On 1/20/12 2:34 PM, Jan Kara wrote:

m_active_trans counter is racy wrt filesystem freezing. The patch replaces it
with generic counter of running transactions which is properly synchronized
with filesystem freezing. Things are a bit more complex because we need to log
a dummy transaction and free block counters after the filesystem is frozen so
we need to pass information to _xfs_trans_alloc() whether the transaction is
part of filesystem freezing or not.

Signed-off-by: Jan Kara <jack@suse.cz>
---

...

 xfs_trans_t *
 _xfs_trans_alloc(
 	xfs_mount_t	*mp,
 	uint		type,
-	uint		memflags)
+	uint		memflags,
+	bool		freezing)
 {
 	xfs_trans_t	*tp;
 
-	atomic_inc(&mp->m_active_trans);
-
+	if (!freezing)
+		sb_start_write(mp->m_super, SB_FREEZE_TRANS);
+	else
+		WARN_ON(xfs_test_for_freeze(mp) != SB_FREEZE_TRANS);

Hm this could be an issue because for both the umount path and the 
freeze / xfs_quiesce_attr path, we call xfs_log_sbcount which sends
"true" for freezing and we'll trip up here because we won't be
in SB_FREEZE_TRANS during umount.

I think we have to push the flag all the way up to xfs_log_sbcount
callers?

-Eric