Re: [PATCH 14/17] xfs: Use new syncing helper

[Jan Kara <jack@suse.cz>]

On Wed 19-08-09 12:33:15, Christoph Hellwig wrote:
> On Wed, Aug 19, 2009 at 06:04:41PM +0200, Jan Kara wrote:
> > diff --git a/fs/xfs/linux-2.6/xfs_lrw.c b/fs/xfs/linux-2.6/xfs_lrw.c
> > index 7078974..aeb5a39 100644
> > --- a/fs/xfs/linux-2.6/xfs_lrw.c
> > +++ b/fs/xfs/linux-2.6/xfs_lrw.c
> > @@ -817,7 +817,7 @@ write_retry:
> >  		xfs_iunlock(xip, iolock);
> >  		if (need_i_mutex)
> >  			mutex_unlock(&inode->i_mutex);
> > -		error2 = sync_page_range(inode, mapping, pos, ret);
> > +		error2 = generic_write_sync(file, pos, ret);
> 
> I don't think this is very optimal for XFS.  This first starts an
> asynchronous writeout of the range in generic_write_sync,
> then calls into ->fsync which waits for all I/O on the file to finish,
> then forces the log inside xfs_fsync,  then waits for the range again in
> generic_write_sync, and after this code calls into
> xfs_write_sync_logforce which forces to log _again_.
> 
> We should be fine just doing your new filemap_fdatawrite_range helper
> here and let xfs_write_sync_logforce do the work.  Long term we should
> just get rid of this stupid submit writes before syncing the iode, then
> wait crap which is a pain for all modern filesystems.
  OK, I'll happily change this. I was just doing the straightforward
conversion and sync_page_range() essentially did what generic_write_sync()
does since it called fdatawrite(), write_inode_now(inode, 1) (here XFS was
forced to wait for pages and force the log), fdatawait().

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR