On Tue, Dec 07, 2010 at 02:44:56PM +1100, Dave Chinner wrote:
> On Mon, Dec 06, 2010 at 09:33:08AM -0500, Christoph Hellwig wrote:
> > While the patch looks good for the ail lock contetion removal, I don't
> > quite like the model with the double iteration over the log item list
> > on the buffer. What do you think about the following plan:
> >
> > (1) merge xfs_istale_done into xfs_iflush_done by checking for
> > XFS_ISTALE
> > (2) convert not only the inode log item completion to your new scheme,
> > but also the dquots
> > (3) replace xfs_buf_do_callbacks with a callback in the buffer, which
> > now points to the inode and dqout routines, or calls the completion
> > for the only items in "normal" buf items.
>
> Seems like a reasonable approach. However, what I'd prefer to do is
> make these changes as a separate set of changes on top of this patch
> series rather than try to integrate them into the existing series.
> If there are problems, that should make it more bisectable. Do you
> have any concerns with such an approach?
Sounds fine, although I think getting rid of xfs_istale_done might be
worth doing before this patch.
|