> On Friday 03 April 2009 03:55:05 Christoph Hellwig wrote: > > On Fri, Apr 03, 2009 at 03:47:20AM +1100, Nick Piggin wrote: > > > Well they now are quite well filesystem defined. We no longer take > > > the page lock before calling them. Not saying it's perfect, but if > > > the backing fs is just using a known subset of ones that work > > > (like loop does). > > > > The page lock doesn't matter. What matters is locks protecting the > > io. Like the XFS iolock or cluster locks in the cluster filesystems, > > and you will get silent data corruption that way. > > Hmm, I can see i_mutex being a problem, but can't see how a filesystem > takes any other locks down that chain? Yes, i_mutex is one problem. Then filesystems may take other locks in their ->aio_write callbacks - as Christoph mentioned, for example OCFS2 has to do some network messaging to synchronize nodes in the cluster accessing the file. I could imagine some clever filesystem doing more clever locking than just one i_mutex covering the whole file... > Naturally a random in-kernel user misses other important things, so yes > a simple write sounds like the best option. Definitely. IMO it's hard to get the locking right without calling ->aio_write callback. Honza -- Jan Kara <jack@xxxxxxx> SuSE CR Labs -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html