Hi,

  while reading through the checkpointing code I've realized that we
actually have to send a barrier before each update of journal superblock
after checkpointing. Attached patch does this. Just I'm not sure whether
the performance cost won't be too big. In principle, we could make this
more lightweight by using the fact that transaction commit also sends the
barrier. So we could check before sending a barrier for transaction commit
whether we are slowly running out of journal space and if so whether some
transaction isn't already checkpointed. If yes, we can happily submit
update of journal superblock after the barrier. In case journal is decently
large this should solve the checkpointing problem without introducing
noticeable overhead...

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR