may occur in xfs_inode_item_precommit() after we
+		 * have marked this buffer stale.  If this buffer was not in
+		 * memory before xfs_ifree_cluster() started, it will not be
+		 * marked XBF_DONE and this will cause problems later in
+		 * xfs_inode_item_precommit() when we trip over a (stale, !done)
+		 * buffer to attached to the transaction.
+		 *
+		 * Hence we have to mark the buffer as XFS_DONE here. This is
+		 * safe because we are also marking the buffer as XBF_STALE and
+		 * XFS_BLI_STALE. That means it will never be dispatched for
+		 * IO and it won't be unlocked until the cluster freeing has
+		 * been committed to the journal and the buffer unpinned. If it
+		 * is written, we want to know about it, and we want it to
+		 * fail. We can acheive this by adding a write verifier to the
+		 * buffer.
 		 */
+		bp->b_flags |= XBF_DONE;
 		bp->b_ops = &xfs_inode_buf_ops;
 
 		/*