xfs: implement optimized fdatasync

Allow us to track the difference between timestamp and size updates
by using mark_inode_dirty from the I/O completion code, and checking
the VFS inode flags in xfs_file_fsync.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Alex Elder <aelder@sgi.com>

1 files changed, 12 insertions, 11 deletions

diff --git a/fs/xfs/linux-2.6/xfs_file.c b/fs/xfs/linux-2.6/xfs_file.c
index 6c283b7be8ab..43f9554adaac 100644
--- a/fs/xfs/linux-2.6/xfs_file.c
+++ b/fs/xfs/linux-2.6/xfs_file.c
@@ -97,16 +97,6 @@ xfs_iozero(
         return (-status);
 }
 
-/*
- * We ignore the datasync flag here because a datasync is effectively
- * identical to an fsync. That is, datasync implies that we need to write
- * only the metadata needed to be able to access the data that is written
- * if we crash after the call completes. Hence if we are writing beyond
- * EOF we have to log the inode size change as well, which makes it a
- * full fsync. If we don't write beyond EOF, the inode core will be
- * clean in memory and so we don't need to log the inode, just like
- * fsync.
- */
 STATIC int
 xfs_file_fsync(
         struct file             *file,
@@ -139,7 +129,18 @@ xfs_file_fsync(
          */
         xfs_ilock(ip, XFS_ILOCK_SHARED);
 
-        if (ip->i_update_core) {
+        /*
+         * First check if the VFS inode is marked dirty.  All the dirtying
+         * of non-transactional updates no goes through mark_inode_dirty*,
+         * which allows us to distinguish beteeen pure timestamp updates
+         * and i_size updates which need to be caught for fdatasync.
+         * After that also theck for the dirty state in the XFS inode, which
+         * might gets cleared when the inode gets written out via the AIL
+         * or xfs_iflush_cluster.
+         */
+        if (((dentry->d_inode->i_state & I_DIRTY_DATASYNC) ||
+            ((dentry->d_inode->i_state & I_DIRTY_SYNC) && !datasync)) &&
+            ip->i_update_core) {
                 /*
                  * Kick off a transaction to log the inode core to get the
                  * updates.  The sync transaction will also force the log.