Re: [PATCH 02/11] writeback: switch to per-bdi threads forflushing data

From: Jens Axboe
Date: Wed May 20 2009 - 08:49:11 EST


On Wed, May 20 2009, Christoph Hellwig wrote:
> On Wed, May 20, 2009 at 02:16:30PM +0200, Jens Axboe wrote:
> > It's a fine rule, I agree ;-)
> >
> > I'll take another look at this when splitting the sync paths.
>
> Btw, there has been quite a bit of work on the higher level sync code in
> the VFS tree, and I have some TODO list items for the lower level sync
> code. The most important one would be splitting data and metadata
> writeback.
>
> Currently __sync_single_inode first calls do_writepages to write back
> the data, then write_inode to potentially write the metadata and then
> finally filemap_fdatawait to wait for the inode to be completed.
>
> Now for one thing doing the data wait after the metadata writeout is
> wrong for all those filesystems performing some kind of metadata updates
> in the I/O completion handler, and e.g. XFS has to work around this
> by doing a wait by itself in it's write_inode handler.
>
> Second inodes are usually clustered together, so if a filesystem can
> issue multiple dirty inodes at the same time performance will be much
> better.
>
> So an optimal sync could would first issue data I/O for all inodes it
> wants to write back, then wait for the data I/O to finish and finally
> write out the inodes in big clusters.
>
> I'm not quite sure when we'll get to that, just making sure we don't
> work against this direction anywhere.
>
> And yeah, I really need to take a detailed look at the current
> incarnation of your patchset :)

Please do, I'm particularly interested in the possibility of having
multiple inode placements. Would it be feasible to have the inode
backing be differentiated by type (eg data or meta-data)?

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/