[dm-devel] mirroring: [patch 7 of 8] device failure tolerance

[Jonathan E Brassow <jbrassow redhat com>]

This patch is necessary _only_ to enable cluster capable mirrors.  It 
also requires changes to multipath if it is to be included.  You could 
safely running mirroring on a single machine without this patch.

In cluster mirroring, if a device failure occurs, an end_io can not be 
issued until a suspend has taken place.  This is because another 
machine could still be able to see the device that the first machine 
got a write failure on, and read it - getting inconsistent results.  
Because the mirror is waiting for a suspend before issuing the 
end_io's, the lock_fs call in dm_suspend must come after the 
presuspend.

 brassow

diff -urN linux-2.6.12-002/drivers/md/dm.c 
linux-2.6.12-003/drivers/md/dm.c
--- linux-2.6.12-002/drivers/md/dm.c	2005-06-17 14:48:29.000000000 -0500
+++ linux-2.6.12-003/drivers/md/dm.c	2005-06-28 16:05:45.118194654 -0500
@@ -1055,14 +1055,14 @@
 	if (test_bit(DMF_BLOCK_IO, &md->flags))
 		goto out_read_unlock;

-	error = __lock_fs(md);
-	if (error)
-		goto out_read_unlock;
-
 	map = dm_get_table(md);
 	if (map)
 		dm_table_presuspend_targets(map);

+	error = __lock_fs(md);
+	if (error)
+		goto out_read_unlock;
+
 	up_read(&md->lock);

 	/*