Re: ext2 + -osync: not as easy as it seems

["Fernando Luis Vázquez Cao" <fernando@oss.ntt.co.jp>]

On Wed, 2009-01-14 at 11:59 -0500, Theodore Tso wrote: 
> On Wed, Jan 14, 2009 at 03:37:56PM +0100, Jan Kara wrote:
> > > Um, we have that already; the sync_inode() followed by
> > > blkdev_issue_flush() is the path taken by fdatasync(), I do believe.
> >
> >   Maybe ext4-patch-queue changes that area but in Linus's tree I see:
> > 
> >   if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
> >          goto out;
> > 
> >   So if we just overwrite some data, we send them to disk via fdatawrite()
> > and then we quickly bail out from ext4_sync_file() without doing
> > blkdev_issue_flush().
> 
> So you're thinking about fdatawrite() being called by some code path
> other than ext4_sync_file() before we call fsync()?  Yeah, that could
> happen....  I think that will only happen if the file is opened
> O_SYNC, but that raises another issue, which is that we're not forcing
> a flush for writes when the file is opened O_SYNC.

Hi Jan, Ted

Is something like the patch below what you had in mind?

--

From: Fernando Luis Vazquez Cao <fernando@oss.ntt.co.jp>
Subject: ext3: call blkdev_issue_flush on fsync

To ensure that bits are truly on-disk after an fsync or fdatasync, we
should call blkdev_issue_flush if barriers are supported.

Signed-off-by: Fernando Luis Vazquez Cao <fernando@oss.ntt.co.jp>
---

--- linux-2.6.29-rc1-orig/fs/ext4/fsync.c	2008-12-25 08:26:37.000000000 +0900
+++ linux-2.6.29-rc1/fs/ext4/fsync.c	2009-01-15 21:03:19.000000000 +0900
@@ -48,6 +48,7 @@ int ext4_sync_file(struct file *file, st
 {
 	struct inode *inode = dentry->d_inode;
 	journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
+	unsigned long i_state = inode->i_state;
 	int ret = 0;
 
 	J_ASSERT(ext4_journal_current_handle() == NULL);
@@ -79,22 +80,35 @@ int ext4_sync_file(struct file *file, st
 		goto out;
 	}
 
-	if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
-		goto out;
+	if (datasync && !(i_state & I_DIRTY_DATASYNC))
+		goto flush_blkdev;
 
 	/*
 	 * The VFS has written the file data.  If the inode is unaltered
 	 * then we need not start a commit.
 	 */
-	if (inode->i_state & (I_DIRTY_SYNC|I_DIRTY_DATASYNC)) {
+	if (i_state & (I_DIRTY_SYNC|I_DIRTY_DATASYNC)) {
 		struct writeback_control wbc = {
 			.sync_mode = WB_SYNC_ALL,
 			.nr_to_write = 0, /* sys_fsync did this */
 		};
 		ret = sync_inode(inode, &wbc);
-		if (journal && (journal->j_flags & JBD2_BARRIER))
-			blkdev_issue_flush(inode->i_sb->s_bdev, NULL);
+		/*
+		 * When there are no blocks attached to the journal transaction
+		 * some optimizations are possible, but if there were dirty
+		 * pages sync_inode() should have ensured that all data gets
+		 * actually written to disk. Thus, we can skip
+		 * blkdev_issue_flush() below.
+		 */
+		if (!(i_state & I_DIRTY_PAGES))
+			goto flush_blkdev;
 	}
+
+	goto out;
+
+flush_blkdev:
+	if (journal && (journal->j_flags & JBD2_BARRIER))
+		blkdev_issue_flush(inode->i_sb->s_bdev, NULL);
 out:
 	return ret;
 }