On Fri 01-02-19 23:08:11, Theodore Y. Ts'o wrote: > On Fri, Feb 01, 2019 at 10:21:20PM +0100, Jan Kara wrote: > > On Thu 31-01-19 23:42:19, Theodore Ts'o wrote: > > > This reverts commit ad211f3e94b314a910d4af03178a0b52a7d1ee0a. > > > > > > As Jan Kara pointed out, this change was unsafe since it means we lose > > > the call to sync_mapping_buffers() in the nojournal case. The > > > original point of the commit was avoid taking the inode mutex (since > > > it causes a lockdep warning in generic/113); but we need the mutex in > > > order to call sync_mapping_buffers(). > > > > Actually, I don't think sync_mapping_buffers() needs inode mutex (i_rwsem > > these days). It uses blkdev_mapping->private_lock for synchronization of > > operations on the list of buffers and fsync_buffers_list() seems to be > > pretty careful about races with mark_buffer_dirty_inode(). So why do you > > think we need i_rwsem? > > Hmm, I think you're right. I wonder if we can therefore remove the > inode_lock() in __generic_file_fsync() then... What do you think? That's actually a good question. I was thinking about why we have inode_lock() in __generic_file_fsync(). The only reason I could come up with is that when fsync(2) races with write(2) or truncate(2), with inode_lock() in __generic_file_fsync() you will either get old or new metadata state on disk. Without inode_lock() you could get some intermediate metadata state and thus after a crash may not be able to see even the old data. We are here on the thin ice of how good data consistency do we provide after a crash for non-journalling filesystems. It is never going to be perfect but this change would seem like a noticeable regression to me. What do you think? Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
