On Fri, Oct 31, 2008 at 04:31:23PM -0400, Christoph Hellwig wrote: > On Fri, Oct 31, 2008 at 11:12:49AM +1100, Dave Chinner wrote: > > Right - that's exactly where we should be going with this, I think. > > I'd suggest two callouts, perhaps: ->sync_data and ->sync_metadata. > > The freeze code can then still operate in two stages, and we can > > also use then for separating data and inode writeback in pdflush.... > > > > FWIW, I mentioned doing this sort of thing here: > > > > http://xfs.org/index.php/Improving_inode_Caching#Avoiding_the_Generic_pdflush_Code > > > > I think I'll look at redoing do_sync() to provide a custom sync > > method before trying to fix XFS.... > > And you weren't the first to thing of this. Reiser4 for example > has bad a patch forever to turn sync_sb_inodes into a filesystem method, > and I think something similar is what we want. When talking about > syncing we basically want a few things: > > - sync out data, either async (from pdflush) or sync > (from sync, freeze, remount ro or unmount) > - sync out metadata (from pdflush), either async or sync > (from sync, freeze, remount ro or unmount) Effectively, yes. Currently we iterate inodes for data and "metadata" sync, and the only other concept is writing superblocks. I think most filesystems have more types of metadata than this, so it makes sense for sync to work on abstracts sync as data and metadata rather than data, inodes and superblocks... > and then we want pdflush / sync / etc call into it. If we are doing > this correctly this would also avoid having our own xfssyncd. Yes, though we'd need to change a couple of the functions that xfssynd does to pdflush operations... > And as we found out it's not just sync that gets it wrong, it's also > fsync (which isn't part of the above picture as it's per-inode) that > gets this utterly wrong, as well as all kinds of syncs, not just the > unmount one. Async writeback (write_inode()) has the same problem as fsync - writing the inode before waiting for data I/O to complete - which means we've got to jump through hoops in the filesystem to avoid blocking on inodes that can't be immediately flushed, and often we end up writing the inode multiple times and having to issue log forces whenw e shouldn't need to. Effectively we have to tell the VFS to "try again later" the entire time data is being flushed before we can write the inode and it's exceedingly inefficient..... > Combine this with the other data integrity issues Nick > found in write_cache_pages I come to the conclusion that this whole area > needs some profound audit and re-architecture urgently. It's looking more and more that way, isn't it? Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- 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
