From: Jeff Layton <jlayton@redhat.com>
Date: Thu, 14 May 2009 12:19:28 -0400
Subject: [fs] nfs: fix an f_mode/f_flags confusion in write.c
Message-id: 1242317968-10572-1-git-send-email-jlayton@redhat.com
O-Subject: [RHEL5.4 PATCH] BZ#490181: NFS: Fix an f_mode/f_flags confusion in fs/nfs/write.c
Bugzilla: 490181
RH-Acked-by: Josef Bacik <josef@redhat.com>

From: Trond Myklebust <Trond.Myklebust@netapp.com>

Another rather obvious bug that was fixed upstream a while ago, but
that we never picked up in RHEL. O_SYNC is stored in filp->f_flags,
not f_mode.

The existing check is equivalent to:

		!(file->f_mode & S_IFIFO)

...so this check is effectively non-existent. The result, I think, is
that the kernel extends writes to cover the entire page even when
O_SYNC is set. I believe this just has performance implications. If
an app is doing a bunch of writes to a file opened with O_SYNC, it's
regularly flushing the entire page contents rather than just the piece
that needs to be written.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

diff --git a/fs/nfs/write.c b/fs/nfs/write.c
index 31c3f23..e8c8faf 100644
--- a/fs/nfs/write.c
+++ b/fs/nfs/write.c
@@ -875,7 +875,7 @@ int nfs_updatepage(struct file *file, struct page *page,
 	 */
 	if (nfs_write_pageuptodate(page, inode) &&
 			inode->i_flock == NULL &&
-			!(file->f_mode & O_SYNC)) {
+			!(file->f_flags & O_SYNC)) {
 		loff_t end_offs = i_size_read(inode) - 1;
 		unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;